Part Number Hot Search : 
CMBT8550 SP568 24C25 1215S TDA8000 LXT908 X24C45S S1616
Product Description
Full Text Search
 

To Download AMD64 Datasheet File
  If you can't view the Datasheet, Please click here to try to view without PDF Reader .  
　
 

  Datasheet File OCR Text:
 AMD64 Technology
AMD64 Architecture Programmer's Manual Volume 2: System Programming
Publication No. 24593
Revision 3.09
Date September 2003
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
(c) 2002, 2003 Advanced Micro Devices, Inc. All rights reserved. The contents of this document are provided in connection with Advanced Micro Devices, Inc. ("AMD") products. AMD makes no representations or warranties with respect to the accuracy or completeness of the contents of this publication and reserves the right to make changes to specifications and product descriptions at any time without notice. No license, whether express, implied, arising by estoppel or otherwise, to any intellectual property rights is granted by this publication. Except as set forth in AMD's Standard Terms and Conditions of Sale, AMD assumes no liability whatsoever, and disclaims any express or implied warranty, relating to its products including, but not limited to, the implied warranty of merchantability, fitness for a particular purpose, or infringement of any intellectual property right. AMD's products are not designed, intended, authorized or warranted for use as components in systems intended for surgical implant into the body, or in other applications intended to support or sustain life, or in any other application in which the failure of AMD's product could create a situation where personal injury, death, or severe property or environmental damage may occur. AMD reserves the right to discontinue or make changes to its products at any time without notice.
Trademarks AMD, the AMD Arrow logo, AMD Athlon, AMD Opteron and combinations thereof, 3DNow!, nX586, and nX686 are trademarks, and AMD-K6 is a registered trademark of Advanced Micro Devices, Inc. MMX is a trademark and Pentium is a registered trademark of Intel Corporation. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Contents
Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xiii Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix Revision History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xxiii About This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xxiii Audience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xxiii Contact Information. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xxiii Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiv Definitions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxv Related Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxxvi 1 System-Programming Overview . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Memory Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Memory Addressing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Memory Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Canonical Address Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Memory Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Paging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Mixing Segmentation and Paging . . . . . . . . . . . . . . . . . . . . . . 10 Real Addressing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Operating Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 Compatibility Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 Legacy Modes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 System Management Mode (SMM) . . . . . . . . . . . . . . . . . . . . . 17 System Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 System-Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 Additional System-Programming Facilities . . . . . . . . . . . . . . 24 Hardware Multitasking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Machine Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 Software Debugging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 Performance Monitoring. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 Operating Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Legacy Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 System-Management Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.2
1.3
1.4 1.5 1.6 1.7
2
x86 and AMD64 Architecture Differences . . . . . . . . . . . . . . . 29
2.1
Contents
iii
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
2.2
2.3 2.4
2.5
2.6
2.7 2.8
Memory Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Memory Addressing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Page Translation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 Protection Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 General-Purpose Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 128-Bit Media Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Flags Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Instruction Pointer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Stack Pointer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Control Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Debug Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Extended Feature Register (EFER) . . . . . . . . . . . . . . . . . . . . 37 Memory Type Range Registers (MTRRs) . . . . . . . . . . . . . . . . 37 Other Model-Specific Registers (MSRs) . . . . . . . . . . . . . . . . . 37 Instruction Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 REX Prefixes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 Segment-Override Prefixes in 64-Bit Mode. . . . . . . . . . . . . . . 38 Operands and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 Address Calculations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 Instructions that Reference RSP . . . . . . . . . . . . . . . . . . . . . . . 39 Branches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 NOP Instruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Single-Byte INC and DEC Instructions . . . . . . . . . . . . . . . . . . 43 MOVSXD Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Invalid Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 FXSAVE and FXRSTOR Instructions . . . . . . . . . . . . . . . . . . . 45 Interrupts and Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Interrupt Descriptor Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Stack Frame Pushes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Stack Switching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 IRET Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 Task-Priority Register (CR8) . . . . . . . . . . . . . . . . . . . . . . . . . . 49 New Exception Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 Hardware Task Switching . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 Long-Mode vs. Legacy-Mode Differences . . . . . . . . . . . . . . . . 50 System-Control Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 CR0 Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 CR2 and CR3 Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 CR4 Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 CR1 and CR5-CR7 Registers . . . . . . . . . . . . . . . . . . . . . . . . . . 64 64-Bit-Mode Extended Control Registers . . . . . . . . . . . . . . . . 64 CR8 (Task Priority Register, TPR). . . . . . . . . . . . . . . . . . . . . . 64 RFLAGS Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 Extended Feature Enable Register (EFER) . . . . . . . . . . . . . . 69
3
System Resources. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.1
iv
Contents
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
3.2
3.3
Model-Specific Registers (MSRs) . . . . . . . . . . . . . . . . . . . . . . 71 System Configuration Register (SYSCFG) . . . . . . . . . . . . . . . 73 System-Linkage Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 Memory-Typing Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Debug-Extension Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 Performance-Monitoring Registers . . . . . . . . . . . . . . . . . . . . . 76 Machine-Check Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 Processor Feature Identification . . . . . . . . . . . . . . . . . . . . . . . 78 Real Mode Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Virtual-8086 Mode Segmentation . . . . . . . . . . . . . . . . . . . . . . 80 Protected Mode Segmented-Memory Models. . . . . . . . . . . . . 81 Multi-Segmented Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 Flat-Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 Segmentation in 64-bit mode . . . . . . . . . . . . . . . . . . . . . . . . . . 82 Segmentation Data Structures and Registers . . . . . . . . . . . . 82 Segment Selectors and Registers . . . . . . . . . . . . . . . . . . . . . . 84 Segment Selectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 Segment Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 Segment Registers in 64-bit Mode . . . . . . . . . . . . . . . . . . . . . . 87 Descriptor Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 Global Descriptor Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 Global Descriptor-Table Register . . . . . . . . . . . . . . . . . . . . . . 91 Local Descriptor Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 Local Descriptor-Table Register . . . . . . . . . . . . . . . . . . . . . . . 93 Interrupt Descriptor Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 Interrupt Descriptor-Table Register . . . . . . . . . . . . . . . . . . . . 97 Legacy Segment Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . 97 Descriptor Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 Code-Segment Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 Data-Segment Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 System Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 Gate Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 Long-Mode Segment Descriptors . . . . . . . . . . . . . . . . . . . . . . 108 Code-Segment Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 Data-Segment Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 System Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Gate Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 Long Mode Descriptor Summary . . . . . . . . . . . . . . . . . . . . . . 116 Segment-Protection Overview . . . . . . . . . . . . . . . . . . . . . . . . 118 Privilege-Level Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Privilege-Level Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 Data-Access Privilege Checks . . . . . . . . . . . . . . . . . . . . . . . . 121 Accessing Data Segments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Accessing Stack Segments . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 Control-Transfer Privilege Checks . . . . . . . . . . . . . . . . . . . . 124 Direct Control Transfers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4
Segmented Virtual Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.1 4.2 4.3
4.4 4.5
4.6
4.7
4.8
4.9
4.10
4.11
Contents
v
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
4.12 4.13
Control Transfers Through Call Gates. . . . . . . . . . . . . . . . . . 129 Return Control Transfers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 Limit Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 Determining Limit Violations . . . . . . . . . . . . . . . . . . . . . . . . 140 Type Checks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 Type Checks in Legacy and Compatibility Modes . . . . . . . . 142 Long Mode Type Check Differences . . . . . . . . . . . . . . . . . . . 143 Page Translation Overview . . . . . . . . . . . . . . . . . . . . . . . . . . 146 Page-Translation Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 Page-Translation Enable (PG) Bit . . . . . . . . . . . . . . . . . . . . . 149 Physical-Address Extensions (PAE) Bit. . . . . . . . . . . . . . . . . 149 Page-Size Extensions (PSE) Bit . . . . . . . . . . . . . . . . . . . . . . . 149 Page-Directory Page Size (PS) Bit . . . . . . . . . . . . . . . . . . . . . 150 Legacy-Mode Page Translation . . . . . . . . . . . . . . . . . . . . . . . 150 CR3 Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Normal (Non-PAE) Paging . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 PAE Paging. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 Long-Mode Page Translation . . . . . . . . . . . . . . . . . . . . . . . . . 160 Canonical Address Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 CR3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 4-Kbyte Page Translation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 2-Mbyte Page Translation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 Page-Translation-Table Entry Fields . . . . . . . . . . . . . . . . . . 168 Field Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 Translation-Lookaside Buffer (TLB) . . . . . . . . . . . . . . . . . . . 172 Global Pages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 TLB Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 Page-Protection Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 No Execute (NX) Bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 User/Supervisor (U/S) Bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 Read/Write (R/W) Bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Write Protect (CR0.WP) Bit . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Protection Across Paging Hierarchy . . . . . . . . . . . . . . . . . . . 176 Access to User Pages when CR0.WP=1 . . . . . . . . . . . . . . . . . 178 Effects of Segment Protection . . . . . . . . . . . . . . . . . . . . . . . . 178 Fast System Call and Return . . . . . . . . . . . . . . . . . . . . . . . . . 181 SYSCALL and SYSRET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 SYSENTER and SYSEXIT (Legacy Mode Only) . . . . . . . . . 184 SWAPGS Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 System Status and Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 Processor Feature Identification (CPUID) . . . . . . . . . . . . . . 186 Accessing Control Registers. . . . . . . . . . . . . . . . . . . . . . . . . . 186 Accessing the RFLAGs Register . . . . . . . . . . . . . . . . . . . . . . 187 Accessing Debug Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . 187
5
Page Translation and Protection . . . . . . . . . . . . . . . . . . . . . . 145
5.1
5.2
5.3
5.4 5.5
5.6
5.7 5.8
6
System-Management Instructions . . . . . . . . . . . . . . . . . . . . . 179
6.1
6.2
vi
Contents
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
6.3
6.4
6.5 6.6
Accessing Model-Specific Registers . . . . . . . . . . . . . . . . . . . 188 Segment Register and Descriptor Register Access . . . . . . . 188 Accessing Segment Registers. . . . . . . . . . . . . . . . . . . . . . . . . 188 Accessing Descriptor-Table Registers . . . . . . . . . . . . . . . . . . 189 Protection Checking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 Checking Access Rights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 Checking Segment Limits. . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 Checking Read/Write Rights . . . . . . . . . . . . . . . . . . . . . . . . . 190 Adjusting Access Rights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 Processor Halt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 Cache and TLB Management . . . . . . . . . . . . . . . . . . . . . . . . . 191 Cache Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 TLB Invalidation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 Memory-Access Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 Read Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 Write Ordering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 Read/Write Barriers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 Memory Coherency and Protocol . . . . . . . . . . . . . . . . . . . . . 199 Special Coherency Considerations . . . . . . . . . . . . . . . . . . . . 201 Memory Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 Buffering and Combining Memory Writes . . . . . . . . . . . . . . 205 Write Buffering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 Write Combining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 Memory Caches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208 Cache Organization and Operation . . . . . . . . . . . . . . . . . . . . 209 Cache Control Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . 212 Cache and Memory Management Instructions . . . . . . . . . . . 215 Serializing Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 Memory-Type Range Registers . . . . . . . . . . . . . . . . . . . . . . . 217 MTRR Type Fields. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 MTRRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 Using MTRRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 MTRRs and Page Cache Controls . . . . . . . . . . . . . . . . . . . . . 227 MTRRs in Multi-Processing Environments. . . . . . . . . . . . . . 229 Page-Attribute Table Mechanism . . . . . . . . . . . . . . . . . . . . . 230 PAT Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230 PAT Indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 Identifying PAT Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 PAT Accesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 Combined Effect of MTRRs and PAT . . . . . . . . . . . . . . . . . . 233 Memory-Mapped I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234 Extended Fixed-Range MTRR Type-Field Encodings . . . . . 235 IORRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 IORR Overlapping. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239 Top of Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
7
Memory System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.1
7.2 7.3 7.4
7.5
7.6
7.7
7.8
Contents
vii
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
8
Exceptions and Interrupts. . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
8.1 General Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 Precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 Instruction Restart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 Types of Exceptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245 Masking External Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . 245 Masking Floating-Point and Media Instructions . . . . . . . . . 246 Disabling Exceptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 #DE--Divide-by-Zero-Error Exception (Vector 0) . . . . . . . . 250 #DB--Debug Exception (Vector 1) . . . . . . . . . . . . . . . . . . . . 250 NMI--Non-Maskable-Interrupt Exception (Vector 2) . . . . . 252 #BP--Breakpoint Exception (Vector 3) . . . . . . . . . . . . . . . . . 252 #OF--Overflow Exception (Vector 4) . . . . . . . . . . . . . . . . . . 252 #BR--Bound-Range Exception (Vector 5). . . . . . . . . . . . . . . 253 #UD--Invalid-Opcode Exception (Vector 6) . . . . . . . . . . . . . 253 #NM--Device-Not-Available Exception (Vector 7). . . . . . . . 254 #DF--Double-Fault Exception (Vector 8) . . . . . . . . . . . . . . . 254 Coprocessor-Segment-Overrun Exception (Vector 9) . . . . . 256 #TS--Invalid-TSS Exception (Vector 10) . . . . . . . . . . . . . . . 256 #NP--Segment-Not-Present Exception (Vector 11) . . . . . . . 257 #SS--Stack Exception (Vector 12) . . . . . . . . . . . . . . . . . . . . . 258 #GP--General-Protection Exception (Vector 13) . . . . . . . . . 259 #PF--Page-Fault Exception (Vector 14) . . . . . . . . . . . . . . . . 261 #MF--x87 Floating-Point Exception-Pending (Vector 16). . 262 #AC--Alignment-Check Exception (Vector 17) . . . . . . . . . . 263 #MC--Machine-Check Exception (Vector 18) . . . . . . . . . . . 264 #XF--SIMD Floating-Point Exception (Vector 19) . . . . . . . 265 User-Defined Interrupts (Vectors 32-255) . . . . . . . . . . . . . . 266 Exceptions During a Task Switch . . . . . . . . . . . . . . . . . . . . . 267 Error Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 Selector-Error Code. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 Page-Fault Error Code. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268 Priorities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 Floating-Point Exception Priorities . . . . . . . . . . . . . . . . . . . . 271 External Interrupt Priorities . . . . . . . . . . . . . . . . . . . . . . . . . 272 Real-Mode Interrupt Control Transfers . . . . . . . . . . . . . . . . 274 Legacy Protected-Mode Interrupt Control Transfers . . . . . 276 Locating the Interrupt Handler . . . . . . . . . . . . . . . . . . . . . . . 276 Interrupt To Same Privilege . . . . . . . . . . . . . . . . . . . . . . . . . . 277 Interrupt To Higher Privilege . . . . . . . . . . . . . . . . . . . . . . . . 278 Privilege Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 Returning From Interrupt Procedures . . . . . . . . . . . . . . . . . 283 Virtual-8086 Mode Interrupt Control Transfers. . . . . . . . . . 284 Protected-Mode Handler Control Transfer . . . . . . . . . . . . . . 285 Virtual-8086 Handler Control Transfer . . . . . . . . . . . . . . . . . 286 Long-Mode Interrupt Control Transfers . . . . . . . . . . . . . . . . 287
8.2
8.3 8.4
8.5
8.6 8.7
8.8
8.9
viii
Contents
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
8.10
Interrupt Gates and Trap Gates . . . . . . . . . . . . . . . . . . . . . . . 287 Locating the Interrupt Handler . . . . . . . . . . . . . . . . . . . . . . . 288 Interrupt Stack Frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 Interrupt-Stack Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292 Returning From Interrupt Procedures . . . . . . . . . . . . . . . . . 294 Virtual Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295 Virtual-8086 Mode Extensions . . . . . . . . . . . . . . . . . . . . . . . . 296 Protected Mode Virtual Interrupts . . . . . . . . . . . . . . . . . . . . 300 Effect of Instructions that Modify EFLAGS.IF . . . . . . . . . . 300 Determining Machine-Check Support. . . . . . . . . . . . . . . . . . 305 Machine-Check Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 Error Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 Machine Check MSRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307 Global Status and Control Registers . . . . . . . . . . . . . . . . . . . 308 Error-Reporting Register Banks . . . . . . . . . . . . . . . . . . . . . . 310 Error Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312 Initializing the Machine-Check Mechanism . . . . . . . . . . . . . 315 Using Machine Check Features . . . . . . . . . . . . . . . . . . . . . . . 316 Handling Machine Check Exceptions . . . . . . . . . . . . . . . . . . 316 Reporting Correctable Machine Check Errors . . . . . . . . . . . 318 SMM Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322 SMM Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322 SMRAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322 SMBASE Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323 SMRAM State-Save Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324 SMM-Revision Identifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330 Using SMM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331 System-Management Interrupt (SMI) . . . . . . . . . . . . . . . . . . 331 SMM Operating-Environment . . . . . . . . . . . . . . . . . . . . . . . . 331 Exceptions and Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . 332 Invalidating the Caches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334 Saving Additional Processor State. . . . . . . . . . . . . . . . . . . . . 334 Operating in Protected Mode and Long Mode . . . . . . . . . . . 335 Auto-Halt Restart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 I/O Instruction Restart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336 Leaving SMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337 Overview of System-Software Considerations . . . . . . . . . . . 339 Determining Media and x87 Feature Support . . . . . . . . . . . 339 Enabling 128-Bit Media Instructions . . . . . . . . . . . . . . . . . . . 340 Media and x87 Processor State . . . . . . . . . . . . . . . . . . . . . . . 341 128-Bit Media State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341 64-Bit Media State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
9
Machine Check Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . 305
9.1 9.2 9.3
9.4 9.5
10
System-Management Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
10.1
10.2
10.3
11
128-Bit, 64-Bit, and x87 Programming . . . . . . . . . . . . . . . . . . 339
11.1 11.2 11.3 11.4
Contents
ix
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
x87 State. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344 Saving Media and x87 Processor State . . . . . . . . . . . . . . . . . 346
12
Task Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
12.1 12.2 Hardware Multitasking Overview . . . . . . . . . . . . . . . . . . . . . 359 Task-Management Resources. . . . . . . . . . . . . . . . . . . . . . . . . 360 TSS Selector. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362 TSS Descriptor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362 Task Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363 Legacy Task-State Segment . . . . . . . . . . . . . . . . . . . . . . . . . . 365 64-Bit Task State Segment . . . . . . . . . . . . . . . . . . . . . . . . . . . 370 Task Gate Descriptor (Legacy Mode Only) . . . . . . . . . . . . . . 373 Hardware Task-Management in Legacy Mode . . . . . . . . . . . 373 Task Memory-Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373 Switching Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375 Task Switches Using Task Gates . . . . . . . . . . . . . . . . . . . . . . 377 Nesting Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379 Software-Debug Resources. . . . . . . . . . . . . . . . . . . . . . . . . . . 384 Debug Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385 Breakpoints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394 Setting Breakpoints. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394 Using Breakpoints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396 Breakpoint Instruction (INT3) . . . . . . . . . . . . . . . . . . . . . . . . 401 Control-Transfer Breakpoint Features. . . . . . . . . . . . . . . . . . 401 Performance Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 403 Performance Counters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404 Performance Event-Select Registers . . . . . . . . . . . . . . . . . . . 405 Using Performance Counters . . . . . . . . . . . . . . . . . . . . . . . . . 408 Time-Stamp Counter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408 Reset and Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411 Built-In Self Test (BIST) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411 Clock Multiplier Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 412 Processor Initialization State . . . . . . . . . . . . . . . . . . . . . . . . . 412 Multiple Processor Initialization . . . . . . . . . . . . . . . . . . . . . . 415 Fetching the First Instruction. . . . . . . . . . . . . . . . . . . . . . . . . 415 Hardware Configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416 Processor Implementation Information . . . . . . . . . . . . . . . . 416 Enabling Internal Caches . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416 Initializing Media and x87 Processor State . . . . . . . . . . . . . 417 Model-Specific Initialization . . . . . . . . . . . . . . . . . . . . . . . . . 419 Initializing Real Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420 Initializing Protected Mode . . . . . . . . . . . . . . . . . . . . . . . . . . 421 Initializing Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422 Enabling and Activating Long Mode. . . . . . . . . . . . . . . . . . . 424
12.3
13
Debug and Performance Resources . . . . . . . . . . . . . . . . . . . . 383
13.1 13.2
13.3
14
Processor Initialization and Long-Mode Activation . . . . . . . 411
14.1
14.2
14.3 14.4 14.5 14.6
x
Contents
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
14.7 14.8
Activating Long Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425 Consistency Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426 Updating System Descriptor Table References . . . . . . . . . . 426 Relocating Page-Translation Tables. . . . . . . . . . . . . . . . . . . . 427 Leaving Long Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427 Long-Mode Initialization Example . . . . . . . . . . . . . . . . . . . . 428 MSR Cross-Reference by MSR Address . . . . . . . . . . . . . . . . 435 System-Software MSRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439 Memory-Typing MSRs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440 Machine-Check MSRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443 Software-Debug MSRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444 Performance-Monitoring MSRs . . . . . . . . . . . . . . . . . . . . . . . 445
Appendix A
MSR Cross-Reference. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
A.1 A.2 A.3 A.4 A.5 A.6
Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Contents
xi
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
xii
Contents
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Figures
Figure 1-1. Segmented-Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Figure 1-2. Flat Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Figure 1-3. Paged Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 Figure 1-4. 64-Bit Flat, Paged-Memory Model . . . . . . . . . . . . . . . . . . . . . . . 11 Figure 1-5. Real-Address Memory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Figure 1-6. Operating Modes of the AMD64 Architecture . . . . . . . . . . . . . 14 Figure 1-7. System Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Figure 1-8. System-Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Figure 3-1. Control Register 0 (CR0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 Figure 3-2. Control Register 2 (CR2)--Legacy-Mode. . . . . . . . . . . . . . . . . . 59 Figure 3-3. Control Register 2 (CR2)--Long Mode . . . . . . . . . . . . . . . . . . . 59 Figure 3-4. Control Register 3 (CR3)--Legacy-Mode Non-PAE Paging . . . 59 Figure 3-5. Control Register 3 (CR3)--Legacy-Mode PAE Paging . . . . . . . 59 Figure 3-6. Control Register 3 (CR3)--Long Mode . . . . . . . . . . . . . . . . . . . 60 Figure 3-7. Control Register 4 (CR4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 Figure 3-8. RFLAGS Register. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Figure 3-9. Extended Feature Enable Register (EFER) . . . . . . . . . . . . . . . 70 Figure 3-10.AMD64 Architecture Model-Specific Registers . . . . . . . . . . . . 73 Figure 3-11.System-Configuration Register (SYSCFG) . . . . . . . . . . . . . . . . 74 Figure 4-1. Segmentation Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . 83 Figure 4-2. Segment and Descriptor-Table Registers . . . . . . . . . . . . . . . . . 84 Figure 4-3. Segment Selector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 Figure 4-4. Segment-Register Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 Figure 4-5. FS and GS Segment-Register Format--64-Bit Mode . . . . . . . . 88 Figure 4-6. Global and Local Descriptor-Table Access . . . . . . . . . . . . . . . . 91 Figure 4-7. GDTR and IDTR Format--Legacy Modes . . . . . . . . . . . . . . . . . 91 Figure 4-8. GDTR and IDTR Format--Long Mode. . . . . . . . . . . . . . . . . . . . 92 Figure 4-9. Relationship between the LDT and GDT. . . . . . . . . . . . . . . . . . 93 Figure 4-10.LDTR Format--Legacy Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 Figure 4-11.LDTR Format--Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 Figure 4-12.Indexing an IDT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Figures
xiii
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Figure 4-13.Generic Segment Descriptor--Legacy Mode. . . . . . . . . . . . . . . 98 Figure 4-14.Code-Segment Descriptor--Legacy Mode . . . . . . . . . . . . . . . . 101 Figure 4-15.Data-Segment Descriptor--Legacy Mode . . . . . . . . . . . . . . . . 103 Figure 4-16.LDT and TSS Descriptor--Legacy/Compatibility Modes . . . . 106 Figure 4-17.Call-Gate Descriptor--Legacy Mode . . . . . . . . . . . . . . . . . . . . 107 Figure 4-18.Interrupt-Gate and Trap-Gate Descriptors--Legacy Mode . . 107 Figure 4-19.Task-Gate Descriptor--Legacy Mode. . . . . . . . . . . . . . . . . . . . 107 Figure 4-20.Code-Segment Descriptor--Long Mode . . . . . . . . . . . . . . . . . . 109 Figure 4-21.Data-Segment Descriptor--Long Mode . . . . . . . . . . . . . . . . . . 110 Figure 4-22.System-Segment Descriptor--64-Bit Mode . . . . . . . . . . . . . . . 113 Figure 4-23.Call-Gate Descriptor--Long Mode . . . . . . . . . . . . . . . . . . . . . . 114 Figure 4-24.Interrupt-Gate and Trap-Gate Descriptors--Long Mode . . . . 115 Figure 4-25.Privilege-Level Relationships . . . . . . . . . . . . . . . . . . . . . . . . . . 120 Figure 4-26.Data-Access Privilege-Check Examples. . . . . . . . . . . . . . . . . . 122 Figure 4-27.Stack-Access Privilege-Check Examples . . . . . . . . . . . . . . . . . 124 Figure 4-28.Nonconforming Code-Segment Privilege-Check Examples . . 127 Figure 4-29.Conforming Code-Segment Privilege-Check Examples . . . . . 129 Figure 4-30.Legacy-Mode Call-Gate Transfer Mechanism . . . . . . . . . . . . . 130 Figure 4-31.Long-Mode Call-Gate Access Mechanism . . . . . . . . . . . . . . . . 131 Figure 4-32.Privilege-Check Examples for Call Gates . . . . . . . . . . . . . . . . 133 Figure 4-33.Legacy-Mode 32-Bit Stack Switch, with Parameters . . . . . . . 135 Figure 4-34.32-Bit Stack Switch, No Parameters--Legacy Mode. . . . . . . . 136 Figure 4-35.Stack Switch--Long Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 Figure 5-1. Virtual to Physical Address Translation--Long Mode . . . . . . 147 Figure 5-2. Control Register 3 (CR3)--Non-PAE Paging Legacy-Mode . . 151 Figure 5-3. Control Register 3 (CR3)--PAE Paging Legacy-Mode . . . . . . 151 Figure 5-4. 4-Kbyte Non-PAE Page Translation--Legacy Mode . . . . . . . . 153 Figure 5-5. 4-Kbyte PDE--Non-PAE Paging Legacy-Mode . . . . . . . . . . . . 153 Figure 5-6. 4-Kbyte PTE--Non-PAE Paging Legacy-Mode . . . . . . . . . . . . 154 Figure 5-7. 4-Mbyte Page Translation--Non-PAE Paging Legacy-Mode . 155 Figure 5-8. 4-Mbyte PDE--Non-PAE Paging Legacy-Mode . . . . . . . . . . . . 155 Figure 5-9. 4-Kbyte PAE Page Translation--Legacy Mode . . . . . . . . . . . . 157 Figure 5-10.4-Kbyte PDPE--PAE Paging Legacy-Mode . . . . . . . . . . . . . . . 158
xiv
Figures
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Figure 5-11.4-Kbyte PDE--PAE Paging Legacy-Mode . . . . . . . . . . . . . . . . 158 Figure 5-12.4-Kbyte PTE--PAE Paging Legacy-Mode . . . . . . . . . . . . . . . . 158 Figure 5-13.2-Mbyte PAE Page Translation--Legacy Mode . . . . . . . . . . . 159 Figure 5-14.2-Mbyte PDPE--PAE Paging Legacy-Mode. . . . . . . . . . . . . . . 160 Figure 5-15.2-Mbyte PDE--PAE Paging Legacy-Mode . . . . . . . . . . . . . . . . 160 Figure 5-16.Control Register 3 (CR3)--Long Mode . . . . . . . . . . . . . . . . . . 161 Figure 5-17.4-Kbyte Page Translation--Long Mode . . . . . . . . . . . . . . . . . . 163 Figure 5-18.4-Kbyte PML4E--Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . 164 Figure 5-19.4-Kbyte PDPE--Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 Figure 5-20.4-Kbyte PDE--Long Mode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 Figure 5-21.4-Kbyte PTE--Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 Figure 5-22.2-Mbyte Page Translation--Long Mode. . . . . . . . . . . . . . . . . . 166 Figure 5-23.2-Mbyte PML4E--Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Figure 5-24.2-Mbyte PDPE--Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Figure 5-25.2-Mbyte PDE--Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Figure 6-1. STAR, LSTAR, CSTAR, and MASK MSRs. . . . . . . . . . . . . . . . 183 Figure 6-2. SYSENTER_CS, SYSENTER_ESP, SYSENTER_EIP MSRs . 185 Figure 7-1. Processor and Memory System . . . . . . . . . . . . . . . . . . . . . . . . . 194 Figure 7-2. MOESI State Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 Figure 7-3. Cache Organization Example . . . . . . . . . . . . . . . . . . . . . . . . . . 210 Figure 7-4. MTRR Mapping of Physical Memory . . . . . . . . . . . . . . . . . . . . 220 Figure 7-5. Fixed-Range MTRR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 Figure 7-6. MTRRphysBasen Register. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 Figure 7-7. MTRRphysMaskn Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 Figure 7-8. MTRR defType Register Format . . . . . . . . . . . . . . . . . . . . . . . 225 Figure 7-9. MTRR Capability Register Format . . . . . . . . . . . . . . . . . . . . . 227 Figure 7-10.PAT Register. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230 Figure 7-11.Extended MTRR Type-Field Format (Fixed-Range MTRRs) 235 Figure 7-12.IORRBasen Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238 Figure 7-13.IORRMaskn Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239 Figure 7-14.Memory Organization Using Top-of-Memory Registers . . . . . 240 Figure 7-15.Top-of-Memory Registers (TOP_MEM, TOP_MEM2) . . . . . . . 241 Figure 8-1. Control Register 2 (CR2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Figures
xv
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Figure 8-2. Selector Error Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 Figure 8-3. Page-Fault Error Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268 Figure 8-4. Task Priority Register (CR8) . . . . . . . . . . . . . . . . . . . . . . . . . . 273 Figure 8-5. Real-Mode Interrupt Control Transfer . . . . . . . . . . . . . . . . . . 274 Figure 8-6. Stack After Interrupt in Real Mode . . . . . . . . . . . . . . . . . . . . . 275 Figure 8-7. Protected-Mode Interrupt Control Transfer . . . . . . . . . . . . . . 277 Figure 8-8. Stack After Interrupt to Same Privilege Level . . . . . . . . . . . . 278 Figure 8-9. Stack After Interrupt to Higher Privilege . . . . . . . . . . . . . . . . 280 Figure 8-10.Privilege-Check Examples for Interrupts . . . . . . . . . . . . . . . . 282 Figure 8-11.Stack After Virtual-8086 Mode Interrupt to Protected Mode 286 Figure 8-12.Long-Mode Interrupt Control Transfer . . . . . . . . . . . . . . . . . . 289 Figure 8-13.Long-Mode Stack After Interrupt--Same Privilege . . . . . . . . 291 Figure 8-14.Long-Mode Stack After Interrupt--Higher Privilege. . . . . . . 292 Figure 8-15.Long-Mode IST Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294 Figure 9-1. MCG_CAP Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308 Figure 9-2. MCG_STATUS Register. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 Figure 9-3. MCG_CTL Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310 Figure 9-4. MCi_CTL Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311 Figure 9-5. MCi_STATUS Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312 Figure 10-1.Default SMRAM Memory Map . . . . . . . . . . . . . . . . . . . . . . . . . 323 Figure 10-2.SMBASE Register. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324 Figure 10-3.SMM-Revision Identifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330 Figure 11-1.128-Bit Media-Instruction State . . . . . . . . . . . . . . . . . . . . . . . . 342 Figure 11-2.64-Bit Media-Instruction State . . . . . . . . . . . . . . . . . . . . . . . . . 343 Figure 11-3.x87-Instruction State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 Figure 11-4.FSAVE/FNSAVE Image (32-Bit, Protected Mode) . . . . . . . . . 348 Figure 11-5.FSAVE/FNSAVE Image (32-Bit, Real/Virtual-8086 Modes) . 349 Figure 11-6.FSAVE/FNSAVE Image (16-Bit, Protected Mode) . . . . . . . . . 350 Figure 11-7.FSAVE/FNSAVE Image (16-Bit, Real/Virtual-8086 Modes) . 351 Figure 11-8.FXSAVE and FXRSTOR Image (Long Mode) . . . . . . . . . . . . . 353 Figure 11-9.FXSAVE and FXRSTOR Image (Legacy Mode) . . . . . . . . . . . 354 Figure 12-1.Task-Management Resources . . . . . . . . . . . . . . . . . . . . . . . . . . 361 Figure 12-2.Task-Segment Selector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
xvi
Figures
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Figure 12-3.TR Format, Legacy Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364 Figure 12-4.TR Format, Long Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364 Figure 12-5.Relationship between the TSS and GDT . . . . . . . . . . . . . . . . . 365 Figure 12-6.Legacy 32-bit TSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 Figure 12-7.I/O-Permission Bitmap Example . . . . . . . . . . . . . . . . . . . . . . . . 370 Figure 12-8.Long Mode TSS Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372 Figure 12-9.Task-Gate Descriptor, Legacy Mode Only . . . . . . . . . . . . . . . . 373 Figure 12-10.Privilege-Check Examples for Task Gates . . . . . . . . . . . . . . . 379 Figure 13-1.Address-Breakpoint Registers (DR0-DR3) . . . . . . . . . . . . . . . 386 Figure 13-2.Debug-Status Register (DR6) . . . . . . . . . . . . . . . . . . . . . . . . . . 387 Figure 13-3.Debug-Control Register (DR7) . . . . . . . . . . . . . . . . . . . . . . . . . 388 Figure 13-4.Debug-Control MSR (DebugCtlMSR). . . . . . . . . . . . . . . . . . . . 391 Figure 13-5.Control-Transfer Recording MSRs . . . . . . . . . . . . . . . . . . . . . . 393 Figure 13-6.Performance Counter (PerfCtrn) . . . . . . . . . . . . . . . . . . . . . . . 404 Figure 13-7.Performance Event-Select Register (PerfEvtSeln) . . . . . . . . 406 Figure 13-8.Time-Stamp Counter (TSC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
Figures
xvii
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
xviii
Figures
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Tables
Table 1-1. Table 1-2. Table 2-1. Table 2-2. Table 2-3. Table 2-4. Table 2-5. Table 2-6. Table 4-1. Table 4-2. Table 4-3. Table 4-4. Table 4-5. Table 4-6. Table 4-7. Table 5-1. Table 5-2. Table 5-3. Table 6-1. Table 7-1. Table 7-2. Table 7-3. Table 7-4. Table 7-5. Table 7-6. Table 7-7. Table 7-8. Table 7-9. Table 8-1. Table 8-2. Table 8-3. Table 8-4. Operating Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Interrupts and Exceptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Instructions That Reference RSP . . . . . . . . . . . . . . . . . . . . . . . . 40 64-Bit Mode Near Branches, Default 64-Bit Operand Size . . . . 41 Invalid Instructions in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . 44 Invalid Instructions in Long Mode . . . . . . . . . . . . . . . . . . . . . . . 45 Reassigned Instructions in 64-Bit Mode. . . . . . . . . . . . . . . . . . . 45 Differences Between Long Mode and Legacy Mode. . . . . . . . . 51 Segment Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 Descriptor Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 Code-Segment Descriptor Types. . . . . . . . . . . . . . . . . . . . . . . . 102 Data-Segment Descriptor Types . . . . . . . . . . . . . . . . . . . . . . . . 104 System-Segment Descriptor Types (S=0)--Legacy Mode . . . 105 System-Segment Descriptor Types--Long Mode . . . . . . . . . . 112 Descriptor-Entry Field Changes in Long Mode. . . . . . . . . . . . 117 Supported Paging Alternatives (CR0.PG=1) . . . . . . . . . . . . . . 148 Physical-Page Protection, CR0.WP=0 . . . . . . . . . . . . . . . . . . . 177 Effect of CR0.WP=1 on Supervisor Page Access . . . . . . . . . . 178 System-Management Instructions . . . . . . . . . . . . . . . . . . . . . . 179 Memory Access by Memory Type . . . . . . . . . . . . . . . . . . . . . . . 205 Caching Policy by Memory Type . . . . . . . . . . . . . . . . . . . . . . . 205 AMD64 Architecture Cache-Operating Modes . . . . . . . . . . . . 213 MTRR Type Field Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . 219 Fixed-Range MTRR Address Ranges . . . . . . . . . . . . . . . . . . . . 222 Combined MTRR and Page-Level Memory Type with Unmodified PAT MSR. . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 PAT Type Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 PAT-Register PA-Field Indexing . . . . . . . . . . . . . . . . . . . . . . . 232 Combined Effect of MTRR and PAT Memory Types . . . . . . . 234 Interrupt-Vector Source and Cause . . . . . . . . . . . . . . . . . . . . . 248 Interrupt-Vector Classification. . . . . . . . . . . . . . . . . . . . . . . . . 249 Double-Fault Exception Conditions . . . . . . . . . . . . . . . . . . . . . 255 Invalid-TSS Exception Conditions . . . . . . . . . . . . . . . . . . . . . . 257
Table 7-10. Extended Fixed-Range MTRR Type Encodings . . . . . . . . . . . 237
Tables
xix
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 8-5. Table 8-6. Table 8-7. Table 8-8. Table 8-9.
Stack Exception Error Codes . . . . . . . . . . . . . . . . . . . . . . . . . . 259 General-Protection Exception Conditions . . . . . . . . . . . . . . . . 260 Data-Type Alignment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264 Simultaneous Interrupt Priorities . . . . . . . . . . . . . . . . . . . . . . 270 Simultaneous Floating-Point Exception Priorities . . . . . . . . . 272
Table 8-10. Virtual-8086 Mode Interrupt Mechanisms. . . . . . . . . . . . . . . . 285 Table 8-11. Effect of Instructions that Modify the IF Bit. . . . . . . . . . . . . . 301 Table 10-1. AMD64 Architecture SMM State-Save Area . . . . . . . . . . . . . . 325 Table 10-2. Legacy SMM State-Save Area (Not used by AMD64 Architecture) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328 Table 10-3. SMM Register Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . 332 Table 11-1. Deriving FSAVE Tag Field from FXSAVE Tag Field . . . . . . 357 Table 12-1. Effects of Task Nesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380 Table 13-1. Breakpoint-Setting Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 395 Table 13-2. Breakpoint Location by Condition . . . . . . . . . . . . . . . . . . . . . . 397 Table 13-3. Operating-System Mode and User Mode Bits . . . . . . . . . . . . . 406 Table 14-1. Initial Processor State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412 Table 14-2. Initial State of Segment-Register Attributes . . . . . . . . . . . . . 415 Table 14-3. x87 Floating-Point State Initialization . . . . . . . . . . . . . . . . . . . 418 Table 14-4. Processor Operating Modes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 424 Table 14-5. Long-Mode Consistency Checks . . . . . . . . . . . . . . . . . . . . . . . . 426 Table A-1. MSRs of the AMD64 Architecture . . . . . . . . . . . . . . . . . . . . . . 435 Table A-2. System-Software MSR Cross-Reference . . . . . . . . . . . . . . . . . 439 Table A-3. Memory-Typing MSR Cross-Reference . . . . . . . . . . . . . . . . . . 441 Table A-4. Machine-Check MSR Cross-Reference. . . . . . . . . . . . . . . . . . . 443 Table A-5. Software-Debug MSR Cross-Reference . . . . . . . . . . . . . . . . . . 444 Table A-6. Performance-Monitoring MSR Cross-Reference. . . . . . . . . . . 445
xx
Tables
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Revision History
Date September 2003 Revision 3.09 Clarified terms in section on FXSAVE/FXSTOR. Corrected several minor errors of omission. Documentation of CR0.NW bit has been corrected. Several register diagrams and figure labels have been corrected. Description of shared cache lines has been clarified in Section 7.2. Made numerous small grammatical changes and factual clarifications. Added Revision History. Description
April 2003
3.08
September 2002
3.07
Chapter :
xxi
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
xxii
Chapter :
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Preface
About This Book
This book is part of a multivolume work entitled the AMD64 Architecture Programmer's Manual. This table lists each volume and its order number.
Title Volume 1, Application Programming Volume 2, System Programming Volume 3, General-Purpose and System Instructions Volume 4, 128-Bit Media Instructions Volume 5, 64-Bit Media and x87 Floating-Point Instructions Order No. 24592 24593 24594 26568 26569
Audience
This volume (Volume 2) is intended for programmers writing operating systems, loaders, linkers, device drivers, or system utilities. It assumes an understanding of AMD64 architecture application-level programming as described in Volume 1. This volume describes the AMD64 architecture's resources and functions that are managed by system software, including operating-mode control, memory management, interrupts and exceptions, task and state-change management, systemmanagement mode (including power management), multiprocessor support, debugging, and processor initialization. Application-programming topics are described in Volume 1. Details about each instruction are described in volumes 3, 4, and 5.
Contact Information
To submit questions or comments concerning this document, c o n t a c t o u r t e ch n i c a l d o c u m e n t a t i o n s t a f f a t AMD64.Feedback@amd.com.
Preface
xxiii
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Organization
This volume begins with an overview of system programming and differences between the x86 and AMD64 architectures. This is followed by chapters that describe the following details of system programming: System Resources--The system registers and processor ID (CPUID) functions. Segmented Virtual Memory--The segmented-memory models supported by the architecture and their associated data structures and protection checks. Page Translation and Protection--The page-translation functions supported by the architecture and their associated data structures and protection checks. System-Management Instructions--The instructions used to manage system functions. Memory System--The memory-system hierarchy and its resources and protocols, including memory-characterization, caching, and buffering functions. Exceptions and Interrupts--Details about the types and causes of exceptions and interrupts, and the methods of transferring control during these events. Machine-Check Mechanism--The resources and functions that support detection and handling of machine-check errors. System-Management Mode--The resources and functions that support system-management mode (SMM), including power-management functions. 128-Bit, 64-Bit, and x87 Programming--The resources and functions that support use (by application software) and state-saving (by the operation system) of the 128-bit media, 64-bit media, and x87 floating-point instructions. Multiple-Processor Management--The features of the instruction set and the system resources and functions that support multiprocessing environments. Debug and Performance Resources--The system resources and functions that support software debugging and performance monitoring.
xxiv
Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Legacy Task Management--Support for the legacy hardware multitasking functions, including register resources and data structures. Processor Initialization and Long-Mode Activation--The methods by which system software initializes and changes operating modes. Mixing Code Across Operating Modes--Things to remember when running programs in different operating modes. There are appendices describing details of model-specific registers (MSRs) and machine-check implementations. Definitions assumed throughout this volume are listed below. The index at the end of this volume cross-references topics within the volume. For other topics relating to the AMD64 architecture, see the tables of contents and indexes of the other volumes.
Definitions
Some of the following definitions assume a knowledge of the legacy x86 architecture. See "Related Documents" on page xxxvi for descriptions of the legacy x86 architecture. Terms and Notation 1011b A binary value--in this example, a 4-bit value. F0EAh A hexadecimal value--in this example a 2-byte value. [1,2) A range that includes the left-most value (in this case, 1) but excludes the right-most value (in this case, 2). 7-4 A bit range, from bit 7 to 4, inclusive. The high-order bit is shown first. 128-bit media instructions Instructions that use the 128-bit XMM registers. These are a combination of the SSE and SSE2 instruction sets. 64-bit media instructions Instructions that use the 64-bit MMXTM registers. These are primarily a combination of MMX and 3DNow!TM instruction Preface xxv
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
sets, with some additional instructions from the SSE and SSE2 instruction sets. 16-bit mode Legacy mode or compatibility mode in which a 16-bit address size is active. See legacy mode and compatibility mode. 32-bit mode Legacy mode or compatibility mode in which a 32-bit address size is active. See legacy mode and compatibility mode. 64-bit mode A submode of long mode. In 64-bit mode, the default address size is 64 bits and new features, such as register extensions, are supported for system and application software. #GP(0) Notation indicating a general-protection exception (#GP) with error code of 0. absolute Said of a displacement that references the base of a code segment rather than an instruction pointer. Contrast with relative. biased exponent The sum of a floating-point value's exponent and a constant bias for a particular floating-point data type. The bias makes the range of the biased exponent always positive, which allows reciprocation without overflow. byte Eight bits. clear To write a bit value of 0. Compare set. compatibility mode A submode of long mode. In compatibility mode, the default address siz e is 32 bits, and legacy 16-bit and 32-bit applications run without modification.
xxvi
Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
commit To irreversibly write, in program order, an instruction's result to software-visible storage, such as a register (including flags), the data cache, an internal write buffer, or memory. CPL Current privilege level. CR0-CR4 A register range, from register CR0 through CR4, inclusive, with the low-order register first. CR0.PE = 1 Notation indicating that the PE bit of the CR0 register has a value of 1. direct Referencing a memory location whose address is included in the instruction's syntax as an immediate operand. The address may be an absolute or relative address. Compare indirect. dirty data Data held in the processor's caches or internal buffers that is more recent than the copy held in main memory. displacement A signed value that is added to the base of a segment (absolute addressing) or an instruction pointer (relative addressing). Same as offset. doubleword Two words, or four bytes, or 32 bits. double quadword Eight words, or 16 bytes, or 128 bits. Also called octword. DS:rSI The contents of a memory location whose segment address is in the DS register and whose offset relative to that segment is in the rSI register.
Preface
xxvii
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
EFER.LME = 0 Notation indicating that the LME bit of the EFER register has a value of 0. effective address size The address size for the current instruction after accounting for the default address size and any address-size override prefix. effective operand size Th e o p e ra n d s i z e for t h e c u r re n t i n s t r u c t i o n a f t e r accounting for the default operand size and any operandsize override prefix. element See vector. exception An abnormal condition that occurs as the result of executing an instruction. The processor's response to an exception depends on the type of the exception. For all exceptions except 128-bit media SIMD floating-point exceptions and x87 floating-point exceptions, control is transferred to the handler (or service routine) for that exception, as defined by the exception's vector. For floating-point exceptions defined by the IEEE 754 standard, there are both masked and unmasked responses. When unmasked, the exception handler is called, and when masked, a default response is provided instead of calling the handler. FF /0 Notation indicating that FF is the first byte of an opcode, and a subopcode in the ModR/M byte has a value of 0. flush An often ambiguous term meaning (1) writeback, if modified, and invalidate, as in "flush the cache line," or (2) invalidate, as in "flush the pipeline," or (3) change a value, as in "flush to zero." GDT Global descriptor table. IDT Interrupt descriptor table. xxviii Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
IGN Ignore. Field is ignored. indirect Referencing a memory location whose address is in a register or other memory location. The address may be an absolute or relative address. Compare direct. IRB The virtual-8086 mode interrupt-redirection bitmap. IST The long-mode interrupt-stack table. IVT The real-address mode interrupt-vector table. LDT Local descriptor table. legacy x86 The legacy x86 architecture. See "Related Documents" on page xxxvi for descriptions of the legacy x86 architecture. legacy mode An operating mode of the AMD64 architecture in which existing 16-bit and 32-bit applications and operating systems run without modification. A processor implementation of the AMD64 architecture can run in either long mode or legacy mode. Legacy mode has three submodes, real mode, protected mode, and virtual-8086 mode. long mode An operating mode unique to the AMD64 architecture. A processor implementation of the AMD64 architecture can run in either long mode or legacy mode. Long mode has two submodes, 64-bit mode and compatibility mode. lsb Least-significant bit. LSB Least-significant byte.
Preface
xxix
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
main memory Physical memory, such as RAM and ROM (but not cache memory) that is installed in a particular computer system. mask (1) A control bit that prevents the occurrence of a floatingpoint exception from invoking an exception-handling routine. (2) A field of bits used for a control purpose. MBZ Must be zero. If software attempts to set an MBZ bit to 1, a general-protection exception (#GP) occurs. memory Unless otherwise specified, main memory. ModRM A byte following an instruction opcode that specifies address calculation based on mode (Mod), register (R), and memory (M) variables. moffset A 16, 32, or 64-bit offset that specifies a memory operand directly, without using a ModRM or SIB byte. msb Most-significant bit. MSB Most-significant byte. multimedia instructions A combination of 128-bit media instructions and 64-bit media instructions. octword Same as double quadword. offset Same as displacement. overflow The condition in which a floating-point number is larger in magnitude than the largest, finite, positive or negative
xxx
Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
number that can be represented in the data-type format being used. packed See vector. PAE Physical-address extensions. physical memory Actual memory, consisting of main memory and cache. probe A check for an address in a processor's caches or internal buffers. External probes originate outside the processor, and internal probes originate within the processor. protected mode A submode of legacy mode. quadword Four words, or eight bytes, or 64 bits. RAZ Read as zero (0), regardless of what is written. real-address mode See real mode. real mode A short name for real-address mode, a submode of legacy mode. relative Referencing with a displacement (also called offset) from an instruction pointer rather than the base of a code segment. Contrast with absolute. REX An instruction prefix that specifies a 64-bit operand size and provides access to additional registers. RIP-relative addressing Addressing relative to the 64-bit RIP instruction pointer.
Preface
xxxi
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
set To write a bit value of 1. Compare clear. SIB A byte following an instruction opcode that specifies address calculation based on scale (S), index (I), and base (B). SIMD Single instruction, multiple data. See vector. SSE Streaming SIMD extensions instruction set. See 128-bit media instructions and 64-bit media instructions. SSE2 Extensions to the SSE instruction set. See 128-bit media instructions and 64-bit media instructions. sticky bit A bit that is set or cleared by hardware and that remains in that state until explicitly changed by software. TOP The x87 top-of-stack pointer. TPR Task-priority register (CR8). TSS Task-state segment. underflow The condition in which a floating-point number is smaller in magnitude than the smallest nonzero, positive or negative number that can be represented in the data-type format being used. vector (1) A set of integer or floating-point values, called elements, that are packed into a single operand. Most of the 128-bit and 64-bit media instructions use vectors as operands. Vectors are also called packed or SIMD (single-instruction multiple-data) operands.
xxxii
Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
(2) An index into an interrupt descriptor table (IDT), used to access exception handlers. Compare exception. virtual-8086 mode A submode of legacy mode. word Two bytes, or 16 bits. x86 See legacy x86. Registers In the following list of registers, the names are used to refer either to a given register or to the contents of that register: AH-DH The high 8-bit AH, BH, CH, and DH registers. Compare AL-DL. AL-DL The low 8-bit AL, BL, CL, and DL registers. Compare AH-DH. AL-r15B The low 8-bit AL, BL, CL, DL, SIL, DIL, BPL, SPL, and R8B-R15B registers, available in 64-bit mode. BP Base pointer register. CRn Control register number n. CS Code segment register. eAX-eSP The 16-bit AX, BX, CX, DX, DI, SI, BP, and SP registers or the 32-bit EAX, EBX, ECX, EDX, EDI, ESI, EBP, and ESP registers. Compare rAX-rSP. EBP Extended base pointer register. EFER Extended features enable register.
Preface
xxxiii
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
eFLAGS 16-bit or 32-bit flags register. Compare rFLAGS. EFLAGS 32-bit (extended) flags register. eIP 16-bit or 32-bit instruction-pointer register. Compare rIP. EIP 32-bit (extended) instruction-pointer register. FLAGS 16-bit flags register. GDTR Global descriptor table register. GPRs General-purpose registers. For the 16-bit data size, these are AX, BX, CX, DX, DI, SI, BP, and SP. For the 32-bit data size, these are EAX, EBX, ECX, EDX, EDI, ESI, EBP, and ESP. For the 64-bit data size, these include RAX, RBX, RCX, RDX, RDI, RSI, RBP, RSP, and R8-R15. IDTR Interrupt descriptor table register. IP 16-bit instruction-pointer register. LDTR Local descriptor table register. MSR Model-specific register. r8-r15 The 8-bit R8B-R15B registers, or the 16-bit R8W-R15W registers, or the 32-bit R8D-R15D registers, or the 64-bit R8-R15 registers. rAX-rSP The 16-bit AX, BX, CX, DX, DI, SI, BP, and SP registers, or the 32-bit EAX, EBX, ECX, EDX, EDI, ESI, EBP, and ESP registers, or the 64-bit RAX, RBX, RCX, RDX, RDI, RSI, xxxiv Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
RBP, and RSP registers. Replace the placeholder r with nothing for 16-bit size, "E" for 32-bit size, or "R" for 64-bit size. RAX 64-bit version of the EAX register. RAZ Read as zero (0), regardless of what is written. RBP 64-bit version of the EBP register. RBX 64-bit version of the EBX register. RCX 64-bit version of the ECX register. RDI 64-bit version of the EDI register. RDX 64-bit version of the EDX register. rFLAGS 16-bit, 32-bit, or 64-bit flags register. Compare RFLAGS. RFLAGS 64-bit flags register. Compare rFLAGS. rIP 16-bit, 32-bit, or 64-bit instruction-pointer register. Compare RIP. RIP 64-bit instruction-pointer register. RSI 64-bit version of the ESI register. RSP 64-bit version of the ESP register. SP Stack pointer register. Preface xxxv
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
SS Stack segment register. TPR Task priority register, a new register introduced in the AMD64 architecture to speed interrupt management. TR Task register. Endian Order The x86 and AMD64 architectures address memory using littleendian byte-ordering. Multibyte values are stored with their least-significant byte at the lowest byte address, and they are illustrated with their least significant byte at the right side. Strings are illustrated in reverse order, because the addresses of their bytes increase from right to left.
Related Documents
Peter Abel, IBM PC Assembly Language and Programming, Prentice-Hall, Englewood Cliffs, NJ, 1995. Rakesh Agarwal, 80x86 Architecture & Programming: Volume II, Prentice-Hall, Englewood Cliffs, NJ, 1991. AMD data sheets and application notes for particular hardware implementations of the AMD64 architecture. AMD, AMD-K6TM MMXTM Enhanced Processor Multimedia Technology, Sunnyvale, CA, 2000. AMD, 3DNow!TM Technology Manual, Sunnyvale, CA, 2000. AMD, AMD Extensions to the 3DNow!TM and MMXTM Instruction Sets, Sunnyvale, CA, 2000. AMD, SYSCALL and SYSRET Instruction Specification Application Note, Sunnyvale, CA, 1998. Don Anderson and Tom Shanley, Pentium Processor System Architecture, Addison-Wesley, New York, 1995. Nabajyoti Barkakati and Randall Hyde, Microsoft Macro Assembler Bible, Sams, Carmel, Indiana, 1992. Barry B. Brey, 8086/8088, 80286, 80386, and 80486 Assembly Language Programming, Macmillan Publishing Co., New York, 1994.
xxxvi
Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Barry B. Brey, Programming the 80286, 80386, 80486, and Pentium Based Personal Computer, Prentice-Hall, Englewood Cliffs, NJ, 1995. Ralf Brown and Jim Kyle, PC Interrupts, Addison-Wesley, New York, 1994. Penn Brumm and Don Brumm, 80386/80486 Assembly Language Programming, Windcrest McGraw-Hill, 1993. Geoff Chappell, DOS Internals, Addison-Wesley, New York, 1994. Chips and Technologies, Inc. Super386 DX Programmer's Reference Manual, Chips and Technologies, Inc., San Jose, 1992. John Crawford and Patrick Gelsinger, Programming the 80386, Sybex, San Francisco, 1987. Cyrix Corporation, 5x86 Processor BIOS Writer's Guide, Cyrix Corporation, Richardson, TX, 1995. Cyrix Corporation, M1 Processor Data Book, Cyrix Corporation, Richardson, TX, 1996. Cyrix Corporation, MX Processor MMX Extension Opcode Table, Cyrix Corporation, Richardson, TX, 1996. Cyrix Corporation, MX Processor Data Book, Cyrix Corporation, Richardson, TX, 1997. Ray Duncan, Extending DOS: A Programmer's Guide to Protected-Mode DOS, Addison Wesley, NY, 1991. William B. Giles, Assembly Language Programming for the Intel 80xxx Family, Macmillan, New York, 1991. Frank van Gilluwe, The Undocumented PC, Addison-Wesley, New York, 1994. John L. Hennessy and David A. Patterson, Computer Architecture, Morgan Kaufmann Publishers, San Mateo, CA, 1996. Thom Hogan, The Programmer's PC Sourcebook, Microsoft Press, Redmond, WA, 1991. Hal Katircioglu, Inside the 486, Pentium, and Pentium Pro, Peer-to-Peer Communications, Menlo Park, CA, 1997. IBM Corporation, 486SLC Microprocessor Data Sheet, IBM Corporation, Essex Junction, VT, 1993. IBM Corporation, 486SLC2 Microprocessor Data Sheet, IBM Corporation, Essex Junction, VT, 1993. Preface xxxvii
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
IBM Corporation, 80486DX2 Processor Floating Point Instructions, IBM Corporation, Essex Junction, VT, 1995. IBM Corporation, 80486DX2 Processor BIOS Writer's Guide, IBM Corporation, Essex Junction, VT, 1995. IBM Corporation, Blue Lightening 486DX2 Data Book, IBM Corporation, Essex Junction, VT, 1994. Institute of Electrical and Electronics Engineers, IEEE Standard for Binary Floating-Point Arithmetic, ANSI/IEEE Std 754-1985. Institute of Electrical and Electronics Engineers, IEEE Standard for Radix-Independent Floating-Point Arithmetic, ANSI/IEEE Std 854-1987. Muhammad Ali Mazidi and Janice Gillispie Mazidi, 80X86 IBM PC and Compatible Computers, Prentice-Hall, Englewood Cliffs, NJ, 1997. Hans-Peter Messmer, The Indispensable Pentium Book, Addison-Wesley, New York, 1995. Karen Miller, An Assembly Language Introduction to Computer Architecture: Using the Intel Pentium, Oxford University Press, New York, 1999. Stephen Morse, Eric Isaacson, and Douglas Albert, The 80386/387 Architecture, John Wiley & Sons, New York, 1987. NexGen Inc., Nx586TM Processor Data Book, NexGen Inc., Milpitas, CA, 1993. NexGen Inc., Nx686TM Processor Data Book, NexGen Inc., Milpitas, CA, 1994. Bipin Patwardhan, Introduction to the Streaming SIMD Extensions in the Pentium(R) III, www.x86.org/articles/sse_pt1/ simd1.htm, June, 2000. Peter Norton, Peter Aitken, and Richard Wilton, PC Programmer's Bible, Microsoft Press, Redmond, WA, 1993. PharLap 386|ASM Reference Manual, Pharlap, Cambridge MA, 1993. PharLap TNT DOS-Extender Reference Manual, Pharlap, Cambridge MA, 1995. Sen-Cuo Ro and Sheau-Chuen Her, i386/i486 Advanced Programming, Van Nostrand Reinhold, New York, 1993. Jeffrey P. Royer, Introduction to Protected Mode Programming, course materials for an onsite class, 1992. xxxviii Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Tom Shanley, Protected Mode System Architecture, Addison Wesley, NY, 1996. SGS-Thomson Corporation, 80486DX Processor SMM Programming Manual, SGS-Thomson Corporation, 1995. Walter A. Triebel, The 80386DX Microprocessor, PrenticeHall, Englewood Cliffs, NJ, 1992. John Wharton, The Complete x86, MicroDesign Resources, Sebastopol, California, 1994. Web sites and newsgroups: - www.amd.com - news.comp.arch - news.comp.lang.asm.x86 - news.intel.microprocessors - news.microsoft
Preface
xxxix
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
xl
Preface
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
1
System-Programming Overview
Th i s e n t i re vo l u m e i s i n t e n d e d fo r s y s t e m - s o f t wa re developers--programmers writing operating systems, loaders, linkers, device drivers, or utilities that require access to system resources. These system resources are generally available only to software running at the highest-privilege level (CPL=0), also referred to as privileged software. Privilege levels and their interactions are fully described in "Segment-Protection Overview" on page 118. This chapter introduces the basic features and capabilities of the AMD64 architecture that are available to system-software developers. The concepts include: The supported address forms and how memory is organized. How memory-management hardware makes use of the various address forms to access memory. The processor operating modes, and how the memorymanagement hardware supports each of those modes. The system-control registers used to manage system resources. The interrupt and exception mechanism, and how it is used to interrupt program execution and to report errors. Additional, miscellaneous features available to system software, including support for hardware multitasking, reporting machine-check exceptions, debugging software problems, and optimizing software performance. Many of the legacy features and capabilities are enhanced by the AMD64 architecture to support 64-bit operating systems and applications, while providing backward-compatibility with existing software.
1.1
Memory Model
The AMD64 architecture memory model is designed to allow system software to manage application software and associated data in a secure fashion. The memory model is backwardcompatible with the legacy memory model. Hardwaretranslation mechanisms are provided to map addresses between virtual-memory space and physical-memory space. The
Chapter 1: System-Programming Overview
1
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
translation mechanisms allow system software to relocate applications and data transparently, either anywhere in physical-memory space, or in areas on the system hard drive managed by the operating system. In long mode, the AMD64 architecture implements a flatmemory model. In legacy mode, the architecture implements all legacy memory models. 1.1.1 Memory Addressing The AMD64 architecture supports address relocation. To do this, several types of addresses are needed to completely describe memory organization. Specifically, four types of addresses are defined by the AMD64 architecture: Logical addresses Effective addresses, or segment offsets, which are a portion of the logical address. Linear (virtual) addresses Physical addresses Logical Addresses. A l o g i c a l a d d r e s s i s a re f e re n c e i n t o a segmented-address space. It is comprised of the segment selector and the effective address. Notationally, a logical address is represented as
Logical Address = Segment Selector : Offset
The segment selector specifies an entry in either the global or local descriptor table. The specified descriptor-table entry describes the segment location in virtual-address space, its size, and other characteristics. The effective address is used as an offset into the segment specified by the selector. Logical addresses are often referred to as far pointers. Far pointers are used in software addressing when the segment reference must be explicit (i.e., a reference to a segment outside the current segment). Effective Addresses. The offset into a memory segment is referred to as an effective address (see "Segmentation" on page 6 for a description of segmented memory). Effective addresses are formed by adding together elements comprising a base value, a scaled-index value, and a displacement value. The effectiveaddress computation is represented by the equation
Effective Address = Base + (Scale x Index) + Displacement
2
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The elements of an effective-address computation are defined as follows: Base--A value stored in any general-purpose register. Scale--A positive value of 1, 2, 4, or 8. Index--A two's-complement value stored in any generalpurpose register. Displacement--An 8-bit, 16-bit, or 32-bit two's-complement value encoded as part of the instruction. Effective addresses are often referred to as near pointers. A near pointer is used when the segment selector is known implicitly or when the flat-memory model is used. Long mode defines a 64-bit effective-address length. If a processor implementation does not support the full 64-bit virtual-address space, the effective address must be in canonical form (see "Canonical Address Form" on page 5). Linear (Virtual) Addresses. The segment-selector portion of a logical address specifies a segment-descriptor entry in either the global or local descriptor table. The specified segmentdescriptor entry contains the segment-base address, which is the starting location of the segment in linear-address space. A linear address is formed by adding the segment-base address to the effective address (segment offset), which creates a reference to any byte location within the supported linearaddress space. Linear addresses are often referred to as virtual addresses, and both terms are used interchangeably throughout this document.
Linear Address = Segment Base Address + Effective Address
When the flat-memory model is used--as in 64-bit mode--a segment-base address is treated as 0. In this case, the linear address is identical to the effective address. In long mode, linear addresses must be in canonical address form, as described in "Canonical Address Form" on page 5. Physical Addresses. A physical address is a reference into the physical-address space, typically main memory. Physical addresses are translated from virtual addresses using paget ra n s l a t i o n m e ch a n i s m s . S e e " Pa g i n g " o n p a g e 8 fo r information on how the paging mechanism is used for virtualaddress to physical-address translation. When the paging
Chapter 1: System-Programming Overview
3
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
mechanism is not enabled, the virtual (linear) address is used as the physical address. 1.1.2 Memory Organization The AMD64 architecture organizes memory into virtual memory and physical memory. Virtual-memory and physical-memory spaces can be (and usually are) different in size. Generally, the virtual-address space is much larger than physical-address memory. System software relocates applications and data between physical memory and the system hard disk to make it appear that much more memory is available than really exists. System software then uses the hardware memory-management mechanisms to map the larger virtual-address space into the smaller physical-address space. Virtual Memory. Software uses virtual addresses to access locations within the virtual-memory space. System software is responsible for managing the relocation of applications and d a t a i n v i r t u a l - m e m o ry s pa c e u s i n g s e g m e n t - m e m o ry management. System software is also responsible for mapping virtual memory to physical memory through the use of page translation. The AMD64 architecture supports different virtualmemory sizes using the following address-translation modes: Protected Mode--This mode supports 4 gigabytes of virtualaddress space using 32-bit virtual addresses. Long Mode--This mode supports 16 exabytes of virtualaddress space using 64-bit virtual addresses. Physical Memory. Physical addresses are used to directly access main memory. For a particular computer system, the size of the available physical-address space is equal to the amount of main memory installed in the system. The maximum amount of physical memory accessible depends on the processor implementation and on the address-translation mode. The AMD64 architecture supports varying physical-memory sizes using the following address-translation modes: Real-Address Mode--This mode, also called real mode, supports 1 megabyte of physical-address space using 20-bit physical addresses. This address-translation mode is described in "Real Addressing" on page 11. Real mode is available only from legacy mode (see "Legacy Modes" on page 16). Legacy Protected Mode--This mode supports several different address-space sizes, depending on the translation 4 Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
mechanism used and whether extensions to those mechanisms are enabled. Legacy protected mode supports 4 gigabytes of physicaladdress space using 32-bit physical addresses. Both segment translation (see "Segmentation" on page 6) and page translation (see "Paging" on page 8) can be used to access the physical address space, when the processor is running in legacy protected mode. When the physical-address size extensions are enabled (see "Physical-Address Extensions (PAE) Bit" on page 149), the page-translation mechanism can be extended to support 52bit physical addresses. 52-bit physical addresses allow up to 4 petabytes of physical-address space to be supported. (Currently, the AMD64 architecture supports 40-bit addresses in this mode, allowing up to 1 terabyte of physicaladdress space to be supported. Long Mode--This mode is unique to the AMD64 architecture. This mode supports up to 4 petabytes of physical-address space using 52-bit physical addresses. Long mode requires the use of page-translation and the physical-address size extensions (PAE). 1.1.3 Canonical Address Form Long mode defines 64 bits of virtual-address space, but processor implementations can support less. Although some processor implementations do not use all 64 bits of the virtual address, they all check bits 63 through the most-significant implemented bit to see if those bits are all zeros or all ones. An address that complies with this property is in canonical address form. In most cases, a virtual-memory reference that is not in canonical form causes a general-protection exception (#GP) to occur. However, implied stack references where the stack address is not in canonical form causes a stack exception (#SS) to occur. Implied stack references include all push and pop instructions, and any instruction using RSP or RBP as a base register. By checking canonical-address form, the AMD64 architecture prevents software from exploiting unused high bits of pointers for other purposes. Software complying with canonical-address form on a specific processor implementation can run unchanged on long-mode implementations supporting larger virtual-address spaces.
Chapter 1: System-Programming Overview
5
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
1.2
Memory Management
Memory management consists of the methods by which a d d re s s e s g e n e ra t e d by s o f t wa re a re t ra n s l a t e d by segmentation and/or paging into addresses in physical memory. Memory management is not visible to application software. It is handled by the system software and processor hardware.
1.2.1 Segmentation
Segmentation was originally created as a method by which system software could isolate software processes (tasks), and the data used by those processes, from one another in an effort to increase the reliability of systems running multiple processes simultaneously. The AMD64 architecture is designed to support all forms of legacy segmentation. However, most modern system software does not use the segmentation features available in the legacy x86 architecture. Instead, system software typically handles program and data isolation using page-level protection. For this reason, the AMD64 architecture dispenses with multiple segments in 64-bit mode and, instead, uses a flat-memory model. The elimination of segmentation allows new 64-bit system software to be coded more simply, and it supports more efficient management of multi-processing than is possible in the legacy x86 architecture. Segmentation is, however, used in compatibility mode and legacy mode. Here, segmentation is a form of base memoryaddressing that allows software and data to be relocated in virtual-address space off of an arbitrary base address. Software and data can be relocated in virtual-address space using one or more variable-siz ed memory segments. The legacy x86 architecture provides several methods of restricting access to segments from other segments so that software and data can be protected from interfering with each other. In compatibility and legacy modes, up to 16,383 unique segments can be defined. The base-address value, segment size (called a limit), protection, and other attributes for each segment are contained in a data structure called a segment descriptor. Collections of segment descriptors are held in descriptor tables. Specific segment descriptors are referenced or selected from the descriptor table using a segment selector register. Six segment-selector registers are available, providing access to as many as six segments at a time.
6
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Fi g u re 1 -1 s h ow s a n ex a m p l e o f s e g m e n t e d m e m o ry. Segmentation is described in Chapter 4, "Segmented Virtual Memory."
Virtual Address Space
Effective Address
Descriptor Table Selectors
CS DS ES FS GS SS Base Limit Segment Base Limit Virtual Address
Segment
513-201.eps
Figure 1-1.
Segmented-Memory Model Flat Segmentation. One special case of segmented memory is the flat-memory model. In the legacy flat-memory model, all segment-base addresses have a value of 0, and the segment limits are fixed at 4 Gbytes. Segmentation cannot be disabled but use of the flat-memory model effectively disables segment translation. The result is a virtual address that equals the effective address. Figure 1-2 on page 8 shows an example of the flat-memory model.
Chapter 1: System-Programming Overview
7
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Software running in 64-bit mode automatically uses the flatmemory model. In 64-bit mode, the segment base is treated as if it were 0, and the segment limit is ignored. This allows an effective addresses to access the full virtual-address space supported by the processor.
Virtual Address Space
Effective Address
Virtual Address
Flat Segment
513-202.eps
Figure 1-2. Flat Memory Model 1.2.2 Paging Paging allows software and data to be relocated in physicaladdress space using fixed-size blocks called physical pages. The legacy x86 architecture supports three different physical-page sizes of 4 Kbytes, 2 Mbytes, and 4 Mbytes. As with segment translation, access to physical pages by lesser-privileged software can be restricted. Page translation uses a hierarchical data structure called a page-translation table to translate virtual pages into physicalpages. The number of levels in the translation-table hierarchy can be as few as one or as many as four, depending on the physical-page size and processor operating mode. Translation tables are aligned on 4-Kbyte boundaries. Physical pages must be aligned on 4-Kbyte, 2-Mbyte, or 4-Mbyte boundaries, depending on the physical-page size. 8 Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Each table in the translation hierarchy is indexed by a portion of the virtual-address bits. The entry referenced by the table index contains a pointer to the base address of the next-lowerlevel table in the translation hierarchy. In the case of the lowestlevel table, its entry points to the physical-page base address. The physical page is then indexed by the least-significant bits of the virtual address to yield the physical address. Figure 1-3 shows an example of paged memory with three levels in the translation-table hierarchy. Paging is described in Chapter 5, "Page Translation and Protection."
Virtual Address
Physical Address Space
Physical Address Table 1 Table 2 Table 3
Page Translation Tables
Physical Page
Page Table Base Address
513-203.eps
Figure 1-3. Paged Memory Model Software running in long mode is required to have page translation enabled. Chapter 1: System-Programming Overview 9
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
1.2.3 Mixing Segmentation and Paging
Memory-management software can combine the use of segmented memory and paged memory. Because segmentation cannot be disabled, paged-memory management requires some minimum initialization of the segmentation resources. Paging c a n b e c o m p l e t e l y d i s a b l e d , s o s e g m e n t e d - m e m o ry management does not require initialization of the paging resources. Segments can range in size from a single byte to 4 Gbytes in length. It is therefore possible to map multiple segments to a single physical page and to map multiple physical pages to a single segment. Alignment between segment and physical-page boundaries is not required, but memory-management software is simplified when segment and physical-page boundaries are aligned. The simplest, most efficient method of memory management is the flat-memory model. In the flat-memory model, all segment base addresses have a value of 0 and the segment limits are fixed at 4 Gbytes. The segmentation mechanism is still used each time a memory reference is made, but because virtual addresses are identical to effective addresses in this model, the segmentation mechanism is effectively ignored. Translation of virtual (or effective) addresses to physical addresses takes place using the paging mechanism only. Because 64-bit mode disables segmentation, it uses a flat, paged-memory model for memory management. The 4 Gbyte segment limit is ignored in 64-bit mode. Figure 1-4 on page 11 shows an example of this model.
10
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Virtual Address Space
Physical Address Space
Effective Address
Virtual Address
Physical Address
Page Translation Tables
Page Frame
Flat Segment Page Table Base Address
513-204.eps
Figure 1-4. 64-Bit Flat, Paged-Memory Model 1.2.4 Real Addressing Real addressing is a legacy-mode form of address translation used in real mode. This simplified form of address translation is backward compatible with 8086-processor effective-to-physical address translation. In this mode, 16-bit effective addresses are mapped to 20-bit physical addresses, providing a 1-Mbyte physical-address space. Segment selectors are used in real-address translation, but not as an index into a descriptor table. Instead, the 16-bit segmentselector value is shifted left by 4 bits to form a 20-bit segmentbase address. The 16-bit effective address is added to this 20-bit segment base address to yield a 20-bit physical address. If the sum of the segment base and effective address carries over into bit 20, that bit can be optionally truncated to mimic the 20-bit
Chapter 1: System-Programming Overview
11
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
address wrapping of the 8086 processor by using the A20M# input signal to mask the A20 address bit. Real-address translation supports a 1-Mbyte physical-address space using up to 64K segments aligned on 16-byte boundaries. Each segment is exactly 64K bytes long. Figure 1-5 shows an example of real-address translation.
Selectors
CS DS ES
15 0
FS GS SS
Effective Address
19
0
19
0
0000
Effective Address
Selector
0000
+
19
0
Physical Address
513-205.eps
Figure 1-5. Real-Address Memory Model
1.3
Operating Modes
The legacy x86 architecture provides four operating modes or e nv i ro n m e n t s t h a t s u p p o r t va ry i n g fo r m s o f m e m o ry management, virtual-memory and physical-memory sizes, and protection: Real Mode. Protected Mode. Virtual-8086 Mode. System Management Mode.
12
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The AMD64 architecture supports all these legacy modes, and it adds a new operating mode called long mode. Table 1-1 shows the differences between long mode and legacy mode. Software can move between all supported operating modes as shown in Figure 1-6 on page 14. Each operating mode is described in the following sections. Table 1-1. Operating Modes
System Software Required Application Recompile Required Defaults1 Address Size (bits) 64 32 16 32 Legacy 32bit OS 16 no 16 Legacy 16bit OS 16 16 32 16 no 32 Operand Size (bits) Register Extensions2 Maximum GPR Width (bits) 64
Mode
64-Bit Mode Long Mode3 Compatibility Mode Protected Mode Legacy Mode Virtual-8086 Mode Real Mode
Note:
yes New 64-bit OS no
32
yes
no
32
32
1. Defaults can be overridden in most modes using an instruction prefix or system control bit. 2. Register extensions includes eight new GPRs and eight new XMM registers (also called SSE registers). 3. Long mode supports only x86 protected mode. It does not support x86 real mode or virtual-8086 mode.
Chapter 1: System-Programming Overview
13
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Long Mode
64-bit Mode CS.L=1 CS.L=0 CS.L=0 EFER.LME=1, CR4.PAE=1 then CR0.PG=1 RSM CR0.PG=0 then EFER.LME=0 RSM SMI# EFLAGS.VM=0 Protected Mode EFLAGS.VM=1 Reset Virtual 8086 Mode SMI# SMI# Compatibility Mode RSM
SMI#
RSM CR0.PE=1 Reset CR0.PE=0 Reset
System Management Mode Reset
Real Mode
SMI# RSM
513-206.eps
Figure 1-6. Operating Modes of the AMD64 Architecture 1.3.1 Long Mode Long mode consists of two submodes: 64-bit mode and compatibility mode. 64-bit mode supports several new features, including the ability to address 64-bit virtual-address space. Compatibility mode provides binary compatibility with existing 16-bit and 32-bit applications when running on 64-bit system software. Throughout this document, references to long mode refer collectively to both 64-bit mode and compatibility mode. If a function is specific to either 64-bit mode or compatibility mode, then those specific names are used instead of the name long mode. 14 Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Before enabling and activating long mode, system software must first enable protected mode. The process of enabling and activating long mode is described in Chapter 14, "Processor Initialization and Long-Mode Activation." Long mode features are described throughout this document, where applicable. 1.3.2 64-Bit Mode 64-bit mode, a submode of long mode, provides support for 64bit system software and applications by adding the following new features: 64-bit virtual addresses (processor implementations can have fewer). Register extensions through a new instruction prefix (REX): - Adds eight GPRs (R8-R15). - Widens GPRs to 64 bits. - Adds eight 128-bit streaming SIMD extension (SSE) registers (XMM8-XMM15). 64-bit instruction pointer (RIP). New RIP-relative data-addressing mode. Flat-segment address space with single code, data, and stack space. The mode is enabled by the system software on an individual code-segment basis. Although code segments are used to enable and disable 64-bit mode, the legacy segmentation mechanism is largely disabled. Page translation is required for memory management purposes. Because 64-bit mode supports a 64-bit virtual-address space, it requires 64-bit system software and development tools. In 64-bit mode, the default address size is 64 bits, and the default operand size is 32 bits. The defaults can be overridden on an instruction-by-instruction basis using instruction prefixes. A new REX prefix is introduced for specifying a 64-bit operand size and the new registers. 1.3.3 Compatibility Mode Compatibility mode, a submode of long mode, allows system software to implement binary compatibility with existing 16-bit and 32-bit x86 applications. It allows these applications to run, without recompilation, under 64-bit system software in long mode, as shown in Table 1-1 on page 13. In compatibility mode, applications can only access the first 4 Gbytes of virtual-address space. Standard x86 instruction Chapter 1: System-Programming Overview 15
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
prefixes toggle between 16-bit and 32-bit address and operand sizes. Compatibility mode, like 64-bit mode, is enabled by system software on an individual code-segment basis. Unlike 64-bit mode, however, segmentation functions the same as in the legacy-x86 architecture, using 16-bit or 32-bit protected-mode semantics. From an application viewpoint, compatibility mode looks like a legacy protected-mode environment. From a system-software viewpoint, the long-mode mechanisms are used for address translation, interrupt and exception handling, and system data-structures. 1.3.4 Legacy Modes Legacy mode consists of three submodes: real mode, protected mode, and virtual-8086 mode. Protected mode can be either paged or unpaged. Legacy mode preserves binary compatibility not only with existing x86 16-bit and 32-bit applications but also with existing x86 16-bit and 32-bit system software. Real Mode. In this mode, also called real-address mode, the processor supports a physical-memory space of 1 Mbyte and operand sizes of 16 bits (default) or 32 bits (with instruction prefixes). Interrupt handling and address generation are nearly identical to the 80286 processor's real mode. Paging is not supported. All software runs at privilege level 0. Real mode is entered after reset or processor power-up. The mode is not supported when the processor is operating in long mode because long mode requires that paged protected mode be enabled. Protected Mode. In this mode, the processor supports virtualmemory and physical-memory spaces of 4 Gbytes and operand sizes of 16 or 32 bits. All segment translation, segment protection, and hardware multitasking functions are available. System software can use segmentation to relocate effective addresses in virtual-address space. If paging is not enabled, virtual addresses are equal to physical addresses. Paging can be optionally enabled to allow translation of virtual addresses to physical addresses and to use the page-based memoryprotection mechanisms. In protected mode, software runs at privilege levels 0, 1, 2, or 3. Typically, application software runs at privilege level 3, the system software runs at privilege levels 0 and 1, and privilege
16
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
level 2 is available to system software for other uses. The 16-bit version of this mode was first introduced in the 80286 processor. Virtual-8086 Mode. Virtual-8086 mode allows system software to run 16-bit real-mode software on a virtualized-8086 processor. In this mode, software written for the 8086, 8088, 80186, or 80188 processor can run as a privilege-level-3 task under protected mode. The processor supports a virtual-memory space of 1 Mbytes and operand sizes of 16 bits (default) or 32 bits (with instruction prefixes), and it uses real-mode address translation. Virtual-8086 mode is enabled by setting the virtual-machine bit in the EFLAGS register (EFLAGS.VM). EFLAGS.VM can only be set or cleared when the EFLAGS register is loaded from the TSS as a result of a task switch, or by executing an IRET instruction from privileged software. The POPF instruction cannot be used to set or clear the EFLAGS.VM bit. Virtual-8086 mode is not supported when the processor is operating in long mode. When long mode is enabled, any attempt to enable virtual-8086 mode is silently ignored. 1.3.5 System Management Mode (SMM) System management mode (SMM) is an operating mode designed for system-control activities that are typically t ra n s p a re n t t o c o nve n t i o n a l s y s t e m s o f t wa re . Powe r management is one popular use for system management mode. SMM is primarily targeted for use by the basic input-output system (BIOS) and specialized low-level device drivers. The code and data for SMM are stored in the SMM memory area, which is isolated from main memory by the SMM output signal. SMM is entered by way of a system management interrupt (SMI). Upon recognizing an SMI, the processor enters SMM and switches to a separate address space where the SMM handler is located and executes. In SMM, the processor supports realmode addressing with 4 Gbyte segment limits and default operand, address, and stack sizes of 16 bits (prefixes can be used to override these defaults).
1.4
System Registers
Figure 1-7 on page 19 shows the system registers defined for the AMD64 architecture. System software uses these registers to, a m o n g o t h e r t h i n g s , ma n a g e t h e p ro c e s s o r o p e ra t in g
Chapter 1: System-Programming Overview
17
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
environment, define system resource characteristics, and to monitor software execution. With the exception of the RFLAGS register, system registers can be read and written only from privileged software. Except for the descriptor-table registers and task register, the AMD64 architecture defines all system registers to be 64 bits wide. The descriptor table and task registers are defined by the AMD64 architecture to include 64-bit base-address fields, in addition to their other fields. As shown in Figure 1-7 on page 19, the system registers include: Control Registers--These registers are used to control system operation and some system features. See "System-Control Registers" on page 53 for details. System-Flags Register--The RFLAGS register contains system-status flags and masks. It is also used to enable virtual-8086 mode and to control application access to I/O devices and interrupts. See "RFLAGS Register" on page 64 for details. Descriptor-Table Registers--These registers contain the location and size of descriptor tables stored in memory. Descriptor tables hold segmentation data structures used in protected mode. See "Descriptor Tables" on page 89 for details. Task Register--The task register contains the location and size in memory of the task-state segment. The hardwaremultitasking mechanism uses the task-state segment to hold state information for a given task. The TSS also holds other data, such as the inner-level stack pointers used when changing to a higher privilege level. See "Task Register" on page 363 for details. Debug Registers--Debug registers are used to control the software-debug mechanism, and to report information back to a debug utility or application. See "Debug Registers" on page 385 for details.
18
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Control Registers CR0 CR2 CR3 CR4 CR8
Extended-Feature-Enable Register EFER System-Configuration Register SYSCFG System-Linkage Registers
Memory-Typing Registers MTRRcap MTRRdefType MTRRphysBasen MTRRphysMaskn MTRRfixn PAT TOP_MEM TOP_MEM2 Performance-Monitoring Registers TSC PerfEvtSeln PerfCtrn Machine-Check Registers MCG_CAP MCG_STAT MCG_CTL MCi_CTL MCi_STATUS MCi_ADDR MCi_MISC
System-Flags Register RFLAGS
STAR LSTAR CSTAR SFMASK FS.base GS.base KernelGSbase SYSENTER_CS SYSENTER_ESP SYSENTER_EIP Debug-Extension Registers
Debug Registers DR0 DR1 DR2 DR3 DR6 DR7
Descriptor-Table Registers GDTR IDTR LDTR
DebugCtlMSR LastBranchFromIP LastBranchToIP LastIntFromIP LastIntToIP
Task Register TR
Model-Specific Registers
513-260.eps
Figure 1-7. System Registers Also defined as system registers are a number of model-specific registers included in the AMD64 architectural definition, and shown in Figure 1-7: Extended-Feature-Enable Register--The EFER register is used to enable and report status on special features not controlled by the CRn control registers. In particular, EFER Chapter 1: System-Programming Overview 19
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
is used to control activation of long mode. See "Extended Feature Enable Register (EFER)" on page 69 for more information. System-Configuration Register--The SYSCFG register is used to enable and configure system-bus features. See "System Configuration Register (SYSCFG)" on page 73 for more information. System-Linkage Registers--These registers are used by system-linkage instructions to specify operating-system entry points, stack locations, and pointers into system-data structures. See "Fast System Call and Return" on page 181 for details. Memory-Typing Registers--Memory-typing registers can be used to characterize (type) system memory. Typing memory gives system software control over how instructions and data are cached, and how memory reads and writes are ordered. See "MTRRs" on page 219 for details. Debug-Extension Registers--These registers control additional software-debug reporting features. See "Debug Registers" on page 385 for details. Performance-Monitoring Registers--Performance-monitoring registers are used to count processor and system events, or the duration of events. See "Performance Optimization" on page 403 for more information. Machine-Check Registers--The machine-check registers control the response of the processor to non-recoverable failures. They are also used to report information on such failures back to system utilities designed to respond to such failures. See "Machine Check MSRs" on page 307 for more information.
1.5
System-Data Structures
Figure 1-8 on page 21 shows the system-data structures defined for the AMD64 architecture. System-data structures are created and maintained by system software for use by the processor when running in protected mode. A processor running in protected mode uses these data structures to manage memory and protection, and to store program-state information when an interrupt or task switch occurs.
20
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Segment Descriptors (Contained in Descriptor Tables) Code Stack Data Gate Task-State Segment Local-Descriptor Table
Task-State Segment
Descriptor Tables Global-Descriptor Table
Descriptor Descriptor
Interrupt-Descriptor Table
Gate Descriptor Gate Descriptor
Local-Descriptor Table
Descriptor Descriptor
...
Descriptor
...
Gate Descriptor
...
Descriptor
Page-Translation Tables Page-Map Level-4 Page-Directory Pointer Page Directory Page Table
513-261.eps
Figure 1-8. System-Data Structures As shown in Figure 1-8, the system-data structures include: Descriptors--A descriptor provides information about a segment to the processor, such as its location, size and privilege level. A special type of descriptor, called a gate, is used to provide a code selector and entry point for a software routine. Any number of descriptors can be defined, but system software must at a minimum create a descriptor for the currently executing code segment and stack segment. See "Legacy Segment Descriptors" on page 97, and "Long-
Chapter 1: System-Programming Overview
21
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Mode Segment Descriptors" on page 108 for complete information on descriptors. Descriptor Tables--As the name implies, descriptor tables hold descriptors. The global-descriptor table holds descriptors available to all programs, while a localdescriptor table holds descriptors used by a single program. The interrupt-descriptor table holds only gate descriptors used by interrupt handlers. System software must initialize the global-descriptor and interrupt-descriptor tables, while use of the local-descriptor table is optional. See "Descriptor Tables" on page 89 for more information. Task-State Segment--The task-state segment is a special segment for holding processor-state information for a specific program, or task. It also contains the stack pointers used when switching to more-privileged programs. The hardware multitasking mechanism uses the state information in the segment when suspending and resuming a task. Calls and interrupts that switch stacks cause the stack pointers to be read from the task-state segment. System software must create at least one task-state segment, even if hardware multitasking is not used. See "Legacy TaskState Segment" on page 365, and "64-Bit Task State Segment" on page 370 for details. Page-Translation Tables--Use of page translation is optional in protected mode, but it is required in long mode. A fourlevel page-translation data structure is provided to allow long-mode operating systems to translate a 64-bit virtualaddress space into a 52-bit physical-address space. Legacy protected mode can use two- or three-level page-translation data structures. See "Page Translation Overview" on page 146 for more information on page translation.
1.6
Interrupts
The AMD64 architecture provides a mechanism for the processor to automatically suspend (interrupt) software execution and transfer control to an interrupt handler when an interrupt or exception occurs. An interrupt handler is privileged software designed to identify and respond to the cause of an interrupt or exception, and return control back to the interrupted software. Interrupts can be caused when system hardware signals an interrupt condition using one of the external-interrupt signals on the processor. Interrupts can also
22
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
be caused by software that executes an interrupt instruction. Exceptions occur when the processor detects an abnormal condition as a result of executing an instruction. The term "interrupts" as used throughout this volume includes both interrupts and exceptions when the distinction is unnecessary. System software not only sets up the interrupt handlers, but it must also create and initialize the data structures the processor uses to execute an interrupt handler when an interrupt occurs. The data structures include the code-segment descriptors for the interrupt-handler software and any data-segment descriptors for data and stack accesses. Interrupt-gate descriptors must also be supplied. Interrupt gates point to interrupt-handler code-segment descriptors, and the entry point in an interrupt handler. Interrupt gates are stored in the interrupt-descriptor table. The code-segment and data-segment descriptors are stored in the global-descriptor table and, optionally, the local-descriptor table. When an interrupt occurs, the processor uses the interrupt vector to find the appropriate interrupt gate in the interruptdescriptor table. The gate points to the interrupt-handler code segment and entry point, and the processor transfers control to that location. Before invoking the interrupt handler, the processor saves information required to return to the interrupted program. For details on how the processor transfers control to interrupt handlers, see "Legacy Protected-Mode Interrupt Control Transfers" on page 276, and "Long-Mode Interrupt Control Transfers" on page 287. Table 1-2 on page 24 shows the supported interrupts and exceptions, ordered by their vector number. Refer to "Vectors" on page 247 for a complete description of each interrupt, and a description of the interrupt mechanism.
Chapter 1: System-Programming Overview
23
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 1-2.
Vector 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 0-255 Any
Interrupts and Exceptions
Description Integer Divide-by-Zero Exception Debug Exception Non-Maskable-Interrupt Breakpoint Exception (INT 3) Overflow Exception (INTO instruction) Bound-Range Exception (BOUND instruction) Invalid-Opcode Exception Device-Not-Available Exception Double-Fault Exception Coprocessor-Segment-Overrun Exception (reserved in AMD64) Invalid-TSS Exception Segment-Not-Present Exception Stack Exception General-Protection Exception Page-Fault Exception (Reserved) x87 Floating-Point Exception Alignment-Check Exception Machine-Check Exception SIMD Floating-Point Exception Interrupt Instructions Hardware Maskable Interrupts
1.7
Additional System-Programming Facilities
A task is any program that the processor can execute, suspend, and later resume executing at the point of suspension. During the time a task is suspended, other tasks are allowed to execute. Chapter 1: System-Programming Overview
1.7.1 Hardware Multitasking
24
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Each task has its own execution space, consisting of a code segment, data segments, and a stack segment for each privilege l eve l . Ta s k s c a n a l s o h ave t h e i r ow n v i r t u a l - m e m o ry environment managed by the page-translation mechanism. The state information defining this execution space is stored in the task-state segment (TSS) maintained for each task. S u p p o r t f o r h a rd wa re m u l t i t a s k i n g i s p r ov i d e d by implementations of the AMD64 architecture when software is running in legacy mode. Hardware multitasking provides automated mechanisms for switching tasks, saving the execution state of the suspended task, and restoring the ex e c u t i o n s t a t e o f t h e res u m e d t as k . Whe n h a rdwa re multitasking is used to switch tasks, the processor takes the following actions: The processor automatically suspends execution of the task, allowing any executing instructions to complete and save their results. The execution state of a task is saved in the task TSS. The execution state of a new task is loaded into the processor from its TSS. The processor begins executing the new task at the location specified in the new task TSS. Use of hardware-multitasking features is optional in legacy mode. Generally, modern operating systems do not use the hardware-multitasking features, and instead perform task management entirely in software. Long mode does not support hardware multitasking at all. Whether hardware multitasking is used or not, system software must create and initialize at least one task-state segment datastructure. This requirement holds for both long-mode and legacy-mode software. The single task-state segment holds critical pieces of the task execution environment and is referenced during certain control transfers. Detailed information on hardware multitasking is available in Chapter 12, "Task Management," along with a full description of the requirements that must be met in initializing a task-state segment when hardware multitasking is not used. 1.7.2 Machine Check Implementations of the AMD64 architecture support the machine-check exception. This exception is useful in system 25
Chapter 1: System-Programming Overview
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
applications with stringent requirements for reliability, availability, and serviceability. The exception allows specialized system-software utilities to report hardware errors that are generally severe and non-recoverable. Providing the capability to report such errors can allow complex system problems to be pinpointed rapidly. The machine-check exception is described in Chapter 9, "Machine Check Mechanism." Much of the error-reporting c a p ab i l i t i e s i s i m p l e m e n t a t i o n d e p e n d e n t . Fo r m o re information, developers of machine-check error-reporting software should also refer to the BIOS writer's guide for a specific implementation. 1.7.3 Software Debugging A software-debugging mechanism is provided in hardware to help software developers quickly isolate programming errors. This capability can be used to debug system software and application software alike. Only privileged software can access the debugging facilities. Generally, software-debug support is provided by a privileged application program rather than by the operating system itself. The facilities supported by the AMD64 architecture allow debugging software to perform the following: Set breakpoints on specific instructions within a program. Set breakpoints on an instruction-address match. Set breakpoints on a data-address match. Set breakpoints on specific I/O-port addresses. Set breakpoints to occur on task switches when hardware multitasking is used. Single step an application instruction-by-instruction. Single step only branches and interrupts. Record a history of branches and interrupts taken by a program. The debugging facilities are fully described in "Software-Debug Resources" on page 384. Some processors provide additional, implementation-specific debug support. For more information, re f e r t o t h e B I O S w r i t e r 's g u i d e fo r t h e s p e c i f i c implementation. 1.7.4 Performance Monitoring 26 For many software developers, the ability to identify and eliminate performance bottlenecks from a program is nearly as Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
i m p o r t a n t a s q u i ck ly i s o l a t i n g p rog ra m m i n g e r ro rs . Implementations of the AMD64 architecture provide hardware performance-monitoring resources that can be used by special software applications to identify such bottlenecks. Nonprivileged software can access the performance monitoring facilities, but only if privileged software grants that access. The performance-monitoring facilities allow the counting of events, or the duration of events. Performance-analysis software can use the data to calculate the frequency of certain events, or the time spent performing specific activities. That information can be used to suggest areas for improvement and the types of optimizations that are helpful. The performance-monitoring facilities are fully described in "Performance Optimization" on page 403. The specific events that can be monitored are generally implementation specific. For more information, refer to the BIOS writer's guide for the specific implementation.
Chapter 1: System-Programming Overview
27
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
28
Chapter 1: System-Programming Overview
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
2
x86 and AMD64 Architecture Differences
The AMD64 architecture is designed to provide full binary compatibility with all previous AMD implementations of the x86 architecture. This chapter summarizes the new features and architectural enhancements introduced by the AMD64 architecture, and compares those features and enhancements with previous AMD x86 processors. Most of the new capabilities introduced by the AMD64 architecture are available only in long mode (64-bit mode, compatibility mode, or both). However, some of the new capabilities are also available in legacy mode, and are mentioned where appropriate. The material throughout this chapter assumes the reader has a solid understanding of the x86 architecture. For those who are unfamiliar with the x86 architecture, please read the remainder of this volume before reading this chapter.
2.1
Operating Modes
See "Operating Modes" on page 12 for a complete description of the operating modes supported by the AMD64 architecture.
2.1.1 Long Mode
The AMD64 architecture introduces long mode and its two submodes: 64-bit mode and compatibility mode. 64-Bit Mode. 64-bit mode provides full support for 64-bit system software and applications. The new features introduced in support of 64-bit mode are summarized throughout this chapter. To use 64-bit mode, a 64-bit operating system and tool chain are required. Compatibility Mode. Compatibility mode allows 64-bit operating systems to implement binary compatibility with existing 16-bit and 32-bit x86 applications. It allows these applications to run, without recompilation, under control of a 64-bit operating system in long mode. The architectural enhancements i n t ro d u c e d by t h e A M D 6 4 a rch i t e c t u re t h a t s u p p o r t compatibility mode are summarized throughout this chapter. Unsupported Modes. Long mode does not support the following two operating modes:
Chapter 2: x86 and AMD64 Architecture Differences
29
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Virtual-8086 Mode--The virtual-8086 mode bit (EFLAGS.VM) is ignored when the processor is running in long mode. When long mode is enabled, any attempt to enable virtual-8086 mode is silently ignored. System software must leave long mode in order to use virtual-8086 mode. Real Mode--Real mode is not supported when the processor is operating in long mode because long mode requires that protected mode be enabled. 2.1.2 Legacy Mode The AMD64 architecture supports a pure x86 legacy mode, which preserves binary compatibility not only with existing 16bit and 32-bit applications but also with existing 16-bit and 32bit operating systems. Legacy mode supports real mode, protected mode, and virtual-8086 mode. A reset always places the processor in legacy mode (real mode), and the processor continues to run in legacy mode until system software activates long mode. New features added by the AMD64 architecture that are supported in legacy mode are summarized in this chapter. The AMD64 architecture supports system-management mode (SMM). SMM can be entered from both long mode and legacy mode, and SMM can return directly to either mode. The following differences exist between the support of SMM in the AMD64 architecture and the SMM support found in previous processor generations: The SMRAM state-save area format is changed to hold the 64-bit processor state. This state-save area format is used regardless of whether SMM is entered from long mode or legacy mode. The auto-halt restart and I/O-instruction restart entries in the SMRAM state-save area are one byte instead of two bytes. The initial processor state upon entering SMM is expanded to reflect the 64-bit nature of the processor. New conditions exist that can cause a processor shutdown while exiting SMM. SMRAM caching considerations are modified because the legacy FLUSH# external signal (writeback, if modified, and invalidate) is not supported on implementations of the AMD64 architecture.
2.1.3 SystemManagement Mode
30
Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
See Chapter 10, "System-Management Mode," for more information on the SMM differences.
2.2
Memory Model
The AMD64 architecture provides enhancements to the legacy memory model to support very large physical-memory and virtual-memory spaces while in long mode. Some of this expanded support for physical memory is available in legacy mode.
2.2.1 Memory Addressing
Virtual-Memory Addressing. Virtual-memory support is expanded to 64 address bits in long mode. This allows up to 16 exabytes of virtual-address space to be accessed. The virtual-address space supported in legacy mode is unchanged. Physical-Memory Addressing. Physical-memory support is expanded to 52 address bits in long mode and legacy mode. This allows up to 4 petabytes of physical memory to be accessed. The expanded physical-memory support is achieved by using paging and the page-size extensions. Implementations can support fewer than 52 physical-address bits. The first implementation of the AMD64 architecture, for example, supports 40-bit physical addressing in both long mode and legacy mode. Effective Addressing. The effective-address length is expanded to 64 bits in long mode. An effective-address calculation uses 64bit base and index registers, and sign-extends 8-bit and 32-bit displacements to 64 bits. In legacy mode, effective addresses remain 32 bits long.
2.2.2 Page Translation
The AMD64 architecture defines an expanded page-translation mechanism supporting translation of a 64-bit virtual address to a 52-bit physical address. See "Long-Mode Page Translation" on page 160 for detailed information on the enhancements to page translation in the AMD64 architecture. The enhancements are summarized below. Physical-Address Extensions (PAE). The AMD64 architecture requires physical-address extensions to be enabled (CR4.PAE=1) before long mode is entered. When PAE is enabled, all paging datastructures are 64 bits, allowing references into the full 52-bit physical-address space supported by the architecture.
Chapter 2: x86 and AMD64 Architecture Differences
31
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Page-Size Extensions (PSE). Page-size extensions (CR4.PSE) are ignored in long mode. Long mode does not support the 4-Mbyte page size enabled by page-size extensions. Long mode does, however, support 4-Kbyte and 2-Mbyte page sizes. Paging Data Structures. The AMD64 architecture extends the pagetranslation data structures in support of long mode. The extensions are: Page-map level-4 (PML4)--Long mode defines a new pagetranslation data structure, the PML4 table. The PML4 table sits at the top of the page-translation hierarchy and references PDP tables. Page-directory pointer (PDP)--The PDP tables in long mode are expanded from 4 entries to 512 entries each. Page-directory pointer entry (PDPE)--Previously undefined fields within the legacy-mode PDPE are defined by the AMD64 architecture. CR3 Register. The CR3 register is expanded to 64 bits for use in long-mode page translation. When long mode is active, the CR3 register references the base address of the PML4 table. In legacy mode, the upper 32 bits of CR3 are masked by the processor to support legacy page translation. CR3 references the PDP base-address when physical-address extensions are enabled, or the page-directory table base-address when physical-address extensions are disabled. Legacy-Mode Enhancements. L e g a cy - m o d e s o f t wa re c a n t a ke advantage of the enhancements made to the physical-address extension (PAE) support and page-size extension (PSE) support. The four-level page translation mechanism introduced by long mode is not available to legacy-mode software. PAE--When physical-address extensions are enabled (CR4.PAE=1), the AMD64 architecture allows legacy-mode software to load up to 52-bit (maximum size) physical addresses into the PDE and PTE. (Addresses are expanded to the maximum physical address size supported by the implementation.) PSE--The use of page-size extensions allows legacy mode software to define 4-Mbyte pages using the 32-bit pagetranslation tables. When page-size extensions are enabled (CR4.PSE=1), the AMD64 architecture enhances the 4Mbyte PDE to support 40 physical-address bits. 32 Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
See "Legacy-Mode Page Translation" on page 150 for more information on these enhancements. 2.2.3 Segmentation In long mode, the effects of segmentation depend on whether the processor is running in compatibility mode or 64-bit mode: In compatibility mode, segmentation functions just as it does in legacy mode, using legacy 16-bit or 32-bit protected mode semantics. 64-bit mode requires a flat-memory model for creating a flat 64-bit virtual-address space. Much of the segmentation capability present in legacy mode and compatibility mode is disabled when the processor is running in 64-bit mode. The differences in the segmentation model as defined by the AMD64 architecture are summarized in the following sections. See Chapter 4, "Segmented Virtual Memory," for a thorough description of these differences. Descriptor-Table Registers. In long mode, the base-address portion of the descriptor-table registers (GDTR, IDTR, LDTR, and TR) are expanded to 64 bits. The full 64-bit base address can only be loaded by software when the processor is running in 64-bit mode (using the LGDT, LIDT, LLDT, and LTR instructions, respectively). However, the full 64-bit base address is used by a processor running in compatibility mode (in addition to 64-bit mode) when making a reference into a descriptor table. A processor running in legacy mode can only load the low 32 bits of the base address, and the high 32 bits are ignored when references are made to the descriptor tables. Code-Segment Descriptors. The AMD64 architecture defines a new code-segment descriptor attribute, L (long). In compatibility mode, the processor treats code-segment descriptors as it does in legacy mode, with the exception that the processor recognizes the L attribute. If a code descriptor with L=1 is l o a d e d i n c o m p a t i b i l i t y m o d e , t h e p ro c e s s o r l e ave s compatibility mode and enters 64-bit mode. In legacy mode, the L attribute is reserved. The following differences exist for code-segment descriptors in 64-bit mode only: The CS base-address field is ignored by the processor. The CS limit field is ignored by the processor. Chapter 2: x86 and AMD64 Architecture Differences 33
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Only the L (long), D (default size), and DPL (descriptorprivilege level) fields are used by the processor in 64-bit mode. All remaining attributes are ignored. Data-Segment Descriptors. The following differences exist for datasegment descriptors in 64-bit mode only: The DS, ES, and SS descriptor base-address fields are ignored by the processor. The FS and GS descriptor base-address fields are expanded to 64 bits and used in effective-address calculations. The 64 bits of base address are mapped to model-specific registers (MSRs), and can only be loaded using the WRMSR instruction. The limit fields and attribute fields of all data-segment descriptors (DS, ES, FS, GS, and SS) are ignored by the processor. In compatibility mode, the processor treats data-segment descriptors as it does in legacy mode. Compatibility mode ignores the high 32 bits of base address in the FS and GS segment descriptors when calculating an effective address. System-Segment Descriptors. In 64-bit mode only, The LDT and TSS system-segment descriptor formats are expanded by 64 bits, allowing them to hold 64-bit base addresses. LLDT and LTR instructions can be used to load these descriptors into the LDTR and TR registers, respectively, from 64-bit mode. In compatibility mode and legacy mode, the formats of the LDT and TSS system-segment descriptors are unchanged. Also, unlike code-segment and data-segment descriptors, systemsegment descriptor limits are checked by the processor in long mode. Some legacy mode LDT and TSS type-field encodings are illegal in long mode (both compatibility mode and 64-bit mode), and others are redefined to new types. See "System Descriptors" on page 111 for additional information. Gate Descriptors. The following differences exist between gate descriptors in long mode (both compatibility mode and 64-bit mode) and in legacy mode: In long mode, all 32-bit gate descriptors are redefined as 64bit gate descriptors, and are expanded to hold 64-bit offsets. 34 Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The length of a gate descriptor in long mode is therefore 128 bits (16 bytes), versus the 64 bits (8 bytes) in legacy mode. Some type-field encodings are illegal in long mode, and others are redefined to new types. See "Gate Descriptors" on page 113 for additional information. The interrupt-gate and trap-gate descriptors define a new field, called the interrupt-stack table (IST) field.
2.3
Protection Checks
The AMD64 architecture makes the following changes to the protection mechanism in long mode: The page-protection-check mechanism is expanded in long mode to include the U/S and R/W protection bits stored in the PML4 entries and PDP entries. Several system-segment types and gate-descriptor types that are legal in legacy mode are illegal in long mode (compatibility mode and 64-bit mode) and fail type checks when used in long mode. Segment-limit checks are disabled in 64-bit mode for the CS, DS, ES, FS, GS, and SS segments. Segment-limit checks remain enabled for the LDT, GDT, IDT and TSS system segments. All segment-limit checks are performed in compatibility mode. Code and data segments used in 64-bit mode are treated as both readable and writable. See "Page-Protection Checks" on page 174 and "SegmentProtection Overview" on page 118 for detailed information on the protection-check changes.
2.4
Registers
The AMD64 architecture adds additional registers to the architecture, and in many cases expands the size of existing registers to 64 bits. The 80-bit floating-point stack registers and their overlaid 64-bit MMXTM registers are not modified by the AMD64 architecture.
2.4.1 General-Purpose Registers
In 64-bit mode, the general-purpose registers (GPRs) are 64 bits wide, and eight additional GPRs are available. The GPRs are: 35
Chapter 2: x86 and AMD64 Architecture Differences
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
RAX, RBX, RCX, RDX, RDI, RSI, RBP, RSP, and the new R8-R15 registers. To access the full 64-bit operand size, or the new R8-R15 registers, an instruction must include a new REX instruction-prefix byte (see "REX Prefixes" on page 37 for a summary of this prefix). In compatibility and legacy modes, the GPRs consist only of the eight legacy 32-bit registers. All legacy rules apply for determining operand size. 2.4.2 128-Bit Media Registers In 64-bit mode, eight additional 128-bit XMM registers are available, XMM8-XMM15. A REX instruction prefix is used to access these registers. In compatibility and legacy modes, the XMM registers consist of the eight 128-bit legacy registers, XMM0-XMM7. The flags register is expanded to 64 bits, and is called RFLAGS. All 64 bits can be accessed in 64-bit mode, but the upper 32 bits are reserved and always read back as zeros. Compatibility mode and legacy mode can read and write only the lower-32 bits of RFLAGS (the legacy EFLAGS). In long mode, the instruction pointer is extended to 64 bits, to support 64-bit code offsets. This 64-bit instruction pointer is called RIP. In 64-bit mode, the size of the stack pointer, RSP, is always 64 bits. The stack siz e is not controlled by a bit in the SS descriptor, as it is in compatibility or legacy mode, nor can it be overridden by an instruction prefix. Address-size overrides are ignored for implicit stack references. The AMD64 architecture defines several enhancements to the control registers (CRn). In long mode, all control registers are expanded to 64 bits, although the entire 64 bits can be read and written only from 64-bit mode. A new control register, the taskpriority register (CR8 or TPR) is added, and can be read and written from 64-bit mode. Last, the function of the page-enable bit (CR0.PG) is expanded. When long mode is enabled, the PG bit is used to activate and deactivate long mode. In long mode, all debug registers are expanded to 64 bits, although the entire 64 bits can be read and written only from 64-bit mode. Expanded register encodings for the decode registers allow up to eight new registers to be defined Chapter 2: x86 and AMD64 Architecture Differences
2.4.3 Flags Register
2.4.4 Instruction Pointer 2.4.5 Stack Pointer
2.4.6 Control Registers
2.4.7 Debug Registers
36
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
(DR8-DR15), although presently those registers are not supported by the AMD64 architecture. 2.4.8 Extended Feature Register (EFER) 2.4.9 Memory Type Range Registers (MTRRs) The EFER is expanded by the AMD64 architecture to include a long-mode-enable bit (LME), and a long-mode-active bit (LMA). These new bits can be accessed from legacy mode and long mode. The legacy MTRRs are architecturally defined as 64 bits, and can accommodate the maximum 52-bit physical address allowed by the AMD64 architecture. From both long mode and legacy mode, implementations of the AMD64 architecture reference the entire 52-bit physical-address value stored in the MTRRs. Long mode and legacy mode system software can update all 64 bits of the MTRRs to manage the expanded physical-address space. Several other MSRs have fields holding physical addresses. Examples include the APIC-base register and top-of-memory register. Generally, any model-specific register that contains a physical address is defined architecturally to be 64 bits wide, and can accommodate the maximum physical-address size defined by the AMD64 architecture. When physical addresses are read from MSRs by the processor, the entire value is read regardless of the operating mode. In legacy implementations, the high-order MSR bits are reserved, and software must write those values with zeros. In legacy mode on AMD64 architecture implementations, software can read and write all supported high-order MSR bits.
2.4.10 Other ModelSpecific Registers (MSRs)
2.5
Instruction Set
REX prefixes are a new family of instruction-prefix bytes used in 64-bit mode to: Specify the new GPRs and XMM registers. Specify a 64-bit operand size. Specify additional control registers. One additional control register, CR8, is defined in 64-bit mode. Specify additional debug registers (although none are currently defined). Not all instructions require a REX prefix. The prefix is necessary only if an instruction references one of the extended
2.5.1 REX Prefixes
Chapter 2: x86 and AMD64 Architecture Differences
37
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
registers or uses a 64-bit operand. If a REX prefix is used when it has no meaning, it is ignored. Default 64-Bit Operand Size. I n 6 4 - b i t m o d e , t wo g r o u p s o f instructions have a default operand size of 64 bits and thus do not need a REX prefix for this operand size: Near branches. All instructions, except far branches, that implicitly reference the RSP. See "Instructions that Reference RSP" on page 39 for additional information. 2.5.2 SegmentOverride Prefixes in 64-Bit Mode In 64-bit mode, the DS, ES, SS, and CS segment-override prefixes have no effect. These four prefixes are no longer treated as segment-override prefixes in the context of multipleprefix rules. Instead, they are treated as null prefixes. The FS and GS segment-override prefixes are treated as segment-override prefixes in 64-bit mode. Use of the FS and GS prefixes cause their respective segment bases to be added to the effective address calculation. See "FS and GS Registers in 64-Bit Mode" on page 88 for additional information on using these segment registers. 2.5.3 Operands and Results The AMD64 architecture provides support for using 64-bit operands and generating 64-bit results when operating in 64-bit mode. See "Operands" in Volume 1 for details. Operand-Size Overrides. In 64-bit mode, the default operand size is 32 bits. A REX prefix can be used to specify a 64-bit operand size. Software uses a legacy operand-size (66h) prefix to toggle to 16-bit operand size. The REX prefix takes precedence over the legacy operand-size prefix. Zero Extension of Results. In 64-bit mode, when performing 32-bit operations with a GPR destination, the processor zero-extends the 32-bit result into the full 64-bit destination. Both 8-bit and 16-bit operations on GPRs preserve all unwritten upper bits of the destination GPR. This is consistent with legacy 16-bit and 32-bit semantics for partial-width results. 2.5.4 Address Calculations The AMD64 architecture modifies aspects of effective-address calculation to support 64-bit mode. These changes are s u m m a r i z e d i n t h e fo l l ow i n g s e c t i o n s . S e e " M e m o ry Addressing" in Volume 1 for details.
38
Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Address-Size Overrides. In 64-bit mode, the default-address size is 64 bits. The address size can be overridden to 32 bits by using the address-size prefix (67h). 16-bit addresses are not supported in 64-bit mode. In compatibility mode and legacy mode, address-size overrides function the same as in x86 legacy architecture. Displacements and Immediates. G e n e ra l ly, d i s p l a c e m e n t a n d immediate values in 64-bit mode are not extended to 64 bits. They are still limited to 32 bits and are sign extended during effective-address calculations. In 64-bit mode, however, support is provided for some 64-bit displacement and immediate forms of the MOV instruction. Zero Extending 16-Bit and 32-Bit Addresses. A l l 1 6 - b i t a n d 3 2 - b i t address calculations are zero-extended in long mode to form 64bit addresses. Address calculations are first truncated to the effective-address size of the current mode (64-bit mode or compatibility mode), as overridden by any address-size prefix. The result is then zero-extended to the full 64-bit address width. RIP-Relative Addressing. A new addressing form, RIP-relative (instruction-pointer relative) addressing, is implemented in 64bit mode. The effective address is formed by adding the displacement to the 64-bit RIP of the next instruction. 2.5.5 Instructions that Reference RSP With the exception of far branches, all instructions that implicitly reference the 64-bit stack pointer, RSP, default to a 64-bit operand size in 64-bit mode (see Table 2-1 on page 40 for a listing). Pushes and pops of 32-bit stack values are not possible in 64-bit mode with these instructions, but they can be overridden to 16 bits.
Chapter 2: x86 and AMD64 Architecture Differences
39
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 2-1. Instructions That Reference RSP
Mnemonic ENTER LEAVE POP reg/mem POP reg POP FS POP GS POPF, POPFD, POPFQ PUSH imm32 PUSH imm8 PUSH reg/mem PUSH reg PUSH FS PUSH GS PUSHF, PUSHFD, PUSHFQ Opcode (hex) C8 C9 8F/0 58-5F 0F A1 0F A9 9D 68 6A FF/6 50-57 0F A0 0F A8 9C Description Create Procedure Stack Frame Delete Procedure Stack Frame Pop Stack (register or memory) Pop Stack (register) Pop Stack into FS Segment Register Pop Stack into GS Segment Register Pop to rFLAGS Word, Doubleword, or Quadword Push onto Stack (sign-extended doubleword) Push onto Stack (sign-extended byte) Push onto Stack (register or memory) Push onto Stack (register) Push FS Segment Register onto Stack Push GS Segment Register onto Stack Push rFLAGS Word, Doubleword, or Quadword onto Stack
2.5.6 Branches
The AMD64 architecture expands two branching mechanisms to accommodate branches in the full 64-bit virtual-address space: In 64-bit mode, near-branch semantics are redefined. In both 64-bit and compatibility modes, a 64-bit call-gate descriptor is defined for far calls. In addition, enhancements are made to the legacy SYSCALL and SYSRET instructions. Near Branches. In 64-bit mode, the operand size for all near branches defaults to 64 bits (see Table 2-2 on page 41 for a listing). Therefore, these instructions update the full 64-bit RIP without the need for a REX operand-size prefix. The following aspects of near branches default to 64 bits:
40
Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Truncation of the instruction pointer. Size of a stack pop or stack push, resulting from a CALL or RET. Size of a stack-pointer increment or decrement, resulting from a CALL or RET. Size of operand fetched by indirect-branch operand size. The operand size for near branches can be overridden to 16 bits in 64-bit mode. Table 2-2. 64-Bit Mode Near Branches, Default 64-Bit Operand Size
Mnemonic CALL Jcc JMP LOOP LOOPcc RET Opcode (hex) E8, FF/2 many E9, EB, FF/4 E2 E0, E1 C3, C2 Description Call Procedure Near Jump Conditional Near Jump Near Loop Loop Conditional Return From Call (near)
The address size of near branches is not forced in 64-bit mode. Such addresses are 64 bits by default, but they can be overridden to 32 bits by a prefix. The size of the displacement field for relative branches is still limited to 32 bits. Far Branches Through Long-Mode Call Gates. Long mode redefines the 32-bit call-gate descriptor type as a 64-bit call-gate descriptor and expands the call-gate descriptor size to hold a 64-bit offset. The long-mode call-gate descriptor allows far branches to reference any location in the supported virtual-address space. In long mode, the call-gate mechanism is changed as follows: In long mode, CALL and JMP instructions that reference call-gates must reference 64-bit call gates. A 64-bit call-gate descriptor must reference a 64-bit codesegment. When a control transfer is made through a 64-bit call gate, the 64-bit target address is read from the 64-bit call-gate Chapter 2: x86 and AMD64 Architecture Differences 41
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
descriptor. The base address in the target code-segment descriptor is ignored. Stack Switching. Automatic stack switching is also modified when a control transfer occurs through a call gate in long mode: The target-stack pointer read from the TSS is a 64-bit RSP value. The SS register is loaded with a null selector. Setting the new SS selector to null allows nested control transfers in 64bit mode to be handled properly. The SS.RPL value is updated to remain consistent with the newly loaded CPL value. The size of pushes onto the new stack is modified to accommodate the 64-bit RIP and RSP values. Automatic parameter copying is not supported in long mode. Far Returns. In long mode, far returns can load a null SS selector from the stack under the following conditions: The target operating mode is 64-bit mode. The target CPL<3. Allowing RET to load SS with a null selector under these conditions makes it possible for the processor to unnest far CALLs (and interrupts) in long mode. Task Gates. C o n t rol t ra n s f e rs t h rough task gates are not supported in long mode. Branches to 64-Bit Offsets. Because immediate values are generally limited to 32 bits, the only way a full 64-bit absolute RIP can be specified in 64-bit mode is with an indirect branch. For this reason, direct forms of far branches are eliminated from the instruction set in 64-bit mode. SYSCALL and SYSRET Instructions. The AMD64 architecture expands the function of the legacy SYSCALL and SYSRET instructions in long mode. In addition, two new STAR registers, LSTAR and CSTAR, are provided to hold the 64-bit target RIP for the instructions when they are executed in long mode. The legacy STAR register is not expanded in long mode. See "SYSCALL and SYSRET" on page 182 for additional information.
42
Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
SWAPGS Instruction. The AMD64 architecture provides the SWAPGS instruction as a fast method for system software to load a pointer to system data-structures. SWAPGS is valid only in 64-bit mode. An undefined-opcode exception (#UD) occurs if software attempts to execute SWAPGS in legacy mode or compatibility mode. See "SWAPGS Instruction" on page 185 for additional information. SYSENTER and SYSEXIT Instructions. The SYSENTER and SYSEXIT instructions are invalid in long mode, and result in an invalid opcode exception (#UD) if software attempts to use them. Software should use the SYSCALL and SYSRET instructions when running in long mode. See "SYSENTER and SYSEXIT (Legacy Mode Only)" on page 184 for additional information. 2.5.7 NOP Instruction The legacy x86 architecture commonly uses opcode 90h as a one-byte NOP. In 64-bit mode, the processor treats opcode 90h specially in order to preserve this NOP definition. This is necessary because opcode 90h is actually the XCHG EAX, EAX instruction in the legacy architecture. Without special handling in 64-bit mode, the instruction would not be a true no-operation. Therefore, in 64-bit mode the processor treats opcode 90h (the legacy XCHG EAX, EAX instruction) as a true NOP, regardless of a REX operand-size prefix. This special handling does not apply to the two-byte ModRM form of the XCHG instruction. Unless a 64-bit operand size is specified using a REX prefix byte, using the two-byte form of XCHG to exchange a register with itself does not result in a nooperation, because the default operation size is 32 bits in 64-bit mode. 2.5.8 Single-Byte INC and DEC Instructions In 64-bit mode, the legacy encodings for the 16 single-byte INC and DEC instructions (one for each of the eight GPRs) are used to encode the REX prefix values. The functionality of these INC and DEC instructions is still available, however, using the ModRM forms of those instructions (opcodes FF /0 and FF /1). See "Single-Byte INC and DEC Instructions in 64-Bit Mode" in Volume 3 for additional information. MOVSXD is a new instruction in 64-bit mode (the legacy ARPL instruction opcode, 63h, is reassigned as the MOVSXD opcode). It reads a fixed-size 32-bit source operand from a register or memory and (if a REX prefix is used with the instruction) signextends the value to 64 bits. MOVSXD is analogous to the 43
2.5.9 MOVSXD Instruction
Chapter 2: x86 and AMD64 Architecture Differences
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
MOVSX instruction, which sign-extends a byte to a word or a word to a doubleword, depending on the effective operand size. See "General-Purpose Instruction Reference" in Volume 3 for additional information. 2.5.10 Invalid Instructions Table 2-3 lists instructions that are illegal in 64-bit mode. Table 2-4 on page 45 lists instructions that are invalid in long mode (both compatibility mode and 64-bit mode). Attempted use of these instructions causes an invalid-opcode exception (#UD) to occur. Table 2-3. Invalid Instructions in 64-Bit Mode
Mnemonic AAA AAD AAM AAS BOUND CALL (far) DAA DAS INTO JMP (far) LDS LES POP DS POP ES POP SS POPA, POPAD PUSH CS PUSH DS PUSH ES Opcode (hex) 37 D5 D4 3F 62 9A 27 2F CE EA C5 C4 1F 07 17 61 0E 1E 06 Description ASCII Adjust After Addition ASCII Adjust Before Division ASCII Adjust After Multiply ASCII Adjust After Subtraction Check Array Bounds Procedure Call Far (absolute) Decimal Adjust after Addition Decimal Adjust after Subtraction Interrupt to Overflow Vector Jump Far (absolute) Load DS Segment Register Load ES Segment Register Pop Stack into DS Segment Pop Stack into ES Segment Pop Stack into SS Segment Pop All to GPR Words or Doublewords Push CS Segment Selector onto Stack Push DS Segment Selector onto Stack Push ES Segment Selector onto Stack
44
Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 2-3. Invalid Instructions in 64-Bit Mode (continued)
Mnemonic PUSH SS PUSHA, PUSHAD Redundant Grp1 (undocumented) SALC (undocumented) Opcode (hex) 16 60 82 D6 Description Push SS Segment Selector onto Stack Push All GPR Words or Doublewords onto Stack Redundant encoding of group1 Eb,Ib opcodes Set AL According to CF
Table 2-4. Invalid Instructions in Long Mode
Mnemonic SYSENTER SYSEXIT Opcode (hex) 0F 34 0F 35 System Call System Return Description
Table 2-5 lists the instructions that are no longer valid in 64-bit mode because their opcodes have been reassigned. The reassigned opcodes are used in 64-bit mode as REX instruction prefixes. Table 2-5. Reassigned Instructions in 64-Bit Mode
Mnemonic Opcode (hex) 63 Description Opcode for MOVSXD instruction in 64-bit mode. In all other modes, this the Adjust Requestor Privilege Level instruction opcode. Decrement by 1, Increment by 1. Two-byte versions of DEC and INC are still valid.
ARPL
DEC and INC
40-4F
2.5.11 FXSAVE and FXRSTOR Instructions
The FXSAVE and FXRSTOR instructions are used to save and restore the entire 128-bit media, 64-bit media, and x87 instruction-set environment during a context switch. The AMD64 architecture modifies the memory format used by these instructions in order to save and restore the full 64-bit instruction and data pointers, as well as the XMM8-XMM15 45
Chapter 2: x86 and AMD64 Architecture Differences
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
registers. Selection of the 32-bit legacy format or the expanded 64-bit format is accomplished by using the corresponding operand size with the FXSAVE and FXRSTOR instructions. When 64-bit software executes an FXSAVE and FXRSTOR with a 32-bit operand size (no operand-size override) the 32-bit legacy format is used. When 64-bit software executes an FXSAVE and FXRSTOR with a 64-bit operand size, the 64-bit format is used. If the fast-FXSAVE/FXRSTOR (FFXSR) feature is enabled in EFER, FXSAVE and FXRSTOR do not save or restore the XMM0-XMM15 registers when executed in 64-bit mode at CPL 0. The x87 environment and MXCSR are saved whether fast-FXSAVE/FXRSTOR is enabled or not. Software can use CPUID to determine whether the fast-FXSAVE/FXRSTOR feature is available (CPUID function 8000_0001h, EDX bit 25). The fast-FXSAVE/FXRSTOR feature doesn't have any effect on FXSAVE/FXRSTOR in non 64-bit mode or when CPL > 0.
2.6
Interrupts and Exceptions
When a processor is running in long mode, an interrupt or exception causes the processor to enter 64-bit mode. All longmode interrupt handlers must be 64-bit code. The AMD64 architecture expands the legacy interrupt-processing and exception-processing mechanism to support handling of interrupts by 64-bit operating systems and applications. The changes are summarized in the following sections. See "LongMode Interrupt Control Transfers" on page 287 for detailed information on these changes.
2.6.1 Interrupt Descriptor Table
The long-mode interrupt-descriptor table (IDT) must contain 64-bit mode interrupt-gate or trap-gate descriptors for all interrupts or exceptions that can occur while the processor is running in long mode. Task gates cannot be used in the longmode IDT, because control transfers through task gates are not supported in long mode. In long mode, the IDT index is formed by scaling the interrupt vector by 16. In legacy protected mode, the IDT is indexed by scaling the interrupt vector by eight. In legacy mode, the size of an IDT entry (16 bits or 32 bits) determines the size of interrupt-stack-frame pushes, and SS:eSP is pushed only on a CPL change. In long mode, the size of interrupt stack-frame pushes is fixed at eight bytes, because Chapter 2: x86 and AMD64 Architecture Differences
2.6.2 Stack Frame Pushes
46
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
interrupts are handled in 64-bit mode. Long mode interrupts also cause SS:RSP to be pushed unconditionally, rather than pushing only on a CPL change. 2.6.3 Stack Switching Legacy mode provides a mechanism to automatically switch stack frames in response to an interrupt. In long mode, a slightly modified version of the legacy stack-switching mechanism is implemented, and an alternative stack-switching mechanism--called the interrupt stack table (IST)--is supported. Long-Mode Stack Switches. When stacks are switched as part of a long-mode privilege-level change resulting from an interrupt, the following occurs: The target-stack pointer read from the TSS is a 64-bit RSP value. The SS register is loaded with a null selector. Setting the new SS selector to null allows nested control transfers in 64bit mode to be handled properly. The SS.RPL value is cleared to 0. The old SS and RSP are saved on the new stack. Interrupt Stack Table. In long mode, a new interrupt stack table (IST) mechanism is available as an alternative to the modified legacy stack-switching mechanism. The IST mechanism unconditionally switches stacks when it is enabled. It can be enabled for individual interrupt vectors using a field in the IDT entry. This allows mixing interrupt vectors that use the modified legacy mechanism with vectors that use the IST mechanism. The IST pointers are stored in the long-mode TSS. The IST mechanism is only available when long mode is enabled. 2.6.4 IRET Instruction In compatibility mode, IRET pops SS:eSP off the stack only if there is a CPL change. This allows legacy applications to run properly in compatibility mode when using the IRET instruction. In 64-bit mode, IRET unconditionally pops SS:eSP off of the interrupt stack frame, even if the CPL does not change. This is done because the original interrupt always pushes SS:RSP. Because interrupt stack-frame pushes are always eight bytes in long mode, an IRET from a long-mode interrupt handler (64-bit code) must pop eight-byte items off the stack. This is Chapter 2: x86 and AMD64 Architecture Differences 47
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
accomplished by preceding the IRET with a 64-bit REX operand-size prefix.
48
Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
In long mode, an IRET can load a null SS selector from the stack under the following conditions: The target operating mode is 64-bit mode. The target CPL<3. Allowing IRET to load SS with a null selector under these conditions makes it possible for the processor to unnest interrupts (and far CALLs) in long mode. 2.6.5 Task-Priority Register (CR8) The AMD64 architecture allows software to define up to 15 external interrupt-priority classes. Priority classes are numbered from 1 to 15, with priority-class 1 being the lowest and priority-class 15 the highest. A new control register (CR8) is introduced by the AMD64 architecture for managing priority classes. This register, also called the task-priority register (TPR), uses the four low-order bits for specifying a task priority. How external interrupts are organized into these priority classes is implementation dependent. See "External Interrupt Priorities" on page 272 for information on this feature. 2.6.6 New Exception Conditions The AMD64 architecture defines a number of new conditions that can cause an exception to occur when the processor is running in long mode. Many of the conditions occur when software attempts to use an address that is not in canonical form. See "Vectors" on page 247 for information on the new exception conditions that can occur in long mode.
2.7
Hardware Task Switching
The legacy hardware task-switch mechanism is disabled when the processor is running in long mode. However, long mode requires system software to create data structures for a single task--the long-mode task. TSS Descriptors--A new TSS-descriptor type, the 64-bit TSS type, is defined for use in long mode. It is the only valid TSS type that can be used in long mode, and it must be loaded into the TR by executing the LTR instruction in 64-bit mode. See "TSS Descriptor" on page 362 for additional information. Task Gates--Because the legacy task-switch mechanism is not supported in long mode, software cannot use task gates in
Chapter 2: x86 and AMD64 Architecture Differences
49
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
long mode. Any attempt to transfer control to another task through a task gate causes a general-protection exception (#GP) to occur. Task-State Segment--A 64-bit task state segment (TSS) is defined for use in long mode. This new TSS format contains 64-bit stack pointers (RSP) for privilege levels 0-2, interrupt-stack-table (IST) pointers, and the I/O-map base address. See "64-Bit Task State Segment" on page 370 for additional information.
2.8
Long-Mode vs. Legacy-Mode Differences
Table 2-6 on page 51 summarizes several major systemprogramming differences between 64-bit mode and legacy protected mode. The third column indicates whether the difference also applies to compatibility mode. "Differences B e t we e n L o n g M o d e a n d L e g a cy M o d e " i n Vo l u m e 3 summarizes the application-programming model differences.
50
Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 2-6.
Differences Between Long Mode and Legacy Mode
Subject 64-Bit Mode Difference Real and virtual-8086 modes not supported Task switching not supported 64-bit virtual addresses Applies To Compatibility Mode? Yes Yes No Yes
x86 Modes Task Switching
Addressing
4-level paging structures PAE must always be enabled CS, DS, ES, SS segment bases are ignored
Loaded Segment (Usage during memory reference)
CS, DS, ES, FS, GS, SS segment limits are ignored DS, ES, FS, GS attribute are ignored CS, DS, ES, SS Segment prefixes are ignored All pushes are 8 bytes
No
Exception and Interrupt Handling
IDT entries are expanded to 16 bytes SS is not changed for stack switch SS:RSP is pushed unconditionally All pushes are 8 bytes 16-bit call gates are illegal
Yes
Call Gates
32-bit call gate type is redefined as 64-bit call gate and is expanded to 16 bytes SS is not changed for stack switch
Yes
System-Descriptor Registers System-Descriptor Table Entries and PseudoDescriptors
GDT, IDT, LDT, TR base registers expanded to 64 bits LGDT and LIDT use expanded 10-byte pseudo-descriptors LLDT and LTR use expanded 16-byte table entries
Yes No
Chapter 2: x86 and AMD64 Architecture Differences
51
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
52
Chapter 2: x86 and AMD64 Architecture Differences
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
3
System Resources
The operating system manages the software-execution environment and general system operation through the use of system resources. These resources consist of system registers (control registers and model-specific registers) and system-data structures (memory-management and protection tables). The system-control registers are described in detail in this chapter; many of the features they control are described elsewhere in this volume. The model-specific registers supported by the AMD64 architecture are introduced in this chapter. Because of their complexity, system-data structures are described in separate chapters. Refer to the following chapters for detailed information on these data structures: Descriptors and descriptor tables are described in "Segmentation Data Structures and Registers" on page 82. Page-translation tables are described in "Legacy-Mode Page Translation" on page 150 and "Long-Mode Page Translation" on page 160. The task-state segment is described in "Legacy Task-State Segment" on page 365 and "64-Bit Task State Segment" on page 370. Not all processor implementations are required to support all possible features. The last section in this chapter addresses processor-feature identification. System software uses the capabilities described in that section to determine which features are supported so that the appropriate service routines are loaded.
3.1
System-Control Registers
The registers that control the AMD64 architecture operating environment include: CR0--Provides operating-mode controls and some processorfeature controls. CR2--This register is used by the page-translation mechanism. It is loaded by the processor with the page-fault virtual address when a page-fault exception occurs.
Chapter 3: System Resources
53
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
CR3--This register is also used by the page-translation mechanism. It contains the base address of the highest-level page-translation table, and also contains cache controls for the specified table. CR4--This register contains additional controls for various operating-mode features. CR8--This new register, accessible in 64-bit mode using the REX prefix, is introduced by the AMD64 architecture. CR8 is used to prioritize external interrupts and is referred to as the task-priority register (TPR). RFLAGS--This register contains processor-status and processor-control fields. The status and control fields are used primarily in the management of virtual-8086 mode, hardware multitasking, and interrupts. EFER--This model-specific register contains status and controls for additional features not managed by the CR0 and CR4 registers. Included in this register are the long-mode enable and activation controls introduced by the AMD64 architecture. Control registers CR1, CR5-CR7, and CR9-CR15 are reserved. In legacy mode, all control registers and RFLAGS are 32 bits. The EFER register is 64 bits in all modes. The AMD64 architecture expands all 32-bit system-control registers to 64 bits. In 64-bit mode, the MOV CRn instructions read or write all 64 bits of these registers (operand-size prefixes are ignored). In compatibility and legacy modes, control-register writes fill the low 32 bits with data and the high 32 bits with zeros, and control-register reads return only the low 32 bits. In 64-bit mode, the high 32 bits of CR0 and CR4 are reserved and must be written with zeros. Writing a 1 to any of the high 32 bits results in a general-protection exception, #GP(0). All 64 bits of CR2 are writable. However, the MOV CRn instructions do not check that addresses written to CR2 are within the virtualaddress limitations of the processor implementation. All CR3 bits are writable, except for unimplemented physical address bits, which must be cleared to 0. The upper 32 bits of RFLAGS are always read as zero by the processor. Attempts to load the upper 32 bits of RFLAGS with anything other than zero are ignored by the processor.
54
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
3.1.1 CR0 Register
The CR0 register is shown in Figure 3-1. The legacy CR0 register is identical to the low 32 bits of the register shown in Figure 3-1 (CR0 bits 31-0).
32 Reserved, MBZ
63
31
P G
30
C D
29
N W
28 Reserved
19
18
A M
17 r
16
W P
15 Reserved
6
5
N E
4
E T
3
T S
2
E M
1
M P
0
P E
Bits Mnemonic 63-32 Reserved 31 PG 30 CD 29 NW 28-19 Reserved 18 AM 17 Reserved 16 WP 15-6 Reserved 5 NE 4 ET 3 TS 2 EM 1 MP 0 PE
Description Reserved, Must be Zero Paging Cache Disable Not Writethrough Reserved Alignment Mask Reserved Write Protect Reserved Numeric Error Extension Type Task Switched Emulation Monitor Coprocessor Protection Enabled
R/W R/W R/W R/W R/W R/W R/W R R/W R/W R/W R/W
Figure 3-1.
Control Register 0 (CR0) The functions of the CR0 control bits are (unless otherwise noted, all bits are read/write): Protected-Mode Enable (PE) Bit. Bit 0. Software enables protected mode by setting PE to 1, and disables protected mode by clearing PE to 0. When the processor is running in protected mode, segment-protection mechanisms are enabled. S e e " S e g m e n t - P ro t e c t i o n O ve rv i ew " o n p a g e 1 1 8 fo r information on the segment-protection mechanisms. Monitor Coprocessor (MP) Bit. Bit 1. Software uses the MP bit with the task-switched control bit (CR0.TS) to control whether execution of the WAIT/FWAIT instruction causes a device-notavailable exception (#NM) to occur, as follows: If both the monitor-coprocessor and task-switched bits are set (CR0.MP=1 and CR0.TS=1), then executing the
Chapter 3: System Resources
55
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
WAIT/FWAIT instruction causes a device-not-available exception (#NM). If either the monitor-coprocessor or task-switched bits are clear (CR0.MP=0 or CR0.TS=0), then executing the WAIT/FWAIT instruction proceeds normally. Software typically should set MP to 1 if the processor implementation supports x87 instructions. This allows the CR0.TS bit to completely control when the x87-instruction context is saved as a result of a task switch. Emulate Coprocessor (EM) Bit. B i t 2 . S o f t wa re f o rc e s a l l x 8 7 instructions to cause a device-not-available exception (#NM) by setting EM to 1. Likewise, setting EM to 1 forces an invalidopcode exception (#UD) when an attempt is made to execute any of the 64-bit or 128-bit media instructions. The exception handlers can emulate these instruction types if desired. Setting the EM bit to 1 does not cause an #NM exception when the WAIT/FWAIT instruction is executed. Task Switched (TS) Bit. Bit 3. When an attempt is made to execute an x87 or media instruction while TS=1, a device-not-available exception (#NM) occurs. Software can use this mechanism-- sometimes referred to as "lazy context-switching"--to save the unit contexts before executing the next instruction of those types. As a result, the x87 and media instruction-unit contexts are saved only when necessary as a result of a task switch. When a hardware task switch occurs, TS is automatically set to 1. System software that implements software task-switching rather than using the hardware task-switch mechanism can still use the TS bit to control x87 and media instruction-unit context saves. In this case, the task-management software uses a MOV CR0 instruction to explicitly set the TS bit to 1 during a task switch. Software can clear the TS bit by either executing the CLTS instruction or by writing to the CR0 register directly. Long-mode system software can use this approach even though the hardware task-switch mechanism is not supported in long mode. The CR0.MP bit controls whether the WAIT/FWAIT instruction causes an #NM exception when TS=1. Extension Type (ET) Bit. B i t 4 , re a d - o n l y. I n s o m e e a r l y x 8 6 processors, software set ET to 1 to indicate support of the 56 Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
387DX math-coprocessor instruction set. This bit is now reserved and forced to 1 by the processor. Software cannot clear this bit to 0. Numeric Error (NE) Bit. Bit 5. Clearing the NE bit to 0 disables internal control of x87 floating-point exceptions and enables external control. When NE is cleared to 0, the IGNNE# input signal controls whether x87 floating-point exceptions are ignored: When IGNNE# is 1, x87 floating-point exceptions are ignored. When IGNNE# is 0, x87 floating-point exceptions are reported by setting the FERR# input signal to 1. External logic can use the FERR# signal as an external interrupt. When NE is set to 1, internal control over x87 floating-point exception reporting is enabled and the external reporting mechanism is disabled. It is recommended that software set NE to 1. This enables optimal performance in handling x87 floating-point exceptions. Write Protect (WP) Bit. Bit 16. Read-only pages are protected from supervisor-level writes when the WP bit is set to 1. When WP is cleared to 0, supervisor software can write into read-only pages. See "Page-Protection Checks" on page 174 for information on the page-protection mechanism. Alignment Mask (AM) Bit. Bit 18. Software enables automatic a l i g n m e n t ch e ck i n g by s e t t i n g t h e A M b i t t o 1 w h e n eFLAGS.AC=1. Alignment checking can be disabled by clearing either AM or eFLAGS.AC to 0. When automatic alignment checking is enabled and CPL=3, a memory reference to an unaligned operand causes an alignment-check exception (#AC). Not Writethrough (NW) Bit. Bit 29. Ignored. This bit can be set to 1 or cleared to 0, but its value is ignored. The NW bit exists only for legacy purposes. Cache Disable (CD) Bit. Bit 30. When CD is cleared to 0, the internal caches are enabled. When CD is set to 1, no new data or instructions are brought into the internal caches. However, the processor still accesses the internal caches when CD=1 under the following situations:
Chapter 3: System Resources
57
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Reads that hit in an internal cache cause the data to be read from the internal cache that reported the hit. Writes that hit in an internal cache cause the cache line that reported the hit to be written back to memory and invalidated in the cache. Cache misses do not affect the internal caches when CD=1. Software can prevent cache access by writing back and invalidating the caches before setting CD to 1 (this avoids caching the instructions that set CD to 1). Setting CD to 1 also causes the processor to ignore the pagelevel cache-control bits (PWT and PCD) when paging is enabled. These bits are located in the page-translation tables and CR3 register. See "Page-Level Writethrough (PWT) Bit" on page 170 and "Page-Level Cache Disable (PCD) Bit" on page 170 for information on page-level cache control. See "Memory Caches" on page 208 for information on the internal caches. Paging Enable (PG) Bit. Bit 31. Software enables page translation by setting PG to 1, and disables page translation by clearing PG to 0. Page translation cannot be enabled unless the processor is in protected mode (CR0.PE=1). If software attempts to set PG to 1 when PE is cleared to 0, the processor causes a generalprotection exception (#GP). See "Page Translation Overview" on page 146 for information on the page-translation mechanism. Reserved Bits. Bits 28-19, 17, 15-6, and 63-32. When writing the CR0 register, software should set the values of reserved bits to the values found during the previous CR0 read. No attempt should be made to change reserved bits, and software should never rely on the values of reserved bits. In long mode, bits 63-32 are reserved and must be written with zero, otherwise a #GP occurs. 3.1.2 CR2 and CR3 Registers The CR2 (page-fault linear address) register, shown in Figures 3-2 and 3-3, and the CR3 (page-translation-table base address) register, shown in Figures 3-4, 3-5, and 3-6, are used only by the page-translation mechanism.
58
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
31 Page-Fault Virtual Address
0
Figure 3-2.
63
Control Register 2 (CR2)--Legacy-Mode
32 Page-Fault Virtual Address
31 Page-Fault Virtual Address
0
Figure 3-3.
Control Register 2 (CR2)--Long Mode See "CR2 Register" on page 261 for a description of the CR2 register. The CR3 register is used to point to the base address of the highest-level page-translation table.
31 Page-Directory-Table Base Address
12
11
Reserved
5
4
P C D
3
P W T
2
0
Reserved
Figure 3-4.
Control Register 3 (CR3)--Legacy-Mode Non-PAE Paging
31 Page-Directory-Pointer-Table Base Address
5
4
P C D
3
P W T
2
0
Reserved
Figure 3-5.
Control Register 3 (CR3)--Legacy-Mode PAE Paging
Chapter 3: System Resources
59
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
63 Reserved
52
51 Page-Map Level-4 Table Base Address
32
31 Page-Map Level-4 Table Base Address
12
11 Reserved
5
4
P C D
3
P W T
2
0
Reserved
Figure 3-6.
Control Register 3 (CR3)--Long Mode The legacy CR3 register is described in "CR3 Register" on page 151, and the long-mode CR3 register is described in "CR3" on page 161.
3.1.3 CR4 Register
The CR4 register is shown in Figure 3-7. In legacy mode, the CR4 register is identical to the low 32 bits of the register shown in Figure 3-7 (CR4 bits 31-0). The features controlled by the bits in the CR4 register are model-specific extensions. Except for the performance-counter extensions (PCE) feature, software can use the CPUID instruction to verify that each feature is supported before using that feature.
32 Reserved, MBZ
63
31 Reserved, MBZ
11
10
9
8
7
P G E
6
M C E
5
P A E
4
P S E
3
D E
2
T S D
1
P V I
0
V M E
O P OSF S C XSR X E
Bits Mnemonic 63-11 Reserved 10 OSXMMEXCPT 9 OSFXSR 8 7 6 5 4 3 2 1 0 PCE PGE MCE PAE PSE DE TSD PVI VME
Description Reserved, Must be Zero Operating System Unmasked Exception Support Operating System FXSAVE/FXRSTOR Support Performance-Monitoring Counter Enable Page-Global Enable Machine Check Enable Physical-Address Extension Page Size Extensions Debugging Extensions Time Stamp Disable Protected-Mode Virtual Interrupts Virtual-8086 Mode Extensions
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Figure 3-7.
Control Register 4 (CR4)
60
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The function of the CR4 control bits are (all bits are read/write): Virtual-8086 Mode Extensions (VME) Bit. B i t 0 . S e t t i n g V M E t o 1 enables hardware-supported performance enhancements for software running in virtual-8086 mode. Clearing VME to 0 disables this support. The enhancements enabled when VME=1 include: Virtualized, maskable, external-interrupt control and notification using the VIF and VIP bits in the rFLAGS register. Virtualizing affects the operation of several instructions that manipulate the rFLAGS.IF bit. Selective intercept of software interrupts (INTn instructions) using the interrupt-redirection bitmap in the TSS. Protected-Mode Virtual Interrupts (PVI) Bit. Bit 1. Setting PVI to 1 enables support for protected-mode virtual interrupts. Clearing PVI to 0 disables this support. When PVI=1, hardware support of two bits in the rFLAGS register, VIF and VIP, is enabled. Only the STI and CLI instructions are affected by enabling PVI. Unlike the case when CR0.VME=1, the interrupt-redirection b i t m a p i n t h e T SS c a n n o t b e u s e d fo r s e l e c t ive I N T n interception. PVI enhancements are also supported in long mode. See "Virtual Interrupts" on page 295 for more information on using PVI. Time-Stamp Disable (TSD) Bit. Bit 2. The TSD bit allows software to control the privilege level at which the time-stamp counter can be read. When TSD is cleared to 0, software running at any privilege level can read the time-stamp counter using the RDTSC instruction. When TSD is set to 1, only software running at privilege-level 0 can execute the RDTSC instruction. Debugging Extensions (DE) Bit. Bit 3. Setting the DE bit to 1 enables the I/O breakpoint capability and enforces treatment of the DR4 and DR5 registers as reserved. Software that accesses DR4 or DR5 when DE=1 causes a invalid opcode exception (#UD). When the DE bit is cleared to 0, I/O breakpoint capabilities are disabled. Software references to the DR4 and DR5 registers are aliased to the DR6 and DR7 registers, respectively.
Chapter 3: System Resources
61
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Page-Size Extensions (PSE) Bit. Bit 4. Setting PSE to 1 enables the use of 4-Mbyte physical pages. With PSE=1, the physical-page size is selected between 4 Kbytes and 4 Mbytes using the pagedirectory entry page-size field (PS). Clearing PSE to 0 disables the use of 4-Mbyte physical pages and restricts all physical pages to 4 Kbytes. The PSE bit has no effect when physical-address extensions are e n a b l e d ( C R 4 . PA E = 1 ) . B e c a u s e l o n g m o d e re q u i re s CR4.PAE=1, the PSE bit is ignored when the processor is running in long mode. See "4-Mbyte Page Translation" on page 154 for more information on 4-Mbyte page translation. Physical-Address Extension (PAE) Bit. Bit 5. Setting PAE to 1 enables the use of physical-address extensions and 2-Mbyte physical pages. Clearing PAE to 0 disables these features. With PAE=1, the page-translation data structures are expanded from 32 bits to 64 bits, allowing the translation of up to 52-bit physical addresses. Also, the physical-page size is selectable between 4 Kbytes and 2 Mbytes using the page-directory-entry page-size field (PS). Long mode requires PAE to be enabled in order to use the 64-bit page-translation data structures to translate 64-bit virtual addresses to 52-bit physical addresses. See "PAE Paging" on page 156 for more information on physical-address extensions. Machine-Check Enable (MCE) Bit. Bit 6. Setting MCE to 1 enables the machine-check exception mechanism. Clearing this bit to 0 disables the mechanism. When enabled, a machine-check exception (#MC) occurs when an uncorrectable machine-check error is encountered. Regardless of whether machine-check exceptions are enabled, the processor records enabled-errors when they occur. Errorreporting is performed by the machine-check error-reporting register banks. Each bank includes a control register for enabling error reporting and a status register for capturing errors. Correctable machine-check errors are also reported, but they do not cause a machine-check exception.
62
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
See Chapter 9, "Machine Check Mechanism," for a description of the machine-check mechanism, the registers used, and the types of errors captured by the mechanism. Page-Global Enable (PGE) Bit. B i t 7 . Wh e n p a g e t ra n s l a t i o n i s enabled, system-software performance can often be improved by making some page translations global to all tasks and procedures. Setting PGE to 1 enables the global-page mechanism. Clearing this bit to 0 disables the mechanism. When PGE is enabled, system software can set the global-page (G) bit in the lowest level of the page-translation hierarchy to 1, indicating that the page translation is global. Page translations marked as global are not invalidated in the TLB when the pagetranslation-table base address (CR3) is updated. When the G bit is cleared, the page translation is not global. All supported physical-page sizes also support the global-page mechanism. See "Global Pages" on page 173 for information on using the global-page mechanism. Performance-Monitoring Counter Enable (PCE) Bit. Bit 8. Setting PCE to 1 allows software running at any privilege level to use the RDPMC instruction. Software uses the RDPMC instruction to read the four performance-monitoring MSRs, PerfCTR[3:0]. Clearing PCE to 0 allows only the most-privileged software (CPL=0) to use the RDPMC instruction. FXSAVE/FXRSTOR Support (OSFXSR) Bit. Bit 9. System software must set the OSFXSR bit to 1 to enable use of the 128-bit media instructions. When this bit is set to 1, it also indicates that system software uses the FXSAVE and FXRSTOR instructions to save and restore the processor state for the x87, 64-bit media, and 128-bit media instructions. Clearing the OSFXSR bit to 0 indicates that 128-bit media instructions cannot be used. Attempts to use those instructions while this bit is clear result in an invalid-opcode exception (#UD). Software can continue to use the FXSAVE/FXRSTOR instructions for saving and restoring the processor state for the x87 and 64-bit media instructions. Unmasked Exception Support (OSXMMEXCPT) Bit. B i t 1 0 . S y s t e m software must set the OSXMMEXCPT bit to 1 when it supports the SIMD floating-point exception (#XF) for handling of unmasked 128-bit media floating-point errors. Clearing the Chapter 3: System Resources 63
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
OSXMMEXCPT bit to 0 indicates the #XF handler is not supported. When OSXMMEXCPT=0, unmasked 128-bit media floating-point exceptions cause an invalid-opcode exception (#UD). See "SIMD Floating-Point Exception Causes" in Volume 1 for more information on 128-bit media unmasked floating-point exceptions. 3.1.4 CR1 and CR5-CR7 Registers 3.1.5 64-Bit-Mode Extended Control Registers Control registers CR1, CR5-CR7, and CR9-CR15 are reserved. Attempts by software to use these registers result in an undefined-opcode exception (#UD). In 64-bit mode, additional encodings for control registers are available. The REX.R bit, in a REX prefix, is used to modify the ModRM reg field when that field encodes a control register, as shown in "REX Prefix-Byte Fields" in Volume 3. These additional encodings enable the processor to address CR8-CR15. One additional control register, CR8, is defined in 64-bit mode for all hardware implementations, as described in "CR8 (Task Priority Register, TPR)," below. Access to the CR9-CR15 registers is implementation-dependent. Any attempt to access an unimplemented register results in an invalid-opcode exception (#UD). 3.1.6 CR8 (Task Priority Register, TPR) The AMD64 architecture introduces a new control register, CR8, defined as the task priority register (TPR). The register is accessible in 64-bit mode using the REX prefix. See "External Interrupt Priorities" on page 272 for a description of the TPR and how system software can use the TPR for controlling external interrupts. The RFLAGS register contains two different types of information: Control bits provide system-software controls and directional information for string operations. Some of these bits can have privilege-level restrictions. Status bits provide information resulting from logical and arithmetic operations. These are written by the processor and can be read by software running at any privilege level. Figure 3-8 on page 65 shows the format of the RFLAGS register. The legacy EFLAGS register is identical to the low 32 bits of the register shown in Figure 3-8 (RFLAGS bits 31-0). The term
3.1.7 RFLAGS Register
64
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
rFLAGS is used to refer to the 16-bit, 32-bit, or 64-bit flags register, depending on context.
63 Reserved, RAZ 32
31 Reserved, RAZ
22
21
I D
20
V I P
19
V I F
18
A C
17
V M
16
R F
15 0
14
N T
13
12
11
O F
10
D F
9
I F
8
T F
7
S F
6
Z F
5 0
4
A F
3 0
2
P F
1 1
0
C F
IOPL
Bits Mnemonic 63-22 Reserved 21 ID 20 VIP 19 VIF 18 AC 17 VM 16 RF 15 Reserved 14 NT 13-12 IOPL 11 OF 10 DF 9 IF 8 TF 7 SF 6 ZF 5 Reserved 4 AF 3 Reserved 2 PF 1 Reserved 0 CF
Description Reserved, Read as Zero ID Flag Virtual Interrupt Pending Virtual Interrupt Flag Alignment Check Virtual-8086 Mode Resume Flag Reserved, Read as Zero Nested Task I/O Privilege Level Overflow Flag Direction Flag Interrupt Flag Trap Flag Sign Flag Zero Flag Reserved, Read as Zero Auxiliary Flag Reserved, Read as Zero Parity Flag Reserved, Read as One Carry Flag
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Figure 3-8.
RFLAGS Register The functions of the RFLAGS control and status bits used by application software are described in "Flags Register" in Volume 1. The functions of RFLAGS system bits are (unless otherwise noted, all bits are read/write): Trap Flag (TF) Bit. Bit 8. Software sets the TF bit to 1 to enable single-step mode during software debug. Clearing this bit to 0 disables single-step mode. When single-step mode is enabled, a debug exception (#DB) occurs after each instruction completes execution. Single stepping begins with the instruction following the instruction that sets TF. Single stepping is disabled (TF=0) when the #DB exception occurs or when any exception or interrupt occurs.
Chapter 3: System Resources
65
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
See "Single Stepping" on page 400 for information on using the single-step mode during debugging. Interrupt Flag (IF) Bit. Bit 9. Software sets the IF bit to 1 to enable maskable interrupts. Clearing this bit to 0 causes the processor to ignore maskable interrupts. The state of the IF bit does not affect the response of a processor to non-maskable interrupts, software-interrupt instructions, or exceptions. The ability to modify the IF bit depends on several factors: The current privilege-level (CPL) The I/O privilege level (RFLAGS.IOPL) Whether or not virtual-8086 mode extensions are enabled (CR4.VME=1) Whether or not protected-mode virtual interrupts are enabled (CR4.PVI=1) See "Masking External Interrupts" on page 245 for information on interrupt masking. See "Accessing the RFLAGs Register" on page 187 for information on the specific instructions used to modify the IF bit. I/O Privilege Level Field (IOPL) Field. B i t s 1 3 - 1 2 . Th e I O P L f i e l d specifies the privilege level required to execute I/O addressspace instructions (i.e., instructions that address the I/O space rather than memory-mapped I/O, such as IN, OUT, INS, OUTS, etc.). For software to execute these instructions, the current privilege-level (CPL) must be equal to or higher than (lower numerical value than) the privilege specified by IOPL (CPL <= IOPL). If the CPL is lower than (higher numerical value than) that specified by the IOPL (CPL > IOPL), the processor causes a general-protection exception (#GP) when software attempts to execute an I/O instruction. See "Protected-Mode I/O" in Volume 1 for information on how IOPL controls access to address-space I/O. Virtual-8086 mode uses IOPL to control virtual interrupts and the IF bit when virtual-8086 mode extensions are enabled ( C R 4 . V M E = 1 ) . Th e p ro t e c t e d - m o d e v i r t u a l - i n t e r r u p t mechanism (PVI) also uses IOPL to control virtual interrupts and the IF bit when PVI is enabled (CR4.PVI=1). See "Virtual Interrupts" on page 295 for information on how IOPL is used by the virtual interrupt mechanism.
66
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Nested Task (NT) Bit. Bit 14, IRET reads the NT bit to determine whether the current task is nested within another task. When NT is set to 1, the current task is nested within another task. When NT is cleared to 0, the current task is at the top level (not nested). The processor sets the NT bit during a task switch resulting from a CALL, interrupt, or exception through a task gate. When an IRET is executed from legacy mode while the NT bit is set, a task switch occurs. See "Task Switches Using Task Gates" on page 377 for information on switching tasks using task gates, and "Nesting Tasks" on page 379 for information on task nesting. Resume Flag (RF) Bit. Bit 16. The RF bit allows an instruction to be restarted following an instruction breakpoint resulting in a debug exception (#DB). This bit prevents multiple debug exceptions from occurring on the same instruction. The processor clears the RF bit after every instruction is successfully executed, except when the instruction is: An IRET that sets the RF bit. JMP, CALL, or INTn through a task gate. In both of the above cases, RF is not cleared to 0 until the next instruction successfully executes. When an exception occurs (or when a string instruction is interrupted), the processor normally sets RF=1 in the rFLAGS image saved on the interrupt stack. However, when a #DB exception occurs as a result of an instruction breakpoint, the processor clears the RF bit to 0 in the interrupt-stack rFLAGS image. For instruction restart to work properly following an instruction breakpoint, the #DB exception handler must set RF to 1 in the interrupt-stack rFLAGS image. When an IRET is later executed to return to the instruction that caused the instructionbreakpoint #DB exception, the set RF bit (RF=1) is loaded from the interrupt-stack rFLAGS image. RF is not cleared by the processor until the instruction causing the #DB exception successfully executes. Virtual-8086 Mode (VM) Bit. Bit 17. Software sets the VM bit to 1 to enable virtual-8086 mode. Software clears the VM bit to 0 to Chapter 3: System Resources 67
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
disable virtual-8086 mode. System software can only change this bit using a task switch or an IRET. It cannot modify the bit using the POPFD instruction. Alignment Check (AC) Bit. Bit 18. Software enables automatic alignment checking by setting the AC bit to 1 when CR0.AM=1. Alignment checking can be disabled by clearing either AC or CR0.AM to 0. When automatic alignment checking is enabled and the current privilege-level (CPL) is 3 (least privileged), a memory reference to an unaligned operand causes an alignment-check exception (#AC). Virtual Interrupt (VIF) Bit. Bit 19. The VIF bit is a virtual image of the RFLAGS.IF bit. It is enabled when either virtual-8086 mode extensions are enabled (CR4.VME=1) or protected-mode v i r t u a l i n t e r r u p t s a re e n ab l e d ( C R 4 . P V I =1 ) , a n d t h e RFLAGS.IOPL field is less than 3. When enabled, instructions that ordinarily would modify the IF bit actually modify the VIF bit with no effect on the RFLAGS.IF bit. System software that supports virtual-8086 mode should enable the VIF bit using CR4.VME. This allows 8086 software to execute instructions that can set and clear the RFLAGS.IF bit without causing an exception. With VIF enabled in virtual-8086 mode, those instructions set and clear the VIF bit instead, giving the appearance to the 8086 software that it is modifying the RFLAGS.IF bit. System software reads the VIF bit to determine whether or not to take the action desired by the 8086 software (enabling or disabling interrupts by setting or clearing the RFLAGS.IF bit). In long mode, the use of the VIF bit is supported when CR4.PVI=1. See "Virtual Interrupts" on page 295 for more information on virtual interrupts. Virtual Interrupt Pending (VIP) Bit. Bit 20. The VIP bit is provided as an extension to both virtual-8086 mode and protected mode. It is used by system software to indicate that an external, maskable interrupt is pending (awaiting) execution by either a virtual-8086 mode or protected-mode interrupt-service routine. S o f t wa re mu s t e n ab l e v i r t u a l - 8 0 8 6 m o d e e x t e n s i o n s ( C R 4 . V M E = 1 ) o r p ro t e c t e d - m o d e v i r t u a l i n t e r r u p t s (CR4.PVI=1) before using VIP. VIP is normally set to 1 by a protected-mode interrupt-service routine that was entered from virtual-8086 mode as a result of 68 Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
an external, maskable interrupt. Before returning to the virtual-8086 mode application, the service routine sets VIP to 1 if EFLAGS.VIF=1. When the virtual-8086 mode application attempts to enable interrupts by clearing EFLAGS.VIF to 0 while VIP=1, a general-protection exception (#GP) occurs. The #GP service routine can then decide whether to allow the virtual-8086 mode service routine to handle the pending external, maskable interrupt. (EFLAGS is specifically referred to in this case because virtual-8086 mode is supported only from legacy mode.) In long mode, the use of the VIP bit is supported when CR4.PVI=1. See "Virtual Interrupts" on page 295 for more information on virtual-8086 mode interrupts and the VIP bit. Processor Feature Identification (ID) Bit. B i t 2 1 . T h e a b i l i t y o f software to modify this bit indicates that the processor implementation supports the CPUID instruction. See "Processor Feature Identification" on page 78 for more information on the CPUID instruction. 3.1.8 Extended Feature Enable Register (EFER) The extended-feature-enable register (EFER) contains control bits that enable additional processor features not controlled by the legacy control registers. The EFER is a model-specific register (MSR) with an address of C000_0080h (see "ModelSpecific Registers (MSRs)" on page 71 for more information on MSRs). It can be read and written only by privileged software. Figure 3-9 on page 70 shows the format of the EFER register.
Chapter 3: System Resources
69
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
63 Reserved, MBZ
32
31
15
14
F F X S R
13
12
11
N X E
10
L M A
9
8
L M E
7
1
0
S C E
Reserved, MBZ
MBZ
r
Reserved, RAZ
Bits Mnemonic 63-15 Reserved, MBZ 14 FFXSR 13-12 Reserved, MBZ 11 NXE 10 LMA 9 Reserved, MBZ 8 LME 7-1 Reserved, RAZ 0 SCE
Description Reserved, Must be Zero Fast FXSAVE/FXRSTOR Reserved, Must be Zero No-Execute Enable Long Mode Active Reserved, Must be Zero Long Mode Enable Reserved, Read as Zero System Call Extensions
R/W
R/W R
R/W
R/W
Figure 3-9.
Extended Feature Enable Register (EFER) The function of the EFER bits are (unless otherwise noted, all bits are read/write): System-Call Extension (SCE) Bit. Bit 0. Setting this bit to 1 enables the SYSCALL and SYSRET instructions. Application software can use these instructions for low-latency system calls and returns in a non-segmented (flat) address space. See "Fast S y s t e m C a l l a n d R e t u r n " o n p a g e 1 8 1 fo r a dd i t i o n a l information. Long Mode Enable (LME) Bit. Bit 8. Setting this bit to 1 enables the processor to activate long mode. Long mode is not activated until software enables paging some time later. When paging is enabled after LME is set to 1, the processor sets the EFER.LMA bit to 1, indicating that long mode is not only enabled but also active. See Chapter 14, "Processor Initialization and LongMode Activation," for more information on activating long mode. Long Mode Active (LMA) Bit. Bit 10, read-only. This bit indicates that long mode is active. The processor sets LMA to 1 when both long mode and paging have been enabled by system software.
70
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
See Chapter 14, "Processor Initialization and Long-Mode Activation," for more information on activating long mode. When LMA=1, the processor is running either in compatibility mode or 64-bit mode, depending on the value of the L bit in a code-segment descriptor, as shown in Figure 1-6 on page 14. When LMA=0, the processor is running in legacy mode. In this mode, the processor behaves like a standard 32-bit x86 processor, with none of the new 64-bit features enabled. No-Execute Enable (NXE) Bit. Bit 11. Setting this bit to 1 enables the no-execute page-protection feature. The feature is disabled when this bit is cleared to 0. See "No Execute (NX) Bit" on page 175 for more information. Before setting NXE, system software should verify the processor supports the feature by examining the extendedfeature flags returned by the CPUID instruction. For more information, see "Function 8000_0001h: Processor Signature and AMD Features" in Volume 3. Fast FXSAVE/FXRSTOR (FFXSR) Bit. B i t 1 4 . S e t t i n g t h i s b i t t o 1 enables the FXSAVE and FXRSTOR instructions to execute faster in 64-bit mode at CPL 0. This is accomplished by not saving or restoring the XMM registers (XMM0-XMM15). The FFXSR bit has no effect when the FXSAVE/FXRSTOR instructions are executed in non 64-bit mode, or when CPL > 0. The FFXSR bit does not affect the save/restore of the legacy x87 floating-point state, or the save/restore of MXCSR. Before setting FFXSR, system software should verify whether this feature is supported by examining the CPUID extended feature flags returned by the CPUID instruction. For more information, see "Function 8000_0001h: Processor Signature and AMD Features" in Volume 3.
3.2
Model-Specific Registers (MSRs)
Processor implementations provide model-specific registers (MSRs) for software control over the unique features supported by that implementation. Software reads and writes MSRs using t h e p r iv i l e g e d R D M S R a n d W R M S R i n s t r u c t i o n s . Implementations of the AMD64 architecture can contain a mixture of two basic MSR types:
Chapter 3: System Resources
71
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Legacy MSRs. The AMD family of processors often share model-specific features with other x86 processor implementations. Where possible, AMD implementations use the same MSRs for the same functions. For example, the memory-typing and debug-extension MSRs are implemented on many AMD and non-AMD processors. AMD model-specific MSRs. There are many MSRs common to the AMD family of processors but not to legacy x86 processors. Where possible, AMD implementations use the same AMD-specific MSRs for the same functions. Every model-specific register, as the name implies, is not necessarily implemented by all members of the AMD family of processors. Appendix A, "MSR Cross-Reference," lists MSRaddress ranges currently used by various AMD and other x86 processors. The AMD64 architecture includes a number of features that are controlled using MSRs. Those MSRs are shown in Figure 3-10. The EFER register--described in "Extended Feature Enable Register (EFER)" on page 69--is also an MSR.
72
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
System-Configuration Register SYSCFG System-Linkage Registers STAR LSTAR CSTAR SFMASK FS.base GS.base KernelGSbase SYSENTER_CS SYSENTER_ESP SYSENTER_EIP Debug-Extension Registers DebugCtlMSR LastBranchFromIP LastBranchToIP LastIntFromIP LastIntToIP
Memory-Typing Registers MTRRcap MTRRdefType MTRRphysBasen MTRRphysMaskn MTRRfixn PAT TOP_MEM TOP_MEM2 Performance-Monitoring Registers TSC PerfEvtSeln PerfCtrn Machine-Check Registers MCG_CAP MCG_STAT MCG_CTL MCi_CTL MCi_STATUS MCi_ADDR MCi_MISC
513-262.eps
Figure 3-10.
AMD64 Architecture Model-Specific Registers
The following sections briefly describe the MSRs in the AMD64 architecture. 3.2.1 System Configuration Register (SYSCFG) The system-configuration register (SYSCFG) contains control bits for enabling and configuring system bus features. SYSCFG is a model-specific register (MSR) with an address of C001_0010h. Figure 3-11 shows the format of the SYSCFG register. Some features are implementation specific, and are described in the BIOS writer's guide for the implementation. Implementation-specific features are not shown in Figure 3-11.
Chapter 3: System Resources
73
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
31 Reserved
22
21
T O M 2
20
M V D M
19
M F D M
18
M F D E
17 Reserved
0
Bits 31-22 21 20 19 18 17-0
Mnemonic Reserved TOM2 MVDM MFDM MFDE Reserved
Description MtrrTom2En MtrrVarDramEn MtrrFixDramModEn MtrrFixDramEn
R/W R/W R/W R/W R/W
Figure 3-11. System-Configuration Register (SYSCFG) The function of the SYSCFG bits are (all bits are read/write unless otherwise noted): MtrrFixDramEn Bit. Bit 18. Setting this bit to 1 enables use of the RdMem and WrMem attributes in the fixed-range MTRR registers. When cleared, these attributes are disabled. The RdMem and WrMem attributes allow system software to define fixed-range IORRs using the fixed-range MTRRs. See "Extended Fixed-Range MTRR Type-Field Encodings" on page 235 for information on using this feature. MtrrFixDramModEn Bit. Bit 19. Setting this bit to 1 allows software to read and write the RdMem and WrMem bits. When cleared, writes do not modify the RdMem and WrMem bits, and reads return 0. See "Extended Fixed-Range MTRR Type-Field Encodings" on page 235 for information on using this feature. MtrrVarDramEn Bit. Bit 20. Setting this bit to 1 enables the TOP_MEM register and the variable-range IORRs. These registers are disabled when the bit is cleared to 0. See "IORRs" on page 237 and "Top of Memory" on page 239 for information on using these features. MtrrTom2En Bit. B i t 2 1 . S e t t i n g t h i s b i t t o 1 e n a b l e s t h e TOP_MEM2 register. The register is disabled when this bit is cleared to 0. See "Top of Memory" on page 239 for information on using this feature. 3.2.2 System-Linkage Registers System-linkage MSRs are used by system software to allow fast control transfers between applications and the operating system. The functions of these registers are: Chapter 3: System Resources
74
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
STAR, LSTAR, CSTAR, and SFMASK Registers. These registers are used to provide mode-dependent linkage information for the SYSCALL and SYSRET instructions. STAR is used in legacy modes, LSTAR in 64-bit mode, and CSTAR in compatibility mode. SFMASK is used by the SYSCALL instruction for rFLAGS in long mode. FS.base and GS.base Registers. These registers allow 64-bit baseaddress values to be specified for the FS and GS segments, for use in 64-bit mode. See "FS and GS Registers in 64-Bit Mode" on page 88 for a description of the special treatment the FS and GS segments receive. KernelGSbase Register. This register is used by the SWAPGS instruction. This instruction exchanges the value located in KernelGSbase with the value located in GS.base. SYSENTERx Registers. The SYSENTER_CS, SYSENTER_ESP, and SYSENTER_EIP registers are used to provide linkage information for the SYSENTER and SYSEXIT instructions. These instructions are only used in legacy mode. The system-linkage instructions and their use of MSRs are described in "Fast System Call and Return" on page 181. 3.2.3 Memory-Typing Registers Memory-typing MSRs are used to characterize, or type, memory. Memory typing allows software to control the cacheability of memory, and determine how accesses to memory are ordered. The memory-typing registers perform the following functions: MTRRcap Register. This register contains information describing the level of MTRR support provided by the processor. MTRRdefType Register. This register establishes the default memory type to be used for physical memory that is not specifically characterized using the fixed-range and variablerange MTRRs. MTRRphysBasen and MTRRphysMaskn Registers. These registers form a register pair that can be used to characterize any address range within the physical-memory space, including all of physical memory. Up to eight address ranges of varying sizes can be characterized using these registers.
Chapter 3: System Resources
75
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
MTRRfixn Registers. These registers are used to characterize fixedsize memory ranges in the first 1 Mbytes of physical-memory space. PAT Register. This register allows memory-type characterization based on the virtual (linear) address. It is an extension to the PCD and PWT memory types supported by the legacy paging mechanism. The PAT mechanism provides the same memorytyping capabilities as the MTRRs, but with the added flexibility provided by the paging mechanism. TOP_MEM and TOP_MEM2 Registers. These top-of-memory registers allow system software to specify physical addresses ranges as memory-mapped I/O locations. Refer to "Memory-Type Range Registers" on page 217 for more information on using these registers. 3.2.4 DebugExtension Registers The debug-extension MSRs provide software-debug capability not available in the legacy debug registers (DR0-DR7). These MSRs allow single stepping and recording of control transfers to take place. The debug-extension registers perform the following functions: DebugCtlMSR Register. This register provides control over controltransfer recording and single stepping, and external-breakpoint reporting and trace messages. LastBranchx and LastExceptionx Registers. Th e four re g i s t e r s , LastBranchToIP, LastBranchFromIP, LastExceptionToIP, and LastExceptionFromIP, are all used to record the source and target of control transfers when branch recording is enabled. Refer to "Control-Transfer Breakpoint Features" on page 401 for more information on using these debug registers. 3.2.5 PerformanceMonitoring Registers The time-stamp counter and performance-monitoring registers are useful in identifying performance bottlenecks. There can be any number of performance counters, each numbered from 0 to n. These registers perform the following functions: TSC Register. This register is used to count processor-clock cycles. It can be read using the RDMSR instruction, or it can be read using the read time-stamp counter instruction, RDTSC. System software can make RDTSC available for use by non-privileged software by clearing the time-stamp disable bit (CR4.TSD) to 0. 76 Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
PerfEvtSeln Registers. These registers are used to specify the events counted by the corresponding performance counter, and to control other aspects of its operation. PerfCtrn Registers. These registers are performance counters that hold a count of processor events or the duration of events, under the control of the corresponding PerfEvtSeln register. Each PerfCtrn register can be read using the RDMSR instruction, or they can be read using the read performancemonitor counter instruction, RDPMC. System software can make RDPMC available for use by non-privileged software by setting the performance-monitor counter enable bit (CR4.PCE) to 1. Refer to "Using Performance Counters" on page 408 for more information on using these registers. 3.2.6 Machine-Check Registers The machine-check registers control the detection and reporting of hardware machine-check errors. The types of errors that can be reported include cache-access errors, load-data and store-data errors, bus-parity errors, and ECC errors. Two types of machine-check MSRs are shown in Figure 3-10 on page 73. The first type is global machine-check registers, which perform the following functions: MCG_CAP Register. This register identifies the machine-check capabilities supported by the processor. MCG_CTL Register. This register provides global control over machine-check-error reporting. MCG_STATUS Register. This register reports global status on detected machine-check errors. The second type is error-reporting register banks, which report on machine-check errors associated with a specific processor unit (or group of processor units). There can be different numbers of register banks for each processor implementation, and each bank is numbered from 0 to i. The registers in each bank perform the following functions: MCi_CTL Registers. These registers control error-reporting. MCi_STATUS Registers. These registers report machine-check errors. Chapter 3: System Resources 77
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
MCi_ADDR Registers. These registers report the machine-check error address. MCi_MISC Registers. These registers report miscellaneous-error information. Refer to "Using Machine Check Features" on page 316 for more information on using these registers.
3.3
Processor Feature Identification
The CPUID instruction provides information about the processor implementation and its capabilities. Software operating at any privilege level can execute the CPUID instruction to collect this information. After the information is collected, software can be tuned to optimize performance and benefit to users. For example, game software can identify and enable the media capabilities of a particular processor implementation. The CPUID instruction supports multiple functions, each p rov i d i n g d i f f e re n t i n fo r m a t i o n ab o u t t h e p ro c e s s o r implementation, including the vendor, model number, revision (stepping), features, cache organization, and name. The multifunction approach allows the CPUID instruction to return a detailed picture of the processor implementation and its capabilities -- more detailed information than could be returned by a single function. This flexibility also allows for the addition of new CPUID functi ons in future processor generations. Function codes are loaded into the EAX register before executing the CPUID instruction. CPUID functions are divided into two types: Standard functions include the earliest features offered in the x86 architecture. Extended functions include newer features of the x86 and AMD64 architectures, such as SSE, SSE2, and 3DNow! instructions, and long mode. See "CPUID" in Volume 3 for details on the operation of this instruction, and the AMD AMD64 Processor Recognition Application Note Addendum (order no. 25481) for information returned by each processor implementation.
78
Chapter 3: System Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
4
Segmented Virtual Memory
The legacy x86 architecture supports a segment-translation mechanism that allows system software to relocate and isolate instructions and data anywhere in the virtual-memory space. A segment is a contiguous block of memory within the linear address space. The size and location of a segment within the linear address space is arbitrary. Instructions and data can be assigned to one or more memory segments, each with its own protection characteristics. The processor hardware enforces the rules dictating whether one segment can access another segment. The segmentation mechanism provides ten segment registers, each of which defines a single segment. Six of these registers (CS, DS, ES, FS, GS, and SS) define user segments. User segments hold software, data, and the stack and can be used by both application software and system software. The remaining four segment registers (GDT, LDT, IDT, and TR) define system segments. System segments contain data structures initialized and used only by system software. Segment registers contain a base address pointing to the starting location of a segment, a limit defining the segment size, and attributes defining the segment-protection characteristics. Although segmentation provides a great deal of flexibility in relocating and protecting software and data, it is often more efficient to handle memory isolation and relocation with a combination of software and hardware paging support. For this re a s o n , m o s t m o d e r n s y s t e m s o f t wa re b y p a s s e s t h e segmentation features. However, segmentation cannot be completely disabled, and an understanding of the segmentation mechanism is important to implementing long-mode system software. In long mode, the effects of segmentation depend on whether the processor is running in compatibility mode or 64-bit mode: In compatibility mode, segmentation functions just as it does in legacy mode, using legacy 16-bit or 32-bit protected mode semantics. 64-bit mode, segmentation is disabled, creating a flat 64-bit virtual-address space. As will be seen, certain functions of
Chapter 4: Segmented Virtual Memory
79
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
some segment registers, particularly the system-segment registers, continue to be used in 64-bit mode.
4.1
Real Mode Segmentation
After reset or power-up, the processor always initially enters real mode. Protected Modes are entered from real mode. As noted in "Real Addressing" on page 11, Real mode (realaddress mode), provides a physical-memory space of 1 Mbyte. In this mode, a 20-bit physical address is determined by shifting a 16-bit segment selectors to the left four bits and adding the 16-bit effective address. Each 64K segment (CS, DS, ES, FS, GS, SS) is aligned on 16-byte boundaries. The segment base is the lowest address in a given segment, and is equal to the segment selector * 16. The POP and MOV instructions can be used to load a (possibly) new segment selector into one of the segment registers. When this occurs, the selector is updated and the selector base is set to selector * 16. The segment limit and segment attributes are unchanged, but are normally 64K (the maximum allowable limit) and read/write data, respectively. On FAR transfers, CS (code segment) selector is updated to the new value, and the CS segment base is set to selector * 16. The CS segment limit and attributes are unchanged, but are usually 64K and read/write, respectively. If the interrupt descriptor table (IDT) is used to find the real mode IDT see "Real-Mode Interrupt Control Transfers" on page 274. The GDT, LDT, and TSS (see below) are not used in real mode.
4.2
Virtual-8086 Mode Segmentation
Virtual-8086 mode supports 16-bit real mode programs running under protected mode (see below). It uses a simple form of memory segmentation, optional paging, and limited protection checking. Programs running in virtual-8086 mode can access up to 1MB of memory space. As with real mode segmentation, each 64K segment (CS, DS, ES, FS, GS, SS) is aligned on 16-byte boundaries. The segment
80
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
base is the lowest address in a given segment, and is equal to the segment selector * 16. The POP and MOV instructions work exactly as in real mode and can be used to load a (possibly) new segment selector into one of the segment registers. When this occurs, the selector is updated and the selector base is set to selector * 16. The segment limit and segment attributes are unchanged, but are normally 64K (the maximum allowable limit) and read/write data, respectively. FAR transfers, with the exception of interrupts and exceptions, operate as in real mode. On FAR transfers, CS (code segment) selector is updated to the new value, and the CS segment base is set to selector * 16. The CS segment limit and attributes are unchanged, but are usually 64K and read/write, respectively. Interrupts and exceptions switch the processor to protected mode. (See Chapter 8, "Exceptions and Interrupts" for more information.)
4.3
Protected Mode Segmented-Memory Models
System software can use the segmentation mechanism to support one of two basic segmented-memory models: a flatm e m o ry m o d e l o r a m u l t i - s e g m e n t e d m o d e l . T h e s e segmentation models are supported in legacy mode and in compatibility mode. Each type of model is described in the following sections.
4.3.1 MultiSegmented Model
In the multi-segmented memory model, each segment register can reference a unique base address with a unique segment size. Segments can be as small as a single byte or as large as 4 Gbytes. When page translation is used, multiple segments can be mapped to a single page and multiple pages can be mapped to a single segment. Figure 1-1 on page 7 shows an example of the multi-segmented model. The multi-segmented memory model provides the greatest level of flexibility for system software using the segmentation mechanism. Compatibility mode allows the multi-segmented model to be used in support of legacy software. However, in compatibility mode, the multi-segmented memory model is restricted to the first 4 Gbytes of virtual-memory space. Access to virtual memory above 4 Gbytes requires the use of 64-bit mode, which does not support segmentation.
Chapter 4: Segmented Virtual Memory
81
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
4.3.2 Flat-Memory Model
The flat-memory model is the simplest form of segmentation to implement. Although segmentation cannot be disabled, the flat-memory model allows system software to bypass most of the segmentation mechanism. In the flat-memory model, all segment-base addresses have a value of 0 and the segment limits are fixed at 4 Gbytes. Clearing the segment-base value to 0 effectively disables segment translation, resulting in a single segment spanning the entire virtual-address space. All segment descriptors reference this single, flat segment. Figure 1-2 on page 8 shows an example of the flat-memory model. In 64-bit mode, segmentation is disabled. The segment-base value is ignored and treated as 0 by the segmentation hardware. Likewise, segment limits and most attributes are ignored. There are a few exceptions. The CS-segment DPL, D, and L attributes are used (respectively) to establish the privilege level for a program, the default operand size, and whether the program is running in 64-bit mode or compatibility mode. The FS and GS segments can be used as additional base registers in address calculations, and those segments can have non-zero baseaddress values. This facilitates addressing thread-local data and certain system-software data structures. See "FS and GS Registers in 64-Bit Mode" on page 88 for details about the FS and GS segments in 64-bit mode. The system-segment registers are always used in 64-bit mode.
4.3.3 Segmentation in 64-bit mode
4.4
Segmentation Data Structures and Registers
Figure 4-1 on page 83 shows the following data structures used by the segmentation mechanism: Segment Descriptors--As the name implies, a segment descriptor describes a segment, including its location in virtual-address space, its size, protection characteristics, and other attributes. Descriptor Tables--Segment descriptors are stored in memory in one of three tables. The global-descriptor table (GDT) holds segment descriptors that can be shared among all tasks. Multiple local-descriptor tables (LDT) can be defined to hold descriptors that are used by specific tasks and are not shared globally. The interrupt-descriptor table (IDT) holds gate descriptors that are used to access the segments where interrupt handlers are located.
82
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Task-State Segment--A task-state segment (TSS) is a special type of system segment that contains task-state information and data structures for each task. For example, a TSS holds a copy of the GPRs and EFLAGS register when a task is suspended. A TSS also holds the pointers to privilegedsoftware stacks. The TSS and task-switch mechanism are described in Chapter 12, "Task Management." Segment Selectors--Descriptors are selected for use from the descriptor tables using a segment selector. A segment selector contains an index into either the GDT or LDT. The IDT is indexed using an interrupt vector, as described in "Legacy Protected-Mode Interrupt Control Transfers" on page 276, and in "Long-Mode Interrupt Control Transfers" on page 287.
Global-Descriptor Table (GDT) Descriptor Descriptor ... Descriptor Segment Selectors Selector 1 Selector 2 ... Selector n Interrupt-Descriptor Table (IDT) Gate Descriptor Gate Descriptor ... Gate Descriptor
513-263.eps
Segment Descriptors Code Stack Data Gate Task-State Segment Local-Descriptor Table
Local-Descriptor Table (LDT) Descriptor Descriptor ... Descriptor
Figure 4-1.
Segmentation Data Structures Figure 4-2 on page 84 shows the registers used by the segmentation mechanism. The registers have the following relationship to the data structures:
Chapter 4: Segmented Virtual Memory
83
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Segment Registers--The six segment registers (CS, DS, ES, FS, GS, and SS) are used to point to the user segments. A segment selector selects a descriptor when it is loaded into one of the segment registers. This causes the processor to automatically load the selected descriptor into a softwareinvisible portion of the segment register. Descriptor-Table Registers--The three descriptor-table registers (GDTR, LDTR, and IDTR) are used to point to the system segments. The descriptor-table registers identify the virtual-memory location and size of the descriptor tables. Task Register (TR)--Describes the location and limit of the current task state segment (TSS).
Code Segment Register CS Data Segment Registers DS ES FS GS
Global-Descriptor-Table Register GDTR
Interrupt-Descriptor-Table Register IDTR
Local-Descriptor-Table Register LDTR Task Register
Stack Segment Register SS
TR
513-264.eps
Figure 4-2.
Segment and Descriptor-Table Registers
A fourth system-segment register, the TR, points to the TSS. The data structures and registers associated with task-state segments are described in "Task-Management Resources" on page 360.
4.5
Segment Selectors and Registers
Segment selectors are pointers to specific entries in the global and local descriptor tables. Figure 4-3 on page 85 shows the segment selector format.
4.5.1 Segment Selectors
84
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
15 SI
3
2
T I
1
0
RPL
Bits 15-3 2 1-0
Mnemonic SI TI RPL
Description Selector Index Table Indicator Requestor Privilege Level
R/W R/W R/W R/W
Figure 4-3.
Segment Selector
The selector format consists of the following fields: Selector Index Field. Bits 15-3. The selector-index field specifies an entry in the descriptor table. Descriptor-table entries are eight bytes long, so the selector index is scaled by 8 to form a byte offset into the descriptor table. The offset is then added to either the global or local descriptor-table base address (as indicated by the table-index bit) to form the descriptor-entry address in virtual-address space. Some descriptor entries in long mode are 16 bytes long rather than 8 bytes (see "Legacy Segment Descriptors" on page 97 for more information on long-mode descriptor-table entries). These expanded descriptors consume two entries in the descriptor table. Long mode, however, continues to scale the selector index by eight to form the descriptor-table offset. It is the responsibility of system software to assign selectors such that they correctly point to the start of an expanded entry. Table Indicator (TI) Bit. Bit 2. The TI bit indicates which table holds the descriptor referenced by the selector index. When TI=0 the GDT is used and when TI=1 the LDT is used. The descriptortable base address is read from the appropriate descriptor-table register and added to the scaled selector index as described above. Requestor Privilege-Level (RPL) Field. Bits 1-0. The RPL represents the privilege level (CPL) the processor is operating under at the time the selector is created. RPL is used in segment privilege-checks to prevent software running at lesser privilege levels from accessing privileged data. See "Data-Access Privilege Checks" on page 121 and "Control-Transfer Privilege Checks" on page 124 for more information on segment privilege-checks. Chapter 4: Segmented Virtual Memory 85
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Null Selector. Null selectors have a selector index of 0 and TI=0, corresponding to the first entry in the GDT. However, null selectors do not reference the first GDT entry but are instead used to invalidate unused segment registers. A generalprotection exception (#GP) occurs if a reference is made to use a segment register containing a null selector in non-64-bit mode. By initializing unused segment registers with null selectors software can trap references to unused segments. Null selectors can only be loaded into the DS, ES, FS and GS data-segment registers, and into the LDTR descriptor-table register. A #GP occurs if software attempts to load the CS register with a null selector or if software attempts to load the SS register with a null selector in non 64-bit mode or at CPL 3. 4.5.2 Segment Registers Six 16-bit segment registers are provided for referencing up to six segments at one time. All software tasks require segment selectors to be loaded in the CS and SS registers. Use of the DS, ES, FS, and GS segments is optional, but nearly all software accesses data and therefore requires a selector in the DS register. Table 4-1 lists the supported segment registers and their functions. Table 4-1. Segment Registers
Segment Register CS DS ES FS GS SS Segment Register Function References code-segment descriptor entry References default data-segment descriptor entry References optional data-segment descriptor entry References optional data-segment descriptor entry References optional data-segment descriptor entry References stack segment descriptor entry
The processor maintains a hidden portion of the segment register in addition to the selector value loaded by software. This hidden portion contains the values found in the descriptortable entry referenced by the segment selector. The processor loads the descriptor-table entry into the hidden portion when the segment register is loaded. By keeping the corresponding
86
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
descriptor-table entry in hardware, performance is optimized for the majority of memory references. Figure 4-4 shows the format of the visible and hidden portions of the segment register. Except for the FS and GS segment base, software cannot directly read or write the hidden portion (shown as gray-shaded boxes in Figure 4-4).
Selector Segment Attributes 32-Bit Segment Limit 32-Bit Segment Base Address
Hidden From Software
513-221.eps
Figure 4-4.
Segment-Register Format
CS Register. The CS register contains the segment selector referencing the current code-segment descriptor entry. All instruction fetches reference the CS descriptor. When a new selector is loaded into the CS register, the current-privilege level (CPL) of the processor is set to that of the CS-segment descriptor-privilege level (DPL). Data-Segment Registers. The DS register contains the segment selector referencing the default data-segment descriptor entry. The SS register contains the stack-segment selector. The ES, FS, and GS registers are optionally loaded with segment selectors referencing other data segments. Data accesses default to referencing the DS descriptor except in the following two cases: The ES descriptor is referenced for string-instruction destinations. The SS descriptor is referenced for stack operations. 4.5.3 Segment Registers in 64-bit Mode CS Register in 64-Bit Mode. In 64-bit mode, most of the hidden portion of the CS register is ignored. Only the L (long), D (default operation size), and DPL (descriptor privilege-level) attributes are recognized by 64-bit mode. Address calculations assume a CS.base value of 0. CS references do not check the 87
Chapter 4: Segmented Virtual Memory
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
CS.limit value, but instead check that the effective address is in canonical form. DS, ES, and SS Registers in 64-Bit Mode. In 64-bit mode, the contents of the ES, DS, and SS segment registers are ignored. All fields (base, limit, and attribute) in the hidden portion of the segment registers are ignored. Address calculations in 64-bit mode that reference the ES, DS, or SS segments are treated as if the segment base is 0. Instead of performing limit checks, the processor checks that all virtualaddress references are in canonical form. Neither enabling and activating long mode nor switching between 64-bit and compatibility modes changes the contents of the visible or hidden portions of the segment registers. These registers remain unchanged during 64-bit mode execution unless explicit segment loads are performed. FS and GS Registers in 64-Bit Mode. Unlike the CS, DS, ES, and SS segments, the FS and GS segment overrides can be used in 64bit mode. When FS and GS segment overrides are used in 64-bit mode, their respective base addresses are used in the effectiveaddress (EA) calculation. The complete EA calculation then becomes (FS or GS).base + base + (scale index) + displacement. The FS.base and GS.base values are also expanded to the full 64-bit virtual-address size, as shown in Figure 4-5. The resulting EA calculation is allowed to wrap across positive and negative addresses.
Selector Segment Attributes 32-Bit Segment Limit 64-Bit Segment Base Address
Hidden from Software and Unused in 64-bit Mode
513-267.eps
Figure 4-5.
FS and GS Segment-Register Format--64-Bit Mode
88
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
In 64-bit mode, FS-segment and GS-segment overrides are not checked for limit or attributes. Instead, the processor checks that all virtual-address references are in canonical form. Segment register-load instructions (MOV to Sreg and POP Sreg) load only a 32-bit base-address value into the hidden portion of the FS and GS segment registers. The base-address bits above the low 32 bits are cleared to 0 as a result of a segment-register load. To allow loading all 64 bits of the base address, the FS.base and GS.base hidden descriptor-register fields are mapped to MSRs. Privileged software (CPL=0) can load the 64-bit base address into FS.base or GS.base using a single WRMSR instruction. The addresses written into the expanded FS.base and GS.base registers must be in canonical form. A WRMSR instruction that attempts to write a non-canonical address to these registers causes a general-protection exception (#GP) to occur. The FS.base MSR address is C000_0100h while the GS.base MSR address is C000_0101h. When in compatibility mode, the FS and GS overrides operate as defined by the legacy x86 architecture regardless of the value loaded into the high 32 bits of the hidden descriptorregister base-address field. Compatibility mode ignores the high 32 bits when calculating an effective address.
4.6
Descriptor Tables
Descriptor tables are used by the segmentation mechanism when protected mode is enabled (CR0.PE=1). These tables hold descriptor entries that describe the location, size, and privilege attributes of a segment. All memory references in protected mode access a descriptor-table entry. As previously mentioned, there are three types of descriptor tables supported by the x86 segmentation mechanism: Global descriptor table (GDT) Local descriptor table (LDT) Interrupt descriptor table (IDT) Software establishes the location of a descriptor table in memory by initializing its corresponding descriptor-table
Chapter 4: Segmented Virtual Memory
89
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
register. The descriptor-table registers and the descriptor tables are described in the following sections. 4.6.1 Global Descriptor Table Protected-mode system software must create a global descriptor table (GDT). The GDT contains code-segment and data-segment descriptor entries (user segments) for segments that can be shared by all tasks. In addition to the user segments, the GDT can also hold gate descriptors and other systemsegment descriptors. System software can store the GDT anywhere in memory and should protect the segment containing the GDT from non-privileged software. Segment selectors point to the GDT when the table-index (TI) bit in the selector is cleared to 0. The selector index portion of the segment selector references a specific entry in the GDT. Figure 4-6 on page 91 shows how the segment selector indexes into the GDT. One special form of a segment selector is the null selector. A null selector points to the first entry in the GDT (the selector index is 0 and TI=0). However, null selectors do not reference memory, so the first GDT entry cannot be used to describe a segment (see "Null Selector" on page 86 for information on using the null selector). The first usable GDT entry is referenced with a selector index of 1.
90
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Selector Index
TI
Segment Selector
Global (TI=0) Local (TI=1) Descriptor Table
+
+
Selector Index
000
Unused in GDT Descriptor Table Base Address Descriptor Table Limit
Global or Local Descriptor-Table Register
513-209.eps
Figure 4-6.
Global and Local Descriptor-Table Access The global descriptor-table register (GDTR) points to the location of the GDT in memory and defines its size. This register is loaded from memory using the LGDT instruction (see "LGDT and LIDT Instructions" on page 189). Figure 4-7 shows the format of the GDTR in legacy mode and compatibility mode.
4.6.2 Global Descriptor-Table Register
16-Bit Descriptor-Table Limit 32-Bit Descriptor-Table Base Address
513-220.eps
Figure 4-7.
GDTR and IDTR Format--Legacy Modes Figure 4-8 on page 92 shows the format of the GDTR in 64-bit mode.
Chapter 4: Segmented Virtual Memory
91
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
16-Bit Descriptor-Table Limit 64-Bit Descriptor-Table Base Address
513-266.eps
Figure 4-8.
GDTR and IDTR Format--Long Mode The GDTR contains 2 fields: Limit. 2 bytes. These bits define the 16-bit limit, or size, of the GDT in bytes. The limit value is added to the base address to yield the ending byte address of the GDT. A general-protection exception (#GP) occurs if software attempts to access a descriptor beyond the GDT limit. The offsets into the descriptor tables are not extended by the AMD64 architecture in support of long mode. Therefore, the GDTR and IDTR limit-field sizes are unchanged from the legacy sizes. The processor does check the limits in long mode during GDT and IDT accesses. Base Address. 8 bytes. The base-address field holds the starting byte address of the GDT in virtual-memory space. The GDT can be located at any byte address in virtual memory, but system software should align the GDT on a doubleword boundary to avoid the potential performance penalties associated with accessing unaligned data. The AMD64 architecture increases the base-address field of the GDTR to 64 bits so that system software running in long mode can locate the GDT anywhere in the 64-bit virtual-address space. The processor ignores the high-order 4 bytes of base address when running in legacy mode.
4.6.3 Local Descriptor Table
Protected-mode system software can optionally create a local descriptor table (LDT) to hold segment descriptors belonging to a single task or even multiple tasks. The LDT typically contains code-segment and data-segment descriptors as well as gate descriptors referenced by the specified task. Like the GDT, system software can store the LDT anywhere in memory and
92
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
should protect the segment containing the LDT from nonprivileged software. Segment selectors point to the LDT when the table-index bit (TI) in the selector is set to 1. The selector index portion of the segment selector references a specific entry in the LDT (see Figure 4-6 on page 91). Unlike the GDT, however, a selector index of 0 references the first entry in the LDT (when TI=1, the selector is not a null selector). LDTs are described by system-segment descriptor entries located in the GDT, and a GDT can contain multiple LDT descriptors. The LDT system-segment descriptor defines the location, size, and privilege rights for the LDT. Figure 4-9 shows the relationship between the LDT and GDT data structures. Loading a null selector into the LDTR is useful if software does not use an LDT. This causes a #GP if an erroneous reference is made to the LDT.
Global Descriptor Table
Local Descriptor Table
LDT Selector LDT Attributes GDT Limit GDT Base Address LDT Limit LDT Base Address
Global Descriptor Table Register
Local Descriptor Table Register
513-208.eps
Figure 4-9.
Relationship between the LDT and GDT The local descriptor-table register (LDTR) points to the location of the LDT in memory, defines its size, and specifies its attributes. The LDTR has two portions. A visible portion holds the LDT selector, and a hidden portion holds the LDT descriptor. When the LDT selector is loaded into the LDTR, the processor automatically loads the LDT descriptor from the GDT 93
4.6.4 Local Descriptor-Table Register
Chapter 4: Segmented Virtual Memory
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
into the hidden portion of the LDTR. The LDTR is loaded in one of two ways: Using the LLDT instruction Instructions" on page 189). (see "LLDT and LTR
Performing a task switch (see "Switching Tasks" on page 375). Figure 4-10 shows the format of the LDTR in legacy mode.
Selector Descriptor Attributes 32-Bit Descriptor-Table Limit 32-Bit Descriptor-Table Base Address
Hidden From Software
513-221.eps
Figure 4-10.
LDTR Format--Legacy Mode
Figure 4-11 shows the format of the LDTR in long mode (both compatibility mode and 64-bit mode).
Selector Descriptor Attributes 32-Bit Descriptor-Table Limit 64-Bit Descriptor-Table Base Address
Hidden From Software
513-267.eps
Figure 4-11. LDTR Format--Long Mode
94
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The LDTR contains four fields: LDT Selector. 2 bytes. These bits are loaded explicitly from the TSS during a task switch, or by using the LLDT instruction. The LDT selector must point to an LDT system-segment descriptor entry in the GDT. If it does not, a general-protection exception (#GP) occurs. The following three fields are loaded automatically from the LDT descriptor in the GDT as a result of loading the LDT selector. The register fields are shown as shaded boxes in Figures 4-10 and 4-11. Base Address. The base-address field holds the starting byte address of the LDT in virtual-memory space. Like the GDT, the LDT can be located anywhere in system memory, but software should align the LDT on a doubleword boundary to avoid performance penalties associated with accessing unaligned data. The AMD64 architecture expands the base-address field of the LDTR to 64 bits so that system software running in long mode can locate an LDT anywhere in the 64-bit virtual-address space. The processor ignores the high-order 32 base-address bits when running in legacy mode. Because the LDTR is loaded from the GDT, the system-segment descriptor format (LDTs are system segments) has been expanded by the AMD64 architecture in support of 64-bit mode. See "Long Mode Descriptor Summary" on page 116 for more information on this expanded format. The high-order base-address bits are only loaded from 64-bit mode using the LLDT instruction (see "LLDT and LTR Instructions" on page 189 for more information on this instruction). Limit. This field defines the limit, or size, of the LDT in bytes. The LDT limit as stored in the LDTR is 32 bits. When the LDT limit is loaded from the GDT descriptor entry, the 20-bit limit field in the descriptor is expanded to 32 bits and scaled based on the value of the descriptor granularity (G) bit. For details on the limit biasing and granularity, see "Granularity (G) Bit" on page 100. If an attempt is made to access a descriptor beyond the LDT limit, a general-protection exception (#GP) occurs. The offsets into the descriptor tables are not extended by the AMD64 architecture in support of long mode. Therefore, the Chapter 4: Segmented Virtual Memory 95
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
LDTR limit-field size is unchanged from the legacy size. The processor does check the LDT limit in long mode during LDT accesses. Attributes. This field holds the descriptor attributes, such as privilege rights, segment presence and segment granularity. 4.6.5 Interrupt Descriptor Table The final type of descriptor table is the interrupt descriptor table (IDT). Multiple IDTs can be maintained by system software. System software selects a specific IDT by loading the interrupt descriptor table register (IDTR) with a pointer to the IDT. As with the GDT and LDT, system software can store the IDT anywhere in memory and should protect the segment containing the IDT from non-privileged software. The IDT can contain only the fo llowing types of gate descriptors: Interrupt gates. Trap gates. Task gates. The use of gate descriptors by the interrupt mechanism is described in Chapter 8, "Exceptions and Interrupts." A general-protection exception (#GP) occurs if the IDT descriptor referenced by an interrupt or exception is not one of the types listed above. IDT entries are selected using the interrupt-vector number rather than a selector value. The interrupt-vector number is scaled by the interrupt-descriptor entry size to form an offset into the IDT. The interrupt-descriptor entry size depends on the processor operating mode as follows: In long mode, interrupt descriptor-table entries are 16 bytes. In legacy mode, interrupt descriptor-table entries are eight bytes. Figure 4-12 on page 97 shows how the interrupt-vector number indexes the IDT.
96
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Interrupt Descriptor Table
+
Interrupt Vector + * Descriptor Entry Size
IDT Base Address
IDT Limit
Interrupt Descriptor Table Register
513-207.eps
Figure 4-12.
Indexing an IDT The interrupt descriptor-table register (IDTR) points to the IDT in memory and defines its size. This register is loaded from memory using the LIDT instruction (see "LGDT and LIDT Instructions" on page 189). The format of the IDTR is identical to that of the GDTR in all modes. Figure 4-7 on page 91 shows the format of the IDTR in legacy mode. Figure 4-8 on page 92 shows the format of the IDTR in long mode. The offsets into the descriptor tables are not extended by the AMD64 architecture in support of long mode. Therefore, the IDTR limit-field size is unchanged from the legacy size. The processor does check the IDT limit in long mode during IDT accesses.
4.6.6 Interrupt Descriptor-Table Register
4.7
Legacy Segment Descriptors
Segment descriptors define, protect, and isolate segments from each other. There are two basic types of descriptors, each of which are used to describe different segment (or gate) types: User Segments--These include code segments and data segments. Stack segments are a type of data segment. System Segments--System segments consist of LDT segments and task-state segments (TSS). Gate descriptors
4.7.1 Descriptor Format
Chapter 4: Segmented Virtual Memory
97
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
are another type of system-segment descriptor. Rather than describing segments, gate descriptors point to program entry points. Figure 4-13 shows the generic format for user-segment and system-segment descriptors. User and system segments are differentiated using the S bit. S=1 indicates a user segment, and S=0 indicates a system segment. Gray shading indicates the field or bit is reserved. The format for a gate descriptor differs from the generic segment descriptor, and is described separately in "Gate Descriptors" on page 106.
31 Base Address 31-24
24 23 22 21 20 19 D G/ B
16 15 14 13 12 11 Type
87 Base Address 23-16
0 +4
A Segment V P DPL S Limit 19-16 L
Base Address 15-0
Segment Limit 15-0
+0
Figure 4-13.
Generic Segment Descriptor--Legacy Mode Figure 4-13 shows the fields in a generic, legacy-mode, 8-byte segment descriptor. In this figure, +0 indicates the address of the descriptor's first byte, and +4 indicates the address of the descriptor's fifth byte. The fields are defined as follows, from least-significant to most-significant bit positions: Segment Limit. Th e 2 0 - b i t s e g m e n t l i m i t i s f o r m e d by concatenating bits 19-16 of byte +4 with bits 15-0 of byte +0. The segment limit defines the segment size, in bytes. The granularity (G) bit controls how the segment-limit field is scaled (see "Granularity (G) Bit" on page 100). For data segments, the expand-down (E) bit determines whether the segment limit defines the lower or upper segment-boundary (see "Expand-Down (E) Bit" on page 103). If software references a segment descriptor with an address beyond the segment limit, a general-protection exception (#GP) occurs. The #GP occurs if any part of the memory reference falls outside the segment limit. For example, a doubleword (4-byte) address reference causes a #GP if one or more bytes are located beyond the segment limit.
98
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Base Address. The 32-bit base address is formed by concatenating bits 31-24 of byte +4 with bits 7-0 of byte +4, and with bits 15-0 of byte +0. The segment-base address field locates the start of a segment in virtual-address space. S Bit and Type Field. Bit 12 of byte +4, and bits 11-8 of byte +4. The S and Type fields, together, specify the descriptor type and its access characteristics. Table 4-2 summarizes the descriptor types by S-field encoding and gives a cross reference to descriptions of the Type-field encodings. Table 4-2.
S Field
Descriptor Types
Descriptor Type LDT Type-Field Encoding
0 (System)
TSS Gate
See Table 4-5 on page 105
1 (User)
Code Data
See Table 4-3 on page 102 See Table 4-4 on page 104
Descriptor Privilege-Level (DPL) Field. Bits 14-13 of byte +4. The DPL field indicates the descriptor-privilege level of the segment. DPL can be set to any value from 0 to 3, with 0 specifying the most privilege and 3 the least privilege. See "Data-Access Privilege Checks" on page 121 and "Control-Transfer Privilege Checks" on page 124 for more information on how the DPL is used during segment privilege-checks. Present (P) Bit. Bit 15 of byte +4. The segment-present bit indicates that the segment referenced by the descriptor is loaded in memory. If a reference is made to a descriptor entry when P=0, a segment-not-present exception (#NP) occurs. This bit is set and cleared by system software and is never altered by the processor. Available To Software (AVL) Bit. Bit 20 of byte +4. This field is available to software, which can write any value to it. The processor does not set or clear this field. Default Operand Size (D/B) Bit. B i t 2 2 o f by t e + 4 . Th e d e f a u l t operand-size bit is found in code-segment and data-segment descriptors but not in system-segment descriptors. Setting this Chapter 4: Segmented Virtual Memory 99
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
bit to 1 indicates a 32-bit default operand size, and clearing it to 0 indicates a 16-bit default size. The effect this bit has on a segment depends on the segment-descriptor type. See "CodeSegment Default-Operand Size (D) Bit" on page 102 for a description of the D bit in code-segment descriptors. "DataSegment Default Operand Size (D/B) Bit" on page 104 describes the D bit in data-segment descriptors, including stack segments, where the bit is referred to as the "B" bit. Granularity (G) Bit. Bit 23 of byte +4. The granularity bit specifies how the segment-limit field is scaled. Clearing the G bit to 0 indicates that the limit field is not scaled. In this case, the limit equals the number of bytes available in the segment. Setting the G bit to 1 indicates that the limit field is scaled by 4 Kbytes (4096 bytes). Here, the limit field equals the number of 4-Kbyte blocks available in the segment. Setting a limit of 0 indicates a 1-byte segment limit when G = 0. Setting the same limit of 0 when G = 1 indicates a segment limit of 4095. Reserved Bits. Generally, software should clear all reserved bits to 0, so they can be defined in future revisions to the AMD64 architecture. 4.7.2 Code-Segment Descriptors Figure 4-14 on page 101 shows the code-segment descriptor format (gray shading indicates the bit is reserved). All software tasks require that a segment selector, referencing a valid codesegment descriptor, is loaded into the CS register. Code segments establish the processor operating mode and execution p r iv i l e g e - l eve l . Th e s e g m e n t s g e n e ra l ly c o n t a i n o n ly instructions and are execute-only, or execute and read-only. Software cannot write into a segment whose selector references a code-segment descriptor.
100
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
31 Base Address 31-24
24 23 22 21 20 19 GD
16 15 14 13 12 11 10 9 8 7 Base Address 23-16
0 +4
A Segment P DPL 1 1 C R A V Limit 19-16 L
Base Address 15-0
Segment Limit 15-0
+0
Figure 4-14.
Code-Segment Descriptor--Legacy Mode Code-segment descriptors have the S bit set to 1, identifying the segments as user segments. Type-field bit 11 differentiates code-segment descriptors (bit 11 set to 1) from data-segment descriptors (bit 11 cleared to 0). The remaining type-field bits (10-8) define the access characteristics for the code-segment, as follows: Conforming (C) Bit. Bit 10 of byte +4. Setting this bit to 1 identifies the code segment as conforming. When control is transferred to a higher-privilege conforming code-segment (C=1) from a lowerprivilege code segment, the processor CPL does not change. Transfers to non-conforming code-segments (C=0) with a higher privilege-level than the CPL can occur only through gate descriptors. See "Control-Transfer Privilege Checks" on page 124 for more information on conforming and nonconforming code-segments. Readable (R) Bit. Bit 9 of byte +4. Setting this bit to 1 indicates the code segment is both executable and readable as data. When this bit is cleared to 0, the code segment is executable, but attempts to read data from the code segment cause a generalprotection exception (#GP) to occur. Accessed (A) Bit. Bit 8 of byte +4. The accessed bit is set to 1 by the processor when the descriptor is copied from the GDT or LDT into the CS register. This bit is only cleared by software. Table 4-3 on page 102 summarizes the code-segment type-field encodings.
Chapter 4: Segmented Virtual Memory
101
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 4-3.
Code-Segment Descriptor Types
Type Field
Hex Value
Bit 11 (Code/Data)
Bit 10 Conforming (C) 0 0 0
Bit 9 Readable (R) 0 0 1 1 0 0 1 1
Bit 8 Accessed (A) 0 1 0 1 0 1 0 1 Execute-Only
Description
8 9 A B C D E F 1
Execute-Only -- Accessed Execute/Readable Execute/Readable -- Accessed Conforming, Execute-Only Conforming, Execute-Only -- Accessed Conforming, Execute/Readable Conforming, Execute/Readable -- Accessed
0 1 1 1 1
Code-Segment Default-Operand Size (D) Bit. Bit 22 of byte +4. In codesegment descriptors, the D bit selects the default operand size and address sizes. In legacy mode, when D=0 the default operand size and address size is 16 bits and when D=1 the default operand size and address size is 32 bits. Instruction prefixes can be used to override the operand size or address size, or both. 4.7.3 Data-Segment Descriptors Figure 4-15 on page 103 shows the data-segment descriptor format. Data segments contain non-executable information and can be accessed as read-only or read/write. They are referenced using the DS, ES, FS, GS, or SS data-segment registers. The DS data-segment register holds the segment selector for the default data segment. The ES, FS and GS data-segment registers hold segment selectors for additional data segments usable by the current software task. The stack segment is a special form of data-segment register. It is referenced using the SS segment register and must be read/write. When loading the SS register, the processor requires that the selector reference a valid, writable data-segment descriptor.
102
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
31 Base Address 31-24
24 23 22 21 20 19 D G/ B
16 15 14 13 12 11 10 9 8 7 Base Address 23-16
0 +4
A Segment P DPL 1 0 E W A V Limit 19-16 L
Base Address 15-0
Segment Limit 15-0
+0
Figure 4-15.
Data-Segment Descriptor--Legacy Mode Data-segment descriptors have the S bit set to 1, identifying them as user segments. Type-field bit 11 differentiates datasegment descriptors (bit 11 cleared to 0) from code-segment descriptors (bit 11 set to 1). The remaining type-field bits (10-8) define the data-segment access characteristics, as follows: Expand-Down (E) Bit. Bit 10 of byte +4. Setting this bit to 1 identifies the data segment as expand-down. In expand-down segments, the segment limit defines the lower segment boundary while the base is the upper boundary. Valid segment offsets in expand-down segments lie in the byte range limit+1 to FFFFh or FFFF_FFFFh, depending on the value of the data segment default operand size (D/B) bit. Expand-down segments are useful for stacks, which grow in the downward direction as elements are pushed onto the stack. The stack pointer, ESP, is decremented by an amount equal to the operand size as a result of executing a PUSH instruction. Clearing the E bit to 0 identifies the data segment as expandup. Valid segment offsets in expand-up segments lie in the byte range 0 to segment limit. Writable (W) Bit. Bit 9 of byte +4. Setting this bit to 1 identifies the data segment as read/write. When this bit is cleared to 0, the segment is read-only. A general-protection exception (#GP) occurs if software attempts to write into a data segment when W=0. Accessed (A) Bit. Bit 8 of byte +4. The accessed bit is set to 1 by the processor when the descriptor is copied from the GDT or LDT into one of the data-segment registers or the stack-segment register. This bit is only cleared by software.
Chapter 4: Segmented Virtual Memory
103
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 4-4 summarizes the data-segment type-field encodings. Table 4-4. Data-Segment Descriptor Types
Type Field Hex Value Bit 10 Bit 11 (Code/Data) ExpandDown (E) 0 0 0 0 0 1 1 1 1 Bit 9 Writable (W) 0 0 1 1 0 0 1 1 Bit 8 Description Accessed (A) 0 1 0 1 0 1 0 1 Read-Only Read-Only -- Accessed Read/Write Read/Write -- Accessed Expand-down, Read-Only Expand-down, Read-Only -- Accessed Expand-down, Read/Write Expand-down, Read/Write -- Accessed
0 1 2 3 4 5 6 7
Data-Segment Default Operand Size (D/B) Bit. Bit 22 of byte +4. For expand-down data segments (E=1), setting D=1 sets the upper bound of the segment at 0_FFFF_FFFFh. Clearing D=0 sets the upper bound of the segment at 0_FFFFh. In the case where a data segment is referenced by the stack selector (SS), the D bit is referred to as the B bit. For stack segments, the B bit sets the default stack size. Setting B=1 establishes a 32-bit stack referenced by the 32-bit ESP register. Clearing B=0 establishes a 16-bit stack referenced by the 16-bit SP register. 4.7.4 System Descriptors There are two general types of system descriptors: systemsegment descriptors and gate descriptors. System-segment descriptors are used to describe the LDT and TSS segments. Gate descriptors do not describe segments, but instead hold pointers to code-segment descriptors. Gate descriptors are used for protected-mode control transfers between less-privileged and more-privileged software.
104
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
System-segment descriptors have the S bit cleared to 0. The type field is used to differentiate the various LDT, TSS, and gate descriptors from one another. Table 4-5 summarizes the system-segment type-field encodings. Table 4-5.
Hex Value 0 1 2 3 4 5 6 7 8 9 A B C D E F
System-Segment Descriptor Types (S=0)--Legacy Mode
Type Field (Bits 11-8) 0000 0001 0010 0011 0100 0101 0110 0111 1000 1001 1010 1011 1100 1101 1110 1111 Reserved (Illegal) Available 16-bit TSS LDT Busy 16-bit TSS 16-bit Call Gate Task Gate 16-bit Interrupt Gate 16-bit Trap Gate Reserved (Illegal) Available 32-bit TSS Reserved (Illegal) Busy 32-bit TSS 32-bit Call Gate Reserved (Illegal) 32-bit Interrupt Gate 32-bit Trap Gate Description
Figure 4-16 on page 106 shows the legacy-mode system-segment descriptor format used for referencing LDT and TSS segments (gray shading indicates the bit is reserved). This format is also used in compatibility mode. The system-segments are used as follows: The LDT typically holds segment descriptors belonging to a single task (see "Local Descriptor Table" on page 92). The TSS is a data structure for holding processor-state information. Processor state is saved in a TSS when a task is Chapter 4: Segmented Virtual Memory 105
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
suspended, and state is restored from the TSS when a task is restarted. System software must create at least one TSS referenced by the task register, TR. See "Legacy Task-State Segment" on page 365 for more information on the TSS.
31 Base Address 31-24 24 23 22 21 20 19 G I G N 16 15 14 13 12 11 Type 87 Base Address 23-16 0 +4
A Segment V P DPL 0 Limit 19-16 L
Base Address 15-0
Segment Limit 15-0
+0
Figure 4-16.
LDT and TSS Descriptor--Legacy/Compatibility Modes Gate descriptors hold pointers to code segments and are used to control access between code segments with different privilege levels. There are four types of gate descriptors: Call Gates--These gates (Figure 4-17 on page 107) are located in the GDT or LDT and are used to control access between code segments in the same task or in different tasks. See "Control Transfers Through Call Gates" on page 129 for information on how call gates are used to control access between code segments operating in the same task. The format of a call-gate descriptor is shown in Figure 4-17. Interrupt Gates and Trap Gates--These gates (Figure 4-18 on page 107) are located in the IDT and are used to control access to interrupt-service routines. "Legacy ProtectedMode Interrupt Control Transfers" on page 276 contains information on using these gates for interrupt-control transfers. The format of interrupt-gate and trap-gate descriptors is shown in Figure 4-17. Task Gates--These gates (Figure 4-19 on page 107) are used to control access between different tasks. They are also used to transfer control to interrupt-service routines if those routines are themselves a separate task. See "TaskManagement Resources" on page 360 for more information on task gates and their use.
4.7.5 Gate Descriptors
106
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
31 Target Code-Segment Offset 31-16
16 15 14 13 12 11 P DPL 0 Type
87654 Reserved IGN Parameter Count
0 +4
Target Code-Segment Selector
Target Code-Segment Offset 15-0
+0
Figure 4-17.
Call-Gate Descriptor--Legacy Mode
31 Target Code-Segment Offset 31-16
16 15 14 13 12 11 P DPL 0 Type
87 Reserved, IGN
0 +4
Target Code-Segment Selector
Target Code-Segment Offset 15-0
+0
Figure 4-18.
Interrupt-Gate and Trap-Gate Descriptors--Legacy Mode
31 Reserved, IGN
16 15 14 13 12 11 P DPL S Type
87 Reserved, IGN
0 +4
TSS Selector
Reserved, IGN
+0
Figure 4-19.
Task-Gate Descriptor--Legacy Mode There are several differences between the gate-descriptor format and the system-segment descriptor format. These differences are described as follows, from least-significant to most-significant bit positions: Target Code-Segment Offset. The 32-bit segment offset is formed by concatenating bits 31-16 of byte +4 with bits 15-0 of byte +0. The segment-offset field specifies the target-procedure entry
Chapter 4: Segmented Virtual Memory
107
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
point (offset) into the segment. This field is loaded into the EIP register as a result of a control transfer using the gate descriptor. Target Code-Segment Selector. Bits 31-16 of byte +0. The segmentselector field identifies the target-procedure segment descriptor, located in either the GDT or LDT. The segment selector is loaded into the CS segment register as a result of a control transfer using the gate descriptor. TSS Selector. Bits 31-16 of byte +0 (task gates only). This field identifies the target-task TSS descriptor, located in any of the three descriptor tables (GDT, LDT, and IDT). Parameter Count (Call Gates Only). Bits 4-0 of byte +4. Legacy-mode call-gate descriptors contain a 5-bit parameter-count field. This field specifies the number of parameters to be copied from the currently-executing program stack to the target program stack during an automatic stack switch. Automatic stack switches are performed by the processor during a control transfer through a call gate to a greater privilege-level. The parameter size depends on the call-gate size as specified in the type field. 32bit call gates copy 4-byte parameters, and 16-bit call gates copy 2-byte parameters. See "Stack Switching" on page 134 for more information on call-gate parameter copying.
4.8
Long-Mode Segment Descriptors
The interpretation of descriptor fields is changed in long mode, and in some cases the format is expanded. The changes depend on the operating mode (compatibility mode or 64-bit mode) and on the descriptor type. The following sections describe the changes.
4.8.1 Code-Segment Descriptors
Code segments continue to exist in long mode. Code segments and their associated descriptors and selectors are needed to establish the processor operating mode as well as execution privilege-level. The new L attribute specifies whether the processor is running in compatibility mode or 64-bit mode (see "Long (L) Attribute Bit" on page 109). Figure 4-20 shows the long-mode code-segment descriptor format. In compatibility mode, the code-segment descriptor is interpreted and behaves just as it does in legacy mode as described in "Code-Segment Descriptors" on page 100. Chapter 4: Segmented Virtual Memory
108
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
In Figure 4-20, gray shading indicates the code-segment descriptor fields that are ignored in 64-bit mode when the descriptor is used during a memory reference. However, the fields are loaded whenever the segment register is loaded in 64bit mode.
31 Base Address 31-24
24 23 22 21 20 19
16 15 14 13 12 11 10 9 8 7 Base Address 23-16
0 +4
A Segment P DPL 1 1 C R A GDLV Limit 19-16 L
Base Address 15-0
Segment Limit 15-0
+0
Figure 4-20.
Code-Segment Descriptor--Long Mode Fields Ignored in 64-Bit Mode. Segmentation is disabled in 64-bit mode, and code segments span all of virtual memory. In this mode, code-segment base addresses are ignored. For the purpose of virtual-address calculations, the base address is treated as if it has a value of zero. Segment-limit checking is not performed, and both the segment-limit field and granularity (G) bit are ignored. Instead, the virtual address is checked to see if it is in canonical-address form. The readable (R) and accessed (A) attributes in the type field are also ignored. Long (L) Attribute Bit. Bit 21 of byte +4. Long mode introduces a new attribute, the long (L) bit, in code-segment descriptors. This bit specifies that the processor is running in 64-bit mode (L=1) or compatibility mode (L=0). When the processor is running in legacy mode, this bit is reserved. Compatibility mode, maintains binary compatibility with legacy 16-bit and 32-bit applications. Compatibility mode is selected on a code-segment basis, and it allows legacy applications to coexist under the same 64-bit system software along with 64-bit applications running in 64-bit mode. System software running in long mode can execute existing 16-bit and 32-bit applications by clearing the L bit of the code-segment descriptor to 0.
Chapter 4: Segmented Virtual Memory
109
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
When L=0, the legacy meaning of the code-segment D bit (see "Code-Segment Default-Operand Size (D) Bit" on page 102)-- and the address-size and operand-size prefixes--are observed. Segmentation is enabled when L=0. From an application viewpoint, the processor is in a legacy 16-bit or 32-bit operating environment (depending on the D bit), even though long mode is activated. If the processor is running in 64-bit mode (L=1), the only valid setting of the D bit is 0. This setting produces a default operand size of 32 bits and a default address size of 64 bits. The combination L=1 and D=1 is reserved for future use. "Instruction Prefixes" in Volume 3 describes the effect of the code-segment L and D bits on default operand and address sizes when long mode is activated. These default sizes can be overridden with operand size, address size, and REX prefixes. 4.8.2 Data-Segment Descriptors Data segments continue to exist in long mode. Figure 4-21 shows the long-mode data-segment descriptor format. In compatibility mode, data-segment descriptors are interpreted and behave just as they do in legacy mode. In Figure 4-21, gray shading indicates the fields that are ignored in 64-bit mode when the descriptor is used during a memory reference. However, the fields are loaded whenever the segment register is loaded in 64-bit mode.
31 Base Address 31-24
24 23 22 21 20 19 D G/ B
16 15 14 13 12 11 10 9 8 7 Base Address 23-16
0 +4
A Segment P DPL 1 0 E W A V Limit 19-16 L
Base Address 15-0
Segment Limit 15-0
+0
Figure 4-21.
Data-Segment Descriptor--Long Mode Fields Ignored in 64-Bit Mode. Segmentation is disabled in 64-bit mode. The interpretation of the segment-base address depends on the segment register used: In data-segment descriptors referenced by the DS, ES and SS segment registers, the base-address field is ignored. For
110
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
the purpose of virtual-address calculations, the base address is treated as if it has a value of zero. Data segments referenced by the FS and GS segment registers receive special treatment in 64-bit mode. For these segments, the base address field is not ignored, and a nonzero value can be used in virtual-address calculations. A 64bit segment-base address can be specified using modelspecific registers. See "FS and GS Registers in 64-Bit Mode" on page 88 for more information. Segment-limit checking is not performed on any data segments in 64-bit mode, and both the segment-limit field and granularity (G) bit are ignored. The D/B bit is unused in 64-bit mode. The expand-down (E), writable (W), and accessed (A) type-field attributes are ignored. A data-segment-descriptor DPL field is ignored in 64-bit mode, and segment-privilege checks are not performed on data segments. System software can use the page-protection mechanisms to isolate and protect data from unauthorized access. 4.8.3 System Descriptors In long mode, the allowable system-descriptor types encoded by the type field are changed. Some descriptor types are modified, and others are illegal. The changes are summarized in Table 4-6 on page 112. An attempt to use an illegal descriptor type causes a general-protection exception (#GP).
Chapter 4: Segmented Virtual Memory
111
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 4-6.
Hex Value 0 1 2 3 4 5 6 7 8 9 A B C D E F
Note:
System-Segment Descriptor Types--Long Mode
Type Field Bit 11 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 Bit 10 0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1 Bit 9 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 Bit 8 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 Available 64-bit TSS Reserved (Illegal) Busy 64-bit TSS 64-bit Call Gate Reserved (Illegal) 64-bit Interrupt Gate 64-bit Trap Gate Reserved (Illegal) Description
Reserved (Illegal) 64-Bit LDT1
1. In 64-bit mode only. In compatibility mode, the type specifies a 32-bit LDT.
In long mode, the modified system-segment descriptor types are: The 32-bit LDT (02h), which is redefined as the 64-bit LDT. The available 32-bit TSS (09h), which is redefined as the available 64-bit TSS. The busy 32-bit TSS (0Bh), which is redefined as the busy 64bit TSS. In 64-bit mode, the LDT and TSS system-segment descriptors are expanded by 64 bits, as shown in Figure 4-22 on page 113. In 112 Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
this figure, gray shading indicates the fields that are ignored in 64-bit mode. Expanding the descriptors allows them to hold 64bit base addresses, so their segments can be located anywhere in the virtual-address space. The expanded descriptor can be loaded into the corresponding descriptor-table register (LDTR or TR) only from 64-bit mode. In compatibility mode, the legacy system-segment descriptor format, shown in Figure 4-16 on page 106, is used. See "LLDT and LTR Instructions" on page 189 for more information.
31
23
20 19
16 15 14 13 12 11 10 9 8 7 00000 Reserved, IGN
0 +12
Reserved, IGN
Base Address 63-32 A Segment V P DPL 0 Limit 19-16 L
+8
Base Address 31-24
G
Type
Base Address 23-16
+4
Base Address 15-0
Segment Limit 15-0
+0
Figure 4-22.
System-Segment Descriptor--64-Bit Mode The 64-bit system-segment base address must be in canonical form. Otherwise, a general-protection exception occurs with a selector error-code, #GP(selector), when the system segment is loaded. System-segment limit values are checked by the processor in both 64-bit and compatibility modes, under the control of the granularity (G) bit. Figure 4-22 shows that bits 12-8 of dword +12 must be cleared to 0. These bits correspond to the S and Type fields in a legacy descriptor. Clearing these bits to 0 corresponds to an illegal type in legacy mode and causes a #GP if an attempt is made to access the upper half of a 64-bit mode system-segment descriptor as a legacy descriptor or as the lower half of a 64-bit mode system-segment descriptor.
4.8.4 Gate Descriptors
As shown in Table 4-6 on page 112, the allowable gatedescriptor types are changed in long mode. Some gate113
Chapter 4: Segmented Virtual Memory
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
descriptor types are modified and others are illegal. The modified gate-descriptor types in long mode are: The 32-bit call gate (0Ch), which is redefined as the 64-bit call gate. The 32-bit interrupt gate (0Eh), which is redefined as the 64bit interrupt gate. The 32-bit trap gate (0Fh), which is redefined as the 64-bit trap gate. In long mode, several gate-descriptor types are illegal. An attempt to use these gates causes a general-protection exception (#GP) to occur. The illegal gate types are: The 16-bit call gate (04h). The task gate (05h). The 16-bit interrupt gate (06h). The 16-bit trap gate (07h). In long mode, gate descriptors are expanded by 64 bits, allowing them to hold 64-bit offsets. The 64-bit call-gate descriptor is shown in Figure 4-23 and the 64-bit interrupt gate and trap gate are shown in Figure 4-24 on page 115. In these figures, gray shading indicates the fields that are ignored in long mode. The interrupt and trap gates contain an additional field, the IST, that is not present in the call gate--see "IST Field (Interrupt and Trap Gates)" on page 116.
31 Reserved, IGN
16 15 14 13 12 11 10 9 8 7 00000 Reserved, IGN
0 +12
Target Offset 63-32
+8
Target Offset 31-16
P DPL S
Type
Reserved, IGN
+4
Target Selector
Target Offset 15-0
+0
Figure 4-23. 114
Call-Gate Descriptor--Long Mode Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
16 15 14 13 12 11 Reserved, IGN
87
32
0 +12
Target Offset 63-32
+8
Target Offset 31-16
P DPL S
Type
Reserved, IGN
IST
+4
Target Selector
Target Offset 15-0
+0
Figure 4-24.
Interrupt-Gate and Trap-Gate Descriptors--Long Mode The target code segment referenced by a long-mode gate descriptor must be a 64-bit code segment (CS.L=1, CS.D=0). If the target is not a 64-bit code segment, a general-protection exception, #GP(error), occurs. The error code reported depends on the gate type: Call gates report the target code-segment selector as the error code. Interrupt and trap gates report the interrupt-vector number as the error code. A general-protection exception, #GP(0), occurs if software attempts to reference a long-mode gate descriptor with a targetsegment offset that is not in canonical form. It is possible for software to store legacy and long mode gate descriptors in the same descriptor table. Figure 4-23 shows that bits 12-8 of byte +12 in a long-mode call gate must be cleared to 0. These bits correspond to the S and Type fields in a legacy call gate. Clearing these bits to 0 corresponds to an illegal type in legacy mode and causes a #GP if an attempt is made to access the upper half of a 64-bit mode call-gate descriptor as a legacy call-gate descriptor. It is not necessary to clear these same bits in a long-mode interrupt gate or trap gate. In long mode, the interruptdescriptor table (IDT) must contain 64-bit interrupt gates or
Chapter 4: Segmented Virtual Memory
115
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
trap gates. The processor automatically indexes the IDT by scaling the interrupt vector by 16. This makes it impossible to access the upper half of a long-mode interrupt gate, or trap gate, as a legacy gate when the processor is running in long mode. IST Field (Interrupt and Trap Gates). Bits 2-0 of byte +4. Long-mode interrupt gate and trap gate descriptors contain a new, 3-bit interrupt-stack-table (IST) field not present in legacy gate descriptors. The IST field is used as an index into the IST portion of a long-mode TSS. If the IST field is not 0, the index references an IST pointer in the TSS, which the processor loads into the RSP register when an interrupt occurs. If the IST index is 0, the processor uses the legacy stack-switching mechanism (with some modifications) when an interrupt occurs. See "Interrupt-Stack Table" on page 292 for more information. Count Field (Call Gates). The count field found in legacy call-gate descriptors is not supported in long-mode call gates. In long mode, the field is reserved and should be cleared to zero. 4.8.5 Long Mode Descriptor Summary System descriptors and gate descriptors are expanded by 64 bits to handle 64-bit base addresses in long mode or 64-bit mode. The mode in which the expansion occurs depends on the purpose served by the descriptor, as follows: Expansion Only In 64-Bit Mode--The system descriptors and pseudo-descriptors that are loaded into the GDTR, IDTR, LDTR, and TR registers are expanded only in 64-bit mode. They are not expanded in compatibility mode. Expansion In Long Mode--Gate descriptors (call gates, interrupt gates, and trap gates) are expanded in long mode (both 64-bit mode and compatibility mode). Task gates and 16-bit gate descriptors are illegal in long mode. The AMD64 architecture redefines several of the descriptorentry fields in support of long mode. The specific change depends on whether the processor is in 64-bit mode or compatibility mode. Table 4-7 on page 117 summarizes the changes in the descriptor entry field when the descriptor entry is loaded into a segment register (as opposed to when the segment register is subsequently used to access memory).
116
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 4-7.
Descriptor Field
Descriptor-Entry Field Changes in Long Mode
Descriptor Type Code Long Mode Compatibility Mode 64-Bit Mode
Limit
Data System
Same as legacy x86
Same as legacy x86
Offset
Gate Code
Expanded to 64 bits
Expanded to 64 bits
Base
Data System
Same as legacy x86
Same as legacy x86
Selector IST1
Gate Gate Code Data
Same as legacy x86 Interrupt and trap gates only. (New for long mode.) Same as legacy x86 Types 02h, 09h, and 0Bh redefined Types 01h and 03h are illegal Types 0Ch, 0Eh, and 0Fh redefined Types 04h-07h are illegal Same as legacy x86
S and Type
System Gate Code
DPL
Data System Gate Code
Same as legacy x86
Same as legacy x86
Present
Data System Gate
Same as legacy x86
Same as legacy x86
Note:
1. Not available (reserved) in legacy mode.
Chapter 4: Segmented Virtual Memory
117
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 4-7.
Descriptor Field
Descriptor-Entry Field Changes in Long Mode (continued)
Descriptor Type Code Data Long Mode Compatibility Mode 64-Bit Mode D=0 Indicates 64-bit address, 32-bit data D=1 Reserved Same as legacy x86 Specifies compatibility mode Specifies 64-bit mode
Default Size
Same as legacy x86
Long1
Code Code
Granularity
Data System Code
Same as legacy x86
Same as legacy x86
Available
Data System
Same as legacy x86
Same as legacy x86
Note:
1. Not available (reserved) in legacy mode.
4.9
Segment-Protection Overview
The AMD64 architecture is designed to fully support the legacy segment-protection mechanism. The segment-protection mechanism provides system software with the ability to restrict program access into other software routines and data. Segment-level protection remains enabled in compatibility mode. 64-bit mode eliminates most type checking, and limit checking is not performed, except on accesses to systemdescriptor tables. The preferred method of implementing memory protection in a long-mode operating system is to rely on the page-protection mechanism as described in "Page-Protection Checks" on page 174. System software still needs to create basic segmentprotection data structures for 64-bit mode. These structures are simplified, however, by the use of the flat-memory model in 64bit mode, and the limited segmentation checks performed when executing in 64-bit mode.
118
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
4.9.1 Privilege-Level Concept
Segment protection is used to isolate and protect programs and data from each other. The segment-protection mechanism supports four privilege levels in protected mode. The privilege levels are designated with a numerical value from 0 to 3, with 0 being the most privileged and 3 being the least privileged. System software typically assigns the privilege levels in the following manner: Privilege-level 0 (most privilege)--This level is used by critical system-software components that require direct access to, and control over, all processor and system resources. This can include BIOS, memory-management functions, and interrupt handlers. Privilege-levels 1 and 2 (moderate privilege)--These levels are used by less-critical system-software services that can access and control a limited scope of processor and system resources. Software running at these privilege levels might include some device drivers and library routines. These software routines can call more-privileged system-software services to perform functions such as memory garbagecollection and file allocation. Privilege-level 3 (least privilege)--This level is used by application software. Software running at privilege-level 3 is normally prevented from directly accessing most processor and system resources. Instead, applications request access to the protected processor and system resources by calling more-privileged service routines to perform the accesses. Figure 4-25 on page 120 shows the relationship of the four privilege levels to each other.
Chapter 4: Segmented Virtual Memory
119
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Memory Management File Allocation Interrupt Handling
Privilege 0 Privilege 1 Privilege 2
513-236.eps
Device-Drivers Library Routines
Privilege 3
Application Programs
Figure 4-25. 4.9.2 Privilege-Level Types
Privilege-Level Relationships
There are three types of privilege levels the processor uses to control access to segments. These are CPL, DPL, and RPL. Current Privilege-Level. The current privilege-level (CPL) is the privilege level at which the processor is currently executing. The CPL is stored in an internal processor register that is invisible to software. Software changes the CPL by performing a control transfer to a different code segment with a new privilege level. Descriptor Privilege-Level. The descriptor privilege-level (DPL) is the privilege level that system software assigns to individual segments. The DPL is used in privilege checks to determine whether software can access the segment referenced by the descriptor. In the case of gate descriptors, the DPL determines whether software can access the descriptor reference by the gate. The DPL is stored in the segment (or gate) descriptor. Requestor Privilege-Level. The requestor privilege-level (RPL) reflects the privilege level of the program that created the selector. The RPL can be used to let a called program know the privilege level of the program that initiated the call. The RPL is stored in the selector used to reference the segment (or gate) descriptor. The following sections describe how the CPL, DPL, and RPL are used by the processor in performing privilege checks on data
120
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
accesses and control transfers. Failure to pass a protection check generally causes an exception to occur.
4.10
Data-Access Privilege Checks
Before loading a data-segment register (DS, ES, FS, or GS) with a segment selector, the processor checks the privilege levels as follows to see if access is allowed: 1. The processor compares the CPL with the RPL in the datasegment selector and determines the effective privilege level for the data access. The processor sets the effective privilege level to the lowest privilege (numerically-higher value) indicated by the comparison. 2. The processor compares the effective privilege level with the DPL in the descriptor-table entry referenced by the segment selector. If the effective privilege level is greater than or equal to (numerically lower-than or equal-to) the DPL, then the processor loads the segment register with the data-segment selector. The processor automatically loads the corresponding descriptor-table entry into the hidden portion of the segment register. If the effective privilege level is lower than (numerically greater-than) the DPL, a general-protection exception (#GP) occurs and the segment register is not loaded. Figure 4-26 on page 122 shows two examples of data-access privilege checks.
4.10.1 Accessing Data Segments
Chapter 4: Segmented Virtual Memory
121
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
CS Data Selector
CPL=3 Max RPL=0
Effective Privilege
3 Access Denied Data Segment
DPL=2 Descriptor Example 1: Privilege Check Fails
CS Data Selector
CPL=0 Max RPL=0
Effective Privilege
0 Access Allowed Data Segment
DPL=2 Descriptor Example 2: Privilege Check Passes
513-229.eps
Figure 4-26.
Data-Access Privilege-Check Examples Example 1 in Figure 4-26 shows a failing data-access privilege check. The effective privilege level is 3 because CPL=3. This value is greater than the descriptor DPL, so access to the data segment is denied. Example 2 in Figure 4-26 shows a passing data-access privilege check. Here, the effective privilege level is 0 because both the CPL and RPL have values of 0. This value is less than the descriptor DPL, so access to the data segment is allowed, and the data-segment register is successfully loaded.
122
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
4.10.2 Accessing Stack Segments
Before loading the stack segment register (SS) with a segment selector, the processor checks the privilege levels as follows to see if access is allowed: 1. The processor checks that the CPL and the stack-selector RPL are equal. If they are not equal, a general-protection exception (#GP) occurs and the SS register is not loaded. 2. The processor compares the CPL with the DPL in the descriptor-table entry referenced by the segment selector. The two values must be equal. If they are not equal, a #GP occurs and the SS register is not loaded. Figure 4-27 on page 124 shows two examples of stack-access privilege checks. In Example 1 the CPL, stack-selector RPL, and stack segment-descriptor DPL are all equal, so access to the stack segment using the SS register is allowed. In Example 2, the stack-selector RPL and stack segment-descriptor DPL are both equal. However, the CPL is not equal to the stack segmentdescriptor DPL, and access to the stack segment through the SS register is denied.
Chapter 4: Segmented Virtual Memory
123
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
CS Stack Selector
CPL=3 = RPL=3 Access Allowed DPL=3 Descriptor Example 1: Privilege Check Passes Stack Segment
CS Stack Selector
CPL=2 = RPL=3 Access Denied DPL=3 Descriptor Example 2: Privilege Check Fails Stack Segment
513-235.eps
Figure 4-27.
Stack-Access Privilege-Check Examples
4.11
Control-Transfer Privilege Checks
Control transfers between code segments (also called far control transfers) cause the processor to perform privilege checks to determine whether the source program is allowed to transfer control to the target program. If the privilege checks pass, access to the target code-segment is granted. When access is granted, the target code-segment selector is loaded into the CS register. The rIP register is updated with the target CS offset taken from either the far-pointer operand or the gate
124
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
descriptor. Privilege checks are not performed during near control transfers because such transfers do not change segments. The following mechanisms can be used by software to perform far control transfers: System-software control transfers using the system-call and system-return instructions. See "SYSCALL and SYSRET" on page 182 and "SYSENTER and SYSEXIT (Legacy Mode Only)" on page 184 for more information on these instructions. SYSCALL and SYSRET are the preferred method of performing control transfers in long mode. SYSENTER and SYSEXIT are not supported in long mode. Direct control transfers using CALL and JMP instructions. These are discussed in the next section, "Direct Control Transfers." Call-gate control transfers using CALL and JMP instructions. These are discussed in "Control Transfers Through Call Gates" on page 129. Return control transfers using the RET instruction. These are discussed in "Return Control Transfers" on page 137. Interrupts and exceptions, including the INTn and IRET instructions. These are discussed in Chapter 8, "Exceptions and Interrupts." Task switches initiated by CALL and JMP instructions. Task switches are discussed in Chapter 12, "Task Management." The hardware task-switch mechanism is not supported in long mode. 4.11.1 Direct Control Transfers A direct control transfer occurs when software executes a farCALL or a far-JMP instruction without using a call gate. The privilege checks and type of access allowed as a result of a direct control transfer depends on whether the target code segment is conforming or nonconforming. The code-segmentdescriptor conforming (C) bit indicates whether or not the target code-segment is conforming (see "Conforming (C) Bit" on page 101 for more information on the conforming bit). Privilege levels are not changed as a result of a direct control transfer. Program stacks are not automatically switched by the processor as they are with privilege-changing control transfers through call gates (see "Stack Switching" on page 134 for more information on automatic stack switching during privilegechanging control transfers). Chapter 4: Segmented Virtual Memory 125
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Nonconforming Code Segments. Soft ware can perform a direct control transfer to a nonconforming code segment only if the target code-segment descriptor DPL and the CPL are equal and the RPL is less than or equal to the CPL. Software must use a c a l l g a t e t o t ra n s f e r c o n t r o l t o a m o re - p r iv i l e g e d , nonconforming code segment (see "Control Transfers Through Call Gates" on page 129 for more information). In far calls and jumps, the far pointer (CS:rIP) references the target code-segment descriptor. Before loading the CS register with a nonconforming code-segment selector, the processor checks as follows to see if access is allowed: 1. DPL = CPL Check--The processor compares the target codesegment descriptor DPL with the currently executing program CPL. If they are equal, the processor performs the next check. If they are not equal, a general-protection exception (#GP) occurs. 2. RPL CPL Check--The processor compares the target codesegment selector RPL with the currently executing program CPL. If the RPL is less than or equal to the CPL, access is allowed. If the RPL is greater than the CPL, a #GP exception occurs. If access is allowed, the processor loads the CS and rIP registers with their new values and begins executing from the target location. The CPL is not changed--the target-CS selector RPL value is disregarded when the selector is loaded into the CS register. Figure 4-28 on page 127 shows three examples of privilege checks performed as a result of a far control transfer to a nonconforming code-segment. In Example 1, access is allowed because CPL = DPL and RPL CPL. In Example 2, access is denied because CPL DPL. In Example 3, access is denied because RPL > CPL.
126
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Code Selector CS
RPL=0 CPL=2 = DPL=2 Descriptor
Access Allowed
Access Allowed ?
Access Allowed
Code Segment
Example 1: Privilege Check Passes
Code Selector CS
RPL=0 CPL=2 = DPL=3 Descriptor
Access Allowed
Access Denied ?
Access Denied
Code Segment
Example 2: Privilege Check Fails
Code Selector CS
RPL=3 CPL=2 = DPL=2 Descriptor
Access Denied
Access Denied ?
Access Allowed
Code Segment
Example 3: Privilege Check Fails
513-230.eps
Figure 4-28.
Nonconforming Code-Segment Privilege-Check Examples
Chapter 4: Segmented Virtual Memory
127
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Conforming Code Segments. O n a d i re c t c o n t ro l t ra n s f e r t o a conforming code segment, the target code-segment descriptor DPL can be lower than (at a greater privilege) the CPL. Before loading the CS register with a conforming code-segment selector, the processor compares the target code-segment descriptor DPL with the currently-executing program CPL. If the DPL is less than or equal to the CPL, access is allowed. If the DPL is greater than the CPL, a #GP exception occurs. On an access to a conforming code segment, the RPL is ignored and not involved in the privilege check. When access is allowed, the processor loads the CS and rIP registers with their new values and begins executing from the target location. The CPL is not changed--the target CSdescriptor DPL value is disregarded when the selector is loaded into the CS register. The target program runs at the same privilege as the program that called it. Figure 4-29 on page 129 shows two examples of privilege checks performed as a result of a direct control transfer to a conforming code segment. In Example 1, access is allowed because the CPL of 3 is greater than the DPL of 0. As the target code selector is loaded into the CS register, the old CPL value of 3 replaces the target-code selector RPL value, and the target program executes with CPL=3. In Example 2, access is denied because CPL < DPL.
128
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Code Selector CS CPL=3 DPL=0 Descriptor Example 1: Privilege Check Passes Access Allowed Code Segment
Code Selector CS CPL=0 DPL=3 Descriptor Example 2: Privilege Check Fails
513-231.eps
Access Denied
Code Segment
Figure 4-29.
Conforming Code-Segment Privilege-Check Examples Control transfers to more-privileged code segments are accomplished through the use of call gates. Call gates are a type of descriptor that contain pointers to code-segment descriptors and control access to those descriptors. System software uses call gates to establish protected entry points into systemservice routines. Transfer Mechanism. The pointer operand of a far-CALL or farJMP instruction consists of two pieces: a code-segment selector (CS) and a code-segment offset (rIP). In a call-gate transfer, the CS selector points to a call-gate descriptor rather than a codesegment descriptor, and the rIP is ignored (but required by the instruction).
4.11.2 Control Transfers Through Call Gates
Chapter 4: Segmented Virtual Memory
129
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Figure 4-30 shows a call-gate control transfer in legacy mode. The call-gate descriptor contains segment-selector and segment-offset fields (see "Gate Descriptors" on page 106 for a detailed description of the call-gate format and fields). These two fields perform the same function as the pointer operand in a direct control-transfer instruction. The segment-selector field points to the target code-segment descriptor, and the segmentoffset field is the instruction-pointer offset into the target codesegment. The code-segment base taken from the code-segment descriptor is added to the offset field in the call-gate descriptor to create the target virtual address (linear address).
Virtual-Address Space Far Pointer
Segment Selector Instruction Offset
Descriptor Table
Call-Gate Descriptor +
Virtual Address
DPL Code-Segment Selector Code-Segment Offset
Code Segment DPL Code-Segment Limit Code-Segment Base
Code-Segment Descriptor
513-233.eps
Figure 4-30.
Legacy-Mode Call-Gate Transfer Mechanism Figure 4-31 on page 131 shows a call-gate control transfer in long mode. The long-mode call-gate descriptor format is expanded by 64 bits to hold a full 64-bit offset into the virtualaddress space. Only long-mode call gates can be referenced in long mode (64-bit mode and compatibility mode). The legacymode 32-bit call-gate types are redefined in long mode as 64-bit types, and 16-bit call-gate types are illegal.
130
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Far Pointer
Segment Selector Instruction Offset
Virtual-Address Space
Descriptor Table
Call-Gate Descriptor
Code-Segment Offset (63:32) DPL Code-Segment Selector Code-Segment Offset (31:0) Virtual Address
DPL
Code-Segment Limit
Code-Segment Base
Code-Segment Descriptor
Unused
Flat Code-Segment
513-234.eps
Figure 4-31.
Long-Mode Call-Gate Access Mechanism A long-mode call gate must reference a 64-bit code-segment descriptor. In 64-bit mode, the code-segment descriptor baseaddress and limit fields are ignored. The target virtual-address is the 64-bit offset field in the expanded call-gate descriptor. Privilege Checks. Before loading the CS register with the codesegment selector located in the call gate, the processor performs three privilege checks. The following checks are performed when either conforming or nonconforming code segments are referenced: 1. The processor compares the CPL with the call-gate DPL from the call-gate descriptor (DPLG). The CPL must be numerically less than or equal to DPLG for this check to pass. In other words, the following expression must be true: CPL DPLG.
Chapter 4: Segmented Virtual Memory
131
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
2. The processor compares the RPL in the call-gate selector with DPLG. The RPL must be numerically less than or equal to DPLG for this check to pass. In other words, the following expression must be true: RPL DPLG. 3. The processor compares the CPL with the target codesegment DPL from the code-segment descriptor (DPLS). The type of comparison varies depending on the type of control transfer. - When a call--or a jump to a conforming code segment--is used to transfer control through a call gate, the CPL must be numerically greater than or equal to DPLS for this check to pass. (This check prevents control transfers to less-privileged programs.) In other words, the following expression must be true: CPL DPLS. - When a JMP instruction is used to transfer control through a call gate to a nonconforming code segment, the CPL must be numerically equal to DPLS for this check to pass. (JMP instructions cannot change CPL.) In other words, the following expression must be true: CPL = DPLS. Figure 4-32 on page 133 shows two examples of call-gate privilege checks. In Example 1, all privilege checks pass as follows: The call-gate DPL (DPLG) is at the lowest privilege (3), specifying that software running at any privilege level (CPL) can access the gate. The selector referencing the call gate passes its privilege check because the RPL is numerically less than or equal to DPLG. The target code segment is at the highest privilege level (DPLS = 0). This means software running at any privilege level can access the target code segment through the call gate.
132
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
CS Call-Gate Selector
CPL=2
RPL=3
DPLG=3 Call-Gate Descriptor DPLS=0 Code-Segment Descriptor Example 1: Privilege Check Passes Code Segment
Access Allowed
CS Call-Gate Selector
CPL=2
RPL=3
DPLG=0 Call-Gate Descriptor DPLS=3 Code-Segment Descriptor Example 2: Privilege Check Fails
513-232.eps
Code Segment
Access Denied
Figure 4-32.
Privilege-Check Examples for Call Gates In Example 2, all privilege checks fail as follows: The call-gate DPL (DPLG) specifies that only software at privilege-level 0 can access the gate. The current program does not have enough privilege to access the call gate because its CPL is 2.
Chapter 4: Segmented Virtual Memory
133
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
The selector referencing the call-gate descriptor does not have enough privilege to complete the reference. Its RPL is numerically greater than DPLG. The target code segment is at a lower privilege (DPLS = 3) than the currently running software (CPL = 2). Transitions from more-privileged software to less-privileged software are not allowed, so this privilege check fails as well. Although all three privilege checks failed in Example 2, failing only one check is sufficient to deny access into the target code segment. Stack Switching. The processor performs an automatic stack switch when a control transfer causes a change in privilege levels to occur. Switching stacks isolates more-privileged software stacks from less-privileged software stacks and provides a mechanism for saving the return pointer back to the program that initiated the call. When switching to more-privileged software, as is done when transferring control using a call gate, the processor uses the corresponding stack pointer (privilege-level 0, 1, or 2) stored in the task-state segment (TSS). The format of the stack pointer stored in the TSS depends on the system-software operating mode: Legacy-mode system software stores a 32-bit ESP value (stack offset) and 16-bit SS selector register value in the TSS for each of three privilege levels 0, 1, and 2. Long-mode system software stores a 64-bit RSP value in the TSS for privilege levels 0, 1, and 2. No SS register value is stored in the TSS because in long mode a call gate must reference a 64-bit code-segment descriptor. 64-bit mode does not use segmentation, and the stack pointer consists solely of the 64-bit RSP. Any value loaded in the SS register is ignored. See "Task-Management Resources" on page 360 for more information on the legacy-mode and long-mode TSS formats. Figure 4-33 on page 135 shows a 32-bit stack in legacy mode before and after the automatic stack switch. This particular example assumes that parameters are passed from the current program to the target program. The process followed by legacy mode in switching stacks and copying parameters is:
134
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
1. The target code-segment DPL is read by the processor and used as an index into the TSS for selecting the new stack pointer (SS:ESP). For example, if DPL=1 the processor selects the SS:ESP for privilege-level 1 from the TSS. 2. The SS and ESP registers are loaded with the new SS:ESP values read from the TSS. 3. The old values of the SS and ESP registers are pushed onto the stack pointed to by the new SS:ESP. 4. The 5-bit count field is read from the call-gate descriptor. 5. The number of parameters specified in the count field (up to 31) are copied from the old stack to the new stack. The size of the parameters copied by the processor depends on the call-gate size: 32-bit call gates copy 4-byte parameters and 16-bit call gates copy 2-byte parameters. 6. The return pointer is pushed onto the stack. The return pointer consists of the current CS-register value and the EIP of the instruction following the calling instruction. 7. The CS register is loaded from the segment-selector field in the call-gate descriptor, and the EIP is loaded from the offset field in the call-gate descriptor. 8. The target program begins executing with the instruction referenced by new CS:EIP.
Old 32-Bit Stack Before CALL
New 32-Bit Stack After CALL
Old SS Old ESP Parameter 1 Parameter 2 ... Parameter n Old CS Old EIP
Stack Switch +(n*4)+12 +(n*4)+8 +(n*4)+4 +(n*4) +8 +4
Parameter 1 Parameter 2 ... Parameter n
+(n-1)*4 +(n-2)*4
Old SS:ESP
New SS:ESP
513-224.eps
Figure 4-33.
Legacy-Mode 32-Bit Stack Switch, with Parameters
Chapter 4: Segmented Virtual Memory
135
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Figure 4-34 shows a 32-bit stack in legacy mode before and after the automatic stack switch when no parameters are passed (count=0). Most software does not use the call-gate descriptor count-field to pass parameters. System software typically defines linkage mechanisms that do not rely on automatic parameter copying.
Old 32-Bit Stack Before CALL
New 32-Bit Stack After CALL
Old SS Old ESP Old CS Old EIP
+12 +8 +4
Old SS:ESP
New SS:ESP
Stack Switch
513-225.eps
Figure 4-34.
32-Bit Stack Switch, No Parameters--Legacy Mode
Figure 4-35 on page 137 shows a long-mode stack switch. In long mode, all call gates must reference 64-bit code-segment descriptors, so a long-mode stack switch uses a 64-bit stack. The process of switching stacks in long mode is similar to switching in legacy mode when no parameters are passed. The process is as follows: 1. The target code-segment DPL is read by the processor and used as an index into the 64-bit TSS for selecting the new stack pointer (RSP). 2. The RSP register is loaded with the new RSP value read from the TSS. The SS register is loaded with a null selector (SS=0). Setting the new SS selector to null allows proper handling of nested control transfers in 64-bit mode. See "Nested Returns to 64-Bit Mode Procedures" on page 139 for additional information. As in legacy mode, it is desirable to keep the stack-segment requestor privilege-level (SS.RPL) equal to the current privilege-level (CPL). When using a call gate to change privilege levels, the SS.RPL is updated to reflect the new CPL. The SS.RPL is restored from the return-target CS.RPL on the subsequent privilege-level-changing far return.
136
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
3. The old values of the SS and RSP registers are pushed onto the stack pointed to by the new RSP. The old SS value is popped on a subsequent far return. This allows system software to set up the SS selector for a compatibility-mode process by executing a RET (or IRET) that changes the privilege level. 4. The return pointer is pushed onto the stack. The return pointer consists of the current CS-register value and the RIP of the instruction following the calling instruction. 5. The CS register is loaded from the segment-selector field in the long-mode call-gate descriptor, and the RIP is loaded from the offset field in the long-mode call-gate descriptor. The target program begins execution with the instruction referenced by the new RIP.
Old 64-Bit Stack Before CALL
New 64-Bit Stack After CALL
Old SS Old RSP Old CS Old RIP
+24 +16 +8
Old SS:RSP
New RSP
(SS=0 + new_CPL)
Stack Switch
Figure 4-35.
Stack Switch--Long Mode
All long-mode stack pushes resulting from a privilege-levelchanging far call are eight-bytes wide and increment the RSP by eight. Long mode ignores the call-gate count field and does not support the automatic parameter-copy feature found in legacy mode. Software can access parameters on the old stack, if necessary, by referencing the old stack segment selector and stack pointer saved on the new process stack. 4.11.3 Return Control Transfers Returns to calling programs can be performed by using the RET instruction. The following types of returns are possible: Near Return--Near returns perform control transfers within the same code segment, so the CS register is unchanged. The
Chapter 4: Segmented Virtual Memory
137
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
new offset is popped off the stack and into the rIP register. No privilege checks are performed. Far Return, Same Privilege--A far return transfers control from one code segment to another. When the original code segment is at the same privilege level as the target code segment, a far pointer (CS:rIP) is popped off the stack and the RPL of the new code segment (CS) is checked. If the requested privilege level (RPL) matches the current privilege level (CPL), then a return is made to the same privilege level. This prevents software from changing the CS value on the stack in an attempt to return to higher-privilege software. Far Return, Less Privilege--Far returns can change privilege levels, but only to a lower-privilege level. In this case a stack switch is performed between the current, higher-privilege program and the lower-privilege return program. The CSregister and rIP-register values are popped off the stack. The lower-privilege stack pointer is also popped off the stack and into the SS register and rSP register. The processor checks both the CS and SS privilege levels to ensure they are equal and at a lesser privilege than the current CS. In the case of nested returns to 64-bit mode, a null selector can be popped into the SS register. See "Nested Returns to 64-Bit Mode Procedures" on page 139. Far returns also check the privilege levels of the DS, ES, FS and GS selector registers. If any of these segment registers have a selector with a higher privilege than the return program, the segment register is loaded with the null selector. Stack Switching. The stack switch performed by a far return to a lower-privilege level reverses the stack switch of a call gate to a higher-privilege level, except that parameters are never automatically copied as part of a return. The process followed by a far-return stack switch in long mode and legacy mode is: 1. The return code-segment RPL is read by the processor from the CS value stored on the stack to determine that a lowerprivilege control transfer is occurring. 2. The return-program instruction pointer is popped off the current-program (higher privilege) stack and loaded into the CS and rIP registers.
138
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
3. The return instruction can include an immediate operand that specifies the number of additional bytes to be popped off of the stack. These bytes may correspond to the parameters pushed onto the stack previously by a call through a call gate containing a non-zero parameter-count field. If the return includes the immediate operand, then the stack pointer is adjusted upward by adding the specified number of bytes to the rSP. 4. The return-program stack pointer is popped off the currentprogram (higher privilege) stack and loaded into the SS and rSP registers. In the case of nested returns to 64-bit mode, a null selector can be popped into the SS register. The operand size of a far return determines the size of stack pops when switching stacks. If a far return is used in 64-bit mode to return from a prior call through a long-mode call gate, the far return must use a 64-bit operand size. The 64-bit operand size allows the far return to properly read the stack established previously by the far call. Nested Returns to 64-Bit Mode Procedures. In long mode, a far call that changes privilege levels causes the SS register to be loaded with a null selector (this is the same action taken by an interrupt in long mode). If the called procedure performs another far call to a higher-privileged procedure, or is interrupted, the null SS selector is pushed onto the stack frame, and another null selector is loaded into the SS register. Using a null selector in this way allows the processor to properly handle returns nested within 64-bit-mode procedures and interrupt handlers. Normally, a RET that pops a null selector into the SS register causes a general-protection exception (#GP) to occur. However, in long mode, the null selector acts as a flag indicating the existence of nested interrupt handlers or other privileged software in 64-bit mode. Long mode allows RET to pop a null selector into SS from the stack under the following conditions: The target mode is 64-bit mode. The target CPL is less than 3. In this case, the processor does not load an SS descriptor, and the null selector is loaded into SS without causing a #GP exception.
Chapter 4: Segmented Virtual Memory
139
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
4.12
Limit Checks
Except in 64-bit mode, limit checks are performed by all instructions that reference memory. Limit checks detect attempts to access memory outside the current segment boundary, attempts at executing instructions outside the current code segment, and indexing outside the current descriptor table. If an instruction fails a limit check, either (1) a general-protection exception occurs for all other segment-limit violations or (2) a stack-fault exception occurs for stacksegment limit violations. In 64-bit mode, segment limits are not checked during accesses to any segment referenced by the CS, DS, ES, FS, GS, and SS selector registers. Instead, the processor checks that the virtual addresses used to reference memory are in canonical-address form. In 64-bit mode, as with legacy mode and compatibility mode, descriptor-table limits are checked.
4.12.1 Determining Limit Violations
To determine segment-limit violations, the processor checks a virtual (linear) address to see if it falls outside the valid range of segment offsets determined by the segment-limit field in the descriptor. If any part of an operand or instruction falls outside the segment-offset range, a limit violation occurs. For example, a doubleword access, two bytes from an upper segment boundary, causes a segment violation because half of the doubleword is outside the segment. Three bits from the descriptor entry are used to control how the segment-limit field is interpreted: the granularity (G) bit, the default operand-size (D) bit, and for data segments, the expanddown (E) bit. See "Legacy Segment Descriptors" on page 97 for a detailed description of each bit. For all segments other than expand-down segments, the minimum segment-offset is 0. The maximum segment-offset depends on the value of the G bit: If G=0 (byte granularity), the maximum allowable segmentoffset is equal to the value of the segment-limit field. If G=1 (4096-byte granularity), the segment-limit field is first scaled by 4096 (1000h). Then 4095 (0FFFh) is added to the scaled value to arrive at the maximum allowable segment-offset, as shown in the following equation:
maximum segment-offset = (limit x 1000h) + 0FFFh
140
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
For example, if the segment-limit field is 0100h, then the maximum allowable segment-offset is (0100h x 1000h) + 0FFFh = 10_1FFFh. In both cases, the maximum segment-size is specified when the descriptor segment-limit field is 0F_FFFFh. Expand-down segments. E x p a n d - d o w n d a t a s e g m e n t s a re supported in legacy mode and compatibility mode but not in 64bit mode. With expand-down data segments, the maximum segment offset depends on the value of the D bit in the datasegment descriptor: If D=0 the maximum segment-offset is 0_FFFFh. If D=1 the maximum segment-offset is 0_FFFF_FFFFh. The minimum allowable segment offset in expand-down segments depends on the value of the G bit: If G=0 (byte granularity), the minimum allowable segment offset is the segment-limit value plus 1. For example, if the segment-limit field is 0100h, then the minimum allowable segment-offset is 0101h. If G=1 (4096-byte granularity), the segment-limit value in the descriptor is first scaled by 4096 (1000h), and then 4095 (0FFFh) is added to the scaled value to arrive at a scaled segment-limit value. The minimum allowable segment-offset is this scaled segment-limit value plus 1, as shown in the following equation:
minimum segment-offset = (limit x 1000) + 0FFFh + 1
For example, if the segment-limit field is 0100h, then the minimum allowable segment-offset is (0100h x 1000h) + 0FFFh + 1 = 10_1000h. For expand-down segments, the maximum segment size is specified when the segment-limit value is 0.
4.13
Type Checks
Type checks prevent software from using descriptors in invalid ways. Failing a type check results in an exception. Type checks are performed using five bits from the descriptor entry: the S bit and the 4-bit Type field. Together, these five bits are used to specify the descriptor type (code, data, segment, or gate) and
Chapter 4: Segmented Virtual Memory
141
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
its access characteristics. See "Legacy Segment Descriptors" on page 97 for a detailed description of the S bit and Type-field encodings. Type checks are performed by the processor in compatibility mode as well as legacy mode. Limited type checks are performed in 64-bit mode. 4.13.1 Type Checks in Legacy and Compatibility Modes The type checks performed in legacy mode and compatibility mode are listed in the following sections. Descriptor-Table Register Loads. L o a d s i n t o t h e L D T R a n d T R descriptor-table registers are checked for the appropriate system-segment type. The LDTR can only be loaded with an LDT descriptor, and the TR only with a TSS descriptor. The checks are performed during any action that causes these registers to be loaded. This includes execution of the LLDT and LTR instructions and during task switches. Segment Register Loads. The following restrictions are placed on the segment-descriptor types that can be loaded into the six user segment registers: Only code segments can be loaded into the CS register. Only writable data segments can be loaded into the SS register. Only the following segment types can be loaded into the DS, ES, FS, or GS registers: - Read-only or read/write data segments. - Readable code segments. These checks are performed during any action that causes the segment registers to be loaded. This includes execution of the MOV segment-register instructions, control transfers, and task switches. Control Transfers. Control transfers (branches and interrupts) place additional restrictions on the segment types that can be referenced during the transfer: The segment-descriptor type referenced by far CALLs and far JMPs must be one of the following: - A code segment - A call gate or a task gate - An available TSS (only allowed in legacy mode) - A task gate (only allowed in legacy mode) 142 Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Only code-segment descriptors can be referenced by callgate, interrupt-gate, and trap-gate descriptors. Only TSS descriptors can be referenced by task-gate descriptors. The link field (selector) in the TSS can only point to a TSS descriptor. This is checked during an IRET control transfer to a task. The far RET and far IRET instructions can only reference code-segment descriptors. The interrupt-descriptor table (IDT), which is referenced during interrupt control transfers, can only contain interrupt gates, trap gates, and task gates. Segment Access. After a segment descriptor is successfully loaded into one of the segment registers, reads and writes into the segments are restricted in the following ways: Writes are not allowed into read-only data-segment types. Writes are not allowed into code-segment types (executable segments). Reads from code-segment types are not allowed if the readable (R) type bit is cleared to 0. These checks are generally performed during execution of instructions that access memory. 4.13.2 Long Mode Type Check Differences Compatibility Mode and 64-bit Mode. Th e fol l ow in g t y p e ch e ck s differ in long mode (64-bit mode and compatibility mode) as compared to legacy mode: System Segments--System-segment types are checked, but the following types that are valid in legacy mode are illegal in long mode: - 16-bit available TSS. - 16-bit busy TSS. - Type-field encoding of 00h in the upper half of a systemsegment descriptor to indicate an illegal type and prevent access as a legacy descriptor. Gates--Gate-descriptor types are checked, but the following types that are valid in legacy mode are illegal in long mode: - 16-bit call gate. - 16-bit interrupt gate. Chapter 4: Segmented Virtual Memory 143
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
-
16-bit trap gate. Task gate.
64-bit Mode. 64-bit mode disables segmentation, and most of the segment-descriptor fields are ignored. The following list identifies situations where type checks in 64-bit mode differ from those in compatibility mode and legacy mode: Code Segments--The readable (R) type bit is ignored in 64-bit mode. None of the legacy type-checks that prevent reads from or writes into code segments are performed in 64-bit mode. Data Segments--Data-segment type attributes are ignored in 64-bit mode. The writable (W) and expand-down (E) type bits are ignored. All data segments are treated as writable.
144
Chapter 4: Segmented Virtual Memory
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
5
Page Translation and Protection
The x86 page-translation mechanism (or simply paging mechanism) enables system software to create separate address spaces for each process or application. These address spaces are known as virtual-address spaces. System software uses the paging mechanism to selectively map individual pages of physical memory into the virtual-address space using a set of hierarchical address-translation tables known collectively as page tables. The paging mechanism and the page tables are used to provide each process with its own private region of physical memory for storing its code and data. Processes can be protected from each other by isolating them within the virtual-address space. A process cannot access physical memory that is not mapped into its virtual-address space by system software. System software can use the paging mechanism to selectively map physical-memory pages into multiple virtual-address spaces. Mapping physical pages in this manner allows them to be shared by multiple processes and applications. The physical pages can be configured by the page tables to allow read-only access. This prevents applications from altering the pages and ensures their integrity for use by all applications. Shared mapping is typically used to allow access of sharedlibrary routines by multiple applications. A read-only copy of the library routine is mapped to each application virtualaddress space, but only a single copy of the library routine is present in physical memory. This capability also allows a copy of the operating-system kernel and various device drivers to reside within the application address space. Applications are provided with efficient access to system services without requiring costly address-space switches. The system-software portion of the address space necessarily includes system-only data areas that must be protected from accesses by applications. System software uses the page tables to protect this memory by designating the pages as supervisor pages. Such pages are only accessible by system software. Finally, system software can use the paging mechanism to map multiple, large virtual-address spaces into a much smaller amount of physical memory. Each application can use the entire
Chapter 5: Page Translation and Protection
145
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
32-bit or 64-bit virtual-address space. System software actively maps the most-frequently-used virtual-memory pages into the available pool of physical-memory pages. The least-frequentlyused virtual-memory pages are swapped out to the hard drive. This process is known as demand-paged virtual memory.
5.1
Page Translation Overview
The x86 architecture provides support for translating 32-bit virtual addresses into 32-bit physical addresses (larger physical addresses, such as 36-bit or 40-bit addresses, are supported as a special mode). The AMD64 architecture enhances this support to allow translation of 64-bit virtual addresses into 52-bit physical addresses, although processor implementations can support smaller virtual-address and physical-address spaces. Virtual addresses are translated to physical addresses through hierarchical translation tables created and managed by system software. Each table contains a set of entries that point to the next-lower table in the translation hierarchy. A single table at one level of the hierarchy can have hundreds of entries, each of which points to a unique table at the next-lower hierarchical level. Each lower-level table can in turn have hundreds of entries pointing to tables further down the hierarchy. The lowest-level table in the hierarchy points to the translated physical page. Figure 5-1 on page 147 shows an overview of the pagetranslation hierarchy used in long mode. Legacy mode paging uses a subset of this translation hierarchy (the page-map level-4 table does not exist in legacy mode and the PDP table may or may not be used, depending on which paging mode is enabled). As this figure shows, a virtual address is divided into fields, each of which is used as an offset into a translation table. The complete translation chain is made up of all table entries referenced by the virtual-address fields. The lowest-order virtual-address bits are used as the byte offset into the physical page.
146
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
63
64-Bit Virtual Address
Sign Extension Page Map Level-4 Offset Page Directory Pointer Offset Page Directory Offset Page Table Offset Physical Page Offset
0
PDPE
PTE PML4E PDE
Physical Address
Page Map Level 4 Table
Page Directory Pointer Table
Page Directory Table
Page Table
Physical Page Frame
Page Map Base Register
CR3
513-200.eps
Figure 5-1.
Virtual to Physical Address Translation--Long Mode Legacy page translation offers a variety of alternatives in translating virtual addresses to physical addresses. Three
Chapter 5: Page Translation and Protection
147
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
physical-page sizes of 4 Kbytes, 2 Mbytes, and 4 Mbytes are available. Virtual addresses are 32 bits long, and physical addresses up to the supported physical-address size can be used. The AMD64 architecture enhances the legacy translation support by allowing virtual addresses of up to 64 bits long to be translated into physical addresses of up to 52 bits long. Currently, the AMD64 architecture defines a mechanism for translating 48-bit virtual addresses to 52-bit physical addresses. The mechanism used to translate a full 64-bit virtual address is reserved and will be described in a future AMD64 architectural specification. 5.1.1 Page-Translation Options The form of page-translation support available to software depends on which paging features are enabled. Four controls are available for selecting the various paging alternatives: Page-Translation Enable (CR0.PG) Physical-Address Extensions (CR4.PAE) Page-Size Extensions (CR4.PSE) Long-Mode Active (EFER.LMA) Not all paging alternatives are available in all modes. Table 5-1 summarizes the paging support available in each mode. Table 5-1. Supported Paging Alternatives (CR0.PG=1)
PhysicalAddress Extensions (CR4.PAE) Page-Size Extensions (CR4.PSE) PageDirectory Page Size PDE.PS=0 Must be enabled - PDE.PS=1 PDE.PS=0 PDE.PS=1 - PDE.PS=0 PDE.PS=1 2 Mbyte 4 Kbyte 2 Mbyte 4 Kbyte 4 Kbyte 4 Mbyte 32-bit 52-bit 52-bit 32-bit 32-bit 40-bit Resulting PhysicalPage Size 4 Kbyte 64-bit 52-bit Maximum Virtual Address Maximum Physical Address
Mode
Long Mode
64-Bit Mode Compatibility Mode
Enabled Legacy Mode Disabled
- Disabled Enabled
148
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
5.1.2 Page-Translation Enable (PG) Bit
Page translation is controlled by the PG bit in CR0 (bit 31). When CR0.PG is set to 1, page translation is enabled. When CR0.PG is cleared to 0, page translation is disabled. The AMD64 architecture uses CR0.PG to activate and deactivate long mode when long mode is enabled. See "Enabling and Activating Long Mode" on page 424 for more information.
5.1.3 Physical-Address Extensions (PAE) Bit
Physical-address extensions are controlled by the PAE bit in CR4 (bit 5). When CR4.PAE is set to 1, physical-address extensions are enabled. When CR4.PAE is cleared to 0, physical-address extensions are disabled. Setting CR4.PAE=1 enables virtual addresses to be translated into physical addresses up to 52 bits long. This is accomplished by doubling the size of paging data-structure entries from 32 bits to 64 bits to accommodate the larger physical baseaddresses for physical-pages. PAE must be enabled before activating long mode. See "Enabling and Activating Long Mode" on page 424.
5.1.4 Page-Size Extensions (PSE) Bit
Page-size extensions are controlled by the PSE bit in CR4 (bit 4). Setting CR4.PSE to 1 allows operating-system software to use 4Mbyte physical pages in the translation process. The 4Mbyte physical pages can be mixed with standard 4-Kbyte physical pages or replace them entirely. The selection of physical-page size is made on a page-directory-entry basis. See "Page Size (PS) Bit" on page 170 for more information on physical-page size selection. When CR4.PSE is cleared to 0, page-size extensions are disabled. The choice of 2 Mbyte or 4 Mbyte as the large physical-page size depends on the value of CR4.PSE and CR4.PAE, as follows: If physical-address extensions are enabled (CR4.PAE=1), the large physical-page size is 2 Mbytes, regardless of the value of CR4.PSE. If physical-address extensions are disabled (CR4.PAE=0) and CR4.PSE=1, the large physical-page size is 4 Mbytes. If both CR4.PAE=0 and CR4.PSE=0, the only available page size is 4 Kbytes. The value of CR4.PSE is ignored when long mode is active. This is because physical-address extensions must be enabled in long
Chapter 5: Page Translation and Protection
149
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
mode, and the only available page sizes are 4 Kbytes and 2 Mbytes. In legacy mode, physical addresses up to 40 bits long can be translated from 32-bit virtual addresses using 32-bit paging data-structure entries when 4-Mbyte physical-page sizes are selected. In this special case, CR4.PSE=1 and CR4.PAE=0. See "4-Mbyte Page Translation" on page 154 for a description of the 4-Mbyte PDE that supports 40-bit physical-address translation. Th e 4 0 - b i t p hy s i c a l - a dd re s s c a p ab i l i t y i s a n A M D 6 4 architecture enhancement over the similar capability available in the legacy x86 architecture. 5.1.5 Page-Directory Page Size (PS) Bit The page directory is one of the data structures used in page translation (see Figure 5-1 on page 147). The page-size (PS) bit in the PDE (bit 7, referred to as PDE.PS) selects between standard 4-Kbyte physical-page sizes and larger (2-Mbyte or 4Mbyte) physical-page sizes. When PDE.PS is set to 1, large physical pages are used, and the PDE becomes the lowest level of the translation hierarchy. The size of the large page is determined by the values of CR4.PAE and CR4.PSE, as shown in Figure 5-1 on page 148. When PDE.PS is cleared to 0, standard 4-Kbyte physical pages are used, and the PTE is the lowest level of the translation hierarchy.
5.2
Legacy-Mode Page Translation
Legacy mode supports two forms of translation: Normal (non-PAE) Paging--This is used when physicaladdress extensions are disabled (CR4.PAE=0). Entries in the page translation table are 32 bits and are used to translate 32-bit virtual addresses into physical addresses as large as 40 bits. PAE Paging--This is used when physical-address extensions are enabled (CR4.PAE=1). Entries in the page translation table are 64 bits and are used to translate 32-bit virtual addresses into physical addresses as large as 52 bits. Legacy paging uses up to three levels of page-translation tables, depending on the paging form used and the physical-page size. Entries within each table are selected using virtual-address bit fields. The legacy page-translation tables are:
150
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Page Table--Each page-table entry (PTE) points to a physical page. If 4-Kbyte pages are used, the page table is the lowest level of the page-translation hierarchy. PTEs are not used when translating 2-Mbyte or 4-Mbyte pages. Page Directory--If 4-Kbyte pages are used, each pagedirectory entry (PDE) points to a page table. If 2-Mbyte or 4-Mbyte pages are used, a PDE is the lowest level of the page-translation hierarchy and points to a physical page. In non-PAE paging, the page directory is the highest level of the translation hierarchy. Page-Directory Pointer--Each page-directory pointer entry (PDPE) points to a page directory. Page-directory pointers are only used in PAE paging (CR4.PAE=1), and are the highest level in the legacy page-translation hierarchy. The translation-table-entry formats and how they are used in the various forms of legacy page translation are described beginning on page 152. 5.2.1 CR3 Register The CR3 register is used to point to the base address of the highest-level page-translation table. The base address is either the page-directory pointer table or the page directory table. The CR3 register format depends on the form of paging being used. Figure 5-2 shows the CR3 format when normal (non-PAE) paging is used (CR4.PAE=0). Figure 5-3 shows the CR3 format when PAE paging is used (CR4.PAE=1).
12 Page-Directory-Table Base Address 11
Reserved
31
5
4
P C D
3
P W T
2
0
Reserved
Figure 5-2.
Control Register 3 (CR3)--Non-PAE Paging Legacy-Mode
31 Page-Directory-Pointer-Table Base Address
5
4
P C D
3
P W T
2
0
Reserved
Figure 5-3.
Control Register 3 (CR3)--PAE Paging Legacy-Mode The CR3 register fields for legacy-mode paging are:
Chapter 5: Page Translation and Protection
151
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table Base Address Field. This field points to the starting physical address of the highest-level page-translation table. The size of this field depends on the form of paging used: Normal (Non-PAE) Paging (CR4.PAE=0)--This 20-bit field occupies bits 31-12, and points to the base address of the page-directory table. The page-directory table is aligned on a 4-Kbyte boundary, with the low-order 12 address bits (11-0) assumed to be 0. This yields a total base-address size of 32 bits. PAE Paging (CR4.PAE=1)--This field is 27 bits and occupies bits 31-5. The CR3 register points to the base address of the page-directory-pointer table. The page-directory-pointer table is aligned on a 32-byte boundary, with the low 5 address bits (4-0) assumed to be 0. Page-Level Writethrough (PWT) Bit. Bit 3. Page-level writethrough indicates whether the highest-level page-translation table has a writeback or writethrough caching policy. When PWT=0, the table has a writeback caching policy. When PWT=1, the table has a writethrough caching policy. Page-Level Cache Disable (PCD) Bit. Bit 4. Page-level cache disable indicates whether the highest-level page-translation table is cacheable. When PCD=0, the table is cacheable. When PCD=1, the table is not cacheable. Reserved Bits. Reserved fields should be cleared to 0 by software when writing CR3. 5.2.2 Normal (NonPAE) Paging Non-PAE paging (CR4.PAE=0) supports 4-Kbyte and 4-Mbyte physical pages, as described in the following sections. 4-Kbyte Page Translation. 4-Kbyte physical-page translation is performed by dividing the 32-bit virtual address into three fields. Each of the upper two fields are used as an index into a two-level page-translation hierarchy. The virtual-address fields are used as follows, and are shown in Figure 5-4 on page 153: Bits 31-22 index into the 1024-entry page-directory table. Bits 21-12 index into the 1024-entry page table. Bits 11-0 provide the byte offset into the physical page.
152
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Virtual Address
31 Page-Directory Offset
10 PageDirectory Table
22 21 Page-Table Offset
10
12 11 Page Offset
12
0
Page Table
4 Kbyte Physical Page 32
PTE
32 PDE
Physical Address
31 Page-Directory Base
12
CR3
Figure 5-4.
4-Kbyte Non-PAE Page Translation--Legacy Mode
Figure 5-5 shows the format of the PDE (page-directory entry), and Figure 5-6 on page 154 shows the format of the PTE (pagetable entry). Each table occupies 4 Kbytes and can hold 1024 of the 32-bit table entries. The fields within these table entries are described in "Page-Translation-Table Entry Fields" on page 168. Figure 5-5 shows bit 7 cleared to 0. This bit is the page-size bit (PS), and specifies a 4-Kbyte physical-page translation.
31 Page-Table Base Address
12 11 AVL
9876543210 I I PPUR G0GACW/ / P N N DTSW
Figure 5-5.
4-Kbyte PDE--Non-PAE Paging Legacy-Mode
Chapter 5: Page Translation and Protection
153
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
31 Physical-Page Base Address
12 11 AVL
9876543210 P PPUR GADACW/ / P T DTSW
Figure 5-6.
4-Kbyte PTE--Non-PAE Paging Legacy-Mode 4-Mbyte Page Translation. 4 - M b y t e p a g e t ra n s l a t i o n i s o n l y supported when page-size extensions are enabled (CR4.PSE=1) and physical-address extensions are disabled (CR4.PAE=0). PSE defines a page-size bit in the 32-bit PDE format (PDE.PS). This bit is used by the processor during page translation to support both 4-Mbyte and 4-Kbyte pages. 4-Mbyte pages are selected when PDE.PS is set to 1, and the PDE points directly to a 4-Mbyte physical page. PTEs are not used in a 4-Mbyte page translation. If PDE.PS is cleared to 0, or if 4-Mbyte page translation is disabled, the PDE points to a PTE. 4-Mbyte page translation is performed by dividing the 32-bit virtual address into two fields. Each field is used as an index into a single-level page-translation hierarchy. The virtualaddress fields are used as follows, and are shown in Figure 5-7 on page 155: Bits 31-22 index into the 1024-entry page-directory table. Bits 21-0 provide the byte offset into the physical page.
154
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Virtual Address
31 Page-Directory Offset
10 PageDirectory Table
22 21 Page Offset
22 4 Mbyte Physical Page
0
40 PDE
Physical Address
31 Page-Directory Base
12
CR3
Figure 5-7.
4-Mbyte Page Translation--Non-PAE Paging Legacy-Mode
The AMD64 architecture modifies the legacy 32-bit PDE format in PSE mode to increase physical-address size support to 40 bits. This increase in address size is accomplished by using bits 20-13 to hold eight additional high-order physical-address bits. Bit 21 is reserved and must be cleared to 0. Figure 5-8 shows the format of the PDE when PSE mode is enabled. The physical-page base-address bits are contained in a split field. The high-order, physical-page base-address bits 39-32 are located in PDE[20:13], and physical-page baseaddress bits 31-22 are located in PDE[31:22].
31
22 21 20
13 12 11 AVL
9876543210 PPUR G1DACW/ / P DTSW
Physical-Page Base Address [31:22] 0
P Physical-Page Base Address A [39:32] T
Figure 5-8.
4-Mbyte PDE--Non-PAE Paging Legacy-Mode 155
Chapter 5: Page Translation and Protection
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
5.2.3 PAE Paging
PAE paging is used when physical-address extensions are enabled (CR4.PAE=1). PAE paging doubles the size of pagetranslation table entries to 64 bits so that the table entries can hold larger physical addresses (up to 52 bits). The size of each table remains 4 Kbytes, which means each table can hold 512 of the 64-bit entries. PAE paging also introduces a third-level page-translation table, known as the page-directory-pointer table (PDP). The size of large pages in PAE-paging mode is 2 MBytes rather than 4 Mbytes. PAE uses the page-directory page-size bit (PDE.PS) to allow selection between 4-Kbyte and 2-Mbyte page sizes. PAE automatically uses the page-size bit, so the value of CR4.PSE is ignored by PAE paging. 4-Kbyte Page Translation. With PAE paging, 4-Kbyte physical-page translation is performed by dividing the 32-bit virtual address into four fields, each of the upper three fields is used as an index into a 3-level page-translation hierarchy. The virtualaddress fields are described as follows and are shown in Figure 5-9 on page 157: Bits 31-30 index into a 4-entry page-directory-pointer table. Bits 29-21 index into the 512-entry page-directory table. Bits 20-12 index into the 512-entry page table. Bits 11-0 provide the byte offset into the physical page.
156
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Virtual Address
31 30 29
Page-DirectoryPointer Offset
2
21 20 Page-Directory Offset
9 PageDirectory Table
12 11 Page-Table Offset
9
0 Page Offset
12 4 Kbyte Physical Page
PageDirectoryPointer Table 52* PDPE
Page Table
PTE
52*
52* PDE
Physical Address
31 Page-Directory-Pointer Base
5
CR3
*This is an architectural limit. A given processor implementation may support fewer bits.
Figure 5-9.
4-Kbyte PAE Page Translation--Legacy Mode Figures 5-10 through 5-12 on page 158 show the legacy-mode 4-Kbyte translation-table formats: Figure 5-10 shows the PDPE (page-directory-pointer entry) format. Figure 5-11 shows the PDE (page-directory entry) format. Figure 5-12 shows the PTE (page-table entry) format. The fields within these table entries are described in "PageTranslation-Table Entry Fields" on page 168. Figure 5-11 shows the PDE.PS bit cleared to 0 (bit 7), specifying a 4-Kbyte physical-page translation.
Chapter 5: Page Translation and Protection
157
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
63 Reserved, MBZ 31
52 51 Page-Directory Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Page-Directory Base Address AVL
98 Reserved, MBZ
543210 PP C W MBZ P DT
Figure 5-10.
4-Kbyte PDPE--PAE Paging Legacy-Mode
63 62 N X 31 Reserved, MBZ
52 51 Page-Table Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Page-Table Base Address AVL
9876543210 I I PPUR G0GACW/ / P N N DTSW
Figure 5-11. 4-Kbyte PDE--PAE Paging Legacy-Mode
63 62 N X 31 Reserved, MBZ
52 51 Physical-Page Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Physical-Page Base Address AVL
9876543210 P PPUR GADACW/ / P T DTSW
Figure 5-12.
4-Kbyte PTE--PAE Paging Legacy-Mode 2-Mbyte Page Translation. 2-Mbyte page translation is performed by dividing the 32-bit virtual address into three fields. Each field is used as an index into a 2-level page-translation
158
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
hierarchy. The virtual-address fields are described as follows and are shown in Figure 5-13: Bits 31-30 index into the 4-entry page-directory-pointer table. Bits 29-21 index into the 512-entry page-directory table. Bits 20-0 provide the byte offset into the physical page.
Virtual Address
31 30 29
Page-DirectoryPointer Offset
2
21 20 Page-Directory Offset
9 PageDirectory Table
0 Page Offset
21 2 Mbyte Physical Page
PageDirectoryPointer Table 52* PDPE
52* PDE
Physical Address
31 Page-Directory-Pointer Base Register
5
*This is an architectural limit. A given processor implementation may support fewer bits.
CR3
Figure 5-13.
2-Mbyte PAE Page Translation--Legacy Mode
Figure 5-14 on page 160 shows the format of the PDPE (pagedirectory-pointer entry) and Figure 5-15 on page 160 shows the format of the PDE (page-directory entry). PTEs are not used in 2-Mbyte page translations. Figure 5-15 on page 160 shows the PDE.PS bit set to 1 (bit 7), specifying a 2-Mbyte physical-page translation.
Chapter 5: Page Translation and Protection
159
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
63 Reserved, MBZ 31
52 51 Page-Directory Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Page-Directory Base Address AVL
98 Reserved, MBZ
543210 PP C W MBZ P DT
Figure 5-14.
2-Mbyte PDPE--PAE Paging Legacy-Mode
63 62 N X 31 Physical-Page Base Address Reserved, MBZ
52 51 Physical-Page Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
21 20 Reserved, MBZ
13 12 11 P A T AVL
9876543210 PPUR G1DACW/ / P DTSW
Figure 5-15.
2-Mbyte PDE--PAE Paging Legacy-Mode
5.3
Long-Mode Page Translation
Long-mode page translation requires the use of physicaladdress extensions (PAE). Before activating long mode, PAE must be enabled by setting CR4.PAE to 1. Activating long mode before enabling PAE causes a general-protection exception (#GP) to occur. The PAE-paging data structures support mapping of 64-bit virtual addresses into 52-bit physical addresses. PAE expands the size of legacy page-directory entries (PDEs) and page-table entries (PTEs) from 32 bits to 64 bits, allowing physical-address sizes of greater than 32 bits. The AMD64 architecture enhances the page-directory-pointer entry (PDPE) by defining previously reserved bits for access and protection control. A new translation table is added to PAE
160
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
paging, called the page-map level-4 (PML4). The PML4 table precedes the PDP table in the page-translation hierarchy. Because PAE is always enabled in long mode, the PS bit in the page directory entry (PDE.PS) selects between 4-Kbyte and 2Mbyte page sizes, and the CR4.PSE bit is ignored. 5.3.1 Canonical Address Form The AMD64 architecture requires implementations supporting fewer than the full 64-bit virtual address to ensure that those addresses are in canonical form. An address is in canonical form if the address bits from the most-significant implemented bit up to bit 63 are all ones or all zeros. If the addresses of all bytes in a virtual-memory reference are not in canonical form, the processor generates a general-protection exception (#GP) or a stack fault (#SS) as appropriate. In long mode, the CR3 register is used to point to the PML4 base address. CR3 is expanded to 64 bits in long mode, allowing the PML4 table to be located anywhere in the 52-bit physicaladdress space. Figure 5-16 shows the long-mode CR3 format.
52 Reserved 51 Page-Map Level-4 Table Base Address (This is an architectural limit. A given implementation may support fewer bits.) 32
5.3.2 CR3
63
31 Page-Map Level-4 Table Base Address
12
11 Reserved
5
4
P C D
3
P W T
2
0
Reserved
Figure 5-16.
Control Register 3 (CR3)--Long Mode The CR3 register fields for long mode are: Table Base Address Field. Bits 51-12. This 40-bit field points to the PML4 base address. The PML4 table is aligned on a 4-Kbyte boundary with the low-order 12 address bits (11-0) assumed to be 0. This yields a total base-address size of 52 bits. System software running on processor implementations supporting less than the full 52-bit physical-address space must clear the unimplemented upper base-address bits to 0. Page-Level Writethrough (PWT) Bit. Bit 3. Page-level writethrough indicates whether the highest-level page-translation table has a writeback or writethrough caching policy. When PWT=0, the
Chapter 5: Page Translation and Protection
161
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
table has a writeback caching policy. When PWT=1, the table has a writethrough caching policy. Page-Level Cache Disable (PCD) Bit. Bit 4. Page-level cache disable indicates whether the highest-level page-translation table is cacheable. When PCD=0, the table is cacheable. When PCD=1, the table is not cacheable. Reserved Bits. Reserved fields should be cleared to 0 by software when writing CR3. 5.3.3 4-Kbyte Page Translation In long mode, 4-Kbyte physical-page translation is performed by dividing the 48-bit virtual address into six fields. Each field is used as an index into the four-level page-translation hierarchy. The virtual-address fields are described as follows, and are shown in Figure 5-17 on page 163: Bits 63-48 are a sign extension of bit 47, as required for canonical-address forms. Bits 47-39 index into the 512-entry page-map level-4 table. Bits 38-30 index into the 512-entry page-directory pointer table. Bits 29-21 index into the 512-entry page-directory table. Bits 20-12 index into the 512-entry page table. Bits 11-0 provide the byte offset into the physical page. Note: The sizes of the sign extension and the PML4 fields depend on the number of virtual address bits supported by the implementation.
162
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Virtual Address
63 Sign Extend 48 47 Page-Map Level-4 Offset (PML4)
9 Page-Map Level-4 Table
39 38
30 29 Page-Directory Offset
9 PageDirectoryPointer Table
21 20 Page-Table Offset
9 PageDirectory Table
12 11 PhysicalPage Offset
12
0
Page-DirectoryPointer Offset
9
Page Table 52*
4 Kbyte Physical Page
PTE 52* 52* PML4E PDE PDPE 52*
Physical Address
51 Page-Map Level-4 Base Address
12
*This is an architectural limit. A given processor implementation may support fewer bits.
CR3
Figure 5-17.
4-Kbyte Page Translation--Long Mode Figures 5-18 through 5-20 on page 164 and Figure 5-21 on page 165 show the long-mode 4-Kbyte translation-table formats: Figure 5-18 shows the PML4E (page-map level-4 entry) format. Figure 5-19 shows the PDPE (page-directory-pointer entry) format. Figure 5-20 shows the PDE (page-directory entry) format. Figure 5-21 shows the PTE (page-table entry) format. The fields within these table entries are described in "PageTranslation-Table Entry Fields" on page 168. Figure 5-20 shows the PDE.PS bit (bit 7) cleared to 0, indicating a 4-Kbyte physical-page translation.
Chapter 5: Page Translation and Protection
163
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
63 62 N X 31 Available
52 51 Page-Directory-Pointer Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Page-Directory-Pointer Base Address AVL
9876543210 I PPUR MBZ G A C W / / P N DTSW
Figure 5-18.
4-Kbyte PML4E--Long Mode
63 62 N X 31 Available
52 51 Page-Directory Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Page-Directory Base Address AVL
9876543210 I PPUR MBZ G A C W / / P N DTSW
Figure 5-19.
4-Kbyte PDPE--Long Mode
63 62 N X 31 Available
52 51 Page-Table Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Page-Table Base Address AVL
9876543210 I I PPUR G0GACW/ / P N N DTSW
Figure 5-20.
4-Kbyte PDE--Long Mode
164
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
63 62 N X 31 Available
52 51 Physical-Page Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Physical-Page Base Address AVL
9876543210 P PPUR GADACW/ / P T DTSW
Figure 5-21.
4-Kbyte PTE--Long Mode In long mode, 2-Mbyte physical-page translation is performed by dividing the 48-bit virtual address into four fields. Each field is used as an index into a three-level page-translation hierarchy. The virtual-address fields are described as follows, and are shown in Figure 5-22 on page 166: Bits 63-48 are a sign extension of bit 47 as required for canonical address forms. Bits 47-39 index into the 512-entry page-map level-4 table. Bits 38-30 index into the 512-entry page-directory-pointer table. Bits 29-21 index into the 512-entry page-directory table. Bits 20-0 provide the byte offset into the physical page.
5.3.4 2-Mbyte Page Translation
Chapter 5: Page Translation and Protection
165
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Virtual Address
63 Sign Extend 48 47 39 38 30 29 21 20 Page-Map Page-DirectoryPage-Directory Level-4 Table Offset Pointer Offset Offset (PML4)
9 Page-Map Level-4 Table 9 PageDirectoryPointer Table 9 PageDirectory Table
0 Page Offset
21 2 Mbyte Physical Page
52* 52* PML4E PDE PDPE 52* Physical Address
51 Page-Map Level-4 Base Address
12
*This is an archtectural limit. A given processor implementation may support fewer bits.
CR3
Figure 5-22.
2-Mbyte Page Translation--Long Mode Figures 5-23 through 5-25 on page 167 show the long-mode 2-Mbyte translation-table formats (the formats are identical to those used for 4-Kbyte page translations and are repeated here for clarity): Figure 5-23 shows the PML4E (page-map level-4 entry) format. Figure 5-24 shows the PDPE (page-directory-pointer entry) format. Figure 5-25 shows the PDE (page-directory entry) format. The fields within these table entries are described in "PageTranslation-Table Entry Fields" on page 168. PTEs are not used in 2-Mbyte page translations. Figure 5-25 shows the PDE.PS bit (bit 7) set to 1, indicating a 2-Mbyte physical-page translation.
166
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
63 62 N X 31 Available
52 51 Page-Directory-Pointer Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Page-Directory-Pointer Base Address AVL
9876543210 I PPUR MBZ G A C W / / P N DTSW
Figure 5-23.
2-Mbyte PML4E--Long Mode
63 62 N X 31 Available
52 51 Page-Directory Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
12 11 Page-Directory Base Address AVL
9876543210 I PPUR MBZ G A C W / / P N DTSW
Figure 5-24.
2-Mbyte PDPE--Long Mode
63 N X 31 Page-Table Base Address Available
52 51 Page-Table Base Address
(This is an architectural limit. A given implementation may support fewer bits.)
32
21 20 Reserved, MBZ
13 12 11 P A T AVL
9876543210 PPUR G1DACW/ / P DTSW
Figure 5-25.
2-Mbyte PDE--Long Mode
Chapter 5: Page Translation and Protection
167
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
5.4
Page-Translation-Table Entry Fields
The page-translation-table entries contain control and informational fields used in the management of the virtualmemory environment. Most fields are common across all translation table entries and modes and occupy the same bit locations. However, some fields are located in different bit positions depending on the page translation hierarchical level, and other fields have different sizes depending on which physical-page size, physical-address size, and operating mode are selected. Although these fields can differ in bit position or size, their meaning is consistent across all levels of the page translation hierarchy and in all operating modes.
5.4.1 Field Definitions
The following sections describe each field within the pagetranslation table entries. Translation-Table Base Address Field. The translation-table baseaddress field points to the physical base address of the nextlower-level table in the page-translation hierarchy. Page datastructure tables are always aligned on 4-Kbyte boundaries, so only the address bits above bit 11 are stored in the translationtable base-address field. Bits 11-0 are assumed to be 0. The size of the field depends on the mode: In normal (non-PAE) paging (CR4.PAE=0), this field specifies a 32-bit physical address. In PAE paging (CR4.PAE=1), this field specifies a 52-bit physical address. 52 bits correspond to the maximum physical-address size a l l ow e d by t h e A M D 6 4 a rc h i t e c t u re . I f a p ro c e s s o r implementation supports fewer than the full 52-bit physical address, software must clear the unimplemented high-order translation-table base-address bits to 0. For example, if a processor implementation supports a 40-bit physical-address size, software must clear bits 51-40 when writing a translationtable base-address field in a page data-structure entry. Physical-Page Base Address Field. The physical-page base-address field points to the base address of the translated physical page. This field is found only in the lowest level of the pagetranslation hierarchy. The size of the field depends on the mode:
168
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
In normal (non-PAE) paging (CR4.PAE=0), this field specifies a 32-bit base address for a physical page. In PAE paging (CR4.PAE=1), this field specifies a 52-bit base address for a physical page. Physical pages can be 4 Kbytes, 2 Mbytes, or 4 Mbytes, and they are always aligned on an address boundary corresponding to the physical-page length. For example, a 2-Mbyte physical page is always aligned on a 2-Mbyte address boundary. Because of this alignment, the low-order address bits are assumed to be 0, as follows: 4-Kbyte pages, bits 11-0 are assumed 0. 2-Mbyte pages, bits 20-0 are assumed 0. 4-Mbyte pages, bits 21-0 are assumed 0. Present (P) Bit. Bit 0. This bit indicates whether the pagetranslation table or physical page is loaded in physical memory. When the P bit is cleared to 0, the table or physical page is not loaded in physical memory. When the P bit is set to 1, the table or physical page is loaded in physical memory. Software clears this bit to 0 to indicate a page table or physical page is not loaded in physical memory. A page-fault exception (#PF) occurs if an attempt is made to access a table or page when the P bit is 0. System software is responsible for loading the missing table or page into memory and setting the P bit to 1. When the P bit is 0, indicating a not-present page, all remaining bits in the page data-structure entry are available to software. Read/Write (R/W) Bit. Bit 1. This bit controls read/write access to all physical pages mapped by the table entry. For example, a page-map level-4 R/W bit controls read/write access to all 128M (512 x 512 x 512) physical pages it maps through the lower-level translation tables. When the R/W bit is cleared to 0, access is restricted to read-only. When the R/W bit is set to 1, both read and write access is allowed. See "Page-Protection Checks" on page 174 for a description of the paging read/write protection mechanism. User/Supervisor (U/S) Bit. Bit 2. This bit controls user (CPL 3) access to all physical pages mapped by the table entry. For example, a page-map level-4 U/S bit controls the access allowed to all 128M (512 x 512 x 512) physical pages it maps through the Chapter 5: Page Translation and Protection 169
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
lower-level translation tables. When the U/S bit is cleared to 0, access is restricted to supervisor level (CPL 0, 1, 2). When the U/S bit is set to 1, both user and supervisor access is allowed. See "Page-Protection Checks" on page 174 for a description of the paging user/supervisor protection mechanism. Page-Level Writethrough (PWT) Bit. Bit 3. This bit indicates whether the page-translation table or physical page to which this entry points has a writeback or writethrough caching policy. When the PWT bit is cleared to 0, the table or physical page has a writeback caching policy. When the PWT bit is set to 1, the table or physical page has a writethrough caching policy. See "Memory Caches" on page 208 for additional information on caching. Page-Level Cache Disable (PCD) Bit. Bit 4. This bit indicates whether the page-translation table or physical page to which this entry points is cacheable. When the PCD bit is cleared to 0, the table or physical page is cacheable. When the PCD bit is set to 1, the table or physical page is not cacheable. See "Memory Caches" on page 208 for additional information on caching. Accessed (A) Bit. Bit 5. This bit indicates whether the pagetranslation table or physical page to which this entry points has been accessed. The A bit is set to 1 by the processor the first time the table or physical page is either read from or written to. The A bit is never cleared by the processor. Instead, software must clear this bit to 0 when it needs to track the frequency of table or physical-page accesses. Dirty (D) Bit. Bit 6. This bit is only present in the lowest level of the page-translation hierarchy. It indicates whether the pagetranslation table or physical page to which this entry points has been written. The D bit is set to 1 by the processor the first time there is a write to the physical page. The D bit is never cleared by the processor. Instead, software must clear this bit to 0 when it needs to track the frequency of physical-page writes. Page Size (PS) Bit. Bit 7. This bit is only present in page-directory entries. It specifies the lowest level of the page-translation hierarchy and the physical-page size. When the PS bit is cleared to 0, the lowest level of the page-translation hierarchy is the page-table entry (PTE), and the physical-page size is 4 Kbytes. When the PS bit is set to 1, the lowest level of the page-
170
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
translation hierarchy is the page-directory entry (PDE)s. The physical-page size is determined as follows: If CR4.PAE=0 and CR4.PSE=1, the physical-page size is 4 Mbytes. If CR4.PAE=1, the physical-page size is 2 Mbytes. See Table 5-1 on page 148 for a description of the relationship between the PS bit, PAE, physical-page sizes, and pagetranslation hierarchy. Global Page (G) Bit. Bit 8. This bit is only present in the lowest level of the page-translation hierarchy. It indicates the physical page is a global page. The TLB entry for a global page (G=1) is not invalidated when CR3 is loaded either explicitly by a MOV CRn instruction or implicitly during a task switch. Use of the G bit requires the page-global enable bit in CR4 to be set to 1 (CR4.PGE=1). See "Global Pages" on page 173 for more information on the global-page mechanism. Available to Software (AVL) Bit. These bits are not interpreted by the processor and are available for use by system software. Page-Attribute Table (PAT) Bit. This bit is only present in the lowest level of the page-translation hierarchy, as follows: If the lowest level is a PTE (PDE.PS=0), PAT occupies bit 7. If the lowest level is a PDE (PDE.PS=1), PAT occupies bit 12. The PAT bit is the high-order bit of a 3-bit index into the PAT register (Figure 7-10 on page 230). The other two bits involved in forming the index are the PCD and PWT bits. Not all processors support the PAT bit by implementing the PAT registers. See "Page-Attribute Table Mechanism" on page 230 for a description of the PAT mechanism and how it is used. No Execute (NX) Bit. Bit 63. This bit is present in the translationtable entries defined for PAE paging, with the exception that the legacy-mode PDPE does not contain this bit. This bit is not supported by non-PAE paging. The NX bit can only be set when the no-execute page-protection feature is enabled by setting EFER.NXE to 1 (see "Extended Fe a t u re E n a b l e R e g i s t e r ( E F E R ) " o n p a g e 6 9 ) . I f EFER.NXE=0, the NX bit is treated as reserved. In this case, a
Chapter 5: Page Translation and Protection
171
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
page-fault exception (#PF) occurs if the NX bit is not cleared to 0. This bit controls the ability to execute code from all physical pages mapped by the table entry. For example, a page-map level-4 NX bit controls the ability to execute code from all 128M (512 x 512 x 512) physical pages it maps through the lower-level translation tables. When the NX bit is cleared to 0, code can be executed from the mapped physical pages. When the NX bit is set to 1, code cannot be executed from the mapped physical pages. See "No Execute (NX) Bit" on page 175 for a description of the no-execute page-protection mechanism. Reserved Bits. Software should clear all reserved bits to 0. If the processor is in long mode, or if page-size and physical-address extensions are enabled in legacy mode, a page-fault exception (#PF) occurs if reserved bits are not cleared to 0.
5.5
Translation-Lookaside Buffer (TLB)
When paging is enabled, every memory access has its virtual address automatically translated into a physical address using the page-translation hierarchy. Translation-lookaside buffers (TLBs), also known as page-translation caches, nearly eliminate the performance penalty associated with page translation. TLBs are special on-chip caches that hold the most-recently used virtual-to-physical address translations. Each memory reference (instruction and data) is checked by the TLB. If the translation is present in the TLB, it is immediately provided to the processor, thus avoiding external memory references for accessing page tables. TLBs take advantage of the principle of locality. That is, if a memory address is referenced, it is likely that nearby memory addresses will be referenced in the near future. In the context of paging, the proximity of memory addresses required for locality can be broad--it is equal to the page size. Thus, it is possible for a large number of addresses to be translated by a small number of page translations. This high degree of locality means that almost all translations are performed using the onchip TLBs. System software is responsible for managing the TLBs when updates are made to the linear-to-physical mapping of addresses. A change to any paging data-structure entry is not
172
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
automatically reflected in the TLB, and hardware snooping of TLBs during memory-reference cycles is not performed. Software must invalidate the TLB entry of a modified translation-table entry so that the change is reflected in subsequent address translations. TLB invalidation is described in "TLB Management" on page 173. Only privileged software running at CPL=0 can manage the TLBs. 5.5.1 Global Pages The processor invalidates the TLB whenever CR3 is loaded either explicitly or implicitly. After the TLB is invalidated, subsequent address references can consume many clock cycles until their translations are cached as new entries in the TLB. Invalidation of TLB entries for frequently-used or critical pages can be avoided by specifying the translations for those pages as global. TLB entries for global pages are not invalidated as a result of a CR3 load. Global pages are invalidated using the INVLPG instruction. Global-page extensions are controlled by setting and clearing the PGE bit in CR4 (bit 7). When CR4.PGE is set to 1, globalpage extensions are enabled. When CR4.PGE is cleared to 0, global-page extensions are disabled. When CR4.PGE=1, setting the global (G) bit in the translation-table entry marks the page as global. The INVLPG instruction ignores the G bit and can be used to invalidate individual global-page entries in the TLB. To invalidate all entries, including global-page entries, disable global-page extensions (CR4.PGE=0). 5.5.2 TLB Management Generally, unless system software modifies the linear-tophysical address mapping, the processor manages the TLB transparently to software. This includes allocating entries and replacing old entries with new entries. Software changes made to paging-data structures are not automatically reflected in the TLB. In these situations, it is necessary for software to invalidate TLB entries so that these changes are immediately propagated to the page-translation mechanism. TLB entries can be explicitly invalidated using operations intended for that purpose or implicitly invalidated as a result of another operation. TLB invalidation has no effect on the associated page-translation tables in memory.
Chapter 5: Page Translation and Protection
173
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Explicit Invalidations. Two mechanisms are provided to explicitly invalidate the TLB: The invalidate TLB entry instruction (INVLPG) can be used to invalidate specific entries within the TLB. This instruction invalidates a page, regardless of whether it is marked as global or not. Updates to the CR3 register cause the entire TLB to be invalidated except for global pages. The CR3 register can be updated with the MOV CR3 instruction. CR3 is also updated during a task switch, with the updated CR3 value read from the TSS of the new task. Implicit Invalidations. The following operations cause the entire TLB to be invalidated, including global pages: Modifying the CR0.PG bit (page enable) or the CR0.PE bit (protected-mode enable). Modifying the CR4.PAE bit (physical-address extensions), the CR4.PSE bit (page-size extensions), or the CR4.PGE bit (page-global enable). Entering SMM as a result of an SMI interrupt. Executing the RSM instruction to return from SMM. Updating a memory-type range register (MTRR) with the WRMSR instruction. External initialization of the processor. External masking of the A20 address bit (asserting the A20M# input signal).
5.6
Page-Protection Checks
Two forms of page-level memory protection are provided by the legacy architecture. The first form of protection prevents nonprivileged (user) code and data from accessing privileged (supervisor) code and data. The second form of protection prevents writes into read-only address spaces. The AMD64 architecture introduces a third form of protection that prevents software from attempting to execute data pages as instructions. All of these forms of protection are available at all levels of the page-translation hierarchy. The processor checks a page for execute permission only when the page translation is loaded into the instruction TLB as a
174
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
result of a page-table walk. The remaining protection checks are performed when a virtual address is translated into a physical address. For those checks, the processor examines the page-level memory-protection bits in the translation tables to determine if the access is allowed. The bits involved in these checks are: User/Supervisor (U/S)--The U/S bit is introduced in "User/Supervisor (U/S) Bit" on page 169. Read/Write (R/W)--The R/W bit is introduced in "Read/Write (R/W) Bit" on page 169. Write-Protect Enable (CR0.WP)--The CR0.WP bit is introduced in "Write Protect (WP) Bit" on page 57. 5.6.1 No Execute (NX) Bit The NX bit in the page-translation tables specifies whether instructions can be executed from the page. This bit is not checked during every instruction fetch. Instead, the NX bits in the page-translation-table entries are checked by the processor when the instruction TLB is loaded with a page translation. The processor attempts to load the translation into the instruction TLB when an instruction fetch misses the TLB. If a set NX bit is detected (indicating the page is not executable), a page-fault exception (#PF) occurs. The no-execute protection check applies to all privilege levels. It does not distinguish between supervisor and user-level accesses. The no-execute protection feature is supported only in PAEpaging mode. It is enabled by setting the NXE bit in the EFER register to 1 (see "Extended Feature Enable Register (EFER)" on page 69). Before setting this bit, system software must verify the processor supports the NX feature by checking the CPUID extended-feature flags (see "Function 8000_0001h: Processor Signature and AMD Features" in Volume 3). 5.6.2 User/Supervisor (U/S) Bit The U/S bit in the page-translation tables determines the privilege level required to access the page. Conceptually, user (non-privileged) pages correspond to a current privilege-level (CPL) of 3, or least-privileged. Supervisor (privileged) pages correspond to a CPL of 0, 1, or 2, all of which are jointly regarded as most-privileged. When the processor is running at a CPL of 0, 1, or 2, it can access both user and supervisor pages. However, when the Chapter 5: Page Translation and Protection 175
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
processor is running at a CPL of 3, it can only access user pages. If an attempt is made to access a supervisor page while the processor is running at CPL=3, a page-fault exception (#PF) occurs. See "Segment-Protection Overview" on page 118 for more information on the protection-ring concept and CPL. 5.6.3 Read/Write (R/W) Bit The R/W bit in the page-translation tables specifies the access type allowed for the page. If R/W=1, the page is read/write. If R/W=0, the page is read-only. A page-fault exception (#PF) occurs if an attempt is made by user software to write to a readonly page. If supervisor software attempts to write a read-only page, the outcome depends on the value of the CR0.WP bit (described below). The ability to write to read-only pages is governed by the processor mode and whether write protection is enabled. If write protection is not enabled, a processor running at CPL 0, 1, or 2 can write to any physical page, even if it is marked as readonly. Enabling write protection prevents supervisor code from writing into read-only pages, including read-only user-level pages. A page-fault exception (#PF) occurs if software attempts to write (at any privilege level) into a read-only page while write protection is enabled.
5.6.4 Write Protect (CR0.WP) Bit
5.7
Protection Across Paging Hierarchy
The privilege level and access type specified at each level of the page-translation hierarchy have a combined effect on the protection of the translated physical page. Enabling and disabling write protection further qualifies the protection effect on the physical page. Table 5-2 on page 177 shows the overall effect that privilege level and access type have on physical-page protection when write protection is disabled (CR0.WP=0). In this case, when any translation-table entry is specified as supervisor level, the physical page is a supervisor page and can only be accessed by software running at CPL 0, 1, or 2. Such a page allows read/write access even if all levels of the page-translation hierarchy specify read-only access.
176
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 5-2.
Physical-Page Protection, CR0.WP=0
Page-DirectoryPointer Entry U/S -- S -- -- U U U U U R/W -- -- -- -- -- R -- -- R/W Page-Directory Entry U/S -- -- S -- U U U U U R/W -- -- -- -- -- -- R -- R/W Page-Table Entry U/S -- -- -- S U U U U U R/W -- -- -- -- -- -- -- R R/W U R/W U R1 S R/W Effective Result on Physical Page U/S R/W
Page-Map Level-4 Entry U/S S -- -- -- U U U U U
Note:
R/W -- -- -- -- R -- -- -- R/W
S = Supervisor Level (CPL=0, 1, or 2), U = User Level (CPL = 3), R = Read-Only Access, R/W = Read/Write Access, -- = Don't Care.
Note:
1. Supervisor-level programs can access these pages as R/W.
If all table entries in the translation hierarchy are specified as user level the physical page is a user page, and both supervisor and user software can access it. In this case the physical page is read-only if any table entry in the translation hierarchy specifies read-only access. All table entries in the translation hierarchy must specify read/write access for the physical page to be read/write. Table 5-3 on page 178 shows the overall effect that privilege level and access type have on physical-page access when write protection is enabled (CR0.WP=1). When any translation-table entry is specified as supervisor level, the physical page is a supervisor page and can only be accessed by supervisor software. In this case, the physical page is read-only if any table entry in the translation hierarchy specifies read-only access. All table entries in the translation hierarchy must specify read/write access for the supervisor page to be read/write.
Chapter 5: Page Translation and Protection
177
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 5-3. Effect of CR0.WP=1 on Supervisor Page Access
Page-Map Level-4 Entry R/W R -- -- -- W
Note:
Page DirectoryPointer Entry R/W -- R -- -- W
Page Directory Entry R/W -- -- R -- W
Page Table Entry R/W -- -- -- R W
Physical Page
R/W
R
W
R = Read-Only Access Type, W = Read/Write Access Type, -- = Don't Care. Physical page is in supervisor mode, as determined by U/S settings in Table 5-2.
5.7.1 Access to User Pages when CR0.WP=1
As shown in Table 5-2 on page 177, read/write access to userlevel pages behaves the same as when write protection is disabled (CR0.WP=0), with one critical difference. When write protection is enabled, supervisor programs cannot write into read-only user pages.
5.8
Effects of Segment Protection
Segment-protection and page-protection checks are performed serially by the processor, with segment-privilege checks performed first, followed by page-protection checks. Pageprotection checks are not performed if a segment-protection violation is found. If a violation is found during either segmentprotection or page-protection checking, an exception occurs and no memory access is performed. Segment-protection violations cause either a general-protection exception (#GP) or a stack exception (#SS) to occur. Page-protection violations cause a page-fault exception (#PF) to occur.
178
Chapter 5: Page Translation and Protection
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
6
System-Management Instructions
System-management instructions provide control over the resources used to manage the processor operating environment. This includes memory management, memory protection, task management, interrupt and exception handling, systemmanagement mode, software debug and performance analysis, and model-specific features. Most instructions used to access these resources are privileged and can only be executed while the processor is running at CPL=0, although some instructions can be executed at any privilege level. Table 6-1 summariz es the instructions used for system management. These include all privileged instructions, instructions whose privilege requirement is under the control of system software, non-privileged instructions that are used primarily by system software, and instructions used to transfer control to system software. Most of the instructions listed in Table 6-1 are summarized in this chapter, although a few are introduced elsewhere in this manual, as indicated in the Reference column of Table 6-1. For details on individual system instructions, see "System Instruction Reference" in Volume 3.
Table 6-1.
Mnemonic ARPL CLI CLTS HLT INT3 INVD INVLPG IRETx
Note:
System-Management Instructions
Name Adjust Requestor Privilege Level Clear Interrupt Flag Clear Task-Switched Flag in CR0 Halt Interrupt to Debug Vector Invalidate Caches Invalidate TLB Entry Interrupt Return (all forms) X X X X X X X Privilege CPL=0 O/S1 Any X Reference "Adjusting Access Rights" on page 190 "CLI and STI Instructions" on page 187 "CLTS Instruction" on page 187 "Processor Halt" on page 190 "Breakpoint Instruction (INT3)" on page 401 "Cache Management" on page 191 "TLB Invalidation" on page 191 "Returning From Interrupt Procedures" on page 283
1. The operating system controls the privilege required to use the instruction.
Chapter 6: System-Management Instructions
179
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 6-1.
Mnemonic LAR LGDT LIDT LLDT LMSW LSL LTR MOV CRn MOV DRn RDMSR RDPMC RDTSC RSM SGDT SIDT SLDT SMSW STI STR SWAPGS SYSCALL
Note:
System-Management Instructions (continued)
Name Load Access-Rights Byte Load Global-Descriptor-Table Register Load Interrupt-Descriptor-Table Register Load Local-Descriptor-Table Register Load Machine-Status Word Load Segment Limit Load Task Register Move to/from Control Registers Move to/from Debug Registers Read Model-Specific Register Read Performance-Monitor Counter Read Time-Stamp Counter Return from System-Management Mode Store Global-Descriptor-Table Register Store Interrupt-Descriptor-Table Register Store Local-Descriptor-Table Register Store Machine-Status Word Set Interrupt Flag Store Task Register Swap GS and KernelGSbase Registers Fast System Call X X X X X X X X X X X X X X X X X X X X Privilege CPL=0 O/S1 Any X Reference "Checking Access Rights" on page 189 "LGDT and LIDT Instructions" on page 189 "LLDT and LTR Instructions" on page 189 "LMSW and SMSW Instructions" on page 186 "Checking Segment Limits" on page 190 "LLDT and LTR Instructions" on page 189 "MOV CRn Instructions" on page 186 "Accessing Debug Registers" on page 187 "RDMSR and WRMSR Instructions" on page 188 "RDPMC Instruction" on page 188 "RDTSC Instruction" on page 188 "Leaving SMM" on page 337 "SGDT and SIDT Instructions" on page 189 "SLDT and STR Instructions" on page 189 "LMSW and SMSW Instructions" on page 186 "CLI and STI Instructions" on page 187 "SLDT and STR Instructions" on page 189 "SWAPGS Instruction" on page 185 "SYSCALL and SYSRET" on page 182
1. The operating system controls the privilege required to use the instruction.
180
Chapter 6: System-Management Instructions
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 6-1.
Mnemonic SYSENTER SYSEXIT SYSRET UD2 VERR VERW WBINVD WRMSR
Note:
System-Management Instructions (continued)
Name System Call System Return Fast System Return Undefined Operation Verify Segment for Reads Verify Segment for Writes Writeback and Invalidate Caches Write Model-Specific Register X X X X X X X Privilege CPL=0 O/S1 Any X Reference "SYSENTER and SYSEXIT (Legacy Mode Only)" on page 184 "SYSCALL and SYSRET" on page 182 "System Instruction Reference" in Volume 3 "Checking Read/Write Rights" on page 190 "Cache Management" on page 191 "RDMSR and WRMSR Instructions" on page 188
1. The operating system controls the privilege required to use the instruction.
The following instructions are summarized in this chapter but are not categorized as system instructions, because of their importance to application programming: The CPUID instruction returns information critical to system software in initializing the operating environment. It is fully described in "Processor Feature Identification" on page 78. The PUSHF and POPF instructions set and clear certain RFLAGS bits depending on the processor operating mode and privilege level. These dependencies are described in "POPF and PUSHF Instructions" on page 187. The MOV, PUSH, and POP instructions can be used to load and store segment registers, as described in "MOV, POP, and PUSH Instructions" on page 188.
6.1
Fast System Call and Return
Operating systems can use both paging and segmentation to implement protected memory models. Segment descriptors provide the necessary memory protection and privilege checking for segment accesses. By setting segment-descriptor fields appropriately, operating systems can enforce access restrictions as needed.
Chapter 6: System-Management Instructions
181
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
A disadvantage of segment-based protection and privilege checking is the overhead associated with loading a new segment selector (and its corresponding descriptor) into a segment register. Even when using the flat-memory model, this overhead still occurs when switching between privilege levels because code segments (CS) and stack segments (SS) are reloaded with different segment descriptors. To initiate a call to the operating system, an application transfers control to the operating system through a gate descriptor (call, interrupt, trap, or task gate). In the past, control was transferred using either a far CALL instruction or a software interrupt. Transferring control through one of these gates is slowed by the segmentation-related overhead, as is the later return using a far RET or IRET instruction. The following checks are performed when control is transferred in this manner: Selectors, gate descriptors, and segment descriptors are in the proper form. Descriptors lie within the bounds of the descriptor tables. Gate descriptors reference the appropriate segment descriptors. The caller, gate, and target privileges all allow the control transfer to take place. The stack created by the call has sufficient properties to allow the transfer to take place. In addition to these call-gate checks, other checks are made involving the task-state segment when a task switch occurs. 6.1.1 SYSCALL and SYSRET SYSCALL and SYSRET Instructions. SYSCALL and SYSRET are lowlatency system call and return instructions. These instructions assume the operating system implements a flat-memory model, which greatly simplifies calls to and returns from the operating system. This simplification comes from eliminating unneeded checks, and by loading pre-determined values into the CS and SS segment registers (both visible and hidden portions). As a result, SYSCALL and SYSRET can take fewer than one-fourth the number of internal clock cycles to complete than the legacy CALL and RET instructions. SYSCALL and SYSRET are particularly well-suited for use in 64-bit mode, which requires implementation of a paged, flat-memory model.
182
Chapter 6: System-Management Instructions
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
SYSCALL and SYSRET require that the code-segment base, limit, and attributes (except for CPL) are consistent for all application and system processes. Only the CPL is allowed to vary. The processor assumes (but does not check) that the SYSCALL target CS has CPL=0 and the SYSRET target CS has CPL=3. For details on the SYSCALL and SYSRET instructions, see "System Instruction Reference" in Volume 3. SYSCALL and SYSRET MSRs. T h e S TA R , L S TA R , a n d C S TA R registers are model-specific registers (MSRs) used to specify the target address of a SYSCALL instruction as well as the CS and SS selectors of the called and returned procedures. The SFMASK register is used in long mode to specify how rFLAGS is handled by these instructions. Figure 6-1 shows the STAR, LSTAR, CSTAR, and SFMASK register formats.
63 STAR
48 47
32 31 32-bit SYSCALL Target EIP
0
C000_0081h SYSRET CS and SS
SYSCALL CS and SS
LSTAR
C000_0082h
Target RIP for 64-Bit-Mode Calling Software
CSTAR C000_0083h
Target RIP for Compatibility-Mode Calling Software
SFMASK C000_0084h
Reserved, RAZ
SYSCALL Flag Mask
Figure 6-1.
STAR, LSTAR, CSTAR, and MASK MSRs STAR--The STAR register has the following fields (unless otherwise noted, all bits are read/write): - SYSRET CS and SS Selectors--Bits 63-48. This field is used to specify both the CS and SS selectors loaded into CS and SS during SYSRET. If SYSRET is returning to 32-bit mode (either legacy or compatibility), this field is copied directly into the CS selector field. If SYSRET is returning to 64-bit mode, the CS selector is set to this field + 16. SS.Sel is set to this field + 8, regardless of the target
Chapter 6: System-Management Instructions
183
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
mode. Because SYSRET always returns to CPL 3, the RPL bits 49-48 should be initialized to 11b. - SYSCALL CS and SS Selectors--Bits 47-32. This field is used to specify both the CS and SS selectors loaded into CS and SS during SYSCALL. This field is copied directly into CS.Sel. SS.Sel is set to this field + 8. Because SYSCALL always switches to CPL 0, the RPL bits 33-32 should be initialized to 00b. - 32-bit SYSCALL Target EIP--Bits 31-0. This is the target EIP of the called procedure. The legacy STAR register is not expanded in long mode to provide a 64-bit target RIP address. Instead, long mode provides two new STAR registers--long STAR (LSTAR) and compatibility STAR (CSTAR)--that hold a 64-bit target RIP. LSTAR and CSTAR--The LSTAR register holds the target RIP of the called procedure in long mode when the calling software is in 64-bit mode. The CSTAR register holds the target RIP of the called procedure in long mode when the calling software is in compatibility mode. The WRMSR instruction is used to load the target RIP into the LSTAR and CSTAR registers. If the RIP written to either of the MSRs is not in canonical form, a #GP fault is generated on the WRMSR instruction. SFMASK--The SFMASK register is used to specify which RFLAGS bits are cleared during a SYSCALL. In long mode, SFMASK is used to specify which RFLAGS bits are cleared when SYSCALL is executed. If a bit in SFMASK is set to 1, the corresponding bit in RFLAGS is cleared to 0. If a bit in SFMASK is cleared to 0, the corresponding rFLAGS bit is not modified. 6.1.2 SYSENTER and SYSEXIT (Legacy Mode Only) SYSENTER and SYSEXIT Instructions. Like SYSCALL and SYSRET, SYSENTER and SYSEXIT are low-latency system call and return instructions designed for use by system and application software implementing a flat-memory model. However, these instructions are illegal in long mode and result in an undefined opcode exception (#UD) if software attempts to use them. Software should use the SYSCALL and SYSRET instructions when running in long mode. SYSENTER and SYSEXIT MSRs. Th re e m o d e l - s p e c i f i c re g i s t e rs (MSRs) are used to specify the target address and stack pointers for the SYSENTER instruction as well as the CS and 184 Chapter 6: System-Management Instructions
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
SS selectors of the called and returned procedures. The register fields are: SYSENTER Target CS--Holds the CS selector of the called procedure. SYSENTER Target ESP--Holds the called-procedure stack pointer. The SS selector is updated automatically to point to the next descriptor entry after the SYSENTER Target CS, and ESP is the offset into that stack segment. SYSENTER Target EIP--Holds the offset into the CS of the called procedure. Figure 6-2 shows the register formats and their corresponding MSR IDs.
63 SYSENTER_CS 174h
32 31
16 15 SYSENTER Target CS
0
SYSENTER_ESP 175h
SYSENTER Target ESP
SYSENTER_EIP 176h
SYSENTER Target EIP
Figure 6-2.
SYSENTER_CS, SYSENTER_ESP, SYSENTER_EIP MSRs The SWAPGS instruction provides a fast method for system software to load a pointer to system data structures. SWAPGS can be used upon entering system-software routines as a result of a SYSCALL instruction or as a result of an interrupt or exception. Before returning to application software, SWAPGS can restore an application data-structure pointer that was replaced by the system data-structure pointer. SWAPGS exchanges the base-address value located in the Ke r n e l G S b a s e m o d e l - s p e c i f i c re g i s t e r ( M S R a dd re s s C000_0102h) with the base-address value located in the hidden portion of the GS selector register (GS.base). This exchange allows the system-kernel software to quickly access kernel data structures by using the GS segment-override prefix during memory references.
6.1.3 SWAPGS Instruction
Chapter 6: System-Management Instructions
185
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
The need for SwapGS arises from the requirement that, upon entry to the OS kernel, the kernel needs to obtain a 64-bit pointer to its essential data structures. When using SYSCALL to implement system calls, no kernel stack exists at the OS entry point. Neither is there a straightforward method to obtain a pointer to kernel structures, from which the kernel stack pointer could be read. Thus, the kernel cannot save GPRs or reference memory. SwapGS does not require any GPR or memory operands, so no registers need to be saved before using it. Similarly, when the OS kernel is entered via an interrupt or exception (where the kernel stack is already set up), SwapGS can be used to quickly get a pointer to the kernel data structures. See "FS and GS Registers in 64-Bit Mode" on page 88 for more information on using the GS.base register in 64-bit mode.
6.2
System Status and Control
System-status and system-control instructions are used to determine the features supported by a processor, gather information about the current execution state, and control the processor operating modes.
6.2.1 Processor Feature Identification (CPUID)
CPUID Instruction. The CPUID instruction provides complete information about the processor implementation and its capabilities. Software operating at any privilege level can execute the CPUID instruction to collect this information. System software normally uses the CPUID instruction to determine which optional features are available so the system can be configured appropriately. The optional features identified by the CPUID instruction are described in "CPUID" in Volume 3. MOV CRn Instructions. The MOV CRn instructions can be used to copy data between the control registers and the generalpurpose registers. These instructions are privileged and cause a general-protection exception (#GP) if non-privileged software attempts to execute them. LMSW and SMSW Instructions. The machine status word is located in CR0 register bits 15-0. The load machine status word (LMSW) instruction writes only the least-significant four status-word bits (CR0[3:0]). All remaining status-word bits (CR0[15:4]) are
6.2.2 Accessing Control Registers
186
Chapter 6: System-Management Instructions
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
left unmodified by the instruction. The instruction is privileged and causes a #GP to occur if non-privileged software attempts to execute it. The store machine status word (SMSW) instruction stores all 16 status-word bits (CR0[15:0]) into the target GPR or memory location. The instruction is not privileged and can be executed by all software. CLTS Instruction. The clear task-switched bit instruction (CLTS) clears CR0.TS to 0. The CR0.TS bit is set to 1 by the processor every time a task switch takes place. The bit is useful to system software in determining when the x87 and multimedia register state should be saved or restored. See "Task Switched (TS) Bit" on page 56 for more information on using CR0.TS to manage x87-instruction state. The CLTS instruction is privileged and causes a #GP to occur if non-privileged software attempts to execute it. 6.2.3 Accessing the RFLAGs Register The RFLAGS register contains both application and system bits. This section describes the instructions used to read and write system bits. Descriptions of instruction effects on application flags can be found in "Flags Register" in Volume 1 and "Instruction Effects on RFLAGS" in Volume 3. POPF and PUSHF Instructions. T h e p o p a n d p u s h R F L AG S instructions are used for moving data between the rFLAGS register and the stack. They are not system-management instructions, but their behavior is mode-dependent. CLI and STI Instructions. The clear interrupt (CLI) and set interrupt ( S T I ) i n s t r u c t i o n s m o d i f y o n ly t h e R F L AG S . I F b i t o r RFLAGS.VIF bit. Clearing rFLAGS.IF to 0 causes the processor to ignore maskable interrupts. Setting RFLAGS.IF to 1 causes the processor to allow maskable interrupts. See "Virtual Interrupts" on page 295 for more information on the operation of these instructions when virtual-8086 mode extensions are enabled (CR4.VME=1). 6.2.4 Accessing Debug Registers The MOV DRn instructions are used to copy data between the debug registers and the general-purpose registers. These instructions are privileged and cause a general-protection exception (#GP) if non-privileged software attempts to execute
Chapter 6: System-Management Instructions
187
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
them. See "Debug Registers" on page 385 for a detailed description of the debug registers. 6.2.5 Accessing Model-Specific Registers RDMSR and WRMSR Instructions. Th e re a d / w r i t e m o d e l - s p e c i f i c register instructions (RDMSR and WRMSR) can be used by privileged software to access the 64-bit MSRs. See "ModelSpecific Registers (MSRs)" on page 71 for details about the MSRs. RDPMC Instruction. The read performance-monitoring counter instruction, RDPMC, is used to read the model-specific performance-monitor registers, PerfCTR[3:0]. RDTSC Instruction. The read time-stamp counter instruction, RDTSC, is used to read the model-specific time-stamp counter (TSC) register.
6.3
Segment Register and Descriptor Register Access
The AMD64 architecture supports the legacy instructions that load and store segment registers and descriptor registers. In some cases the instruction capabilities are expanded to support long mode.
6.3.1 Accessing Segment Registers
MOV, POP, and PUSH Instructions. The MOV and POP instructions can be used to load a selector into a segment register from a general-purpose register or memory (MOV) or from the stack (POP). Any segment register, except the CS register, can be loaded with the MOV and POP instructions. The CS register must be loaded with a far-transfer instruction. All segment register selectors can be stored in a generalpurpose register or memory using the MOV instruction or pushed onto the stack using the PUSH instruction. When a selector is loaded into a segment register, the processor automatically loads the corresponding descriptor-table entry into the hidden portion of the selector register. The hidden portion contains the base address, limit, and segment attributes. Segment-load and segment-store instructions work normally in 64-bit mode. The appropriate entry is read from the system descriptor table (GDT or LDT) and is loaded into the hidden portion of the segment descriptor register. However, the
188
Chapter 6: System-Management Instructions
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
contents of data-segment and stack-segment descriptor registers are ignored, except in the case of the FS and GS segment-register base fields--see "FS and GS Registers in 64Bit Mode" on page 88 for more information. The ability to use segment-load instructions allows a 64-bit o p e ra t i n g s y s t e m t o s e t u p s e g m e n t re g i s t e r s f o r a c o m p a t i b i l i t y - m o d e a p p l i c a t i o n b e fo re sw i t ch i n g t o compatibility mode. 6.3.2 Accessing Descriptor-Table Registers LGDT and LIDT Instructions. The load GDTR (LGDT) and load IDTR (LIDT) instructions load a pseudo-descriptor from memory into the GDTR or IDTR registers, respectively. LLDT and LTR Instructions. The load LDTR (LLDT) and load TR (LTR) instructions load a system-segment descriptor from the GDT into the LDTR and TR segment-descriptor registers (hidden portion), respectively. SGDT and SIDT Instructions. The store GDTR (SGDT) and store IDTR (SIDT) instructions reverse the operation of the LGDT and LIDT instructions. SGDT and SIDT store a pseudo-descriptor from the GDTR or IDTR register into memory. SLDT and STR Instructions. In all modes, the store LDTR (SLDT) and store TR (STR) instructions store the LDT or task selector from the visible portion of the LDTR or TR register into a generalpurpose register or memory, respectively. The hidden portion of the LDTR or TR register is not stored.
6.4
Protection Checking
Several instructions are provided to allow software to determine the outcome of a protection check before performing a memory access that could result in a protection violation. By performing the checks before a memory access, software can avoid violations that result in a general-protection exception (#GP).
6.4.1 Checking Access Rights
LAR Instruction. The load access-rights (LAR) instruction can be used to determine if access to a segment is allowed, based on privilege checks and type checks. The LAR instruction uses a segment-selector in the source operand to reference a descriptor in the GDT or LDT. LAR performs a set of accessrights checks and, if successful, loads the segment-descriptor 189
Chapter 6: System-Management Instructions
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
access rights into the destination register. Software can further examine the access-rights bits to determine if access into the segment is allowed. 6.4.2 Checking Segment Limits LSL Instruction. The load segment-limit (LSL) instruction uses a segment-selector in the source operand to reference a d e s c r i p t o r i n t h e G D T o r L D T. L S L p e r fo r m s a s e t o f preliminary access-rights checks and, if successful, loads the segment-descriptor limit field into the destination register. Software can use the limit value in comparisons with pointer offsets to prevent segment limit violations. VERR and VERW Instructions. The verify read-rights (VERR) and verify write-rights (VERW) can be used to determine if a target code or data segment (not a system segment) can be read or written from the current privilege level (CPL). The source operand for these instructions is a pointer to the segment selector to be tested. If the tested segment (code or data) is readable from the current CPL, the VERR instruction sets RFLAGS.ZF to 1; otherwise, it is cleared to zero. Likewise, if the tested data segment is writable, the VERW instruction sets the RFLAGS.ZF to 1. A code segment cannot be tested for writability. ARPL Instruction. The adjust RPL-field (ARPL) instruction can be used by system software to prevent access into privileged-data segments by lower-privileged software. This can happen if an application passes a selector to system software and the selector RPL is less than (has greater privilege than) the calling-application CPL. To prevent this surrogate access, system software executes ARPL with the following operands: The destination operand is the data-segment selector passed to system software by the application. The source operand is the application code-segment selector (available on the system-software stack as a result of the CALL into system software by the application). ARPL is not supported in 64-bit mode.
6.4.3 Checking Read/Write Rights
6.4.4 Adjusting Access Rights
6.5
Processor Halt
The processor halt instruction (HLT) halts instruction execution, leaving the processor in the halt state. No registers or machine state are modified as a result of executing the HLT instruction.
190
Chapter 6: System-Management Instructions
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The processor remains in the halt state until one of the following occurs: A non-maskable interrupt (NMI). An enabled, maskable interrupt (INTR). Processor reset (RESET). Processor initialization (INIT). System-management interrupt (SMI).
6.6
Cache and TLB Management
Cache-management instructions are used by system software to maintain coherency within the memory hierarchy. Memory coherency and caches are discussed in Chapter 7, "Memory System." Similarly, TLB-management instructions are used to maintain coherency between page translations cached in the TLB and the translation tables maintained by system software in memory. See "Translation-Lookaside Buffer (TLB)" on page 172 for more information.
6.6.1 Cache Management
WBINVD Instruction. The writeback and invalidate (WBINVD) instruction is used to write all modified cache lines to memory so that memory contains the most recent copy of data. After the writes are complete, the instruction invalidates all cache lines. This instruction operates on all caches in the memory hierarchy, including caches that are external to the processor. INVD Instruction. The invalidate (INVD) instruction is used to invalidate all cache lines in all caches in the memory hierarchy. Unlike the WBINVD instruction, no modified cache lines are written to memory. The INVD instruction should only be used in situations where memory coherency is not required.
6.6.2 TLB Invalidation
INVLPG Instruction. T h e i n v a l i d a t e T L B e n t r y ( I N V L P G ) instruction can be used to invalidate specific entries within the TLB. The source operand is a virtual-memory address that specifies the TLB entry to be invalidated. Invalidating a TLB entry does not remove the associated page-table entry from the data cache. See "Translation-Lookaside Buffer (TLB)" on page 172 for more information.
Chapter 6: System-Management Instructions
191
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
192
Chapter 6: System-Management Instructions
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
7
Memory System
This chapter describes: Cache coherency mechanisms Cache control mechanisms Memory typing Memory mapped I/O Memory ordering rules Serializing instructions Figure 7-1 on page 194 shows a conceptual picture of a processor and memory system, and how data and instructions flow between the various components. This diagram is not i n t e n d e d t o re p re s e n t a s p e c i f i c m i c ro a rch i t e c t u ra l implementation but instead is used to illustrate the major memory-system components covered by this chapter.
Chapter 7: Memory System
193
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Main Memory
System Bus Interface
L2 Cache
L1 Instruction Cache
L1 Data Cache
Write-Combining Buffers
Write Buffers
Load/Store Unit
Execution Units
Processor Chip
513-211.eps
Figure 7-1.
Processor and Memory System The memory-system components described in this chapter are shown as unshaded boxes in Figure 7-1. Those items are summarized in the following paragraphs. Main memory is external to the processor chip and is the memory-hierarchy level farthest from the processor execution units. Caches are the memory-hierarchy levels closest to the processor execution units. They are much smaller and much faster than main memory, and can be either internal or external to the
194
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
processor chip. Caches contain copies of the most frequently used instructions and data. By allowing fast access to frequently used data, software can run much faster than if it had to access that data from main memory. Figure 7-1 shows three caches, all internal to the processor: L1 Data Cache--The L1 (level-1) data cache holds the data most recently read or written by the software running on the processor. L1 Instruction Cache--The L1 instruction cache is similar to the L1 data cache except that it holds only the instructions executed most frequently. In some processor implementations, the L1 instruction cache can be combined with the L1 data cache to form a unified L1 cache. L2 Cache--The L2 (level-2) cache is usually several times larger than the L1 caches, but it is also slower. It is common for L2 caches to be implemented as a unified cache containing both instructions and data. Recently used instructions and data that do not fit within the L1 caches can reside in the L2 cache. The L2 cache can be exclusive, meaning it does not cache information contained in the L1 cache. Conversely, inclusive L2 caches contain a copy of the L1-cached information. Memory-read operations from cacheable memory first check the cache to see if the requested information is available. A read hit occurs if the information is available in the cache, and a read miss occurs if the information is not available. Likewise, a write hit occurs if the memory write can be stored in the cache, and a write miss occurs if it cannot be stored in the cache. Caches are divided into fixed-size blocks called cache lines. The cache allocates lines to correspond to regions in memory of the same size as the cache line, aligned on an address boundary equal to the cache-line size. For example, in a cache with 32byte lines, the cache lines are aligned on 32-byte boundaries and byte addresses 0007h and 001Eh are both located in the same cache line. The size of a cache line is implementation dependent. Most implementations have either 32-byte or 64byte cache lines. The process of loading data into a cache is a cache-line fill. Even if only a single byte is requested, all bytes in a cache line are loaded from memory. Typically, a cache-line fill must remove (evict) an existing cache line to make room for the new line Chapter 7: Memory System 195
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
loaded from memory. This process is called cache-line replacement. If the existing cache line was modified before the replacement, the processor performs a cache-line writeback to main memory when it performs the cache-line fill. Cache-line writebacks help maintain coherency (consistency) between the caches and main memory. Internally, the processor can also maintain cache coherency by internally probing (checking) the other caches and write buffers for a more recent version of the requested data. External devices can also check processor caches for more recent versions of data by externally probing the processor. Throughout this document, the term probe is used to refer to external probes, while internal probes are always qualified with the word internal. Write buffers temporarily hold data writes when main memory or the caches are busy with other memory accesses. The existence of write buffers is implementation dependent. Implementations of the architecture can use write-combining buffers if the order and size of non-cacheable writes to main memory is not important to the operation of software. These buffers can combine multiple, individual writes to main memory and transfer the data in fewer bus transactions.
7.1
Memory-Access Ordering
The flexibility in which memory accesses can be ordered is closely related to the flexibility in which a processor implementation can execute and retire instructions. Instruction execution creates results and status and determines whether or not the instruction causes an exception. Instruction retirement commits the results of instruction execution, in program order, to software-visible resources such as memory, caches, writecombining buffers, and registers, or it causes an exception to occur if instruction execution created one. Implementations of the AMD64 architecture retire instructions in program order, but implementations can execute instructions in any order. Implementations can also speculatively execute instructions--executing instructions before knowing they are needed. Internally, implementations manage data reads and writes so that instructions complete in order. However, because implementations can execute instructions out of order and speculatively, the sequence of memory accesses can also be out
196
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
of program order (weakly ordered). Processor implementations adhere to the following rules governing memory accesses, which can be further restricted depending on the memory type being accessed: 7.1.1 Read Ordering Generally, reads do not affect program order because they do not affect the state of software-visible resources. However, some system devices might be sensitive to reads. In such a situation software can map a read-sensitive device to a memory type that enforces strong read-ordering, or use read/write barrier instructions to force strong read-ordering. For cacheable memory types, the following rules govern read ordering: Out-of-order reads are allowed. Out-of-order reads can occur as a result of out-of-order instruction execution or speculative execution. The processor can read memory outof-order to allow out-of-order execution to proceed. Speculative reads are allowed. A speculative read occurs when the processor begins executing a memory-read instruction before it knows the instruction will actually complete. For example, the processor can predict a branch will occur and begin executing instructions following the predicted branch before it knows whether the prediction is valid. When one of the speculative instructions reads data from memory, the read itself is speculative. Reads can be reordered ahead of writes. Reads are generally given a higher priority by the processor than writes because instruction execution stalls if the read data required by an instruction is not immediately available. Allowing reads ahead of writes usually maximizes software performance. A read cannot be reordered ahead of a prior write if the read is from the same location as the prior write. In this case, the read instruction stalls until the write instruction completes execution. The read instruction requires the result of the write instruction for proper software operation. For cacheable memory types, the write data can be forwarded to the read instruction before it is actually written to memory. 7.1.2 Write Ordering Writes affect program order because they affect the state of software-visible resources. The following rules govern write ordering:
Chapter 7: Memory System
197
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Generally, out-of-order writes are not allowed. Write instructions executed out of order cannot commit (write) their result to memory until all previous instructions have completed in program order. The processor can, however, hold the result of an out-of-order write instruction in a private buffer (not visible to software) until that result can be committed to memory. It is possible for writes to write-combining memory types to appear to complete out of order, relative to writes into other memory types. See "Memory Types" on page 202 and "Write Combining" on page 207 for additional information. Speculative writes are not allowed. As with out-of-order writes, speculative write instructions cannot commit their result to memory until all previous instructions have completed in program order. Processors can hold the result in a private buffer (not visible to software) until the result can be committed. Write buffering is allowed. When a write instruction completes and commits its result, that result can be buffered before actually writing the result into a memory location in program order. Although the write buffer itself is not directly accessible by software, the results in the buffer are accessible during memory accesses to the locations that are buffered. For cacheable memory types, the write buffer can be read out-of-order and speculatively read, just like memory. Write combining is allowed. In some situations software can relax the write-ordering rules and allow several writes to be combined into fewer writes to memory. When writecombining is used, it is possible for writes to other memory types to proceed ahead of (out-of-order) memory-combining writes, unless the writes are to the same address. Writecombining should be used only when the order of writes does not affect program order (for example, writes to a graphics frame buffer). 7.1.3 Read/Write Barriers When the order of memory accesses must be strictly enforced, software can use read/write barrier instructions to force reads and writes to proceed in program order. Read/write barrier instructions force all prior reads or writes to complete before subsequent reads or writes are executed. The LFENCE, SFENCE, and MFENCE instructions are provided as dedicated read, write, and read/write barrier instructions (respectively). Chapter 7: Memory System
198
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Serializing instructions, I/O instructions, and locked instructions can also be used as read/write barriers. Table 7-1 on page 205 shows the memory-access ordering possible for each memory type supported by the AMD64 architecture.
7.2
Memory Coherency and Protocol
Implementations that support caching support a cachecoherency protocol for maintaining coherency between main memory and the caches. The cache-coherency protocol is also used to maintain coherency between all processors in a multiprocessor system. The cache-coherency protocol supported by the AMD64 architecture is the MOESI (modified, owned, exclusive, shared, invalid) protocol. The states of the MOESI protocol are: Invalid--A cache line in the invalid state does not hold a valid copy of the data. Valid copies of the data can be either in main memory or another processor cache. Exclusive--A cache line in the exclusive state holds the most recent, correct copy of the data. The copy in main memory is also the most recent, correct copy of the data. No other processor holds a copy of the data. Shared--A cache line in the shared state holds the most recent, correct copy of the data. Other processors in the system may hold copies of the data in the shared state, as well. The copy in main memory is also the most recent, correct copy of the data (since no other processor holds it in owned state). Modified--A cache line in the modified state holds the most recent, correct copy of the data. The copy in main memory is stale (incorrect), and no other processor holds a copy. Owned--A cache line in the owned state holds the most recent, correct copy of the data. The owned state is similar to the shared state in that other processors can hold a copy of the most recent, correct data. Unlike the shared state, however, the copy in main memory can be stale (incorrect). Only one processor can hold the data in the owned state--all other processors must hold the data in the shared state. Figure 7-2 on page 200 shows the general MOESI state transitions possible with various types of memory accesses. This
Chapter 7: Memory System
199
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
is a logical software view, not a hardware view, of how cache-line state transitions. Instruction-execution activity and externalbus transactions can both be used to modify the cache MOESI state in multiprocessing or multi-mastering systems.
Reset INVD, WBINVD
Read Hit
Probe Write Hit
Invalid
Read Miss, Exclusive
Exclusive
iss, Sha
red
Probe W rite Hit
Hit Write
Read M
Pro Wr ite Mi ss (W
be
Wr ite em
Hit
P
e rob
R
ead
Hit
Bm
Pro be te Wri Hit
ory
)
Shared
Probe Read
Hit
Modified
Owned
Read Hit Probe Read Hit
Write Hit
Read Hit Write Hit
Write Hit Read Hit Probe Read Hit
513-212.eps
Figure 7-2.
MOESI State Transitions To maintain memory coherency, external bus masters (typically other processors with their own internal caches) need to acquire the most recent copy of data before caching it
200
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
internally. That copy can be in main memory or in the internal caches of other bus-mastering devices. When an external master has a cache read-miss or write-miss, it probes the other mastering devices to determine whether the most recent copy of data is held in any of their caches. If one of the other mastering devices holds the most recent copy, it provides it to the requesting device. Otherwise, the most recent copy is provided by main memory. There are two general types of bus-master probes: Read probes indicate the external master is requesting the data for read purposes. Write probes indicate the external master is requesting the data for the purpose of modifying it. Referring back to Figure 7-2, the state transitions involving probes are initiated by other processors and external bus masters into the processor. Some read probes are initiated by devices that intend to cache the data. Others, such as those initiated by I/O devices, do not intend to cache the data. Some processor implementations do not change the data MOESI state if the read probe is initiated by a device that does not intend to cache the data. State transitions involving read misses and write misses can cause the processor to generate probes into external bus masters and to read main memory. Read hits do not cause a MOESI-state change. Write hits generally cause a MOESI-state change into the modified state. If the cache line is already in the modified state, a write hit does not change its state. The specific operation of external-bus signals and transactions a n d h ow t h e y i n f l u e n c e a c a ch e M O E S I s t a t e a re implementation dependent. For example, an implementation could convert a write miss to a WB memory type into two separate MOESI-state changes. The first would be a read-miss placing the cache line in the exclusive state. This would be followed by a write hit into the exclusive cache line, changing the cache-line state to modified. 7.2.1 Special Coherency Considerations Chapter 7: Memory System In some cases, data can be modified in a manner that is impossible for the memory-coherency protocol to handle due to the effects of instruction prefetching. In such situations 201
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
software must use serializing instructions and/or cacheinvalidation instructions to guarantee subsequent data accesses are coherent. An example of this type of a situation is a page-table update followed by accesses to the physical pages referenced by the updated page tables. The following sequence of events shows what can happen when software changes the translation of virtual-page A from physical-page M to physical-page N: 1. Software invalidates the TLB entry. The tables that translate virtual-page A to physical-page M are now held only in main memory. They are not cached by the TLB. 2. Software changes the page-table entry for virtual-page A in main memory to point to physical-page N rather than physical-page M. 3. Software accesses data in virtual-page A. During Step 3, software expects the processor to access the data from physical-page N. However, it is possible for the processor to prefetch the data from physical-page M before the page table for virtual-page A is updated in Step 2. This is because the physical-memory references for the page tables are different than the physical-memory references for the data. Because the physical-memory references are different, the processor does not recognize them as requiring coherency checking and believes it is safe to prefetch the data from virtual-page A, which is translated into a read from physical page M. To prevent this problem, software should insert a read/write barrier instruction (typically a serializing instruction) immediately after the page-table update to force subsequent instructions to access data using the correct virtual-page-tophysical-page translation. The serializing instructions INVLPG and MOV CR3, used to flush the TLB, can be used for this purpose.
7.3
Memory Types
The AMD64 architecture defines the following memory types: Uncacheable (UC)--Reads from, and writes to, UC memory are not cacheable. Reads from UC memory cannot be speculative. Write-combining to UC memory is not allowed.
202
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Reads from UC memory cause the write buffers to be written to memory and invalidated. The UC memory type is useful for memory-mapped I/O devices where strict ordering of reads and writes is important. Cache Disable (CD)--The CD memory type is a form of uncacheable memory type that occurs when caches are disabled (CR0.CD=1 and CR0.NW=0). With CD memory, it is possible for the address to be cached due to an earlier cacheable access, or due to two virtual-addresses aliasing to a single physical address. For the L1 data cache and the L2 cache, reads from, and writes to, CD memory that hit the cache cause the cache line to be invalidated before accessing main memory. If the cache line is in the modified state, the line is written to main memory and then invalidated. For the L1 instruction cache, reads from CD memory that hit the cache read the cached instructions rather than access main memory. Reads that miss the cache access main memory and do not cause cache-line replacement. Write-Combining (WC)--Reads from, and writes to, WC memory are not cacheable. Reads from WC memory can be speculative. Writes to this memory type can be combined internally by the processor and written to memory as a single write operation to reduce memory accesses. For example, four word writes to consecutive addresses can be combined by the processor into a single quadword write, resulting in one memory access instead of four. The WC memory type is useful for graphics-display memory buffers where the order of writes is not important. Write-Protect (WP)--Reads from WP memory are cacheable and allocate cache lines on a read miss. Reads from WP memory can be speculative. Writes to WP memory that hit in the cache do not update the cache. Instead, all writes update memory (write to memory), and writes that hit in the cache invalidate the cache line. Write buffering of WP memory is allowed. The WP memory type is useful for shadowed-ROM memory where updates must be immediately visible to all devices that read the shadow locations. Chapter 7: Memory System 203
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Writethrough (WT)--Reads from WT memory are cacheable and allocate cache lines on a read miss. Reads from WT memory can be speculative. All writes to WT memory update main memory, and writes that hit in the cache update the cache line (cache lines remain in the same state after a write that hits a cache line). Writes that miss the cache do not allocate a cache line. Write buffering of WT memory is allowed. Writeback (WB)--Reads from WB memory are cacheable and allocate cache lines on a read miss. Cache lines can be allocated in the shared, exclusive, or modified states. Reads from WB memory can be speculative. All writes that hit in the cache update the cache line and place the cache line in the modified state. Writes that miss the cache allocate a new cache line and place the cache line in the modified state. Writes to main memory only take place during writeback operations. Write buffering of WB memory is allowed. The WB memory type provides the highest-possible performance and is useful for most software and data stored in system memory (DRAM). Table 7-1 on page 205 shows the memory access ordering possible for each memory type supported by the AMD64 architecture. Table 7-2 on page 205 shows the caching policy for the same memory types.
204
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 7-1.
Memory Access by Memory Type
Memory Access Allowed Out-of-Order Memory Type UC/CD no no no no no no no WC yes yes yes yes no yes yes WP yes yes yes no no yes no WT yes yes yes no no yes yes WB yes yes yes no no yes yes
Read
Speculative Reorder Before Write Out-of-Order
Write
Speculative Buffering Combining1
Note:
1. Write-combining buffers are separate from write buffers.
Table 7-2.
Caching Policy by Memory Type
Caching Policy Read Cacheable Write Cacheable Read Allocate Write Allocate Memory Type UC no no no no yes2 CD no no no no yes1 WC no no no no yes2 WP yes no yes no yes3 WT yes yes yes no yes WB yes yes yes yes no
Write Hits Update Memory
Note:
1. For the L1 data cache and the L2 cache, if an access hits the cache, the cache line is invalidated. If the cache line is in the modified state, the line is written to main memory and then invalidated. For the L1 instruction cache, read hits access the cache rather than main memory. 2. The data is not cached, so a cache write hit cannot occur. However, memory is updated. 3. Write hits update memory and invalidate the cache line.
7.4
Buffering and Combining Memory Writes
Writes to memory (main memory and caches) can be stored internally by the processor in write buffers before actually 205
7.4.1 Write Buffering
Chapter 7: Memory System
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
writing the data into a memory location. System performance can be improved by buffering writes, as shown in the following examples: When higher-priority memory transactions, such as reads, compete for memory access with writes, writes can be delayed in favor of reads, which minimizes or eliminates an instruction-execution stall due to a memory-operand read. When the memory is busy, buffering writes while the memory is busy removes the writes from the instructionexecution pipeline, which frees instruction-execution resources. The processor manages the write buffer so that it is transparent to software. Memory accesses check the write buffer, and the processor completes writes into memory from the buffer in program order. Also, the processor completely empties the write buffer by writing the contents to memory as a result of performing any of the following operations: SFENCE Instruction--Executing a store-fence (SFENCE) instruction forces all memory writes before the SFENCE (in program order) to be written into memory before memory writes that follow the SFENCE instruction. The memoryfence (MFENCE) instruction has a similar effect, but it forces the ordering of loads in addition to stores. Serializing Instructions--Executing a serializing instruction forces the processor to retire the serializing instruction (complete both instruction execution and result writeback) before the next instruction is fetched from memory. I/O instructions--Before completing an I/O instruction, all previous reads and writes must be written to memory, and the I/O instruction must complete before completing subsequent reads or writes. Writes to I/O-address space (OUT instruction) are never buffered. Locked Instructions--Before completing a locked instruction (an instruction executed using the LOCK prefix), all previous reads and writes must be written to memory, and the locked instruction must complete before completing subsequent writes. Locked writes are never buffered, although locked reads and writes are cacheable. Interrupts and Exceptions--Interrupts and exceptions are serializing events that force the processor to write all results
206
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
from the write buffer to memory before fetching the first instruction from the interrupt or exception service routine. UC-Memory Reads--UC-memory reads are not reordered ahead of writes. Write buffers can behave similarly to write-combining buffers because multiple writes are collected internally before transferring the data to caches or main memory. See the following section for a description of write combining. 7.4.2 Write Combining Write-combining memory uses a different buffering scheme than write buffering described above. Writes to writecombining (WC) memory can be combined internally by the processor in a buffer for more efficient transfer to main memory at a later time. For example, 16 doubleword writes to consecutive memory addresses can be combined in the WC buffers and transferred to main memory as a single burst operation rather than as individual memory writes. The following instructions perform writes to WC memory: MASKMOVDQU MASKMOVQ MOVNTDQ MOVNTI MOVNTPD MOVNTPS MOVNTQ WC memory is not cacheable. A WC buffer writes its contents only to main memory. The size and number of WC buffers available is implementation dependent. The processor assigns an address range to an empty WC buffer when a WC-memory write occurs. The size and alignment of this address range is equal to the buffer size. All subsequent writes to WC memory that fall within this address range can be stored by the processor in the WC-buffer entry until an event occurs that causes the processor to write the WC buffer to main memory. After the WC buffer is written to main memory, the processor can assign a new address range on a subsequent WC-memory write. Writes to consecutive addresses in WC memory are not required for the processor to combine them. The processor combines any Chapter 7: Memory System 207
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
WC memory write that falls within the active-address range for a buffer. Multiple writes to the same address overwrite each other (in program order) until the WC buffer is written to main memory. It is possible for writes to proceed out of program order when WC memory is used. For example, a write to cacheable memory that follows a write to WC memory can be written into the cache before the WC buffer is written to main memory. For this reason, and the reasons listed in the previous paragraph, software that is sensitive to the order of memory writes should avoid using WC memory. WC buffers are written to main memory under the same conditions as the write buffers, namely when: Executing a store-fence (SFENCE) instruction. Executing a serializing instruction. Executing an I/O instruction. Executing a locked instruction (an instruction executed using the LOCK prefix). An interrupt or exception occurs. WC buffers are also written to main memory when: A subsequent non-write-combining operation has a write address that matches the WC-buffer active-address range. A write to WC memory falls outside the WC-buffer activeaddress range. The existing buffer contents are written to main memory, and a new address range is established for the latest WC write.
7.5
Memory Caches
The AMD64 architecture supports the use of internal and external caches. The size, organization, coherency mechanism, and replacement algorithm for each cache is implementation d e p e n d e n t . G e n e ra l ly, t h e ex i s t e n c e o f t h e c a ch e s i s transparent to both application and system software. In some cases, however, software can use cache-structure information to optimize memory accesses or manage memory coherency. Such software can use the extended-feature functions of the CPUID instruction to gather information on the caching subsystem supported by the processor. For more information on using
208
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
CPUID in this manner, see "Functions 8000_0005h and 8000_0006h: Cache Information" in Volume 3. 7.5.1 Cache Organization and Operation Although the detailed organization of a processor cache depends on the implementation, the general constructs are similar. L1 caches--data and instruction, or unified--and L2 caches usually are implemented as n-way set-associative caches. Figure 7-3 on page 210 shows a typical logical organization of an n-way set-associative cache. The physical implementation of the cache can be quite different.
Chapter 7: Memory System
209
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Way 0
Tag Data Other Tag
Way 1
Data Other
...
Tag
Way n-1
Data Other
Set 0 Set 1 Set 2 Set 3
Line Data 0,2 Line Data 1,2 Line Data n-1,2
...
Set m-1
Miss Miss Hit Miss
=
Hit
=
...
Hit
=
MUX n:1
Data
Cache
Hit Data Physical Address
Tag Field Index Field Offset Field
513-213.eps
Figure 7-3.
Cache Organization Example As shown in Figure 7-3, the cache is organized as an array of cache lines. Each cache line consists of three parts: a cache-data line (a fixed-size copy of a memory block), a tag, and other information. Rows of cache lines in the cache array are sets, and columns of cache lines are ways. In an n-way set-associative cache, each set is a collection of n lines. For example, in a four-
210
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
way set-associative cache, each set is a collection of four cache lines, one from each way. The cache is accessed using the physical address of the data or instruction being referenced. To access data within a cache line, the physical address is used to select the set, way, and byte from the cache. This is accomplished by dividing the physical address into the following three fields: Index--The index field selects the cache set (row) to be examined for a hit. All cache lines within the set (one from each way) are selected by the index field. Tag--The tag field is used to select a specific cache line from the cache set. The physical-address tag field is compared with each cache-line tag in the set. If a match is found, a cache hit is signalled, and the appropriate cache line is selected from the set. If a match is not found, a cache miss is signalled. Offset--The offset field points to the first byte in the cache line corresponding to the memory reference. The referenced data or instruction value is read from (or written to, in the case of memory writes) the selected cache line starting at the location selected by the offset field. In Figure 7-3 on page 210, the physical-address index field is shown selecting Set 2 from the cache. The tag entry for each cache line in the set is compared with the physical-address tag field. The tag entry for Way 1 matches the physical-address tag field, so the cache-line data for Set 2, Way 1 is selected using the n:1 multiplexor. Finally, the physical-address offset field is used to point to the first byte of the referenced data (or instruction) in the selected cache line. Cache lines can contain other information in addition to the data and tags, as shown in Figure 7-3. MOESI state and the state bits associated with the cache-replacement algorithm are typical pieces of information kept with the cache line. Instruction caches can also contain pre-decode or branchprediction information. The type of information stored with the cache line is implementation dependent. Self-Modifying Code. Software that writes into a code segment is classified as self-modifying code. To avoid cache-coherency problems due to self-modifying code, a check is made during data writes to see whether the data-memory location Chapter 7: Memory System 211
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
corresponds to a code-segment memory location. If it does, implementations of the AMD64 architecture invalidate the corresponding instruction-cache line(s) during the data-memory write. Entries in the data cache are not invalidated, and it is possible for the modified instruction to be cached by the data cache following the memory write. A subsequent fetch of the modified instruction goes to main memory to get the coherent version of the instruction. If the data cache holds the most recent copy of the instruction rather than main memory, it provides that copy. The processor determines whether a write is in a code segment by internally probing the instruction cache and prefetched instructions. If the internal probe returns a hit, the instructioncache line and prefetched instructions are invalidated. The internal probes into the instruction cache and prefetch hardware are always performed using the physical address of an instruction in order to avoid potential aliasing problems associated with using virtual (linear) addresses. 7.5.2 Cache Control Mechanisms The AMD64 architecture provides a number of mechanisms for controlling the cacheability of memory. These are described in the following sections. Cache Disable. Bit 30 of the CR0 register is the cache-disable bit, CR0.CD. Caching is enabled when CR0.CD is cleared to 0, and caching is disabled when CR0.CD is set to 1. When caching is disabled, reads and writes access main memory. Software can disable the cache while the cache still holds valid data (or instructions). If a read or write hits the L1 data cache or the L2 cache when CR0.CD=1, the processor does the following: 1. Writes the cache line back if it is in the modified or owned state. 2. Invalidates the cache line. 3. Performs a non-cacheable main-memory access to read or write the data. If an instruction fetch hits the L1 instruction cache when CR0.CD=1, the processor reads the cached instructions rather than access main memory.
212
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The processor also responds to cache probes when CR0.CD=1. Probes that hit the cache cause the processor to perform Step 1. Step 2 (cache-line invalidation) is performed only if the probe is performed on behalf of a memory write or an exclusive read. Writethrough Disable. B i t 2 9 o f t h e C R 0 re g i s t e r i s t h e n o t writethrough disable bit, CR0.NW. In early x86 processors, CR0.NW is used to control cache writethrough behavior, and the combination of CR0.NW and CR0.CD determines the cache operating mode. In early x86 processors, clearing CR0.NW to 0 enables writeback caching for main memory, effectively disabling writethrough caching for main memory. When CR0.NW=0, software can disable writeback caching for specific memory pages or regions by using other cache control mechanisms. When software sets CR0.NW to 1, writeback caching is disabled for main memory, while writethrough caching is enabled. In implementations of the AMD64 architecture, CR0.NW is not used to qualify the cache operating mode established by CR0.CD. Table 7-3 shows the effects of CR0.NW and CR0.CD on the AMD64 architecture cache-operating modes. Table 7-3.
CR0.CD 0 0 1 1
AMD64 Architecture Cache-Operating Modes
CR0.NW 0 1 0 1 Cache Operating Mode Cache enabled with a writeback-caching policy. Invalid setting--causes a general-protection exception (#GP). Cache disabled. See "Cache Disable" on page 212.
Page-Level Cache Disable. Bit 4 of all paging data-structure entries controls page-level cache disable (PCD). When a data-structureentry PCD bit is cleared to 0, the page table or physical page pointed to by that entry is cacheable, as determined by the CR0.CD bit. When the PCD bit is set to 1, the page table or physical page is not cacheable. The PCD bit in the paging datastructure base-register (bit 4 in CR3) controls the cacheability of the highest-level page table in the page-translation hierarchy.
Chapter 7: Memory System
213
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Page-Level Writethrough Enable. Bit 3 of all paging data-structure entries is the page-level writethrough enable control (PWT). When a data-structure-entry PWT bit is cleared to 0, the page table or physical page pointed to by that entry has a writeback caching policy. When the PWT bit is set to 1, the page table or physical page has a writethrough caching policy. The PWT bit in the paging data-structure base-register (bit 3 in CR3) controls the caching policy of the highest-level page table in the pagetranslation hierarchy. The corresponding PCD bit must be cleared to 0 (page caching enabled) for the PWT bit to have an effect. Memory Typing. Two mechanisms are provided for software to control access to and cacheability of specific memory regions: The memory-type range registers (MTRRs) control cacheability based on physical addresses. See "MTRRs" on page 219 for more information on the use of MTRRs. The page-attribute table (PAT) mechanism controls cacheability based on virtual addresses. PAT extends the capabilities provided by the PCD and PWT page-level cache controls. See "Page-Attribute Table Mechanism" on page 230 for more information on the use of the PAT mechanism. System software can combine the use of both the MTRRs and PAT m e c h a n i s m s t o m a x i m i z e c o n t r o l ove r m e m o ry cacheability. If the MTRRs are disabled in implementations that support the M T R R m e ch a n i s m , t h e d e fa u l t m e m o ry t y p e i s s e t t o uncacheable (UC). Memory accesses are not cached even if the caches are enabled by clearing CR0.CD to 0. Cacheable memory types must be established using the MTRRs in order for memory accesses to be cached. Cache Control Precedence. The cache-control mechanisms are used to define the memory type and cacheability of main memory and regions of main memory. Taken together, the most restrictive memory type takes precedence in defining the caching policy of memory. The order of precedence is: 1. Uncacheable (UC) 2. Write-combining (WC)
214
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
3. Write-protected (WP) 4. Writethrough (WT) 5. Writeback (WB) For example, assume a large memory region is designated a writethrough type using the MTRRs. Individual pages within t h a t re g i o n c a n h ave c a ch i n g d i s ab l e d by s e t t i n g t h e appropriate page-table PCD bits. However, no pages within that region can have a writeback caching policy, regardless of the page-table PWT values. 7.5.3 Cache and Memory Management Instructions Data Prefetch. The prefetch instructions are used by software as a hint to the processor that the referenced data is likely to be used in the near future. The processor can preload the cache line containing the data in anticipation of its use. PREFETCH provides a hint that the data is to be read. PREFETCHW provides a hint that the data is to be written. The processor can mark the line as modified if it is preloaded using PREFETCHW. Memory Ordering. Instructions are provided for software to enforce memory ordering (serialization) in weakly-ordered memory types. These instructions are: SFENCE (store fence)--forces all memory writes (stores) preceding the SFENCE (in program order) to be written into memory before memory writes following the SFENCE. LFENCE (load fence)--forces all memory reads (loads) preceding the LFENCE (in program order) to be read from memory before memory reads following the LFENCE. MFENCE (memory fence)--forces all memory accesses (reads and writes) preceding the MFENCE (in program order) to be written into or read from memory before memory accesses following the MFENCE. Cache Line Flush. The CLFLUSH instruction (writeback, if modified, and invalidate) takes the byte memory-address operand (a linear address), and checks to see if the address is cached. If the address is cached, the entire cache line containing the address is invalidated. If any portion of the cache line is dirty (in the modified or owned state), the entire line is written to main memory before it is invalidated. CLFLUSH affects all caches in the memory hierarchy--internal and external to the processor. The checking and invalidation
Chapter 7: Memory System
215
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
process continues until the address has been invalidated in all caches. In most cases, the underlying memory type assigned to the address has no effect on the behavior of this instruction. However, when the underlying memory type for the address is UC or WC (as defined by the MTRRs), the processor does not proceed with checking all caches to see if the address is cached. In both cases, the address is uncacheable, and invalidation is unnecessary. Write-combining buffers are written back to memory if the corresponding physical address falls within the buffer active-address range. Cache Writeback and Invalidate. Unlike the CLFLUSH instruction, the WBINVD instruction operates on the entire cache, rather than a single cache line. The WBINVD instruction first writes back all cache lines that are dirty (in the modified or owned state) to main memory. After writeback is complete, the instruction invalidates all cache lines. The checking and invalidation process continues until all internal caches are invalidated. A special bus cycle is transmitted to higher-level external caches directing them to perform a writeback-andinvalidate operation. Cache Invalidate. The INVD instruction is used to invalidate all cache lines. Unlike the WBINVD instruction, dirty cache lines are not written to main memory. The process continues until all internal caches have been invalidated. A special bus cycle is transmitted to higher-level external caches directing them to perform an invalidation. The INVD instruction should only be used in situations where memory coherency is not required. 7.5.4 Serializing Instructions Serializing instructions force the processor to retire the serializing instruction and all previous instructions before the next instruction is fetched. A serializing instruction is retired when the following operations are complete: The instruction has executed. All registers modified by the instruction are updated. All memory updates performed by the instruction are complete. All data held in the write buffers have been written to memory. 216 Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Serializing instructions can be used as a barrier between m e m o ry a c c e s s e s t o forc e s t ro n g o rd e r in g o f m e m o ry operations. Care should be exercised in using serializing instructions because they modify processor state and affect p rog ra m f l ow. Th e i n s t r u c t i o n s a l s o fo rc e e x e c u t i o n serialization, which can significantly degrade performance. When strongly-ordered memory accesses are required, but execution serialization is not, it is recommended that software use the memory-ordering instructions described on page 215. The following are serializing instructions: Non-Privileged Instructions - CPUID - IRET - RSM Privileged Instructions - MOV CRn - MOV DRn - LGDT, LIDT, LLDT, LTR - SWAPGS - WRMSR - WBINVD, INVD - INVLPG
7.6
Memory-Type Range Registers
The AMD64 architecture supports three mechanisms for software access-control and cacheability-control over memory regions. These mechanisms can be used in place of similar capabilities provided by external chipsets used with early x86 processors. This section describes a control mechanism that uses a set of programmable model-specific registers (MSRs) called the memory-type-range registers (MTRRs). The MTRR mechanism provides system software with the ability to manage hardwaredevice memory mapping. System software can characterize physical-memory regions by type (e.g., ROM, flash, memorymapped I/O) and assign hardware devices to the appropriate physical-memory type.
Chapter 7: Memory System
217
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Another control mechanism is implemented as an extension to the page-translation capability and is called the page attribute t a b l e ( PAT ) . I t i s d e s c r i b e d i n " Pa g e - A t t r i b u t e Tab l e Mechanism" on page 230. Like the MTRRs, PAT provides system software with the ability to manage hardware-device memory mapping. With PAT, however, system software can characterize physical pages and assign virtually-mapped devices to those physical pages using the page-translation mechanism. PAT may be used in conjunction with the MTTR mechanism to maximize flexibility in memory control. Finally, control mechanisms are provided for managing memory-mapped I/O. These mechanisms employ extensions to the MTRRs and a separate feature called the top-of-memory registers. The MTRR extensions include additional MTRR typefield encodings for fixed-range MTRRs and variable-range I/O range registers (IORRs). These mechanisms are described in "Memory-Mapped I/O" on page 234. 7.6.1 MTRR Type Fields The MTRR mechanism provides a means for characterizing physical-address ranges with a memory type (see "Memory Types" on page 202). The MTRRs contain a type field used to specify the memory type in effect for a given physical-address range. There are two variants of the memory type-field encodings: standard and extended. Both the standard and extended encodings use type-field bits 2-0 to specify the memory type. For the standard encodings, bits 7-3 are reserved and must be zero. For the extended encodings, bits 7-5 are reserved, but bits 4-3 are defined as the RdMem and WrMem bits. "Extended Fixed-Range MTRR Type-Field Encodings" on page 235 describes the function of these extended bits and how software enables them. Only the fixed-range MTRRs support the extended type-field encodings. Variable-range MTRRs use the standard encodings. Table 7-4 on page 219 shows the memory types supported by the MTRR mechanism and their encoding in the MTRR type fields referenced throughout this section. Unless the extended typefield encodings are explicitly enabled, the processor uses the type values shown in Table 7-4.
218
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 7-4.
Type Value 00h 01h 04h
MTRR Type Field Encodings
Type Name UC--Uncacheable WC--Write-Combining WT--Writethrough Type Description All accesses are uncacheable. Write combining is not allowed. Speculative accesses are not allowed All accesses are uncacheable. Write combining is allowed. Speculative reads are allowed Reads allocate cache lines on a cache miss. Cache lines are not allocated on a write miss. Write hits update the cache and main memory. Reads allocate cache lines on a cache miss. All writes update main memory. Cache lines are not allocated on a write miss. Write hits invalidate the cache line and update main memory. Reads allocate cache lines on a cache miss, and can allocate to either the shared, exclusive, or modified state. Writes allocate to the modified state on a cache miss.
05h
WP--Write-Protect
06h
WB--Writeback
If the MTRRs are disabled in implementations that support the M T R R m e ch a n i s m , t h e d e fa u l t m e m o ry t y p e i s s e t t o uncacheable (UC). Memory accesses are not cached even if the caches are enabled by clearing CR0.CD to 0. Cacheable memory types must be established using the MTRRs to enable memory accesses to be cached. 7.6.2 MTRRs Both fixed-size and variable-size address ranges are supported by the MTRR mechanism. The fixed-size ranges are restricted to the lower 1 Mbyte of physical-address space, while the variable-size ranges can be located anywhere in the physicaladdress space. Figure 7-4 on page 220 shows an example mapping of physical memory using the fixed-size and variable-size MTRRs. The areas shaded gray are not mapped by the MTRRs. Unmapped areas are set to the software selected default memory type.
Chapter 7: Memory System
219
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Physical Memory
0_FFFF_FFFF_FFFFh
Default (Unmapped) Ranges
Up to 8 Variable Ranges
64 4-Kbyte Ranges 16 16-Kbyte Ranges 8 64-Kbyte Ranges
256 Kbytes 256 Kbytes 512 Kbytes
10_0000h 0F_FFFFh
00_0000h
513-214.eps
Figure 7-4.
MTRR Mapping of Physical Memory
MTRRs are 64-bit model-specific registers (MSRs). They are read using the RDMSR instruction and written using the WRMSR instruction. See "Memory-Typing MSRs" on page 440 for a listing of the MTRR MSR numbers. The following sections describe the types of MTRRs and their function. Fixed-Range MTRRs. T h e f i x e d - ra n g e M T R R s a re u s e d t o characterize the first 1 Mbyte of physical memory. Each fixedrange MTRR contains eight type fields for characterizing a total of eight memory ranges. Fixed-range MTRRs support extended type-field encodings as described in "Extended Fixed-Range MTRR Type-Field Encodings" on page 235. The extended type 220 Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
field allows a fixed-range MTRR to be used as a fixed-range IORR. Figure 7-5 shows the format of a fixed-range MTRR.
63 Type 56 55 Type 48 47 Type 40 39 Type 32
31 Type
24
23 Type
16
15 Type
8
7 Type
0
Figure 7-5.
Fixed-Range MTRR For the purposes of memory characterization, the first 1 Mbyte of physical memory is segmented into a total of 88 nonoverlapping memory ranges, as follows: The 512 Kbytes of memory spanning addresses 00_0000h to 07_FFFFh are segmented into eight 64-Kbyte ranges. A single MTRR is used to characterize this address space. The 256 Kbytes of memory spanning addresses 08_0000h to 0B_FFFFh are segmented into 16 16-Kbyte ranges. Two MTRRs are used to characterize this address space. The 256 Kbytes of memory spanning addresses 0C_0000h to 0F_FFFFh are segmented into 64 4-Kbyte ranges. Eight MTRRs are used to characterize this address space. Table 7-5 on page 222 shows the address ranges corresponding to the type fields within each fixed-range MTRR. The grayshaded heading boxes represent the bit ranges for each type field in a fixed-range MTTR. See Table 7-4 on page 219 for the type-field encodings.
Chapter 7: Memory System
221
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 7-5.
Fixed-Range MTRR Address Ranges
Physical Address Range (in hexadecimal) Register Name MTRRfix64K_00000 MTRRfix16K_80000 MTRRfix16K_A0000 MTRRfix4K_C0000 MTRRfix4K_C8000 MTRRfix4K_D0000 MTRRfix4K_D8000 MTRRfix4K_E0000 MTRRfix4K_E8000 MTRRfix4K_F0000 MTRRfix4K_F8000
63-56 70000- 7FFFF 9C000- 9FFFF BC000- BFFFF C7000- C7FFF CF000- CFFFF D7000- D7FFF DF000- DFFFF E7000- E7FFF EF000- EFFFF F7000- F7FFF FF000- FFFFF
55-48 60000- 6FFFF 98000- 9BFFF B8000- BBFFF C6000- C6FFF CE000- CEFFF D6000- D6FFF DE000- DEFFF E6000- E6FFF EE000- EEFFF F6000- F6FFF FE000- FEFFF
47-40 50000- 5FFFF 94000- 97FFF B4000- B7FFF C5000- C5FFF CD000- CDFFF D5000- D5FFF DD000- DDFFF E5000- E5FFF ED000- EDFFF F5000- F5FFF FD000- FDFFF
39-32 40000- 4FFFF 90000- 93FFF B0000- B3FFF C4000- C4FFF CC000- CCFFF D4000- D4FFF DC000- DCFFF E4000- E4FFF EC000- ECFFF F4000- F4FFF FC000- FCFFF
31-24 30000- 3FFFF 8C000- 8FFFF AC000- AFFFF C3000- C3FFF CB000- CBFFF D3000- D3FFF DB000- DBFFF E3000- E3FFF EB000- EBFFF F3000- F3FFF FB000- FBFFF
23-16 20000- 2FFFF 88000- 8BFFF A8000- ABFFF C2000- C2FFF CA000- CAFFF D2000- D2FFF DA000- DAFFF E2000- E2FFF EA000- EAFFF F2000- F2FFF FA000- FAFFF
15-8 10000- 1FFFF 84000- 87FFF A4000- A7FFF C1000- C1FFF C9000- C9FFF D1000- D1FFF D9000- D9FFF E1000- E1FFF E9000- E9FFF F1000-F 1FFF F9000- F9FFF
7-0 00000- 0FFFF 80000- 83FFF A0000- A3FFF C0000- C0FFF C8000- C8FFF D0000- D0FFF D8000- D8FFF E0000- E0FFF E8000- E8FFF F0000- F0FFF F8000- F8FFF
Variable-Range MTRRs. The variable-range MTRRs can be used to characterize any address range within the physical-memory space, including all of physical memory. Up to eight address ranges of varying sizes can be characterized using the MTRR. Two variable-range MTRRs are used to characterize each address range: MTRRphysBasen and MTRRphysMaskn (n is the address-range number from 0 to 7). For example, address-range 3 i s c h a ra c t e r i z e d u s i n g t h e M T R R p hy s B a s e 3 a n d MTRRphysMask3 register pair.
222
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Figure 7-6 shows the format of the MTRRphysBasen register a n d Fi g u re 7 -7 o n p a g e 2 2 4 s h ow s t h e fo r m a t o f t h e MTRRphysMaskn register. The fields within the register pair are read/write. MTRRphysBasen Registers. The fields in these variable-range MTRRs, shown in Figure 7-6, are: Type--Bits 7-0. The memory type used to characterize the memory range. See Table 7-4 on page 219 for the type-field encodings. Variable-range MTRRs do not support the extended type-field encodings. Range Physical Base-Address (PhysBase)--Bits 51-12. The memory-range base-address in physical-address space. PhysBase is aligned on a 4-Kbyte (or greater) address in the 52-bit physical-address space supported by the AMD64 architecture. PhysBase represents the most-significant 40address bits of the physical address. Physical-address bits 11-0 are assumed to be 0.
63 Reserved, MBZ 52 51 PhysBase (This is an architectural limit. A given implementation may support fewer bits.) 32
31 PhysBase
12
11
8
7 Type
0
Reserved, MBZ
Bits 63-52 51-12 11-8 7-0
Mnemonic Reserved PhysBase Reserved Type
Description Reserved, Must be Zero Range Physical Base Address Reserved, Must be Zero Default Memory Type
R/W R/W R/W
Figure 7-6.
MTRRphysBasen Register MTRRphysMaskn Registers. The fields in these variable-range MTRRs , shown in Figure 7-7 on page 224, are: Valid (V)--Bit 11. Indicates that the MTRR pair is valid (enabled) when set to 1. When the valid bit is cleared to 0 the register pair is not used. Range Physical Mask (PhysMask)--Bits 51-12. The mask value used to specify the memory range. Like PhysBase, PhysMask is aligned on a 4-Kbyte physical-address boundary. Bits 11-0 of PhysMask are assumed to be 0.
Chapter 7: Memory System
223
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
63 Reserved, MBZ
52
51 PhysMask (This is an architectural limit. A given implementation may support fewer bits.)
32
31 PhysMask
12
11 V
10 Reserved, MBZ
0
Bits 63-52 51-12 11 10-0
Mnemonic Reserved PhysMask V Reserved
Description Reserved, Must be Zero Range Physical Mask MTRR Pair Enable (Valid) Reserved, Must be Zero
R/W R/W R/W
Figure 7-7.
MTRRphysMaskn Register PhysMask and PhysBase are used together to determine whether a target physical-address falls within the specified address range. PhysMask is logically ANDed with PhysBase and separately ANDed with the upper 40 bits of the target physicaladdress. If the results of the two operations are identical, the target physical-address falls within the specified memory range. The pseudo-code for the operation is:
MaskBase = PhysMask AND PhysBase MaskTarget = PhysMask AND Target_Address[51:12] if MaskBase = MaskTarget then Target_Address_In_Range else Target_Address_Not_In_Range
Variable Range Size and Alignment. T h e s i z e a n d a l i g n m e n t o f variable memory-ranges (MTRRs) and I/O ranges (IORRs) are restricted as follows: The boundary on which a variable range is aligned must be equal to the range size. For example, a memory range of 16 Mbytes must be aligned on a 16-Mbyte boundary. The range size must be a power of 2 (2n, 52 > n > 11), with a minimum allowable size of 4 Kbytes. For example, 4 Mbytes and 8 Mbytes are allowable memory range sizes, but 6 Mbytes is not allowable. PhysMask and PhysBase Values. S o f t wa re c a n c a l c u l a t e t h e PhysMask value using the following procedure: 1. Subtract the memory-range physical base-address from the upper physical-address of the memory range. 224 Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
2. Subtract the value calculated in Step 1 from the physical memory size. 3. Truncate the lower 12 bits of the result in Step 2 to create the PhysMask value to be loaded into the MTRRphysMaskn register. Truncation is performed by right-shifting the value 12 bits. For example, assume a 32-Mbyte memory range is specified within the 52-bit physical address space, starting at address 200_0000h. The upper address of the range is 3FF_FFFFh. Following the process outlined above yields: 1. 3FF_FFFFh-200_0000h = 1FF_FFFFh 2. F_FFFF_FFFF_FFFF-1FF_FFFFh = F_FFFF_FE00_0000h 3. Right shift (F_FFFF_FE00_0000h) by 12 = FF_FFFF_E000h In this example, the 40-bit value loaded into the PhysMask field is FF_FFFF_E000h. Software must also truncate the lower 12 bits of the physical base-address before loading it into the PhysBase field. In the example above, the 40-bit PhysBase field is 00_0000_2000h. Default-Range MTRRs. Physical addresses that are not within ranges established by fixed-range and variable-range MTRRs are set to a default memory-type using the MTRRdefType register. The format of this register is shown in Figure 7-8.
63 Reserved, MBZ 32
31 Reserved, MBZ
12
11 E
10
9
8
7
Type
0
F Reserved, E MBZ
Bits 63-12 11 10 9-8 7-0
Mnemonic Reserved E FE Reserved Type
Description Reserved, Must be Zero MTRR Enable Fixed Range Enable Reserved, Must be Zero Default Memory Type
R/W R/W R/W R/W
Figure 7-8.
MTRR defType Register Format
Chapter 7: Memory System
225
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
The fields within the MTRRdefType register are read/write. These fields are: Type--Bits 7-0. The default memory-type used to characterize physical-memory space. See Table 7-4 on page 219 for the type-field encodings. The extended typefield encodings are not supported by this register. Fixed-Range Enable (FE)--Bit 10. All fixed-range MTRRs are enabled when FE is set to 1. Clearing FE to 0 disables all fixed-range MTRRs. Setting and clearing FE has no effect on the variable-range MTRRs. The FE bit has no effect unless the E bit is set to 1 (see below). MTRR Enable (E)--Bit 11. This is the MTRR enable bit. All fixed-range and variable-range MTRRs are enabled when E is set to 1. Clearing E to 0 disables all fixed-range and variable-range MTRRs and sets the default memory-type to uncacheable (UC) regardless of the value of the Type field. 7.6.3 Using MTRRs Identifying MTRR Features. S o f t wa re d e t e r m i n e s w h e t h e r a processor supports the MTRR mechanism by executing the CPUID instruction with either standard-function 1 or extendedfunction 8000_0001h. If MTRRs are supported, bit 12 in the EDX register is set to 1 by CPUID. See "Processor Feature Identification" on page 78 for more information on the CPUID instruction. The MTRR capability register (MTRRcap) is a read-only register containing information describing the level of MTRR support provided by the processor. Figure 7-9 on page 227 shows the format of this register. If MTRRs are supported, software can read MTRRcap using the RDMSR instruction. Attempting to write to the MTRRcap register causes a generalprotection exception (#GP).
226
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
63 Reserved
32
31 Reserved
11
10
W C
9 r
8
F I X
7 VCNT
0
Bits 63-11 10 9 8 7-0
Mnemonic Reserved WC Reserved FIX VCNT
Description Reserved Write Combining Reserved Fixed-Range Registers Variable-Range Register Count
R/W R R R
Figure 7-9.
MTRR Capability Register Format The MTRRcap register field are: Variable-Range Register Count (VCNT)--Bits 7-0. The VCNT field contains the number of variable-range register pairs supported by the processor. For example, a processor supporting eight register pairs returns a 08h in this field. Fixed-Range Registers (FIX)--Bit 8. The FIX bit indicates whether or not the fixed-range registers are supported. If the processor returns a 1 in this bit, all fixed-range registers are supported. If the processor returns a 0 in this bit, no fixed-range registers are supported. Write-Combining (WC)--Bit 10. The WC bit indicates whether or not the write-combining memory type is supported. If the processor returns a 1 in this bit, WC memory is supported, otherwise it is not supported.
7.6.4 MTRRs and Page Cache Controls
When paging and the MTRRs are both enabled, the address ranges defined by the MTRR registers can span multiple pages, each of which can characterize memory with different types (using the PCD and PWT page bits). When caching is enabled (CR0.CD=0 and CR0.NW=0), the effective memory-type is determined as follows: 1. If the page is defined as cacheable and writeback (PCD=0 and PWT=0), then the MTRR defines the effective memorytype. 2. If the page is defined as not cacheable (PCD=1), then UC is the effective memory-type.
Chapter 7: Memory System
227
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
3. If the page is defined as cacheable and writethrough (PCD=0 and PWT=1), then the MTRR defines the effective memory-type unless the MTRR specifies WB memory, in which case WT is the effective memory-type. Table 7-6 lists the MTRR and page-level cache-control combinations and their combined effect on the final memorytype, if the PAT register holds the default settings. Table 7-6. Combined MTRR and Page-Level Memory Type with Unmodified PAT MSR
Page PCD Bit -- 0 WC 1 1 WP 0 1 0 1 0 WB 0 1
Note:
MTRR Memory Type UC
Page PWT Bit -- -- 0 1 -- -- -- -- 0 1 --
Effective Memory-Type UC WC WC1 UC WP UC WT UC WB WT UC
WT
1. The effective memory-type resulting from the combination of PCD=1, PWT=0, and an MTRR WC memory type is implementation dependent.
Large Page Sizes. When paging is enabled, software can use large page sizes (2 Mbytes and 4 Mbytes) in addition to the more typical 4-Kbyte page size. When large page sizes are used, it is possible for multiple MTRRs to span the memory range within a single large page. Each MTRR can characterize the regions within the page with different memory types. If this occurs, the effective memory-type used by the processor within the large page is undefined.
228
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Software can avoid the undefined behavior in one of the following ways: Avoid using multiple MTRRs to characterize a single large page. Use multiple 4-Kbyte pages rather than a single large page. If multiple MTRRs must be used within a single large page, software can set the MTRR type fields to the same value. If the multiple MTRRs must have different type-field values, software can set the large page PCD and PWT bits to the most restrictive memory type defined by the multiple MTRRs. Overlapping MTRR Registers. If the address ranges of two or more MTRRs overlap, the following rules are applied to determine the memory type used to characterize the overlapping address range: 1. Fixed-range MTRRs, which characterize only the first 1 Mbyte of physical memory, have precedence over variablerange MTRRs. 2. If two or more variable-range MTRRs overlap, the following rules apply: A. If the memory types are identical, then that memory type is used. B. If at least one of the memory types is UC, the UC memory type is used. C. If at least one of the memory types is WT, and the only other memory type is WB, then the WT memory type is used. D. If the combination of memory types is not listed Steps 1 through 3 immediately above, then the memory type used is undefined. 7.6.5 MTRRs in MultiProcessing Environments In multi-processing environments, the MTRRs located in all processors must characterize memory in the same way. Generally, this means identical values are written to the MTRRs used by the processors. Processor implementations do not check the MTRR settings in other processors to ensure consistency. It is the responsibility of system software to initialize and maintain MTRR consistency across all processors.
Chapter 7: Memory System
229
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
7.7
Page-Attribute Table Mechanism
The page-attribute table (PAT) mechanism extends the pagetable entry format and enhances the capabilities provided by the PCD and PWT page-level cache controls. PAT (and PCD, PWT) allow memory-type characterization based on the virtual (linear) address. The PAT mechanism provides the same memory-typing capabilities as the MTRRs but with the added flexibility of the paging mechanism. Software can use both the PAT and MTRR mechanisms to maximize flexibility in memorytype control.
7.7.1 PAT Register
Like the MTRRs, the PAT register is a 64-bit model-specific register (MSR). The format of the PAT registers is shown in Figure 7-10. See "Memory-Typing MSRs" on page 440 for more information on the PAT MSR number and reset value.
56 PA7 55 Reserved 51 50 PA6 48 47 Reserved 43 42 PA5 40 41 Reserved 35 34 PA4 32
63 Reserved
59
58
31 Reserved
27
26 PA3
24
23 Reserved
19
18 PA2
16
15 Reserved
11
10 PA1
8
7 Reserved
3
2 PA0
0
Figure 7-10.
PAT Register The PAT register contains eight page-attribute (PA) fields, numbered from PA0 to PA7. The PA fields hold the encoding of a memory type, as found in Table 7-7 on page 231. The PAT type-encodings match the MTRR type-encodings, with the exception that PAT adds the 07h encoding. The 07h encoding corresponds to a UC- type. The UC- type (07h) is identical to the UC type (00h) except it can be overridden by an MTRR type of WC. Software can write any supported memory-type encoding into any of the eight PA fields. An attempt to write anything but zeros into the reserved fields causes a general-protection exception (#GP). An attempt to write an unsupported type encoding into a PA field also causes a #GP exception. The PAT register fields are initiated at processor reset to the default values shown in Table 7-8 on page 232.
230
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 7-7. PAT Type Encodings
Type Value 00h 01h Type Name UC--Uncacheable WC--Write-Combining Type Description All accesses are uncacheable. Write combining is not allowed. Speculative accesses are not allowed. All accesses are uncacheable. Write combining is allowed. Speculative reads are allowed. Reads allocate cache lines on a cache miss, but only to the shared state. Cache lines are not allocated on a write miss. Write hits update the cache and main memory. Reads allocate cache lines on a cache miss, but only to the shared state. All writes update main memory. Cache lines are not allocated on a write miss. Write hits invalidate the cache line and update main memory. Reads allocate cache lines on a cache miss, and can allocate to either the shared or exclusive state. Writes allocate to the modified state on a cache miss. All accesses are uncacheable. Write combining is not allowed. Speculative accesses are not allowed. Can be overridden by an MTRR with the WC type.
04h
WT--Writethrough
05h
WP--Write-Protect
06h
WB--Writeback
07h
UC- (UC minus)
7.7.2 PAT Indexing
PA fields in the PAT register are selected using three bits from the page-table entries. These bits are: PAT (page attribute table)--The PAT bit is bit 7 in 4-Kbyte PTEs; it is bit 12 in 2-Mbyte and 4-Mbyte PDEs. Page-table entries that don't have a PAT bit (PML4 entries, for example) assume PAT = 0. PCD (page cache disable)--The PCD bit is bit 4 in all pagetable entries. The PCD from the PTE or PDE is selected depending on the paging mode. PWT (page writethrough)--The PWT bit is bit 3 in all pagetable entries. The PWT from the PTE or PDE is selected depending on the paging mode. Table 7-8 on page 232 shows the various combinations of the PAT, PCD, and PWT bits used to select a PA field within the PAT register. Table 7-8 also shows the default memory-type values established in the PAT register by the processor after a reset. The default values correspond to the memory types established by the PCD and PWT bits alone in processor implementations t h a t d o n o t s u p p o r t t h e PAT m e ch a n i s m . I n s u ch
Chapter 7: Memory System
231
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
implementations, the PAT field in page-table entries is reserved and cleared to 0. See "Page-Translation-Table Entry Fields" on page 168 for more information on the page-table entries. Table 7-8. PAT-Register PA-Field Indexing
PAT Register Field PA0 PA1 PA2 PA3 PA4 PA5 PA6 PA7 Default Memory Type WB WT UC-1 UC WB WT UC-1 UC
Page-Table Entry Bits PAT 0 0 0 0 1 1 1 1
Note:
PCD 0 0 1 1 0 0 1 1
PWT 0 1 0 1 0 1 0 1
1. Can be overridden by WC memory type set by an MTRR.
7.7.3 Identifying PAT Support
Software determines whether a processor supports the PAT mechanism by executing the CPUID instruction with either standard-function 1 or extended-function 8000_0001h. If PAT is supported, bit 16 in the EDX register is set to 1 by CPUID. See "Processor Feature Identification" on page 78 for more information on the CPUID instruction. If PAT is supported by a processor implementation, it is always enabled. The PAT mechanism cannot be disabled by software. Software can effectively avoid using PAT by: Not setting PAT bits in page-table entries to 1. Not modifying the reset values of the PA fields in the PAT register. In this case, memory is characterized using the same types that are used by implementations that do not support PAT.
7.7.4 PAT Accesses
In implementations that support the PAT mechanism, all memory accesses that are translated through the paging mechanism use the PAT index bits to specify a PA field in the Chapter 7: Memory System
232
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
PAT register. The memory type stored in the specified PA field is applied to the memory access. The process is summarized as: 1. A virtual address is calculated as a result of a memory access. 2. The virtual address is translated to a physical address using the page-translation mechanism. 3. The PAT, PCD and PWT bits are read from the corresponding page-table entry during the virtual-address to physical-address translation. 4. The PAT, PCD and PWT bits are used to select a PA field from the PAT register. 5. The memory type is read from the appropriate PA field. 6. The memory type is applied to the physical-memory access using the translated physical address. Page-Translation Table Access. The PAT bit exists only in the PTE (4K paging) or PDEs (2/4 Mbyte paging). In the remaining upper levels (PML4 PDP, 4K PDEs), only the PWT and PCD bits are used to index into the first 4 entries in the PAT register. The resulting memory type is used for the next lower paging level. 7.7.5 Combined Effect of MTRRs and PAT The memory types established by the PAT mechanism can be combined with MTRR-established memory types to form an effective memory-type. The combined effect of MTRR and PAT memory types are shown in Table 7-9 on page 234. In the AMD64 architecture, reserved and undefined combinations of MTRR and PAT memory types result in undefined behavior. If the MTRRs are disabled in implementations that support the M T R R m e ch a n i s m , t h e d e fa u l t m e m o ry t y p e i s s e t t o uncacheable (UC).
Chapter 7: Memory System
233
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 7-9.
Combined Effect of MTRR and PAT Memory Types
MTRR Memory Type UC, WC, WP, WT, WB UC UCWC WP, WT, WB WC -- UC WC WP WP WT WB UC WT WC, WP WT, WB UC WC WB WP WT WB Effective Memory Type UC UC WC UC1 WC UC UC1 WP UC1 WP UC UC1 WT UC WC WP WT WB
PAT Memory Type UC
Note:
1. Previously reserved (undefined) combinations are set to the UC memory type by the AMD64 architecture.
7.8
Memory-Mapped I/O
Processor implementations can independently direct reads and writes to either system memory or memory-mapped I/O. The m e t h o d u s e d fo r d i re c t i n g t h o s e m e m o ry a c c e s s e s i s implementation dependent. In some implementations, separate system-memory and memory-mapped I/O buses can be provided at the processor interface. In other implementations, system
234
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
memory and memory-mapped I/O share common data and address buses, and system logic uses sideband signals from the processor to route accesses appropriately. Refer to AMD data sheets and application notes for more information about p a r t i c u l a r h a rdwa re i m p l e m e n t a t i o n s o f t h e A M D 6 4 architecture. The I/O range registers (IORRs), and the top-of-memory registers allow system software to specify where memory accesses are directed for a given address range. The MTRR extensions are described in the following section. "IORRs" on page 237 describes the IORRs and "Top of Memory" on p a g e 2 3 9 d e s c r i b e s t h e t o p - o f - m e m o ry re g i s t e rs . I n implementations that support these features, the default action taken when the features are disabled is to direct memory accesses to memory-mapped I/O. 7.8.1 Extended FixedRange MTRR TypeField Encodings The fixed-range MTRRs support extensions to the type-field encodings that allow system software to direct memory accesses to system memory or memory-mapped I/O. The extended MTRR type-field encodings use previously reserved bits 4-3 to specify whether reads and writes to a physical-address range are to system memory or to memory-mapped I/O. The format for this encoding is shown in Figure 7-11. The new bits are: WrMem--Bit 3. When set to 1, the processor directs write requests for this physical address range to system memory. When cleared to 0, writes are directed to memory-mapped I/O. RdMem--Bit 4. When set to 1, the processor directs read requests for this physical address range to system memory. When cleared to 0, reads are directed to memory-mapped I/O. The type subfield (bits 2-0) allows the encodings specified in Table 7-4 on page 219 to be used for memory characterization.
7 Reserved
5
4 RdMem
3 WrMem
2 Type
0
Figure 7-11. Extended MTRR Type-Field Format (Fixed-Range MTRRs)
Chapter 7: Memory System
235
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
These extensions are enabled using the following bits in the SYSCFG MSR: MtrrFixDramEn--Bit 18. When set to 1, RdMem and WrMem attributes are enabled. When cleared to 0, these attributes are disabled. When disabled, accesses are directed to memorymapped I/O space. MtrrFixDramModEn--Bit 19. When set to 1, software can read and write the RdMem and WrMem bits. When cleared to 0, writes do not modify the RdMem and WrMem bits, and reads return 0. To use the MTRR extensions, system software must first set MtrrFixDramModEn=1 to allow modification to the RdMem and WrMem bits. After the attribute bits are properly initialized in the fixed-range registers, the extensions can be enabled by setting MtrrFixDramEn=1. RdMem and WrMem allow the processor to independently direct reads and writes to either system memory or memorymapped I/O. The RdMem and WrMem controls are particularly useful when shadowing ROM devices located in memorymapped I/O space. It is often useful to shadow such devices in RAM system memory to improve access performance, but writes into the RAM location can corrupt the shadowed ROM information. The MTRR extensions solve this problem. System software can create the shadow location by setting WrMem=1 and RdMem=0 for the specified memory range and then copy the ROM location into itself. Reads are directed to the memorymapped ROM, but writes go to the same physical addresses in system memory. After the copy is complete, system software can change the bit values to WrMem=0 and RdMem=1. Now reads are directed to the faster copy located in system memory, and writes are directed to memory-mapped ROM. The ROM responds as it would normally to a write, which is to ignore it. Not all combinations of RdMem and WrMem are supported for each memory type encoded by bits 2-0. Table 7-10 on page 237 shows the allowable combinations. The behavior of reserved encoding combinations (shown as gray-shaded cells) is undefined and results in unpredictable behavior.
236
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 7-10.
RdMem
Extended Fixed-Range MTRR Type Encodings
WrMem Type 0 (UC) 1 (WC) Implication or Potential Use UC I/O WC I/O WT I/O WP I/O Reserved Used while creating a shadowed ROM
0
0
4 (WT) 5 (WP) 6 (WB) 0 (UC) 1 (WC)
0
1
4 (WT) 5 (WP) 6 (WB) 0 (UC) 1 (WC) Used to access a shadowed ROM Reserved WP Memory (Can be used to access shadowed ROM) Reserved UC Memory WC Memory WT Memory Reserved WB Memory Reserved
1
0
4 (WT) 5 (WP) 6 (WB) 0 (UC) 1 (WC)
1
1
4 (WT) 5 (WP) 6 (WB)
7.8.2 IORRs
The IORRs operate similarly to the variable-range MTRRs. The IORRs specify whether reads and writes in any physicaladdress range map to system memory or memory-mapped I/O. Up to two address ranges of varying sizes can be controlled using the IORRs. A pair of IORRs are used to control each address range: IORRBasen and IORRMaskn (n is the addressrange number from 0 to 1).
Chapter 7: Memory System
237
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Figure 7-12 on page 238 shows the format of the IORRBasen registers and Figure 7-13 on page 239 shows the format of the IORRMaskn registers. The fields within the register pair are read/write. IORRBasen Registers. The fields in these IORRs are: WrMem--Bit 3. When set to 1, the processor directs write requests for this physical address range to system memory. When cleared to 0, writes are directed to memory-mapped I/O. RdMem--Bit 4. When set to 1, the processor directs read requests for this physical address range to system memory. When cleared to 0, reads are directed to memory-mapped I/O. Range Physical-Base-Address (PhysBase)--Bits 51-12. The memory-range base-address in physical-address space. PhysBase is aligned on a 4-Kbyte (or greater) address in the 52-bit physical-address space supported by the AMD64 architecture. PhysBase represents the most-significant 40address bits of the physical address. Physical-address bits 11-0 are assumed to be 0. The format of these registers is shown in Figure 7-12.
63 Reserved, MBZ 52 51 PhysBase (This is an architectural limit. A given implementation may support fewer bits.) 32
31 PhysBase
12
11 Reserved, MBZ
5
4
R d
3
W r
0
Reserved, MBZ
Bits 63-52 51-12 11-5 4 3 2-0
Mnemonic Reserved PhysBase Reserved Rd Wr Reserved
Description Reserved, Must be Zero Range Physical Base Address Reserved, Must be Zero RdMem Enable WrMem Enable Reserved, Must be Zero
R/W R/W R/W R/W
Figure 7-12.
IORRBasen Register IORRMaskn Registers. The fields in these IORRs are: Valid (V)--Bit 11. Indicates that the IORR pair is valid (enabled) when set to 1. When the valid bit is cleared to 0
238
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
the register pair is not used for memory-mapped I/O control (disabled). Range Physical-Mask (PhysMask)--Bits 51-12. The mask value used to specify the memory range. Like PhysBase, PhysMask is aligned on a 4-Kbyte physical-address boundary. Bits 11-0 of PhysMask are assumed to be 0. The format of these registers is shown in Figure 7-13.
63 Reserved, MBZ 52 51 PhysMask (This is an architectural limit. A given implementation may support fewer bits.) 32
31 PhysMask
12
11 V
10 Reserved, MBZ
0
Bits 63-52 51-12 11 10-0
Mnemonic Reserved PhysMask V Reserved
Description Reserved, Must be Zero Range Physical Mask I/O Register Pair Enable (Valid) Reserved, Must be Zero
R/W R/W R/W
Figure 7-13.
IORRMaskn Register The operation of the PhysMask and PhysBase fields is identical to that of the variable-range MTRRs. See page 224 for a description of this operation.
7.8.3 IORR Overlapping 7.8.4 Top of Memory
The use of overlapping IORRs is not recommended. If overlapping IORRs are specified, the resulting behavior is implementation-dependent. The top-of-memory registers, TOP_MEM and TOP_MEM2, allow system software to specify physical addresses ranges as memory-mapped I/O locations. Processor implementations can direct accesses to memory-mapped I/O differently than system I/O, and the precise method depends on the implementation. System software specifies memory-mapped I/O regions by writing an address into each of the top-of-memory registers. The memory regions specified by the TOP_MEM registers are aligned on 8-Mbyte boundaries as follows: Memory accesses from physical address 0 to one less than the value in TOP_MEM are directed to system memory.
Chapter 7: Memory System
239
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Memory accesses from the physical address specified in TOP_MEM to FFFF_FFFFh are directed to memory-mapped I/O. Memory accesses from physical address 1_0000_0000h to one less than the value in TOP_MEM2 are directed to system memory. Memory accesses from the physical address specified in TOP_MEM2 to the maximum physical address supported by the system are directed to memory-mapped I/O. Figure 7-14 shows how the top-of-memory registers organize memory into separate system-memory and memory-mapped I/O regions.
Physical Memory
Maximum System Memory
Memory-Mapped I/O
TOP_MEM2 TOP_MEM2 - 1
System Memory
4GB 4GB - 1 TOP_MEM TOP_MEM - 1
Memory-Mapped I/O
System Memory
0
513-269.eps
Figure 7-14.
Memory Organization Using Top-of-Memory Registers
Figure 7-15 on page 241 shows the format of the TOP_MEM and TOP_MEM2 registers. Bits 51-23 specify an 8-Mbyte aligned physical address. All remaining bits are reserved and ignored by the processor. System software should clear those bits to zero to maintain compatibility with possible future extensions to the registers. The TOP_MEM registers are model-specific registers. See "Memory-Typing MSRs" on page 440 for information on the MSR address and reset values for these registers. 240 Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
63 Reserved, IGN
52
51 Top-of-Memory Physical Address (This is an architectural limit. A given implementation may support fewer bits.)
32
31 Top-of-Memory Physical Address
23
22 Reserved, IGN
0
Figure 7-15.
Top-of-Memory Registers (TOP_MEM, TOP_MEM2) T h e T O P _ M E M re g i s t e r i s e n a b l e d b y s e t t i n g t h e MtrrVarDramEn bit in the SYSCFG MSR (bit 20) to 1. The TOP_MEM2 register is enabled by setting the MtrrTom2En bit in the SYSCFG MSR (bit 21) to 1. The registers are disabled when their respective enable bits are cleared to 0. When the top-of-memory registers are disabled, memory accesses default to memory-mapped I/O space.
Chapter 7: Memory System
241
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
242
Chapter 7: Memory System
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
8
Exceptions and Interrupts
Exceptions and interrupts force control transfers from the currently-executing program to a system-software service routine that handles the interrupting event. These routines are referred to as exception handlers and interrupt handlers, or collectively as event handlers. Typically, interrupt events can be handled by the service routine transparently to the interrupted program. During the control transfer to the service routine, the processor stops executing the interrupted program and saves its return pointer. The system-software service routine that handles the exception or interrupt is responsible for saving the state of the interrupted program. This allows the processor to restart the interrupted program after system software has handled the event. When an exception or interrupt occurs, the processor uses the interrupt-vector number as an index into the interruptdescriptor table (IDT). An IDT is used in all processor operating modes, including real mode (also called real-address mode), protected mode, and long mode. Exceptions and interrupts come from three general sources: Exceptions occur as a result of software execution errors or other internal-processor errors. Exceptions also occur during non-error situations, such as program single stepping or address-breakpoint detection. Exceptions are considered synchronous events because they are a direct result of executing the interrupted instruction. Software interrupts occur as a result of executing interrupt instructions. Unlike exceptions and external interrupts, software interrupts allow intentional triggering of the interrupt-handling mechanism. Like exceptions, software interrupts are synchronous events. External interrupts are generated by system logic in response to an error or some other event outside the processor. They are reported over the processor bus using external signalling. External interrupts are asynchronous events that occur independently of the interrupted instruction. Throughout this section, the term masking can refer to either disabling or delaying an interrupt. For example, masking external interrupts delays the interrupt, with the processor
Chapter 8: Exceptions and Interrupts
243
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
holding the interrupt as pending until it is unmasked. With floating-point exceptions (128-bit media and x87), masking prevents an interrupt from occurring and causes the processor to perform a default operation on the exception condition.
8.1
General Characteristics
Exceptions and interrupts have several different characteristics that depend on how events are reported and the implications for program restart.
8.1.1 Precision
Precision describes how the exception is related to the interrupted program: Precise exceptions are reported on a predictable instruction boundary. This boundary is generally the first instruction that has not completed when the event occurs. All previous instructions (in program order) are allowed to complete before transferring control to the event handler. The pointer to the instruction boundary is saved automatically by the processor. When the event handler completes execution, it returns to the interrupted program and restarts execution at the interrupted-instruction boundary. Imprecise exceptions are not guaranteed to be reported on a predictable instruction boundary. The boundary can be any instruction that has not completed when the interrupt event occurs. Imprecise events can be considered asynchronous, because the source of the interrupt is not necessarily related to the interrupted instruction. Imprecise exception and interrupt handlers typically collect machine-state information related to the interrupting event for reporting through system-diagnostic software. The interrupted program is not restartable.
8.1.2 Instruction Restart
As mentioned above, precise exceptions are reported on an instruction boundary. The instruction boundary can be reported in one of two locations: Most exceptions report the boundary before the instruction causing the exception. In this case, all previous instructions (in program order) are allowed to complete, but the interrupted instruction is not. No program state is updated as a result of partially executing an interrupted instruction.
244
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Some exceptions report the boundary after the instruction causing the exception. In this case, all previous instructions--including the one executing when the exception occurred--are allowed to complete. Program state can be updated when the reported boundary is after the instruction causing the exception. This is particularly true when the event occurs as a result of a task switch. In this case, the general registers, segment-selector registers, page-base address register, and LDTR are all updated by the hardware task-switch mechanism. The event handler cannot rely on the state of those registers when it begins execution and must be careful in validating the state of the segment-selector registers before restarting the interrupted task. This is not an issue in long mode, however, because the hardware task-switch mechanism is disabled in long mode. 8.1.3 Types of Exceptions There are three types of exceptions, depending on whether they are precise and how they affect program restart: Faults are precise exceptions reported on the boundary before the instruction causing the exception. Generally, faults are caused by an error condition involving the faulted instruction. Any machine-state changes caused by the faulting instruction are discarded so that the instruction can be restarted. The saved rIP points to the faulting instruction. Traps are precise exceptions reported on the boundary following the instruction causing the exception. The trapped instruction is completed by the processor and all state changes are saved. The saved rIP points to the instruction following the faulting instruction. Aborts are imprecise exceptions. Because they are imprecise, aborts typically do not allow reliable program restart. 8.1.4 Masking External Interrupts General Masking Capabilities. Software can mask the occurrence of certain exceptions and interrupts. Masking can delay or even prevent triggering of the exception-handling or interrupthandling mechanism when an interrupt-event occurs. External interrupts are classified as maskable or nonmaskable: Maskable interrupts trigger the interrupt-handling mechanism only when RFLAGS.IF=1. Otherwise they are held pending for as long as the RFLAGS.IF bit is cleared to 0.
Chapter 8: Exceptions and Interrupts
245
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Nonmaskable interrupts (NMI) are unaffected by the value of the rFLAGS.IF bit. However, the occurrence of an NMI masks further NMIs until an IRET instruction is executed. Masking During Stack Switches. The processor delays recognition of maskable external interrupts and debug exceptions during certain instruction sequences that are often used by software to switch stacks. The typical programming sequence used to switch stacks is: 1. Load a stack selector into the SS register. 2. Load a stack offset into the ESP register. If an interrupting event occurs after the selector is loaded but before the stack offset is loaded, the interrupted-program stack pointer is invalid during execution of the interrupt handler. To prevent interrupts from causing stack-pointer problems, the processor does not allow external interrupts or debug exceptions to occur until the instruction immediately following the MOV SS or POP SS instruction completes execution. The recommended method of performing this sequence is to use the LSS instruction. LSS loads both SS and ESP, and the instruction inhibits interrupts until both registers are updated successfully. 8.1.5 Masking Floating-Point and Media Instructions Any x87 floating-point exceptions can be masked and reported later using bits in the x87 floating-point status register (FSW) and the x87 floating-point control register (FCW). The floatingpoint exception-pending exception is used for unmasked x87 floating-point exceptions (see "#MF--x87 Floating-Point Exception-Pending (Vector 16)" on page 262). The SIMD floating-point exception is used for unmasked 128-bit media floating-point exceptions (see "#XF--SIMD FloatingPoint Exception (Vector 19)" on page 265). 128-bit media floating-point exceptions are masked using the MXCSR register. The exception mechanism is not triggered when these exceptions are masked. Instead, the processor handles the exceptions in a default manner. 8.1.6 Disabling Exceptions Disabling an exception prevents the exception condition from being recognized, unlike masking an exception which prevents triggering the exception mechanism after the exception is recognized. Some exceptions can be disabled by system Chapter 8: Exceptions and Interrupts
246
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
software running at CPL=0, using bits in the CR0 register or CR4 register: Alignment-check exception (see "#AC--Alignment-Check Exception (Vector 17)" on page 263). Device-not-available exception (see "#NM--Device-NotAvailable Exception (Vector 7)" on page 254). Machine-check exception (see "#MC--Machine-Check Exception (Vector 18)" on page 264). The debug-exception mechanism provides control over when s p e c i f i c b re a k p o i n t s a re e n ab l e d a n d d i s ab l e d . S e e "Breakpoints" on page 394 for more information on how breakpoint controls are used for triggering the debug-exception mechanism.
8.2
Vectors
Specific exception and interrupt sources are assigned a fixed vector-identification number (also called an "interrupt vector" or simply "vector"). The interrupt vector is used by the interrupt-handling mechanism to locate the system-software service routine assigned to the exception or interrupt. Up to 256 unique interrupt vectors are available. The first 32 vectors are reserved for predefined exception and interrupt conditions. Software-interrupt sources can trigger an interrupt using any available interrupt vector. Table 8-1 on page 248 lists the supported interrupt-vector numbers, the corresponding exception or interrupt name, the mnemonic, the source of the interrupt event, and a summary of the possible causes.
Chapter 8: Exceptions and Interrupts
247
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 8-1.
Vector 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20--31 0--255 0--255
Interrupt-Vector Source and Cause
Exception/Interrupt Divide-by-Zero-Error Debug Non-Maskable-Interrupt Breakpoint Overflow Bound-Range Invalid-Opcode Device-Not-Available Double-Fault Coprocessor-Segment-Overrun Invalid-TSS Segment-Not-Present Stack General-Protection Page-Fault Reserved x87 Floating-Point ExceptionPending Alignment-Check Machine-Check SIMD Floating-Point Reserved External Interrupts (Maskable) Software Interrupts -- -- External interrupts INTn instruction #MF #AC #MC #XF Mnemonic #DE #DB -- #BP #OF #BR #UD #NM #DF -- #TS #NP #SS #GP #PF Cause DIV, IDIV, AAM instructions Instruction accesses and data accesses External NMI signal INT3 instruction INTO instruction BOUND instruction Invalid instructions x87 instructions Exception during the handling of another exception or interrupt Unsupported (Reserved) Task-state segment access and task switch Segment register loads SS register loads and stack references Memory accesses and protection checks Memory accesses when paging enabled -- x87 floating-point instructions Misaligned memory accesses Model specific 128-bit media floating-point instructions --
Table 8-2 on page 249 shows how each interrupt vector is classified. Reserved interrupt vectors are indicated by the grayshaded rows. 248 Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 8-2.
Interrupt-Vector Classification
Vector 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Note:
Interrupt (Exception) Divide-by-Zero-Error Debug Non-Maskable-Interrupt Breakpoint Overflow Bound-Range Invalid-Opcode Device-Not-Available Double-Fault Coprocessor-Segment-Overrun Invalid-TSS Segment-Not-Present Stack General-Protection Page-Fault Reserved x87 Floating-Point ExceptionPending Alignment-Check Machine-Check SIMD Floating-Point
Type Fault Fault or Trap -- Trap
Precise yes --
Class2 Contributory
yes Fault
Benign
Abort
no
Contributory Fault yes Benign or Contributory
Fault Abort Fault
no yes no yes Benign
1. External interrupts are not classified by type or whether or not they are precise. 2. See "#DF--Double-Fault Exception (Vector 8)" on page 254 for a definition of benign and contributory classes.
Chapter 8: Exceptions and Interrupts
249
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 8-2.
Interrupt-Vector Classification (continued)
Vector 20--31 0--255 0--255
Note:
Interrupt (Exception) Reserved External Interrupts (Maskable) Software Interrupts
Type
Precise
Class2
--1
--1
Benign
1. External interrupts are not classified by type or whether or not they are precise. 2. See "#DF--Double-Fault Exception (Vector 8)" on page 254 for a definition of benign and contributory classes.
The following sections describe each interrupt in detail. The format of the error code reported by each interrupt is described in "Error Codes" on page 267. 8.2.1 #DE--Divide-byZero-Error Exception (Vector 0) A #DE exception occurs when the denominator of a DIV instruction or an IDIV instruction is 0. A #DE also occurs if the result is too large to be represented in the destination. #DE cannot be disabled. Error Code Returned. None. Program Restart. #DE is a fault-type exception. The saved instruction pointer points to the instruction that caused the #DE. 8.2.2 #DB--Debug Exception (Vector 1) When the debug-exception mechanism is enabled, a #DB exception can occur under any of the following circumstances: Instruction execution. Instruction single stepping. Data read. Data write. I/O read. I/O write. Task switch. Debug-register access, or general detect fault (debug register access when DR7.GD=1). Executing the INT1 instruction (opcode 0F1h).
250
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
#DB conditions are enabled and disabled using the debugcontrol register, DR7 and RFLAGS.TF. Each #DB condition is described in more detail in "Breakpoints" on page 394. Error Code Returned. None. #DB information is returned in the debug-status register, DR6. Program Restart. #DB can be either a fault-type or trap-type exception. In the following cases, the saved instruction pointer points to the instruction that caused the #DB: Instruction execution. Invalid debug-register access, or general detect. In all other cases, the instruction that caused the #DB is completed, and the saved instruction pointer points to the instruction after the one that caused the #DB. The RFLAGS.RF bit can be used to restart an instruction following an instruction breakpoint resulting in a #DB. In most cases, the processor clears RFLAGS.RF to 0 after every instruction is successfully executed. However, in the case of the IRET, JMP, CALL, and INTn (through a task gate) instructions, RFLAGS.RF is not cleared to 0 until the next instruction successfully executes. When a non-debug exception occurs (or when a string instruction is interrupted), the processor normally sets RFLAGS.RF to 1 in the RFLAGS image that is pushed on the interrupt stack. A subsequent IRET back to the interrupted program pops the RFLAGS image off the stack and into the RFLAGS register, with RFLAGS.RF=1. The interrupted instruction executes without causing an instruction breakpoint, after which the processor clears RFLAGS.RF to 0. However, when a #DB exception occurs, the processor clears RFLAGS.RF to 0 in the RFLAGS image that is pushed on the interrupt stack. The #DB handler has two options: Disable the instruction breakpoint completely. Set RFLAGS.RF to 1 in the interrupt-stack rFLAGS image. The instruction breakpoint condition is ignored immediately after the IRET, but reoccurs if the instruction address is accessed later, as can occur in a program loop.
Chapter 8: Exceptions and Interrupts
251
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
8.2.3 NMI--NonMaskable-Interrupt Exception (Vector 2)
An NMI exception occurs as a result of system logic signalling a non-maskable interrupt to the processor. Error Code Returned. None. Program Restart. NMI is an interrupt. The processor recognizes an NMI at an instruction boundary. The saved instruction pointer points to the instruction immediately following the boundary where the NMI was recognized. Masking. NMI cannot be masked. However, when an NMI is recognized by the processor, recognition of subsequent NMIs are disabled until an IRET instruction is executed.
8.2.4 #BP-- Breakpoint Exception (Vector 3)
A #BP exception occurs when an INT3 instruction is executed. The INT3 is normally used by debug software to set instruction breakpoints by replacing instruction-opcode bytes with the INT3 opcode. #BP cannot be disabled. Error Code Returned. None. Program Restart. # B P i s a t ra p - t y p e ex c e p t i o n . Th e s ave d i n s t r u c t i o n p o i n t e r p o i n t s t o t h e byt e a f t e r t h e I N T 3 instruction. This location can be the start of the next instruction. However, if the INT3 is used to replace the first opcode bytes of an instruction, the restart location is likely to be in the middle of an instruction. In the latter case, the debug software must replace the IN T3 byte with the correct instruction byte. The saved RIP instruction pointer must then be decremented by one before returning to the interrupted program. This allows the program to be restarted correctly on the interrupted-instruction boundary.
8.2.5 #OF--Overflow Exception (Vector 4)
An #OF exception occurs as a result of executing an INTO instruction while the overflow bit in RFLAGS is set to 1 (RFLAGS.OF=1). #OF cannot be disabled. Error Code Returned. None. Program Restart. # O F i s a t ra p - t y p e ex c e p t i o n . Th e s ave d instruction pointer points to the instruction following the INTO instruction that caused the #OF.
252
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
8.2.6 #BR--BoundRange Exception (Vector 5)
A #BR exception can occur as a result of executing the BOUND instruction. The BOUND instruction compares an array index (first operand) with the lower bounds and upper bounds of an array (second operand). If the array index is not within the array boundary, the #BR occurs. #BR cannot be disabled. Error Code Returned. None. Program Restart. #BR is a fault-type exception. The saved instruction pointer points to the BOUND instruction that caused the #BR.
8.2.7 #UD--InvalidOpcode Exception (Vector 6)
A #UD exception occurs when an attempt is made to execute an invalid or undefined opcode. The validity of an opcode often depends on the processor operating mode. A #UD occurs under the following conditions: Execution of any reserved or undefined opcode in any mode. Execution of the UD2 instruction. Use of the LOCK prefix on an instruction that cannot be locked. Use of the LOCK prefix on a lockable instruction with a nonmemory target location. Execution of an instruction with an invalid-operand type. Execution of the SYSENTER or SYSEXIT instructions in long mode. Execution of any of the following instructions in 64-bit mode: AAA, AAD, AAM, AAS, BOUND, CALL (opcode 9A), DAA, DAS, DEC, INC, INTO, JMP (opcode EA), LDS, LES, POP (DS, ES, SS), POPA, PUSH (CS, DS, ES, SS), PUSHA, SALC. Execution of the ARPL, LAR, LLDT, LSL, LTR, SLDT, STR, VERR, or VERW instructions when protected mode is not enabled, or when virtual-8086 mode is enabled. Execution of any 128-bit media instruction (uses XMM registers), or 64-bit media instruction (uses MMXTM registers) when CR0.NM = 1. Execution of any 128-bit media floating-point instruction (uses XMM registers) that causes a numeric exception when CR4.OSXMMEXCPT = 0. Use of the DR4 or DR5 debug registers when CR4.DE = 1.
Chapter 8: Exceptions and Interrupts
253
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Execution of RSM when not in SMM mode. See the specific instruction description (in the other volumes) for additional information on invalid conditions. #UD cannot be disabled. Error Code Returned. None. Program Restart. #UD is a fault-type exception. The saved instruction pointer points to the instruction that caused the #UD. 8.2.8 #NM--DeviceNot-Available Exception (Vector 7) A #NM exception occurs under any of the following conditions: An FWAIT/WAIT instruction is executed when CR0.MP=1 and CR0.TS=1. Any x87 instruction other than FWAIT is executed when CR0.EM=1. Any x87 instruction is executed when CR0.TS=1. The CR0.MP bit controls whether the FWAIT/WAIT instruction causes an #NM exception when TS=1. Any 128-bit or 64-bit media instruction when CR0.TS=1. #NM can be enabled or disabled under the control of the CR0.MP, CR0.EM, and CR0.TS bits as described above. See "CR0 Register" on page 55 for more information on the CR0 bits used to control the #NM exception. Error Code Returned. None. Program Restart. #NM is a fault-type exception. The saved instruction pointer points to the instruction that caused the #NM. 8.2.9 #DF--DoubleFault Exception (Vector 8) A #DF exception can occur when a second exception occurs during the handling of a prior (first) exception or interrupt handler. Usually, the first and second exceptions can be handled sequentially without resulting in a #DF. In this case, the first exception is considered benign, as it does not harm the ability of the processor to handle the second exception. In some cases, however, the first exception adversely affects the ability of the processor to handle the second exception. These 254 Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
exceptions contribute to the occurrence of a #DF, and are called contributory exceptions. If a contributory exception is followed by another contributory exception, a double-fault exception occurs. Likewise, if a page fault is followed by another page fault or a contributory exception, a double-fault exception occurs. Table 8-3 shows the conditions under which a #DF occurs. Page faults are either benign or contributory, and are listed separately. See the "Class" column in Table 8-2 on page 249 for information on whether an exception is benign or contributory. Table 8-3. Double-Fault Exception Conditions
Second Interrupting Event Divide-by-Zero-Error Exception Invalid-TSS Exception Segment-Not-Present Exception Stack Exception General-Protection Exception Page Fault Exception Divide-by-Zero-Error Exception Invalid-TSS Exception Segment-Not-Present Exception Stack Exception General-Protection Exception
First Interrupting Event Contributory Exceptions * Divide-by-Zero-Error Exception * Invalid-TSS Exception * Segment-Not-Present Exception * Stack Exception * General-Protection Exception
Page Fault Exception
If a third interrupting event occurs while transferring control to the #DF handler, the processor shuts down. Only an NMI, RESET, or INIT can restart the processor in this case. However, if the processor shuts down as it is executing an NMI handler, the processor can only be restarted with RESET or INIT. #DF cannot be disabled. Error Code Returned. Zero. Program Restart. #DF is an abort-type exception. The saved instruction pointer is undefined, and the program cannot be restarted.
Chapter 8: Exceptions and Interrupts
255
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
8.2.10 CoprocessorSegment-Overrun Exception (Vector 9)
This interrupt vector is reserved. It is for a discontinued exception originally used by processors that supported external x87-instruction coprocessors. On those processors, the exception condition is caused by an invalid-segment or invalidpage access on an x87-instruction coprocessor-instruction operand. On current processors, this condition causes a generalprotection exception to occur. Error Code Returned. Not applicable. Program Restart. Not applicable.
8.2.11 #TS--InvalidTSS Exception (Vector 10)
A #TS exception occurs when an invalid reference is made to a segment selector as part of a task switch. A #TS also occurs during a privilege-changing control transfer (through a call gate or an interrupt gate), if a reference is made to an invalid stacksegment selector located in the TSS. Table 8-4 on page 257 lists the conditions under which a #TS occurs and the error code returned by the exception mechanism. #TS cannot be disabled. Error Code Returned. See Table 8-4 on page 257 for a list of error codes returned by the #TS exception. Program Restart. #TS is a fault-type exception. If the exception occurs before loading the segment selectors from the TSS, the saved instruction pointer points to the instruction that caused the #TS. However, most #TS conditions occur due to errors with the loaded segment selectors. When an error is found with a segment selector, the hardware task-switch mechanism completes loading the new task state from the TSS, and then triggers the #TS exception mechanism. In this case, the saved instruction pointer points to the first instruction in the new task. In long mode, a #TS cannot be caused by a task switch, because the hardware task-switch mechanism is disabled. A #TS occurs only as a result of a control transfer through a gate descriptor that results in an invalid stack-segment reference using an SS selector in the TSS. In this case, the saved instruction pointer always points to the control-transfer instruction that caused the #TS.
256
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 8-4.
Invalid-TSS Exception Conditions
Error Condition TSS limit check on a task switch TSS limit check on an inner-level stack pointer LDT does not point to GDT Error Code TSS Selector Index
Selector Reference Task-State Segment
LDT Segment
LDT reference outside GDT GDT entry is not an LDT descriptor LDT descriptor is not present CS reference outside GDT or LDT
LDT Selector Index
Code Segment
Privilege check (conforming DPL > CPL) Privilege check (non-conforming DPL CPL) Type check (CS not executable)
CS Selector Index
Data Segment
Data segment reference outside GDT or LDT Type check (data segment not readable) SS reference outside GDT or LDT
DS, ES, FS or GS Selector Index
Stack Segment
Privilege check (stack segment descriptor DPL CPL) Privilege check (stack segment selector RPL CPL) Type check (stack segment not writable)
SS Selector Index
8.2.12 #NP--SegmentNot-Present Exception (Vector 11)
An #NP occurs when an attempt is made to load a segment or gate with a clear present bit, as described in the following situations: Using the MOV, POP, LDS, LES, LFS, or LGS instructions to load a segment selector (DS, ES, FS, and GS) that references a segment descriptor containing a clear present bit (descriptor.P=0). Far transfer to a CS that is not present. Referencing a gate descriptor containing a clear present bit. Referencing a TSS descriptor containing a clear present bit. This includes attempts to load the TSS descriptor using the LTR instruction.
Chapter 8: Exceptions and Interrupts
257
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Attempting to load a descriptor containing a clear present bit into the LDTR using the LLDT instruction. Loading a segment selector (CS, DS, ES, FS, or GS) as part of a task switch, with the segment descriptor referenced by the segment selector having a clear present bit. In long mode, an #NP cannot be caused by a task switch, because the hardware task-switch mechanism is disabled. When loading a stack-segment selector (SS) that references a descriptor with a clear present bit, a stack exception (#SS) occurs. For information on the #SS exception, see the next section, "#SS--Stack Exception (Vector 12)." #NP cannot be disabled. Error Code Returned. The segment-selector index of the segment descriptor causing the #NP exception. Program Restart. #NP is a fault-type exception. In most cases, the saved instruction pointer points to the instruction that loaded the segment selector resulting in the #NP. See "Exceptions During a Task Switch" on page 267 for a description of the consequences when this exception occurs during a task switch. 8.2.13 #SS--Stack Exception (Vector 12) An #SS exception can occur in the following situations: Implied stack references in which the stack address is not in canonical form. Implied stack references include all push and pop instructions, and any instruction using RSP or RBP as a base register. Attempting to load a stack-segment selector that references a segment descriptor containing a clear present bit (descriptor.P=0). Any stack access that fails the stack-limit check. #SS cannot be disabled. Error Code Returned. The error code depends on the cause of the #SS, as shown in Table 8-5 on page 259:
258
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 8-5.
Stack Exception Error Codes
Stack Exception Cause Error Code SS Selector Index 0 0
Stack-segment descriptor present bit is clear Stack-limit violation Stack reference using a non-canonical address
Program Restart. #SS is a fault-type exception. In most cases, the saved instruction pointer points to the instruction that caused the #SS. See "Exceptions During a Task Switch" on page 267 for a description of the consequences when this exception occurs during a task switch. 8.2.14 #GP--GeneralProtection Exception (Vector 13) Table 8-6 on page 260 describes the general situations that can cause a #GP exception. The table is not an exhaustive, detailed list of #GP conditions, but rather a guide to the situations that can cause a #GP. If an invalid use of an AMD64 architectural feature results in a #GP, the specific cause of the exception is described in detail in the section describing the architectural feature. #GP cannot be disabled. Error Code Returned. As shown in Table 8-6, a selector index is reported as the error code if the #GP is due to a segmentdescriptor access. In all other cases, an error code of 0 is returned. Program Restart. #GP is a fault-type exception. In most cases, the saved instruction pointer points to the instruction that caused the #GP. See "Exceptions During a Task Switch" on page 267 for a description of the consequences when this exception occurs during a task switch.
Chapter 8: Exceptions and Interrupts
259
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 8-6.
General-Protection Exception Conditions
Error Condition Error Code
Any segment privilege-check violation. Any segment limit-check violation, except for the SS selector. Any segment type-check violation, including read, write, and execute violations. Accessing a long mode gate-descriptor containing a non-canonical address. Accessing a 64-bit mode system-descriptor containing a non-canonical address. Loading a non-canonical base address into the GDTR or IDTR in 64-bit mode. Loading the CS or SS register with a null segment selector. Accessing a gate descriptor containing a null segment selector. Accessing memory with a null DS, ES, FS, or GS data-segment selector. In long mode, referencing any 16-bit gate descriptor. In long mode, referencing a call gate or interrupt gate that does not point to a 64-bit code segment. In long mode, referencing any system descriptor that does not have the second type field loaded with 0. Attempting a control transfer to a TSS or task gate in long mode. Referencing an LDT descriptor or TSS descriptor located in an LDT. Control transfer to a busy TSS (except IRET). Using the LTR instruction to load a busy TSS descriptor into the TR register. Referencing virtual memory with a non-canonical address (except stack references). Any long-mode consistency-check violation. Using WRMSR to write a non-canonical address into the FS.base or GS.base register. Using WRMSR to write a non-canonical address into the KernelGSbase register. Using WRMSR to write a non-canonical address into the CSTAR or LSTAR register. Using WRMSR to write read-only MSRs. Writing invalid type encodings into the MTRRs or the PAT MSR. Writing a 1 into any register field that is reserved, must be zero (MBZ). Using a 128-bit media instruction requiring 16-byte alignment with a memory operand not aligned on a 16-byte boundary. 0 Selector Index
260
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 8-6.
General-Protection Exception Conditions (continued)
Error Condition Error Code
Executing privileged instructions when CPL > 0. Enabling paging when protected mode is disabled. Setting CR0.NW=1 while CR0.CD=0. 0
8.2.15 #PF--Page-Fault Exception (Vector 14)
A #PF exception can occur during a memory access in any of the following situations: A page-translation-table entry or physical page involved in translating the memory access is not present in physical memory. This is indicated by a cleared present bit (P=0) in the translation-table entry. An attempt is made by the processor to load the instruction TLB with a translation for a non-executable page. The memory access fails the paging-protection checks (user/supervisor, read/write, or both). A reserved bit in one of the page-translation-table entries is set to 1. A #PF occurs for this reason only when CR4.PSE=1 or CR4.PAE=1. #PF cannot be disabled. CR2 Register. The virtual (linear) address that caused the #PF is stored in the CR2 register. Figure 8-1 shows the legacy 32-bit CR2 register. The CR2 register in the AMD64 architecture is 64 bits long, as shown in Figure 8-1. In AMD64 implementations, only the low-order 32 bits of CR2 can be used in legacy mode. The processor clears the high-order 32 bits of CR2 when it is written in legacy mode either by software or during a page fault.
63 Page-Fault Virtual Address
0
Figure 8-1.
Control Register 2 (CR2) Error Code Returned. The page-fault error code is pushed onto the page-fault exception-handler stack. See "Page-Fault Error Code" on page 268 for a description of this error code.
Chapter 8: Exceptions and Interrupts
261
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Program Restart. #PF is a fault-type exception. In most cases, the saved instruction pointer points to the instruction that caused the #PF. See "Exceptions During a Task Switch" on page 267 for a description of what can happen if this exception occurs during a task switch. 8.2.16 #MF--x87 Floating-Point Exception-Pending (Vector 16) The #MF exception is used to handle unmasked x87 floatingpoint exceptions. An #MF occurs when all of the following conditions are true: CR0.NE=1. An unmasked x87 floating-point exception is pending. This is indicated by an exception bit in the x87 floating-point status-word register being set to 1 The corresponding mask bit in the x87 floating-point control-word register is cleared to 0. The FWAIT/WAIT instruction or any waiting floating-point instruction is executed. If there is an exception mask bit (in the FPU control word) set, the exception is not reported. Instead, the x87-instruction unit responds in a default manner and execution proceeds normally. The x87 floating-point exceptions reported by the #MF exception are (including mnemonics): IE--Invalid-operation exception (also called #I), which is either: - IE alone--Invalid arithmetic-operand exception (also called #IA), or - SF and IE together--x87 Stack-fault exception (also called #IS). DE--Denormalized-operand exception (also called #D). ZE--Zero-divide exception (also called #Z). OE--Overflow exception (also called #O). UE--Underflow exception (also called #U). PE--Precision exception (also called #P or inexact-result exception). Error Code Returned. None. Exception information is provided by the x87 status-word registe r. See "x87 Floating-Point Programming" in Volume 1 for more information on using this register. 262 Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Program Restart. # M F i s a fa u l t - t y p e ex c e p t i o n . Th e # M F exception is not precise, because multiple instructions and exceptions can occur before the #MF handler is invoked. Also, the saved instruction pointer does not point to the instruction that caused the exception resulting in the #MF. Instead, the saved instruction pointer points to the x87 floating-point instruction or FWAIT/WAIT instruction that is about to be executed when the #MF occurs. The address of the last instruction that caused an x87 floating-point exception is in the x87 instruction-pointer register. See "x87 Floating-Point Programming" in Volume 1 for information on accessing this register. Masking. Each type of x87 floating-point exception can be masked by setting the appropriate bits in the x87 control-word register. See "x87 Floating-Point Programming" in Volume 1 for more information on using this register. #MF can also be disabled by clearing the CR0.NE bit to 0. See "Numeric Error (NE) Bit" on page 57 for more information on using this bit. 8.2.17 #AC-- Alignment-Check Exception (Vector 17) An #AC exception occurs when an unaligned-memory data reference is performed while alignment checking is enabled. After a processor reset, #AC exceptions are disabled. Software enables the #AC exception by setting the following register bits: CR0.AM=1. RFLAGS.AC=1. When the above register bits are set, an #AC can occur only when CPL=3. #AC never occurs when CPL < 3. Table 8-7 on page 264 lists the data types and the alignment boundary required to avoid an #AC exception when the mechanism is enabled.
Chapter 8: Exceptions and Interrupts
263
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 8-7. Data-Type Alignment
Supported Data Type Word Doubleword Quadword Bit string 128-bit media 64-bit media Segment selector 32-bit near pointer 32-bit far pointer 48-bit far pointer x87 Floating-point single-precision x87 Floating-point double-precision x87 Floating-point extended-precision x87 Floating-point save areas Required Alignment (Byte Boundary) 2 4 8 2, 4 or 8 (depends on operand size) 16 8 2 4 2 4 4 8 8 2 or 4 (depends on operand size)
Error Code Returned. Zero. Program Restart. #AC is a fault-type exception. The saved instruction pointer points to the instruction that caused the #AC. 8.2.18 #MC--MachineCheck Exception (Vector 18) T h e # M C e x c e p t i o n i s m o d e l s p e c i f i c . P ro c e s s o r implementations are not required to support the #MC exception, and those implementations that do support #MC can vary in how the #MC exception mechanism works. The exception is enabled by setting CR4.MCE to 1. The machine-check architecture can include model-specific masking for controlling the reporting of some errors. Refer to Chapter 9, "Machine Check Mechanism," for more information.
264
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Error Code Returned. None. Error information is provided by model-specific registers (MSRs) defined by the machine-check architecture. Program Restart. #MC is an abort-type exception. There is no reliable way to restart the program. If the EIPV flag (EIP valid) is set in the MCG_Status MSR, the saved CS and rIP point to the instruction that caused the error. If EIP is clear, the CS:rIP of the instruction causing the failure is not known or the machine check is not related to a specific instruction. 8.2.19 #XF--SIMD Floating-Point Exception (Vector 19) The #XF exception is used to handle unmasked 128-bit media floating-point exceptions. A #XF exception occurs when all of the following conditions are true: A 128-bit media floating-point exception occurs. The exception causes the processor to set the appropriate exception-status bit in the MXCSR register to 1. The exception-mask bit in the MXCSR that corresponds to the 128-bit media floating-point exception is clear (=0). CR4.OSXMMEXCPT=1, indicating that the operating system supports handling of 128-bit media floating-point exceptions. The exception-mask bits are used by software to specify the handling of 128-bit media floating-point exceptions. When the corresponding mask bit is cleared to 0, an exception occurs under the control of the CR4.OSXMMEXCPT bit. However, if the mask bit is set to 1, the 128-bit media floating-point unit responds in a default manner and execution proceeds normally. The CR4.OSXMMEXCPT bit specifies the interrupt vector to be taken when an unmasked 128-bit media floating-point exception occurs. When CR4.OSXMMEXCPT=1, the #XF interrupt vector is taken when an exception occurs. When CR4.OSXMMEXCPT=0, the #UD (undefined opcode) interrupt vector is taken when an exception occurs. The 128-bit media floating-point exceptions reported by the #XF exception are (including mnemonics): IE--Invalid-operation exception (also called #I). DE--Denormalized-operand exception (also called #D). ZE--Zero-divide exception (also called #Z). OE--Overflow exception (also called #O). Chapter 8: Exceptions and Interrupts 265
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
UE--Underflow exception (also called #U). PE--Precision exception (also called #P or inexact-result exception). Each type of 128-bit media floating-point exception can be masked by setting the appropriate bits in the MXCSR register. #XF can also be disabled by clearing the CR4.OSXMMEXCPT bit to 0. Error Code Returned. None. Exception information is provided by the 128-bit media floating-point MXCSR register. See "128-Bit Media and Scientific Programming" in Volume 1 for more information on using this register. Program Restart. #XF is a fault-type exception. Unlike the #MF exception, the #XF exception is precise. The saved instruction pointer points to the instruction that caused the #XF. 8.2.20 User-Defined Interrupts (Vectors 32-255) User-defined interrupts can be initiated either by system logic or software. They occur when: System logic signals an external interrupt request to the processor. The signalling mechanism and the method of communicating the interrupt vector to the processor are implementation dependent. Software executes an INTn instruction. The INTn instruction operand provides the interrupt vector number. Both methods can be used to initiate an interrupt into vectors 0 through 255. However, because vectors 0 through 31 are defined or reserved by the AMD64 architecture, software should not use vectors in this range for purposes other than their defined use. Error Code Returned. None. Program Restart. The saved instruction pointer depends on the interrupt source: External interrupts are recognized on instruction boundaries. The saved instruction pointer points to the instruction immediately following the boundary where the external interrupt was recognized. If the interrupt occurs as a result of executing the INTn instruction, the saved instruction pointer points to the instruction after the INTn.
266
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Masking. The ability to mask user-defined interrupts depends on the interrupt source: External interrupts can be masked using the rFLAGS.IF bit. Setting rFLAGS.IF to 1 enables external interrupts, while clearing rFLAGS.IF to 0 inhibits them. Software interrupts (initiated by the INTn instruction) cannot be disabled.
8.3
Exceptions During a Task Switch
An exception can occur during a task switch while loading a segment selector. Page faults can also occur when accessing a TSS. In these cases, the hardware task-switch mechanism completes loading the new task state from the TSS, and then triggers the appropriate exception mechanism. No other checks are performed. When this happens, the saved instruction pointer points to the first instruction in the new task. In long mode, an exception cannot occur during a task switch, because the hardware task-switch mechanism is disabled.
8.4
Error Codes
The processor exception-handling mechanism reports error and status information for some exceptions using an error code. The error code is pushed onto the stack by the exception-mechanism during the control transfer into the exception handler. The error code has two formats: a selector format for most errorreporting exceptions, and a page-fault format for page faults. These formats are described in the following sections.
8.4.1 Selector-Error Code
31
Figure 8-2 shows the format of the selector-error code.
16 Reserved
15 Selector Index
3
2
T I
1
I D T
0
E X T
Figure 8-2.
Selector Error Code
Chapter 8: Exceptions and Interrupts
267
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
The information reported by the selector-error code includes: EXT--Bit 0. If this bit is set to 1, the exception source is external to the processor. If cleared to 0, the exception source is internal to the processor. IDT--Bit 1. If this bit is set to 1, the error-code selector-index field references a gate descriptor located in the interruptdescriptor table (IDT). If cleared to 0, the selector-index field references a descriptor in either the global-descriptor table (GDT) or local-descriptor table (LDT), as indicated by the TI bit. TI--Bit 2. If this bit is set to 1, the error-code selector-index field references a descriptor in the LDT. If cleared to 0, the selector-index field references a descriptor in the GDT. The TI bit is relevant only when the IDT bit is cleared to 0. Selector Index--Bits 15-3. The selector-index field specifies the index into either the GDT, LDT, or IDT, as specified by the IDT and TI bits. Some exceptions return a zero in the selector-error code. 8.4.2 Page-Fault Error Code Figure 8-3 shows the format of the page-fault error code.
31 Reserved
4
I/D
3
R S V
2
1
0
P
U/S R/W
Figure 8-3.
Page-Fault Error Code The information reported by the page-fault error code includes: P--Bit 0. If this bit is cleared to 0, the page fault was caused by a not-present page. If this bit is set to 1, the page fault was caused by a page-protection violation. R/W--Bit 1. If this bit is cleared to 0, the access that caused the page fault is a memory read. If this bit is set to 1, the memory access that caused the page fault was a write. This bit does not necessarily indicate the cause of the page fault was a read or write violation. U/S--Bit 2. If this bit is cleared to 0, an access in supervisor mode (CPL=0, 1, or 2) caused the page fault. If this bit is set
268
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
to 1, an access in user mode (CPL=3) caused the page fault. This bit does not necessarily indicate the cause of the page fault was a privilege violation. RSV--Bit 3. If this bit is set to 1, the page fault is a result of the processor reading a 1 from a reserved field within a page-translation-table entry. This type of page fault occurs only when CR4.PSE=1 or CR4.PAE=1. If this bit is cleared to 0, the page fault was not caused by the processor reading a 1 from a reserved field. I/D--Bit 4. If this bit is set to 1, it indicates that the access that caused the page fault was an instruction fetch. Otherwise, this bit is cleared to 0. This bit is only defined if no-execute feature is enabled (EFER.NXE=1 && CR4.PAE=1).
8.5
Priorities
To allow for consistent handling of multiple-interrupt conditions, simultaneous interrupts are prioritized by the processor. The AMD64 architecture defines priorities between groups of interrupts, and interrupt prioritization within a group is implementation dependent. Table 8-8 on page 270 shows the interrupt priorities defined by the AMD64 architecture. When simultaneous interrupts occur, the processor transfers control to the highest-priority interrupt handler. Lower-priority interrupts from external sources are held pending by the processor, and they are handled after the higher-priority interrupt is handled. Lower-priority interrupts that result from internal sources are discarded. Those interrupts reoccur when the high-priority interrupt handler completes and transfers control back to the interrupted instruction. Software interrupts are discarded as well, and reoccur when the software-interrupt instruction is restarted.
Chapter 8: Exceptions and Interrupts
269
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 8-8.
Interrupt Priority (High) 0
Simultaneous Interrupt Priorities
Interrupt Condition Processor Reset Machine-Check Exception External Processor Initialization (INIT) Interrupt Vector -- 18
1
SMI Interrupt External Clock Stop (Stpclk)
--
2 3 4
Data, and I/O Breakpoint (Debug Register) Single-Step Execution Instruction Trap (rFLAGS.TF=1) Non-Maskable Interrupt Maskable External Interrupt (INTR) Instruction Breakpoint (Debug Register)
1 2 32--255 1 13 14 6 7 13 10 11 12 13 14 16 17 19
5
Code-Segment-Limit Violation Instruction-Fetch Page Fault Invalid Opcode Exception
6
Device-Not-Available Exception Instruction-Length Violation (> 15 Bytes) Invalid-TSS Exception Segment-Not-Present Exception Stack Exception
7
General-Protection Exception Data-Access Page Fault Floating-Point Exception-Pending Exception Alignment-Check Exception SIMD Floating-Point Exception
270
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
8.5.1 Floating-Point Exception Priorities
Floating-point exceptions (128-bit media and x87 floatingpoint) can be handled in one of two ways: Unmasked exceptions are reported in the appropriate floating-point status register, and a software-interrupt handler is invoked. See "#MF--x87 Floating-Point Exception-Pending (Vector 16)" on page 262 and "#XF-- SIMD Floating-Point Exception (Vector 19)" on page 265 for more information on the floating-point interrupts. Masked exceptions are also reported in the appropriate floating-point status register. Instead of transferring control to an interrupt handler, however, the processor handles the exception in a default manner and execution proceeds. If the processor detects more than one exception while executing a single floating-point instruction, it prioritizes the exceptions in a predictable manner. When responding in a default manner to masked exceptions, it is possible that the processor acts only on the high-priority exception and ignores lower-priority exceptions. In the case of vector (SIMD) floatingpoint instructions, priorities are set on sub-operations, not across all operations. For example, if the processor detects and acts on a QNaN operand in one sub-operation, the processor can still detect and act on a denormal operand in another suboperation. When reporting 128-bit media floating-point exceptions before taking an interrupt or handling them in a default manner, the processor first classifies the exceptions as follows: Input exceptions include SNaN operand (#I), invalid operation (#I), denormal operand (#D), or zero-divide (#Z). Using a NaN operand with a maximum, minimum, compare, or convert instruction is also considered an input exception. Output exceptions include numeric overflow (#O), numeric underflow (#U), and precision (#P). Using the above classification, the processor applies the following procedure to report the exceptions: 1. The exceptions for all sub-operations are prioritized. 2. The exception conditions for all sub-operations are logically ORed together to form a single set of exceptions covering all operations. For example, if two sub-operations produce a denormal result, only one denormal exception is reported.
Chapter 8: Exceptions and Interrupts
271
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
3. If the set of exceptions includes any unmasked input exceptions, all input exceptions are reported in MCXSR, and no output exceptions are reported. Otherwise, all input and output exceptions are reported in MCXSR. 4. If any exceptions are unmasked, control is transferred to the appropriate interrupt handler. Table 8-9 lists the priorities for simultaneous floating-point exceptions. Table 8-9.
Exception Priority SNaN Operand (High) 0 NaN Operand of Maximum, Minimum, Compare, and Convert Instructions (Vector Floating-Point) Stack Overflow (x87 Floating-Point) Stack Underflow (x87 Floating-Point) 1 2 3 4 5 (Low) QNaN Operand Invalid Operation (Remaining Conditions) Zero Divide Denormal Operand Numeric Overflow Numeric Underflow Precision -- #I #Z #D #O #U #P #I
Simultaneous Floating-Point Exception Priorities
Exception Condition
8.5.2 External Interrupt Priorities
The AMD64 architecture allows software to define up to 15 external interrupt-priority classes. Priority classes are numbered from 1 to 15, with priority-class 1 being the lowest and priority-class 15 the highest. The organization of these priority classes is implementation dependent. A typical method is to use the upper four bits of the interrupt vector number to define the priority. Thus, interrupt vector 53h has a priority of 5 and interrupt vector 37h has a priority of 3.
272
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
A new control register (CR8) is introduced by the AMD64 architecture for managing priority classes. This register, called the task-priority register (TPR), uses its four low-order bits to specify a task priority. The remaining 60 bits are reserved and must be written with zeros. Figure 8-4 shows the format of the TPR. The TPR is available only in 64-bit mode.
63 Reserved, MBZ
43
0
Task Priority (TPR)
Figure 8-4.
Task Priority Register (CR8) System software can use the TPR register to temporarily block low-priority interrupts from interrupting a high-priority task. T h i s i s a c c o m p l i s h e d by l o a d i n g T P R w i t h a va l u e corresponding to the highest-priority interrupt that is to be blocked. For example, loading TPR with a value of 9 (1001b) blocks all interrupts with a priority class of 9 or less, while allowing all interrupts with a priority class of 10 or more to be recognized. Loading TPR with 0 enables all external interrupts. Loading TPR with 15 (1111b) disables all external interrupts. The TPR is cleared to 0 on reset. System software reads and writes the TPR using a MOV CR8 instruction. The MOV CR8 instruction requires a privilege level of 0. Programs running at any other privilege level cannot read or write the TPR, and an attempt to do so results in a generalprotection exception (#GP). A serializing instruction is not required after loading the TPR, because a new priority level is established when the MOV instruction completes execution. For example, assume two sequential TPR loads are performed, in which a low value is first loaded into TPR and immediately followed by a load of a higher value. Any pending, lower-priority interrupt enabled by the first MOV CR8 is recognized between the two MOVs. The TPR is an architectural abstraction of the interrupt controller (IC), which prioritizes and manages external interrupt delivery to the processor. The IC can be an external
Chapter 8: Exceptions and Interrupts
273
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
system device, or it can be integrated on the chip like the local advanced programmable interrupt controller (APIC). Typically, the IC contains a priority mechanism similar, if not identical to, the TPR. The IC, however, is implementation dependent, and the underlying priority mechanisms are subject to change. The TPR, by contrast, is part of the AMD64 architecture. Effect of IC on TPR. The features of the implementation-specific IC can impact the operation of the TPR. For example, the TPR might affect interrupt delivery only if the IC is enabled. Also, the mapping of an external interrupt to a specific interrupt priority is an implementation-specific behavior of the IC.
8.6
Real-Mode Interrupt Control Transfers
In real mode, the IDT is a table of 4-byte entries, one entry for each of the 256 possible interrupts implemented by the system. The real mode IDT is often referred to as an interrupt-vector table, or IVT. Table entries contain a far pointer (CS:IP pair) to an exception or interrupt handler. The base of the IDT is stored in the IDTR register, which is loaded with a value of 00h during a processor reset. Figure 8-5 shows how the real-mode interrupt handler is located by the interrupt mechanism.
Memory Interrupt-Descriptor Table
Interrupt Vector * +
CS Offset Interrupt Handler
4
IDT Base Address
Interrupt-Descriptor-Table Register
513-239.eps
Figure 8-5.
Real-Mode Interrupt Control Transfer
274
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
When an exception or interrupt occurs in real mode, the processor performs the following: 1. Pushes the FLAGS register (EFLAGS[15:0]) onto the stack. 2. Clears EFLAGS.IF to 0 and EFLAGS.TF to 0. 3. Saves the CS register and IP register (RIP[15:0]) by pushing them onto the stack. 4. Locates the interrupt-handler pointer (CS:IP) in the IDT by scaling the interrupt vector by four and adding the result to the value in the IDTR. 5. Transfers control to the interrupt handler referenced by the CS:IP in the IDT. Figure 8-6 shows the stack after control is transferred to the interrupt handler in real mode.
Interrupt-Handler and Interrupted-Program Stack
Return FLAGS Return CS Return IP
+4 +2 SS:SP
513-243.eps
Figure 8-6.
Stack After Interrupt in Real Mode
An IRET instruction is used to return to the interrupted program. When an IRET is executed, the processor performs the following: 1. Pops the saved CS value off the stack and into the CS register. The saved IP value is popped into RIP[15:0]. 2. Pops the FLAGS value off of the stack and into EFLAGS[15:0]. 3. Execution begins at the saved CS.IP location.
Chapter 8: Exceptions and Interrupts
275
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
8.7
Legacy Protected-Mode Interrupt Control Transfers
In protected mode, the interrupt mechanism transfers control to an exception or interrupt handler through gate descriptors. In protected mode, the IDT is a table of 8-byte gate entries, one for each of the 256 possible interrupt vectors implemented by the system. Three gate types are allowed in the IDT: Interrupt gates. Trap gates. Task gates. If a reference is made to any other descriptor type in the IDT, a general-protection exception (#GP) occurs. Interrupt-gate control transfers are similar to CALLs and JMPs through call gates. The interrupt mechanism uses gates (interrupt, trap, and task) to establish protected entry-points into the exception and interrupt handlers. The remainder of this chapter discusses control transfers through interrupt gates and trap gates. If the gate descriptor in the IDT is a task gate, a TSS-segment selector is referenced, and a task switch occurs. See Chapter 12, "Task Management," for more information on the hardware task-switch mechanism.
8.7.1 Locating the Interrupt Handler
When an exception or interrupt occurs, the processor scales the interrupt-vector number by eight and uses the result as an offset into the IDT. If the gate descriptor referenced by the IDT offset is an interrupt gate or a trap gate, it contains a segmentselector and segment-offset field (see "Legacy Segment Descriptors" on page 97 for a detailed description of the gatedescriptor format and fields). These two fields perform the same function as the pointer operand in a far control-transfer instruction. The gate-descriptor segment-selector field points to the target code-segment descriptor located in either the GDT or LDT. The gate-descriptor segment-offset field is the instruction-pointer offset into the interrupt-handler code segment. The code-segment base taken from the code-segment descriptor is added to the gate-descriptor segment-offset field to create the interrupt-handler virtual address (linear address). Figure 8-7 on page 277 shows how the protected-mode interrupt handler is located by the interrupt mechanism.
276
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Interrupt Descriptor Table
CS Selector
DPL + *
Interrupt Vector
Code-Segment Offset
8
IDT Base Address
IDT Limit
Interrupt-Descriptor-Table Register Virtual-Address Space
Global or Local Descriptor Table
+
Interrupt Handler
CS Limit
DPL
Code-Segment Base
Code Segment
513-240.eps
Figure 8-7.
Protected-Mode Interrupt Control Transfer When a control transfer to an exception or interrupt handler at the same privilege level occurs (through an interrupt gate or a trap gate), the processor performs the following: 1. Pushes the EFLAGS register onto the stack. 2. Clears the TF, NT, RF, and VM bits in EFLAGS to 0.
8.7.2 Interrupt To Same Privilege
Chapter 8: Exceptions and Interrupts
277
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
3. The processor handles EFLAGS.IF based on the gatedescriptor type: - If the gate descriptor is an interrupt gate, EFLAGS.IF is cleared to 0. - If the gate descriptor is a trap gate, EFLAGS.IF is not modified. 4. Saves the return CS register and EIP register (RIP[31:0]) by pushing them onto the stack. The CS value is padded with two bytes to form a doubleword. 5. If the interrupt has an associated error code, the error code is pushed onto the stack. 6. The CS register is loaded from the segment-selector field in the gate descriptor, and the EIP is loaded from the offset field in the gate descriptor. 7. The interrupt handler begins executing with the instruction referenced by new CS:EIP. Figure 8-8 shows the stack after control is transferred to the interrupt handler.
With Error Code
Interrupt-Handler and Interrupted Program Stack
With No Error Code
Return EFLAGS
+12 Return EFLAGS SS:ESP +8
Return CS +8 Return EIP Error Code +4
Return CS +4 Return EIP SS:EIP
513-242.eps
Figure 8-8. 8.7.3 Interrupt To Higher Privilege
Stack After Interrupt to Same Privilege Level
When a control transfer to an exception or interrupt handler running at a higher privilege occurs (numerically lower CPL value), the processor performs a stack switch using the following steps:
278
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
1. The target CPL is read by the processor from the target code-segment DPL and used as an index into the TSS for selecting the new stack pointer (SS:ESP). For example, if the target CPL is 1, the processor selects the SS:ESP for privilege-level 1 from the TSS. 2. Pushes the return stack pointer (old SS:ESP) onto the new stack. The SS value is padded with two bytes to form a doubleword. 3. Pushes the EFLAGS register onto the new stack. 4. Clears the following EFLAGS bits to 0: TF, NT, RF, and VM. 5. The processor handles the EFLAGS.IF bit based on the gate-descriptor type: - If the gate descriptor is an interrupt gate, EFLAGS.IF is cleared to 0. - If the gate descriptor is a trap gate, EFLAGS.IF is not modified. 6. Saves the return-address pointer (CS:EIP) by pushing it onto the stack. The CS value is padded with two bytes to form a doubleword. 7. If the interrupt-vector number has an error code associated with it, the error code is pushed onto the stack. 8. The CS register is loaded from the segment-selector field in the gate descriptor, and the EIP is loaded from the offset field in the gate descriptor. 9. The interrupt handler begins executing with the instruction referenced by new CS:EIP. Figure 8-9 on page 280 shows the new stack after control is transferred to the interrupt handler.
Chapter 8: Exceptions and Interrupts
279
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Interrupt-Handler Stack With Error Code With No Error Code
Return SS Return ESP Return EFLAGS Return CS Return EIP Error Code
+20 +16 +12 +8 +4 New SS:ESP Return SS Return ESP Return EFLAGS Return CS Return EIP +16 +12 +8 +4 ESS:ESP
513-241.eps
Figure 8-9. 8.7.4 Privilege Checks
Stack After Interrupt to Higher Privilege
Before loading the CS register with the interrupt-handler codesegment selector (located in the gate descriptor), the processor performs privilege checks similar to those performed on call gates. The checks are performed when either conforming or nonconforming interrupt handlers are referenced: 1. The processor reads the gate DPL from the interrupt-gate or trap-gate descriptor. The gate DPL is the minimum privilegelevel (numerically-highest value) needed by a program to access the gate. The processor compares the CPL with the gate DPL. The CPL must be numerically less-than or equal-to the gate DPL for this check to pass. 2. The processor compares the CPL with the interrupt-handler code-segment DPL. For this check to pass, the CPL must be numerically greater-than or equal-to the code-segment DPL. This check prevents control transfers to less-privileged interrupt handlers. Unlike call gates, no RPL comparison takes place. This is because the gate descriptor is referenced in the IDT using the interrupt-vector number rather than a selector, and no RPL field exists in the interrupt-vector number.
280
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Exception and interrupt handlers should be made reachable from software running at any privilege level that requires them. If the gate DPL value is too low (requiring more privilege), or the interrupt-handler code-segment DPL is too high (runs at lower privilege), the interrupt control transfer can fail the privilege checks. Setting the gate DPL=3 and interrupt-handler code-segment DPL=0 makes the exception handler or interrupt handler reachable from any privilege level. Figure 8-10 on page 282 shows two examples of interrupt privilege checks. In Example 1, both privilege checks pass: The interrupt-gate DPL is at the lowest privilege (3), which means that software running at any privilege level (CPL) can access the interrupt gate. The interrupt-handler code segment is at the highestprivilege level, as indicated by DPL=0. This means software running at any privilege can enter the interrupt handler through the interrupt gate.
Chapter 8: Exceptions and Interrupts
281
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
CS
CPL=2
Interrupt Vector DPL=3 Gate Descriptor DPL=0 Code Descriptor
Access Allowed
?
Access Allowed Interrupt Handler
Access Allowed
Example 1: Privilege Check Passes
CS
CPL=2
Interrupt Vector DPL=0 Gate Descriptor DPL=3 Code Descriptor
Access Denied
?
Access Denied Interrupt Handler
Access Denied
Example 2: Privilege Check Fails
513-244.eps
Figure 8-10.
Privilege-Check Examples for Interrupts In Example 2, both privilege checks fail: The interrupt-gate DPL specifies that only software running at privilege-level 0 can access the gate. The current program does not have a high enough privilege level to access the interrupt gate, since its CPL is set at 2.
282
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The interrupt handler has a lower privilege (DPL=3) than the currently-running software (CPL=2). Transitions from more-privileged software to less-privileged software are not allowed, so this privilege check fails as well. Although both privilege checks fail, only one such failure is required to deny access to the interrupt handler. 8.7.5 Returning From Interrupt Procedures A return to an interrupted program should be performed using the IRET instruction. An IRET is a far return to a different code segment, with or without a change in privilege level. The actions of an IRET in both cases are described in the following sections. IRET, Same Privilege. Before performing the IRET, the stack pointer must point to the return EIP. If there was an error code pushed onto the stack as a result of the exception or interrupt, that error code should have been popped off the stack earlier by the handler. The IRET reverses the actions of the interrupt mechanism: 1. Pops the return pointer off of the stack, loading both the CS register and EIP register (RIP[31:0]) with the saved values. The return code-segment RPL is read by the processor from the CS value stored on the stack to determine that an equalprivilege control transfer is occurring. 2. Pops the saved EFLAGS image off of the stack and into the EFLAGS register. 3. Transfers control to the return program at the target CS:EIP. IRET, Less Privilege. If an IRET changes privilege levels, the return program must be at a lower privilege than the interrupt handler. The IRET in this case causes a stack switch to occur: 1. The return pointer is popped off of the stack, loading both the CS register and EIP register (RIP[31:0]) with the saved values. The return code-segment RPL is read by the processor from the CS value stored on the stack to determine that a lower-privilege control transfer is occurring. 2. The saved EFLAGS image is popped off of the stack and loaded into the EFLAGS register.
Chapter 8: Exceptions and Interrupts
283
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
3. The return-program stack pointer is popped off of the stack, loading both the SS register and ESP register (RSP[31:0]) with the saved values. 4. Control is transferred to the return program at the target CS:EIP.
8.8
Virtual-8086 Mode Interrupt Control Transfers
This section describes interrupt control transfers as they relate to virtual-8086 mode. Virtual-8086 mode is not supported by long mode. Therefore, the control-transfer mechanism described here is not applicable to long mode. When a software interrupt occurs (not external interrupts, INT1, or INT3) while the processor is running in virtual-8086 mode (EFLAGS.VM=1), the control transfer that occurs depends on three system controls: EFLAGS.IOPL--This field controls interrupt handling based on the CPL. See "I/O Privilege Level Field (IOPL) Field" on page 66 for more information on this field. Setting IOPL<3 redirects the interrupt to the generalprotection exception (#GP) handler. CR4.VME--This bit enables virtual-mode extensions. See "Virtual-8086 Mode Extensions (VME) Bit" on page 61 for more information on this bit. TSS Interrupt-Redirection Bitmap--The TSS interruptredirection bitmap contains 256 bits, one for each possible INTn vector (software interrupt). When CR4.VME=1, the bitmap is used by the processor to direct interrupts to the handler provided by the currently-running 8086 program (bitmap entry is 0), or to the protected-mode operatingsystem interrupt handler (bitmap entry is 1). See "Legacy Task-State Segment" on page 365 for information on the location of this field within the TSS. If IOPL<3, CR4.VME=1, and the corresponding interrupt redirection bitmap entry is 0, the processor uses the virtualinterrupt mechanism. See "Virtual Interrupts" on page 295 for more information on this mechanism. Table 8-10 on page 285 summarizes the actions of the above system controls on interrupts taken when the processor is running in virtual-8086 mode.
284
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 8-10.
Virtual-8086 Mode Interrupt Mechanisms
CR4.VME 0 TSS Interrupt Redirection Bitmap Entry -- 1 0 -- 1 0 Interrupt Mechanism
EFLAGS.IOPL
0, 1, or 2
1 1 0
General-Protection Exception Virtual Interrupt Protected-Mode Handler Virtual 8086 Handler
3
1 1
8.8.1 Protected-Mode Handler Control Transfer
Control transfers to protected-mode handlers from virtual-8086 mode differ from standard protected-mode transfers in several ways. The processor follows these steps in making the control transfer: 1. Reads the CPL=0 stack pointer (SS:ESP) from the TSS. 2. Pushes the GS, FS, DS, and ES selector registers onto the stack. Each push is padded with two bytes to form a doubleword. 3. Clears the GS, FS, DS, and ES selector registers to 0. This places a null selector in each of the four registers 4. Pushes the return stack pointer (old SS:ESP) onto the new stack. The SS value is padded with two bytes to form a doubleword. 5. Pushes the EFLAGS register onto the new stack. 6. Clears the following EFLAGS bits to 0: TF, NT, RF, and VM. 7. Handles EFLAGS.IF based on the gate-descriptor type: - If the gate descriptor is an interrupt gate, EFLAGS.IF is cleared to 0. - If the gate descriptor is a trap gate, EFLAGS.IF is not modified. 8. Pushes the return-address pointer (CS:EIP) onto the stack. The CS value is padded with two bytes to form a doubleword.
Chapter 8: Exceptions and Interrupts
285
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
9. If the interrupt has an associated error code, pushes the error code onto the stack. 10. Loads the segment-selector field from the gate descriptor into the CS register, and loads the offset field from the gate descriptor into the EIP register. 11. Begins execution of the interrupt handler with the instruction referenced by the new CS:EIP. Figure 8-11 shows the new stack after control is transferred to the interrupt handler with an error code.
Interrupt-Handler Stack With No Error Code
With Error Code
Return GS +36 Return FS +32 Return GS +32 Return FS Return ES Return SS Return ESP Return EFLAGS +28
Return DS +28 Return ES +24 Return SS +20 Return ESP Return EFLAGS +16 +12
Return DS +24 +20 +16 +12 +8
Return CS +8 Return EIP Error Code +4 New SS:ESP (From TSS, CPL=0)
Return CS +4 Return EIP SS:ESP
513-249.eps
Figure 8-11. Stack After Virtual-8086 Mode Interrupt to Protected Mode An IRET from privileged protected-mode software (CPL=0) to virtual-8086 mode reverses the stack-build process. After the return pointer, EFLAGS, and return stack-pointer are restored, the processor restores the ES, DS, FS, and GS registers by popping their values off the stack. 8.8.2 Virtual-8086 Handler Control Transfer When a control transfer to an 8086 handler occurs from virtual8086 mode, the processor creates an interrupt-handler stack identical to that created when an interrupt occurs in real mode (see Figure 8-6 on page 275). The processor performs the following actions during a control transfer: Chapter 8: Exceptions and Interrupts
286
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
1. Pushes the FLAGS register (EFLAGS[15:0]) onto the stack. 2. Clears the EFLAGS.IF and EFLAGS.RF bits to 0. 3. Saves the CS register and IP register (RIP[15:0]) by pushing them onto the stack. 4. Locates the interrupt-handler pointer (CS:IP) in the 8086 IDT by scaling the interrupt vector by four and adding the result to the virtual (linear) address 0. The value in the IDTR is not used. 5. Transfers control to the interrupt handler referenced by the CS:IP in the IDT. An IRET from the 8086 handler back to virtual-8086 mode reverses the stack-build process.
8.9
Long-Mode Interrupt Control Transfers
The long-mode architecture expands the legacy interruptm e ch a n i s m t o s u p p o r t 6 4 - b i t o p e ra t i n g s y s t e m s a n d applications. These changes include: All interrupt handlers are 64-bit code and operate in 64-bit mode. The size of an interrupt-stack push is fixed at 64 bits (8 bytes). The interrupt-stack frame is aligned on a 16-byte boundary. The stack pointer, SS:RSP, is pushed unconditionally on interrupts, rather than conditionally based on a change in CPL. The SS selector register is loaded with a null selector as a result of an interrupt, if the CPL changes. The IRET instruction behavior changes, to unconditionally pop SS:RSP, allowing a null SS to be popped. A new interrupt stack-switch mechanism, called the interrupt-stack table or IST, is introduced.
8.9.1 Interrupt Gates and Trap Gates
Only long-mode interrupt and trap gates can be referenced in long mode (64-bit mode and compatibility mode). The legacy 32-bit interrupt-gate and 32-bit trap-gate types (0Eh and 0Fh, as described in "System Descriptors" on page 111) are redefined in long mode as 64-bit interrupt-gate and 64-bit trap-gate types. 32-bit and 16-bit interrupt-gate and trap-gate types do not exist 287
Chapter 8: Exceptions and Interrupts
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
in long mode, and software is prohibited from using task gates. If a reference is made to any gate other than a 64-bit interrupt gate or a 64-bit trap gate, a general-protection exception (#GP) occurs. The long-mode gate types are 16 bytes (128 bits) long. They are an extension of the legacy-mode gate types, allowing a full 64bit segment offset to be stored in the descriptor. See "Legacy Segment Descriptors" on page 97 for a detailed description of the gate-descriptor format and fields. 8.9.2 Locating the Interrupt Handler When an interrupt occurs in long mode, the processor multiplies the interrupt-vector number by 16 and uses the result as an offset into the IDT. The gate descriptor referenced by the IDT offset contains a segment-selector and a 64-bit segment-offset field. The gate-descriptor segment-offset field contains the complete virtual address for the interrupt handler. The gate-descriptor segment-selector field points to the target code-segment descriptor located in either the GDT or LDT. The code-segment descriptor is only used for privilege-checking purposes and for placing the processor in 64-bit mode. The code segment-descriptor base field, limit field, and most attributes are ignored. Figure 8-12 on page 289 shows how the long-mode interrupt handler is located by the interrupt mechanism.
288
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Interrupt-Descriptor Table
Code-Segment Offset CS Selector DPL + *
Interrupt Vector
16
IDT Base Address
IDT Limit
Global- or LocalDescriptor Table
Interrupt-Descriptor-Table Register Virtual-Address Space
CS Limit
DPL
Code-Segment Base
Interrupt Handler
513-245.eps
Figure 8-12.
Long-Mode Interrupt Control Transfer In long mode, the return-program stack pointer (SS:RSP) is always pushed onto the interrupt-handler stack, regardless of whether or not a privilege change occurs. Although the SS register is not used in 64-bit mode, SS is pushed to allow returns into compatibility mode. Pushing SS:RSP unconditionally presents operating systems with a consistent interrupt-stackframe size for all interrupts, except for error codes. Interrupt service-routine entry points that handle interrupts generated by non-error-code interrupts can push an error code on the stack for consistency.
8.9.3 Interrupt Stack Frame
Chapter 8: Exceptions and Interrupts
289
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
In long mode, when a control transfer to an interrupt handler occurs, the processor performs the following: 1. Aligns the new interrupt-stack frame by masking RSP with FFFF_FFFF_FFFF_FFF0h. 2. If IST field in interrupt gate is not 0, reads IST pointer into RSP. 3. If a privilege change occurs, the target DPL is used as an index into the long-mode TSS to select a new stack pointer (RSP). 4. If a privilege change occurs, SS is cleared to zero indicating a null selector. 5. Pushes the return stack pointer (old SS:RSP) onto the new stack. The SS value is padded with six bytes to form a quadword. 6. Pushes the 64-bit RFLAGS register onto the stack. The upper 32 bits of the RFLAGS image on the stack are written as zeros. 7. Clears the TF, NT, and RF bits in RFLAGS bits to 0. 8. Handles the RFLAGS.IF bit according to the gate-descriptor type: - If the gate descriptor is an interrupt gate, RFLAGS.IF is cleared to 0. - If the gate descriptor is a trap gate, RFLAGS.IF is not modified. 9. Pushes the return CS register and RIP register onto the stack. The CS value is padded with six bytes to form a quadword. 10. If the interrupt-vector number has an error code associated with it, pushes the error code onto the stack. The error code is padded with four bytes to form a quadword. 11. Loads the segment-selector field from the gate descriptor into the CS register. The processor checks that the target code-segment is a 64-bit mode code segment. 12. Loads the offset field from the gate descriptor into the target RIP. The interrupt handler begins execution when control is transferred to the instruction referenced by the new RIP.
290
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Figure 8-13 shows the stack after control is transferred to the interrupt handler.
Interrupt-Handler Stack With Error Code With No Error Code
Return SS +40 Return RSP Return RFLAGS +32 +24 Return CS +16 Return RIP Error Code +8 RSP Return RIP Return RSP Return RFLAGS Return SS +32 +24 +16 Return CS +8 RSP
Figure 8-13.
Long-Mode Stack After Interrupt--Same Privilege
Interrupt-Stack Alignment. In legacy mode, the interrupt-stack pointer can be aligned at any address boundary. Long mode, however, aligns the stack on a 16-byte boundary. This alignment is performed by the processor in hardware before pushing items o n t o t h e s t a c k f ra m e . T h e p re v i o u s R S P i s s ave d unconditionally on the new stack by the interrupt mechanism. A subsequent IRET instruction automatically restores the previous RSP. Aligning the stack on a 16-byte boundary allows optimal performance for saving and restoring the 16-byte XMM registers. The interrupt handler can save and restore the XMM registers using the faster 16-byte aligned loads and stores (MOVAPS), rather than unaligned loads and stores (MOVUPS). Although the RSP alignment is always performed in long mode, it is only of consequence when the interrupted program is already running at CPL=0, and it is generally used only within the operating-system kernel. The operating system should put 16-byte aligned RSP values in the TSS for interrupts that change privilege levels.
Chapter 8: Exceptions and Interrupts
291
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Stack Switch. In long mode, the stack-switch mechanism differs slightly from the legacy stack-switch mechanism (see "Interrupt To Higher Privilege" on page 278). When stacks are switched during a long-mode privilege-level change resulting from an interrupt, a new SS descriptor is not loaded from the TSS. Long mode only loads an inner-level RSP from the TSS. However, the SS selector is loaded with a null selector, allowing nested control transfers, including interrupts, to be handled properly in 64-bit mode. The SS.RPL is set to the new CPL value. See "Nested IRETs to 64-Bit Mode Procedures" on page 295 for additional information. The interrupt-handler stack that results from a privilege change in long mode looks identical to a long-mode stack when no privilege change occurs. Figure 8-14 shows the stack after the switch is performed and control is transferred to the interrupt handler.
Interrupt-Handler Stack With Error Code Without Error Code
Return SS +40 Return RSP Return RFLAGS +32 +24 Return CS +16 Return RIP Error Code +8 New RSP (from TSS) SS=0 (if CPL changes) Return RIP Return RSP Return RFLAGS Return SS +32 +24 +16 Return CS +8 New RSP (from TSS) SS=0 (if CPL changes)
Figure 8-14. 8.9.4 Interrupt-Stack Table
Long-Mode Stack After Interrupt--Higher Privilege
In long mode, a new interrupt-stack table (IST) mechanism is introduced as an alternative to the modified legacy stack-switch mechanism described above. The IST mechanism provides a method for specific interrupts, such as NMI, double-fault, and machine-check, to always execute on a known-good stack. In legacy mode, interrupts can use the hardware task-switch Chapter 8: Exceptions and Interrupts
292
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
mechanism to set up a known-good stack by accessing the interrupt service routine through a task gate located in the IDT. However, the hardware task-switch mechanism is not supported in long mode. When enabled, the IST mechanism unconditionally switches stacks. It can be enabled on an individual interrupt-vector basis using a new field in the IDT gate-descriptor entry. This allows some interrupts to use the modified legacy mechanism, and others to use the IST mechanism. The IST mechanism is only available in long mode. The IST mechanism uses new fields in the 64-bit TSS format and the long-mode interrupt-gate and trap-gate descriptors: Figure 12-8 on page 372 shows the format of the 64-bit TSS and the location of the seven IST pointers. The 64-bit TSS offsets from 24h to 5Bh provide space for seven IST pointers, each of which are 64 bits (8 bytes) long. The long-mode interrupt-gate and trap-gate descriptors define a 3-bit IST-index field in bits 2-0 of byte +4. Figure 4-24 on page 115 shows the format of long-mode interrupt-gate and trap-gate descriptors and the location of the IST-index field. To enable the IST mechanism for a specific interrupt, system software stores a non-zero value in the interrupt gate-descriptor IST-index field. If the IST index is zero, the modified legacy stack-switching mechanism (described in the previous section) is used. Figure 8-15 on page 294 shows how the IST mechanism is used to create the interrupt-handler stack. When an interrupt occurs and the IST index is non-zero, the processor uses the index to select the corresponding IST pointer from the TSS. The IST pointer is loaded into the RSP to establish a new stack for the interrupt handler. The SS register is loaded with a null selector if the CPL changes and the SS.RPL is set to the new CPL value. After the stack is loaded, the processor pushes the old stack pointer, RFLAGS, the return pointer, and the error code (if applicable) onto the stack. Control is then transferred to the interrupt handler.
Chapter 8: Exceptions and Interrupts
293
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
64-Bit Interrupt-Handler Stack
Return SS +40
64-Bit TSS Long-Mode Interrupt- or TrapGate Descriptor
Return RSP Return RFLAGS
+32 +24 Return CS +16
Return RIP IST1 : IST7 Error Code
+8
SS=0 RSP
IST
RSP0 : RSP2
513-248.eps
Figure 8-15.
Long-Mode IST Mechanism As with legacy mode, a return to an interrupted program in long mode should be performed using the IRET instruction. However, in long mode, the IRET semantics are different from legacy mode: In 64-bit mode, IRET pops the return-stack pointer unconditionally off the interrupt-stack frame and into the SS:RSP registers. This reverses the action of the long-mode interrupt mechanism, which saves the stack pointer whether or not a privilege change occurs. IRET also allows a null selector to be popped off the stack and into the SS register. See "Nested IRETs to 64-Bit Mode Procedures" on page 295 for additional information. In compatibility mode, IRET behaves as it does in legacy mode. The SS:ESP is popped off the stack only if a control transfer to less privilege (numerically greater CPL) is performed. Otherwise, it is assumed that a stack pointer is not present on the interrupt-handler stack. The long-mode interrupt mechanism always uses a 64-bit stack when saving values for the interrupt handler, and the interrupt handler is always entered in 64-bit mode. To work properly, an
8.9.5 Returning From Interrupt Procedures
294
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
IRET used to exit the 64-bit mode interrupt-handler requires a series of eight-byte pops off the stack. This is accomplished by using a 64-bit operand-size prefix with the IRET instruction. The default stack size assumed by an IRET in 64-bit mode is 32 bits, so a 64-bit REX prefix is needed by 64-bit mode interrupt handlers. Nested IRETs to 64-Bit Mode Procedures. In long mode, an interrupt causes a null selector to be loaded into the SS register if the CPL changes (this is the same action taken by a far CALL in long mode). If the interrupt handler performs a far call, or is itself interrupted, the null SS selector is pushed onto the stack frame, and another null selector is loaded into the SS register. Using a null selector in this way allows the processor to properly handle returns nested within 64-bit-mode procedures and interrupt handlers. The null selector enables the processor to properly handle nested returns to 64-bit mode (which do not use the SS register), and returns to compatibility mode (which do use the SS register). Normally, an IRET that pops a null selector into the SS register causes a general-protection exception (#GP) to occur. However, in long mode, the null selector indicates the existence of nested interrupt handlers and/or privileged software in 64-bit mode. Long mode allows an IRET to pop a null selector into SS from the stack under the following conditions: The target mode is 64-bit mode. The target CPL<3. In this case, the processor does not load an SS descriptor, and the null selector is loaded into SS without causing a #GP exception.
8.10
Virtual Interrupts
The term virtual interrupts includes two classes of extensions to the interrupt-handling mechanism: Virtual-8086 Mode Extensions (VME)--These allow virtual interrupts and interrupt redirection in virtual-8086 mode. VME has no effect on protected-mode programs. Protected-Mode Virtual Interrupts (PVI)--These allow virtual interrupts in protected mode when CPL=3. Interrupt
Chapter 8: Exceptions and Interrupts
295
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
redirection is not available in protected mode. PVI has no effect on virtual-8086-mode programs. Because virtual-8086 mode is not supported in long mode, VME extensions are not supported in long mode. PVI extensions are, however, supported in long mode. 8.10.1 Virtual-8086 Mode Extensions The virtual-8086-mode extensions (VME) enable performance enhancements for 8086 programs running as protected tasks in virtual-8086 mode. These extensions are enabled by setting CR4.VME (bit 0) to 1. The extensions enabled by CR4.VME are: Virtualizing control and notification of maskable external interrupts with the EFLAGS VIF (bit 19) and VIP (bit 20) bits. Selective interception of software interrupts (INTn instructions) using the TSS interrupt redirection bitmap (IRB). Background. Legacy-8086 programs expect to have full access to the EFLAGS interrupt flag (IF) bit, allowing programs to enable and disable maskable external interrupts. When those programs run in virtual-8086 mode under a multitasking protected-mode environment, it can disrupt the operating system if programs enable or disable interrupts for their own purposes. This is particularly true if interrupts associated with one program can occur during execution of another program. For example, a program could request that an area of memory be copied to disk. System software could suspend the program before external hardware uses an interrupt to acknowledge that the block has been copied. System software could subsequently start a second program which enables interrupts. This second program could receive the external interrupt indicating that the memory block of the first program has been copied. If that were to happen, the second program would probably be unprepared to handle the interrupt properly. Access to the IF bit must be managed by system software on a task-by-task basis to prevent corruption of system resources. In order to completely manage the IF bit, system software must be able to interrupt all instructions that can read or write the bit. These instructions include STI, CLI, PUSHF, POPF, INTn, and IRET. These instructions are part of an instruction class that is IOPL-sensitive. The processor takes a general-protection exception (#GP) whenever an IOPL-sensitive instruction is 296 Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
executed and the EFLAGS.IOPL field is less than the CPL. Because all virtual-8086 programs run at CPL=3, system software can interrupt all instructions that modify the IF bit by setting IOPL<3. System software maintains a virtual image of the IF bit for each virtual-8086 program by emulating the actions of IOPLsensitive instructions that modify the IF bit. When an external maskable-interrupt occurs, system software checks the state of the IF image for the current virtual-8086 program to determine whether the program is masking interrupts. If the program is masking interrupts, system software saves the interrupt information until the virtual-8086 program attempts to reenable interrupts. When the virtual-8086 program unmasks interrupts with an IOPL-sensitive instruction, system software traps the action with the #GP handler. The performance of a processor can be significantly degraded by the overhead of trapping and emulating IOPL-sensitive instructions, and the overhead of maintaining images of the IF bit for each virtual-8086 program. This performance loss can be eliminated by running virtual-8086 programs with IOPL set to 3, thus allowing changes to the real IF flag from any privilege level. Unfortunately, this can leave critical system resources unprotected. In addition to the performance problems caused by virtualizing the IF bit, software interrupts (INTn instructions) cannot be masked by the IF bit or virtual copies of the IF bit. The IF bit only affects maskable external interrupts. Software interrupts in virtual-8086 mode are normally directed to the real mode interrupt-vector table (IVT), but it can be desirable to redirect certain interrupts to the protected-mode interrupt-descriptor table (IDT). The virtual-8086-mode extensions are designed to support both external interrupts and software interrupts, with mechanisms that preserve high performance without compromising protection. Virtualization of external interrupts is supported using two bits in the EFLAGS register: the virtual-interrupt flag (VIF) bit and the virtual-interrupt pending (VIP) bit. Redirection of software interrupts is supported using the interrupt-redirection bitmap (IRB) in the TSS. A separate TSS can be created for each virtual-8086 program, allowing system
Chapter 8: Exceptions and Interrupts
297
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
software to control interrupt redirection independently for each virtual-8086 program. VIF and VIP Extensions for External Interrupts. When VME extensions are enabled, the IF-modifying instructions normally trapped by system software are allowed to execute. However, instead of modifying the IF bit, they modifying the EFLAGS VIF bit. This leaves control over maskable interrupts to the system software. It can also be used as an indicator to system software that the virtual-8086 program is able to, or is expecting to, receive external interrupts. When an unmasked external interrupt occurs, the processor transfers control from the virtual-8086 program to a protectedmode interrupt handler. If the interrupt handler determines that the interrupt is for the virtual-8086 program, it can check the state of the VIF bit in the EFLAGS value pushed on the stack for the virtual-8086 program. If the VIF bit is set (indicating the virtual-8086 program attempted to unmask interrupts), system software can allow the interrupt to be handled by the appropriate virtual-8086 interrupt handler. If the VIF bit is clear (indicating the virtual-8086 program attempted to mask interrupts) and the interrupt is for the virtual-8086 program, system software can hold the interrupt pending. System software holds an interrupt pending by saving appropriate information about the interrupt, such as the interrupt vector, and setting the virtual-8086 program's VIP bit in the EFLAGS image on the stack. When the virtual-8086 program later attempts to set IF, the previously set VIP bit causes a general-protection exception (#GP) to occur. System software can then pass the saved interrupt information to the virtual-8086 interrupt handler. To summariz e, when the VM E extensions are enabled (CR4.VME=1), the VIF and VIP bits are set and cleared as follows: VIF Bit--This bit is set and cleared by the processor in virtual-8086 mode in response to an attempt by a virtual8086 program to set and clear the EFLAGS.IF bit. VIF is used by system software to determine whether a maskable external interrupt should be passed on to the virtual-8086 program, emulated by system software, or held pending. VIF is also cleared during software interrupts through interrupt
298
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
gates, with the original VIF value preserved in the EFLAGS image on the stack. VIP Bit--System software sets and clears this bit in the EFLAGS image saved on the stack after an interrupt. It can be set when an interrupt occurs for a virtual-8086 program that has a clear VIF bit. The processor examines the VIP bit when an attempt is made by the virtual-8086 program to set the IF bit. If VIP is set when the program attempts to set IF, a general-protection exception (#GP) occurs before execution of the IF-setting instruction. System software must clear VIP to avoid repeated #GP exceptions when returning to the interrupted instruction. The VIF and VIP bits can be used by system software to minimize the overhead associated with managing maskable external interrupts because virtual copies of the IF flag do not have to be maintained by system software. Instead, VIF and VIP are maintained during context switches along with the remaining EFLAGS bits. Table 8-11 on page 301 shows how the behavior of instructions that modify the IF bit are affected by the VME extensions. Interrupt Redirection of Software Interrupts. In virtual-8 086 mode, software interrupts (INTn instructions) are trapped using a #GP exception handler if the IOPL is less than 3 (the CPL for virtual8086 mode). This allows system software to interrupt and emulate 8086-interrupt handlers. System software can set the IOPL to 3, in which case the INTn instruction is vectored through a gate descriptor in the protected-mode IDT. System software can use the gate to control access to the virtual-8086 mode interrupt-vector table (IVT), or to redirect the interrupt to a protected-mode interrupt handler. When VME extensions are enabled, for INTn instructions to execute normally, vectoring directly to a virtual-8086 interrupt handler through the virtual-8086 IVT (located at address 0 in the virtual-address space of the task). For security or performance reasons, however, it can be necessary to intercept INTn instructions on a vector-specific basis to allow servicing by protected-mode interrupt handlers. This is performed by using the interrupt-redirection bitmap (IRB), located in the TSS and enabled when CR4.VME=1. The IRB is available only in virtual8086 mode.
Chapter 8: Exceptions and Interrupts
299
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Figure 12-6 on page 366 shows the format of the TSS, with the interrupt redirection bitmap located near the top. The IRB contains 256 bits, one for each possible software-interrupt vector. The most-significant bit of the IRB controls interrupt vector 255, and is located immediately before the IOPB base. The least-significant bit of the IRB controls interrupt vector 0. The bits in the IRB function as follows: When set to 1, the INTn instruction behaves as if the VME extensions are not enabled. The interrupt is directed through the IDT to a protected-mode interrupt handler if IOPL=3. If IOPL<3, the INTn causes a #GP exception. When cleared to 0, the INTn instruction is directed through the IVT for the virtual-8086 program to the corresponding virtual-8086 interrupt handler. Only software interrupts can be redirected using the IRB mechanism. External interrupts are asynchronous events that occur outside the context of a virtual-8086 program. Therefore, external interrupts require system-software intervention to determine the appropriate context for the interrupt. The VME extensions described in "VIF and VIP Extensions for External Interrupts" on page 298 are provided to assist system software with external-interrupt intervention. 8.10.2 Protected Mode Virtual Interrupts The protected-mode virtual-interrupt (PVI) bit in CR4 enables support for interrupt virtualization in protected mode. When enabled, the processor maintains program-specific VIF and VIP bits similar to the manner defined by the virtual-8086 mode extensions (VME). However, unlike VME, only the STI and CLI instructions are affected by the PVI extension. When a program is running at CPL=3, it can use STI and CLI to set and clear its copy of the VIF flag without causing a general-protection exception. The last section of Table 8-11 on page 301 describes the behavior of instructions that modify the IF bit when PVI extensions are enabled. The interrupt redirection bitmap (IRB) defined by the VME extensions is not supported by the PVI extensions. 8.10.3 Effect of Instructions that Modify EFLAGS.IF Table 8-11 on page 301 shows how the behavior of instructions that modify the IF bit are affected by the VME and PVI extensions. The table columns specify the following:
300
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Operating Mode--the processor mode in effect when the instruction is executed. Instruction--the IF-modifying instruction. IOPL--the value of the EFLAGS.IOPL field. VIP--the value of the EFLAGS.VIP bit. #GP--indicates whether the conditions in the first four columns cause a general-protection exception (#GP) to occur. Effect on IF Bit--indicates the effect the conditions in the first four columns have on the EFLAGS.IF bit and the image of EFLAGS.IF on the stack. Effect on VIF Bit--indicates the effect the conditions in the first four columns have on the EFLAGS.VIF bit and the image of EFLAGS.VIF on the stack. Table 8-11. Effect of Instructions that Modify the IF Bit
Operating Mode Instruction IOPL VIP #GP CLI Real Mode
CR0.PE=0 EFLAGS.VM=0 CR4.VME=0 CR4.PVI=0
Effect on IF Bit IF = 0 IF = 1 EFLAGS.IF Stack Image = IF
Effect on VIF Bit
STI PUSHF POPF INTn IRET no
IF = EFLAGS.IF stack image EFLAGS.IF Stack Image = IF IF = 0 IF = EFLAGS.IF Stack Image
Notes:
Gray-shaded boxes indicate the bits are unsupported (ignored) in the specified operating mode. "x" indicates the value of the bit is a "don't care". "--" indicates the instruction causes a general-protection exception (#GP).
Notes:
1. If the EFLAGS.IF stack image is 0, no #GP exception occurs and the IRET instruction is executed. 2. If the EFLAGS.IF stack image is 1, the IRET is not executed, and a #GP exception occurs.
Chapter 8: Exceptions and Interrupts
301
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 8-11. Effect of Instructions that Modify the IF Bit (continued)
Operating Mode Instruction IOPL VIP #GP CLI STI Protected Mode
CR0.PE=1 EFLAGS.VM=0 CR4.VME=x CR4.PVI=0
Effect on IF Bit IF = 0 -- IF = 1 -- EFLAGS.IF Stack Image = IF IF = EFLAGS.IF Stack Image No Change
Effect on VIF Bit
CPL no yes no yes
PUSHF POPF INTn gate IRET IRETD CLI STI PUSHF
no x
EFLAGS.IF Stack Image = IF IF = 0 IF = EFLAGS.IF Stack Image
3 <3 3 <3 3 <3 3 <3 3 <3
no yes no yes no yes no yes no yes no yes no yes
IF = 0 -- IF = 1 -- EFLAGS.IF Stack Image = IF -- IF = EFLAGS.IF Stack Image -- EFLAGS.IF Stack Image = IF IF = 0 -- IF = EFLAGS.IF Stack Image -- IF = EFLAGS.IF Stack Image --
Virtual-8086 Mode
CR0.PE=1 EFLAGS.VM=1 CR4.VME=0 CR4.PVI=x
POPF
INTn gate
IRET IRETD
Notes:
3 <3 3 <3
Gray-shaded boxes indicate the bits are unsupported (ignored) in the specified operating mode. "x" indicates the value of the bit is a "don't care". "--" indicates the instruction causes a general-protection exception (#GP).
Notes:
1. If the EFLAGS.IF stack image is 0, no #GP exception occurs and the IRET instruction is executed. 2. If the EFLAGS.IF stack image is 1, the IRET is not executed, and a #GP exception occurs.
302
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 8-11. Effect of Instructions that Modify the IF Bit (continued)
Operating Mode Instruction IOPL VIP #GP CLI 3 <3 3 STI <3 3 <3 3 <3 3 <3 3 <3 3 INTn gate <3 3 IRET <3 3 <3 x 0 1 x no no no
1
Effect on IF Bit IF = 0 No Change IF = 1 No Change EFLAGS.IF Stack Image = IF Not Pushed EFLAGS.IF Stack Image = IF IF = EFLAGS.IF Stack Image No Change IF = EFLAGS.IF Stack Image EFLAGS.IF Stack Image = IF IF = 0 No Change IF = EFLAGS.IF Stack Image No Change No Change IF = EFLAGS.IF Stack Image
Effect on VIF Bit No Change VIF = 0 No Change VIF = 1 -- Not Pushed EFLAGS.IF Stack Image = VIF EFLAGS.VIF Stack Image = VIF -- No Change VIF = EFLAGS.IF Stack Image -- No Change -- No Change EFLAGS.IF Stack Image = VIF VIF = 0 No Change VIF = EFLAGS.IF Stack Image VIF = EFLAGS.IF Stack Image -- VIF = EFLAGS.IF Stack Image --
x x 0 1 x x x 0 1 x
no no no yes no no yes no no yes no yes
PUSHF PUSHFD Virtual-8086 Mode with VME Extensions POPF
CR0.PE=1 EFLAGS.VM=1 CR4.VME=1 CR4.PVI=x
POPFD
x
no
yes2 no yes
IRETD
Notes:
Gray-shaded boxes indicate the bits are unsupported (ignored) in the specified operating mode. "x" indicates the value of the bit is a "don't care". "--" indicates the instruction causes a general-protection exception (#GP).
Notes:
1. If the EFLAGS.IF stack image is 0, no #GP exception occurs and the IRET instruction is executed. 2. If the EFLAGS.IF stack image is 1, the IRET is not executed, and a #GP exception occurs.
Chapter 8: Exceptions and Interrupts
303
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 8-11. Effect of Instructions that Modify the IF Bit (continued)
Operating Mode Instruction IOPL VIP #GP CLI 3 <3 3 Protected Mode with PVI Extensions
CR0.PE=1 EFLAGS.VM=0 CR4.VME=x CR4.PVI=1 CPL=3
Effect on IF Bit IF = 0 No Change IF = 1 No Change
Effect on VIF Bit No Change VIF = 0 No Change VIF = 1 -- Not Pushed EFLAGS.VIF Stack Image = VIF No Change VIF = 0 No Change No Change VIF = EFLAGS.VIF Stack Image
x x 0 1
no no no yes
STI PUSHF PUSHFD POPF POPFD INTn gate IRET IRETD
<3
EFLAGS.IF Stack Image = IF IF = EFLAGS.IF Stack Image x x no EFLAGS.IF Stack Image = IF IF = 0 (if interrupt gate) IF = EFLAGS.IF Stack Image
Notes:
Gray-shaded boxes indicate the bits are unsupported (ignored) in the specified operating mode. "x" indicates the value of the bit is a "don't care". "--" indicates the instruction causes a general-protection exception (#GP).
Notes:
1. If the EFLAGS.IF stack image is 0, no #GP exception occurs and the IRET instruction is executed. 2. If the EFLAGS.IF stack image is 1, the IRET is not executed, and a #GP exception occurs.
304
Chapter 8: Exceptions and Interrupts
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
9
Machine Check Mechanism
The machine-check mechanism allows the processor to detect and report a variety of hardware errors. The types of errors that can be reported include: Cache errors associated with reading and writing data, probing, cache-line fills, and cache-line writebacks. Parity errors associated with the caches and TLBs. ECC errors associated with the caches and DRAM. Bus errors associated with reading and writing on the processor external bus. Software can enable the processor to report machine-check errors through the machine-check exception (for additional information, see "#MC--Machine-Check Exception (Vector 18)" on page 264). Most machine-check error conditions do not allow reliable restarting of interrupted programs. System software instead uses the machine-check mechanism to report the source of hardware problems for possible servicing. The basic machine-check mechanism is consistent across processor implementations, but the error-reporting registers are model specific. Processor implementations are not required to support the mechanism, and those implementations that do support it can vary in how the mechanism works.
9.1
Determining Machine-Check Support
The availability of machine-check registers and support of the machine-check exception is implementation dependent. System software executes the CPUID instruction to determine whether a processor implements these features. After CPUID is executed, the values of the machine-check architecture (MCA) bit and the machine-check exception (MCE) bit loaded in the EDX register indicate whether the processor implements the machine-check registers and the machine-check exception, respectively. See "Processor Feature Identification" on page 78 for more information on using the CPUID instruction to determine the level of machine-check support. Once system software determines that the machine-check registers are available, it must determine the extent of
Chapter 9: Machine Check Mechanism
305
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
processor support for the machine-check mechanism. This is accomplished by reading the machine-check capabilities register (MCG_CAP). See "Machine-Check Global-Capabilities Register" on page 308 for more information on interpreting the MCG_CAP contents. Implementation-specific information concerning the machinecheck mechanism can be found in the BIOS writer's guide for the implementation.
9.2
Machine-Check Errors
The following classes of machine-check errors are defined for the AMD64 architecture: Recoverable--The error has been corrected by the processor. Recoverable errors do not cause a machine check exception (#MC). However, the error is still logged in the machinecheck registers. It is the responsibility of system software to periodically poll the machine-check registers to determine whether recoverable errors have occurred. Fatal/Unrecoverable--The error cannot be corrected by the processor. Unrecoverable errors cause a machine check exception if CR4.MCE is set to 1. In both cases, the contents of the machine-check registers are maintained through a warm reset, which allows errors to be reported even if a reset occurs.
9.2.1 Error Sources
Implementations can detect errors from any number of sources located within the various processor units. Those processor units can include the following: Data-Cache Unit (DC)--Includes the cache structures that hold data and tags, the data TLBs, and cache-probing logic. Instruction-Cache Unit (IC)--Includes the cache structures that hold instructions and tags, the instruction TLBs, and cache-probing logic. Bus Unit (BU)--Includes the L2 cache and any external caches. Load/Store Unit (LS)--Includes logic used to manage loads and stores. Northbridge (NB)--Includes the system-bus interface and DRAM controller.
306
Chapter 9: Machine Check Mechanism
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
A given processor implementation can monitor machine-check errors in sources other than those listed above. The number is implementation-specific and is determined by examining the MCG_CAP register (see "Machine-Check Global-Capabilities Register" on page 308). For further information, see the documentation for particular implementations of the architecture.
9.3
Machine Check MSRs
The AMD64 architecture defines a set of model-specific registers (MSRs) in support of the machine-check mechanism. These registers include: Global-status and global-control registers: - Machine-check global-capabilities register (MCG_CAP). - Machine-check global-status register (MCG_STATUS). - Machine-check global-control register (MCG_CTL). Error-reporting register banks, each containing: - Machine-check control register (MCi_CTL). - Machine-check status register (MCi_STATUS). - Machine-check address register (MCi_ADDR). - Machine-check miscellaneous error-information register (MCi_MISC). Each error-reporting register bank is associated with a specific processor unit (or group of processor units). In some cases, the machine-check handler cannot be invoked due to an error, the error-reporting registers retain their values through a warm reset. (A warm reset is a reset that occurs while the processor is powered up, as opposed to a cold reset, which occurs during power-up.) This allows the BIOS or other systemboot software to recover and report information associated with the error. The RDMSR and WRMSR instructions are used to read and write the machine-check MSRs. See "Machine-Check MSRs" on page 443 for a listing of the machine-check MSR numbers and their reset values. The following sections describe each machine-check MSR and its function.
Chapter 9: Machine Check Mechanism
307
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
9.3.1 Global Status and Control Registers
The global-status and global-control MSRs supported by the m a c h i n e - ch e c k m e ch a n i s m i n c l u d e t h e M C G _ C A P, MCG_STATUS, and MCG_CTL registers. Machine-Check Global-Capabilities Register. Fi g u re 9 -1 s h ow s t h e format of the machine-check global-capabilities register (MCG_CAP). MCG_CAP is a read-only register that specifies the machine-check mechanism capabilities supported by the processor implementation.
63 Reserved
32
31 Reserved
9
8
M C G C P
7 Count
0
Bits 63-9 8 7-0
Mnemonic Reserved MCG_CTL_P Count
Description MCG_CTL_PMCG_CTL register present Number of reporting banks
R/W R R
Figure 9-1.
MCG_CAP Register The fields within the MCG_CAP register are: Count--Bits 7-0. This field specifies how many errorreporting register banks are supported by the processor implementation. MCG_CTL Register Present (MCG_CTL_P)--Bit 8. This bit specifies whether or not the MCG_CTL register is supported by the processor. When the bit is set to 1, the register is supported. When the bit is cleared to 0, the register is unsupported. All remaining bits in the MCG_CAP register are reserved. Writing values to the MCG_CAP register produces undefined results. Machine-Check Global-Status Register. Figure 9-2 on page 309 shows the format of the machine-check global-status register (MCG_STATUS). MCG_STATUS provides basic information about the processor state after the occurrence of a machinecheck error.
308
Chapter 9: Machine Check Mechanism
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
63 Reserved
32
31 Reserved
3
2
M C I P
1
E I P V
0
R I P V
Bits 63-3 2 1 0
Mnemonic Reserved MCIP EIPV RIPV
Description Machine Check In-Progress Error IP Valid Flag Restart IP Valid Flag
R/W R/W R/W R/W
Figure 9-2.
MCG_STATUS Register The fields within the MCG_STATUS register are: Restart-IP Valid (RIPV)--Bit 0. When this bit is set to 1, the interrupted program can be reliably restarted at the instruction addressed by the instruction pointer pushed onto the stack by the machine-check error mechanism. If this bit is cleared to 0, the interrupted program cannot be reliably restarted. Error-IP Valid (EIPV)--Bit 1. When this bit is set to 1, the instruction that is referenced by the instruction pointer pushed onto the stack by the machine-check error mechanism is responsible for the machine-check error. If this bit is cleared to 0, it is possible that the instruction referenced by the instruction pointer is not responsible for the machine-check error. Machine Check In-Progress (MCIP)--Bit 2. When this bit is set to 1, it indicates that a machine-check error is in progress. If another machine-check error occurs while this bit is set, the processor enters the shutdown state. The processor sets this bit whenever a machine check exception is generated. Software is responsible for clearing it after the machine check exception is handled. All remaining bits in the MCG_STATUS register are reserved. Machine-Check Global-Control Register. Figure 9-3 on page 310 shows the format of the machine-check global-control register (MCG_CTL). MCG_CTL is used by software to control reporting machine-check errors from various sources. Each errorreporting register bank supported by the processor is
Chapter 9: Machine Check Mechanism
309
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
controlled by a corresponding enable bit in this register. Setting all bits to 1 in this register enables all error-reporting register banks. The number of controls and how they are used is implementation-specific (for further information, see the documentation for particular implementations of the architecture). The presence of the MCG_CTL register is indicated by the MCG_CAP register MCG_CTL_P bit, described on page 308.
63
E N 6 3
2
1
E N 1
0
E N 0
...
Error-Reporting Register-Bank Enable Bits
...
E N 2
Figure 9-3.
MCG_CTL Register Error-reporting register banks contain the following registers: Machine-check control registers (MCi_CTL). Machine-check status register (MCi_STATUS). Machine-check address register (MCi_ADDR). Machine-check miscellaneous error-information register (MCi_MISC). The i in each register name corresponds to the number of a supported register bank. Each error-reporting register bank is associated with a specific processor unit (or group of processor units). The number of error-reporting register banks is implementation-specific. For more information, see the BIOS a n d Ke r n e l D e v e l o p 's G u i d e f o r A M D A t h l o n TM 6 4 a n d AMD OpteronTM Processors (order# 26094) for particular implementations of the AMD64 architecture. Software reads the MCG_CAP register to determine the number of supported register banks. The first error-reporting register (MC0_CTL) always starts with MSR address 400h, followed by MC0_STATUS (401h), MC0_ADDR (402h), and MC0_MISC (403h). Error-reporting-register MSR addresses are assigned sequentially through the remaining supported register banks. Using this information, software can access all errorreporting registers in an implementation-independent manner. Machine-Check Control Registers. T h e m a ch i n e - ch e ck c o n t r o l registers (MCi_CTL), as shown in Figure 9-4 on page 311, contain an enable bit for each error source within an error-
9.3.2 Error-Reporting Register Banks
310
Chapter 9: Machine Check Mechanism
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
reporting register bank. Setting an enable bit to 1 enables errorreporting for the specific feature controlled by the bit, and clearing the bit to 0 disables error reporting for the feature. For more information, see the BIOS and Kernel Developer's Guide for the AMD AthlonTM 64 and AMD OpteronTM Processors (order# 26094) for particular implementations of the AMD64 architecture.
63
E N 6 3
2
1
E N 1
0
E N 0
...
Error-Reporting Register-Bank Enable Bits
...
E N 2
Figure 9-4.
MCi_CTL Registers Machine-Check Status Registers. Each error-reporting register bank includes a machine-check status register (MCi_STATUS) that the processor uses to report machine-check error information. The machine-check mechanism writes the status-register bits when an error is detected, and sets the valid bit in the register (bit 63) to 1, indicating that the status information is valid. Error reporting for the detected error does not need to be enabled for the processor to write the status register. Error reporting must be enabled for the error to result in a machinecheck exception. Software is responsible for clearing the status register after the exception has been handled. Attempting to write a value other than 0 to an MCi_STATUS register will raise a general protection (#GP) exception. Figure 9-5 shows the format of the MCi_STATUS register.
Chapter 9: Machine Check Mechanism
311
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
63
V A L
62
O V E R
61
U C
60
E N
59
M I S C V
58
A D D R V
57
P C C
56 Other Information
32
31 Model-Specific Error Code
16
15 MCA Error Code
0
Bits Mnemonic Description 63 VAL Valid 62 OVER Status Register Overflow 61 UC Uncorrected Error 60 EN Error Condition Enabled 59 MISCV Miscellaneous-Error Register Valid 58 ADDRV Error-Address Register Valid 57 PCC Processor-Context Corrupt 56-32 Other Information 31-16 Model-Specific Error Code 15-0 MCA Error Code *Applications can only clear this bit to 0.
R/W R/W* R/W* R/W* R/W* R/W* R/W* R/W* R/W* R/W* R/W*
Figure 9-5.
MCi_STATUS Register When a machine-check error occurs, the processor loads an error code into the appropriate MCi_STATUS register MCA error-code field. The MCi_STATUS.VAL bit is set to 1, indicating that the MCi_STATUS register contents are valid. The machine-check mechanism also allows the processor to load a model-specific error code into the MCi_STATUS register. MCA error-codes are used to report errors in the memory hierarchy, the system bus, and the system-interconnection logic. Error-codes are divided into subfields that are used to describe the cause of an error. The information is implementation specific. It includes the location in the memory hierarchy where the error occurred and the type of transaction that caused the error. For further information, see the documentation for particular implementations of the architecture. The fields within the MCi_STATUS register are: MCA Error Code--Bits 15-0. This field encodes information about the error, including: - The type of transaction that caused the error. The memory-hierarchy level involved in the error. - The type of request that caused the error. - Other information concerning the transaction type. See "Error Codes" on page 312 for information on the format and encoding of the MCA error code.
9.3.3 Error Codes
312
Chapter 9: Machine Check Mechanism
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Model-Specific Error Code--Bits 31-16. This field encodes model-specific information about the error. For further information, see the documentation for particular implementations of the architecture. Other Information--Bits 56-32. This field holds modelspecific error information. Software should not rely on the field definitions being consistent between processor implementations. Presently, the bits in this field are defined as: - Bits 44-32--Reserved. - Bit 45--When set to 1, this bit indicates the error is an uncorrectable ECC error. - Bit 46--When set to 1, this bit indicates the error is a correctable ECC error. - Bits 54-47--This field holds the ECC syndrome when an ECC error occurs. - Bits 56-55--Reserved. PCC--Bit 57. When set to 1, this bit indicates that the processor state is likely to be corrupt due to the machinecheck error. In this case, it is possible software cannot restart the processor reliably. When this bit is cleared to 0, the processor state is not corrupted by the machine-check error. If the PCC bit is set in any error bank, the processor will clear RIPV and EIPV in the MCG_Status register. ADDRV--Bit 58. When set to 1, this bit indicates that the address saved in the corresponding error-reporting address register (MCi_ADDR) is valid, and contains the address where the error was detected. When this bit is cleared to 0, MCi_ADDR does not contain a valid error address. MISCV--Bit 59. When set to 1, this bit indicates that additional information about the machine-check error is saved in the corresponding error-reporting miscellaneous register (MCi_MISC). This bit is cleared to 0 when the MCi_MISC registers are not implemented. EN--Bit 60. When set to 1, this bit indicates that the error condition is enabled in the corresponding error-reporting control register (MCi_CTL). Errors disabled by MCi_CTL do not cause a machine-check exception, but the machine-check mechanism can log errors when error reporting is disabled in MCi_CTL.
Chapter 9: Machine Check Mechanism
313
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
UC--Bit 61. When set to 1, this bit indicates that the processor did not correct the error condition. When this bit is cleared to 0, the processor corrected the error condition. OVER--Bit 62. This bit is set to 1 by the processor if the VAL bit is already set to 1 as the processor attempts to load error information into MCi_STATUS. This indicates that the results of a previous machine-check error are still in the MCi_STATUS register. In this situation, the machine-check mechanism handles the contents of MCi_STATUS as follows: - Status for an enabled error replaces status for a disabled error. - Status for an uncorrectable error replaces status for a correctable error. - Status for an enabled uncorrectable error is never replaced. VAL--Bit 63. This bit is set to 1 by the processor if the contents of MCi_STATUS are valid. Software should clear the VAL bit after reading the MCi_STATUS register, otherwise a subsequent machine-check error sets the OVER bit as described above. Machine-Check Address Registers. Each error-reporting register bank includes a machine-check address register (MCi_ADDR) that the processor uses to report the address associated with the machine-check error. The address field can hold either a virtual (linear) or physical address, depending on the type of error. For further information, see the documentation for particular implementations of the architecture. The contents of this register are valid only if the ADDRV bit in the corresponding MCi_STATUS register is set to 1. Machine-Check Miscellaneous-Error Information Registers. E a ch e r ro rreporting register bank can include a machine-check miscellaneous error-information register (MCi_MISC) that the processor uses to report additional information concerning the machine-check error. The contents of this register are valid only if the MISCV bit in the corresponding MCi_STATUS register is set to 1. The miscellaneous error information reported in this register is implementation dependent.
314
Chapter 9: Machine Check Mechanism
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
9.4
Initializing the Machine-Check Mechanism
Following a processor reset, all machine-check error-reporting enable bits are disabled. System software must enable these bits before machine-check errors can be reported. Generally, s y s t e m s o f t wa re s h o u l d i n i t i a l i z e t h e m a ch i n e - ch e ck mechanism using the following process: Execute the CPUID instruction and verify that the processor supports the machine-check exception (MCE) and machinecheck registers (MCA). MCE is supported when EDX bit 7 is set to 1, and MCA is supported when EDX bit 14 is set to 1. Software should not proceed with initializing the machinecheck mechanism if the machine-check registers are not supported. If the machine-check registers are supported, system software should take the following steps: - Check to see if the MCG_CTL_P bit in the MCG_CAP register is set to 1. If it is, then the MCG_CTL register is supported by the processor. If the MCG_CTL register is supported, software should set its enable bits to 1 for the machine-check features it uses. Software can load MCG_CTL with all 1s to enable all machine-check features. - Read the COUNT field from the MCG_CAP register to determine the number of error-reporting register banks supported by the processor. For each error-reporting register bank, software should set the enable bits to 1 in the MCi_CTL register for the error types it wants the processor to report. Software can load each MCi_CTL with all 1s to enable all error-reporting mechanisms. The error-reporting register banks are numbered from 0 to one less than the value found in the MCG_CAP.COUNT field. For example, if the COUNT field indicates five register banks are supported, they are numbered 0 to 4. - For each error-reporting register bank, software should clear all status fields in the MCi_STATUS register by writing all 0s to the register. It is possible that valid error-status is already reported by the MCi_STATUS registers at the time software clears them. The status can reflect fatal errors recorded before a warm reset, or errors recorded during the system powerup and boot process. Before clearing the MCi_STATUS
Chapter 9: Machine Check Mechanism
315
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
registers, software should examine their contents and log any errors found. As a final step in the initialization process, system software should enable the machine-check exception by setting CR4.MCE (bit 6) to 1.
9.5
Using Machine Check Features
System software can detect and handle machine-check errors using two methods: Software can periodically examine the machine-check status registers for reported errors, and log any errors found. Software can enable the machine-check exception (#MC). When an uncorrectable error occurs, the processor immediately transfers control to the machine-check exception handler. In this case, system software provides a machine-check exception handler that, at a minimum, logs detected errors. The exception handler can be designed for a specific processor implementation or can be generalized to work on multiple implementations.
9.5.1 Handling Machine Check Exceptions
The processor uses the interrupt control-transfer mechanism to invoke an exception handler after a machine-check exception occurs. This requires system software to initialize the interruptdescriptor table (IDT) with either an interrupt gate or a trap gate that references the interrupt handler. See "Legacy Protected-Mode Interrupt Control Transfers" on page 276 and "Long-Mode Interrupt Control Transfers" on page 287 for more information on interrupt control transfers. At a minimum, the machine-check exception handler must be capable of logging errors for later examination. This can be a sufficient implementation for some handlers. More thorough exception-handler implementations can analyze the error to determine if it is unrecoverable, and whether it can be recovered in software. Machine-check exception handlers that attempt to correct unrecoverable errors must be thorough in their analysis and their corrective actions. The following guidelines should be used when writing such a handler: All status registers in the error-reporting register banks must be examined to identify the cause or causes of the
316
Chapter 9: Machine Check Mechanism
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
machine-check exception. Read the COUNT field from MCG_CAP to determine the number of status registers supported by the processor. The status registers are numbered from 0 to one less than the value found in the MCG_CAP.COUNT field. For example, if the COUNT field indicates five status registers are supported, they are named MC0_STATUS to MC4_STATUS. Check the valid bit in each status register (MCi_STATUS.VAL). The MCi_STATUS register does not need to be examined when its valid bit is clear. Check the valid MCi_STATUS registers to see if error recovery is possible. Error recovery is not possible when: - The processor-context corrupt bit (MCi_STATUS.PCC) is set to 1. - The error-overflow status bit (MCi_STATUS.OVER) is set to 1. This bit indicates that more than one machine-check error occurred, but only one error is reported by the status register. If error recovery is not possible, the handler should log the error information and return to the operating system. Check the MCi_STATUS.UC bit to see if the processor corrected the error. If UC=1, the processor did not correct the error, and the exception handler must correct the error before restarting the interrupted program. If the handler cannot correct the error, it should log the error information and return to the operating system. When identifying the error condition, portable exception handlers should examine only the MCi_STATUS register MCA error-code field. See "Error Codes" on page 312 for information on interpreting this field. If the MCG_STATUS.RIPV bit is set to 1, the interrupted program can be restarted reliably at the instruction-pointer address pushed onto the exception-handler stack. If RIPV=0, the interrupted program cannot be restarted reliably at that location, although it can be restarted at that location for debugging purposes. When logging errors, particularly those that are not recoverable, check the MCG_STATUS.EIPV bit to see if the instruction-pointer address pushed onto the exceptionhandler stack is related to the machine-check error. If
Chapter 9: Machine Check Mechanism
317
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
EIPV=0, the address is not guaranteed to be related to the error. Before exiting the machine-check handler, be sure to clear MCG_STATUS.MCIP to 0. MCIP indicates a machine-check exception occurred. If this bit is set when another machinecheck exception occurs, the processor enters the shutdown state. When an exception handler is able to, at a minimum, successfully log an error condition, the MCi_STATUS registers should be cleared to 0 before exiting the machinecheck handler. Software is responsible for clearing at least the MCi_STATUS.VAL bits. Additional machine-check exception-handler portability can be added by having the handler use the CPUID instruction to identify the processor and its capabilities. Implementation-specific software can be added to the machine-check exception handler based on the processor information reported by CPUID. 9.5.2 Reporting Correctable Machine Check Errors Machine-check exceptions do not occur if the error is correctable by the processor. If system software wishes to log and report correctable machine-check errors, a system-service routine must be provided to check the contents of the machinecheck status registers for correctable errors. The service routine can be invoked by system software on a periodic basis, or it can be manually invoked by the user as needed. If the processor supports the machine-check registers, a service routine that reports correctable errors should perform the following: Examine each status register (MCi_STATUS) in the errorreporting register banks. For each MCi_STATUS register with a set valid bit (VAL=1), the service routine should: - Save the contents of the MCi_STATUS register. - Save the contents of the corresponding MCi_ADDR register if MCi_STATUS.ADDRV=1. - Save the contents of the corresponding MCi_MISC register if MCi_STATUS.MISCV=1. - Check to see if MCG_STATUS.MCIP=1, which indicates that the machine-check exception handler is in progress. If this is the case, then the machine-check exception handler has called the service routine to log the errors. In 318 Chapter 9: Machine Check Mechanism
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
this situation, the error-logging service routine should determine whether or not the interrupted program is restartable, and report the determination back to the exception handler. The program is not restartable if either of the following is true: - MCi_STATUS.PCC=1, which indicates the processor context is corrupted, or - MCG_STATUS.RIPV=0, which indicates the interrupted program cannot be restarted reliably at the instructionpointer address pushed onto the exception-handler stack. Once the information found in the error-reporting register banks is saved, the MCi_STATUS register should be cleared to 0. This allows the processor to properly report any subsequent errors in the MCi_STATUS registers. The service routine can save the time-stamp counter with each error logged. This can help in determining how frequently errors occur. For further information, see "TimeStamp Counter" on page 408. In multiprocessor configurations, the service routine can save the processor-node identifier. This can help locate a failing multiprocessor-system component, which can then be isolated from the rest of the system. For further information, see the documentation for particular implementations of the architecture.
Chapter 9: Machine Check Mechanism
319
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
320
Chapter 9: Machine Check Mechanism
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
10
System-Management Mode
System-management mode (SMM) is an operating mode designed for system-control activities like power management. Normally, these activities are transparent to conventional operating systems and applications. SMM is used by systemspecific BIOS (basic input-output system) and specialized lowlevel device drivers, rather than the operating system. The SMM interrupt-handling mechanism differs substantially from the standard interrupt-handling mechanism described in Chapter 8, "Exceptions and Interrupts." SMM is entered using a special external interrupt called the system-management interrupt (SMI). After an SMI is received by the processor, the processor saves the processor state in a separate address space, called SMRAM. The SMM-handler software and data structures are also located in the SMRAM space. Interrupts and exceptions that ordinarily cause control transfers to the operating system are disabled when SMM is entered. The processor exits SMM, restores the saved processor state, and resumes normal execution by using a special instruction, RSM. In SMM, address translation is disabled and addressing is similar to real mode. SMM programs can address up to 4 Gbytes of physical memory. See "SMM Operating-Environment" on page 331 for additional information on memory addressing in SMM. The following sections describe the components of the SMM mechanism: "SMM Resources" on page 322--this section describes SMRAM, the SMRAM save-state area used to hold the processor state, and special SMRAM save-state entries used in support of SMM. "Using SMM" on page 331--this section describes the mechanism of entering and exiting SMM. It also describes SMM memory allocation, addressing, and interrupts and exceptions. Of these mechanisms, only the format of the SMRAM save-state area differs between the AMD64 architecture and the legacy architecture.
Chapter 10: System-Management Mode
321
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
10.0.1 SMM Differences
There are functional differences between the SMM support in the AMD64 architecture and the SMM support found in previous architectures. These are: The SMRAM state-save area layout is changed to hold the 64-bit processor state. The initial processor state upon entering SMM is expanded to reflect the 64-bit nature of the processor. New conditions exist that can cause a processor shutdown while in SMM. The auto-halt restart and I/O-instruction restart entries in the SMRAM state-save area are one byte each instead of two bytes each. SMRAM caching considerations are modified because the legacy FLUSH# external signal (writeback, if modified, and invalidate) is not supported on implementations of the AMD64 architecture. Some previous AMD x86 processors saved and restored the CR2 register in the SMRAM state-save area. This register is not saved by the SMM implementation in the AMD64 architecture. SMM handlers that save and restore CR2 must perform the operation in software.
10.1
SMM Resources
The SMM resources supported by the processor consist of SMRAM, the SMRAM state-save area, and special entries within the SMRAM state-save area. In addition to the save-state area, SMRAM includes space for the SMM handler.
10.1.1 SMRAM
SMRAM is the memory-address space accessed by the processor when in SMM. The default size of SMRAM is 64 Kbytes and can range in size between 32 Kbytes and 4 Gbytes. System logic can use physically separate SMRAM and main memory, directing memory transactions to SMRAM after recognizing SMM is entered, and redirecting memory transactions back to system memory after recognizing SMM is exited. When separate SMRAM and main memory are used, the system designer needs to provide a method of mapping SMRAM into main memory so that the SMI handler and data structures can be loaded. Figure 10-1 on page 323 shows the default SMRAM memory map. The default SMRAM code-segment (CS) has a base
322
Chapter 10: System-Management Mode
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
address of 0003_0000h (the base address is automatically scaled by the processor using the CS-selector register, which is set to the value 3000h). This default SMRAM-base address is known as SMBASE. A 64-Kbyte memory region, addressed from 0003_0000h to 0003_FFFFh, makes up the default SMRAM memory space. The top 32 Kbytes (0003_8000h to 0003_FFFFh) must be supported by system logic, with physical memory covering that entire address range. The top 512 bytes (0003_FE00h to 0003_FFFFh) of this address range are the default SMM state-save area. The default entry point for the SMM interrupt handler is located at 0003_8000h.
SMRAM
SMM State-Save Area
0003_FFFFh (SMBASE+FFFFh) 0003_FE00h
SMM Handler 0003_8000h (SMBASE+8000h)
0003_0000h (SMBASE)
513-250.eps
Figure 10-1. Default SMRAM Memory Map 10.1.2 SMBASE Register The format of the SMBASE register is shown in Figure 10-2 on page 324. SMBASE is an internal processor register that holds the value of the SMRAM-base address. SMBASE is set to 30000h after a processor reset.
Chapter 10: System-Management Mode
323
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
31 SMRAM Base Address
0
Figure 10-2.
SMBASE Register In some operating environments, relocation of SMRAM to a higher memory area can provide more low memory for legacy software. SMBASE relocation is supported when the SMM-base relocation bit in the SMM-revision identifier (bit 17) is set to 1. In processors implementing the AMD64 architecture, SMBASE relocation is always supported. Software can only modify SMBASE (relocate the SMRAM-base address) by entering SMM, modifying the SMBASE image stored in the SMRAM state-save area, and exiting SMM. The SMM-handler entry point must be loaded at the new memory location specified by SMBASE+8000h. The next time SMM is entered, the processor saves its state in the new state-save area at SMBASE+0FE00h, and begins executing the SMM handler at SMBASE+8000h. The new SMBASE address is used for every SMM until it is changed, or a hardware reset occurs. When SMBASE is used to relocate SMRAM to an address above 1 Mbyte, 32-bit address-size-override prefixes must be used to access this memory. This is because addressing in SMM behaves as it does in real mode, with a 16-bit default operand size and address size. The values in the 16-bit segment-selector registers are left-shifted four bits to form a 20-bit segment-base address. Without using address-size overrides, the maximum address that can be computed is 10FFEh. Because SMM memory-addressing is similar to real-mode addressing, the SMBASE address must be less than 4 Gbytes. Physical-address extensions (CR4.PAE) should not be enabled in SMM, restricting the SMRAM address space to the range 0h to 0FFFF_FFFFh.
10.1.3 SMRAM StateSave Area
When an SMI occurs, the processor saves its state in the 512byte SMRAM state-save area during the control transfer into SMM. The format of the state-save area defined by the AMD64 Chapter 10: System-Management Mode
324
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
architecture is shown in Table 10-1. This table shows the offsets in the SMRAM state-save area relative to the SMRAM-base address. The state-save area is located between offset 0_FE00h (SMBASE+0_FE00h) and offset 0_FFFFh (SMBASE+0_FFFFh). Software should not modify offsets specified as read-only or reserved, otherwise unpredictable results can occur. Table 10-1. AMD64 Architecture SMM State-Save Area
Contents Selector ES Attributes Limit Base Selector CS Attributes Limit Base Selector SS Attributes Limit Base Selector DS Attributes Limit Base Selector FS Attributes Limit Base Size Word Word Doubleword Quadword Word Word Doubleword Quadword Word Word Doubleword Quadword Word Word Doubleword Quadword Word Word Doubleword Quadword Read-Only Read-Only Read-Only Read-Only Read-Only Allowable Access
Offset (Hex) from SMBASE FE00h FE02h FE04h FE08h FE10h FE12h FE14h FE18h FE20h FE22h FE24h FE28h FE30h FE32h FE34h FE38h FE40h FE42h FE44h FE48h
Note:
1. The offset for the SMM-revision identifier is compatible with previous implementations.
Chapter 10: System-Management Mode
325
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 10-1.
AMD64 Architecture SMM State-Save Area (continued)
Contents Selector GS Attributes Limit Base Reserved Reserved GDTR Limit Reserved Base Selector LDTR Attributes Limit Base Reserved Reserved IDTR Limit Reserved Base Selector TR Attributes Limit Base Reserved I/O Instruction Restart Auto-Halt Restart Size Word Word Doubleword Quadword 2 Bytes Word Word 2 Bytes Quadword Word Word Doubleword Quadword 2 Bytes Word Word 2 Bytes Quadword Word Word Doubleword Quadword 40 Bytes Byte Byte -- Read/Write Read-Only Read-Only Read-Only Read-Only Read-Only Allowable Access
Offset (Hex) from SMBASE FE50h FE52h FE54h FE58h FE60h-FE61h FE62h FE64h FE66h-FE67h FE68h FE70h FE72h FE74h FE78h FE80h-FEB1h FE82h FE84h FEB6h-FEB7h FE88h FE90h FE92h FE94h FE98h FEA0h-FEC7h FEC8h FEC9h
Note:
1. The offset for the SMM-revision identifier is compatible with previous implementations.
326
Chapter 10: System-Management Mode
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 10-1.
AMD64 Architecture SMM State-Save Area (continued)
Contents Reserved EFER Reserved SMM-Revision Identifier1 SMBASE Reserved CR4 CR3 CR0 DR7 DR6 RFLAGS RIP R15 R14 R13 R12 R11 R10 R9 R8 Size 6 Bytes Quadword 36 Bytes Doubleword Doubleword 68 Bytes Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Read/Write Read-Only Read/Write Read-Only Allowable Access -- Read-Only -- Read-Only Read/Write --
Offset (Hex) from SMBASE FECAh--FECFh FED0h FED8h--FEFBh FEFCh FF00h FF04h--FF47h FF48h FF50h FF58h FF60h FF68h FF70h FF78h FF80h FF88h FF90h FF98h FFA0h FFA8h FFB0h FFB8h
Note:
1. The offset for the SMM-revision identifier is compatible with previous implementations.
Chapter 10: System-Management Mode
327
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 10-1.
AMD64 Architecture SMM State-Save Area (continued)
Contents RDI RSI RBP RSP RBX RDX RCX RAX Size Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Read/Write Allowable Access
Offset (Hex) from SMBASE FFC0h FFC8h FFD0h FFD8h FFE0h FFE8h FFF0h FFF8h
Note:
1. The offset for the SMM-revision identifier is compatible with previous implementations.
A number of other registers are not saved or restored automatically by the SMM mechanism. See "Saving Additional Processor State" on page 334 for information on using these registers in SMM. As a reference for legacy processor implementations, the legacy S M M s t a t e - s ave a re a fo r m a t i s s h ow n i n Tab l e 1 0 -2 . Implementations of the AMD64 architecture do not use this format. Table 10-2. Legacy SMM State-Save Area (Not used by AMD64 Architecture)
Offset (Hex) from SMBASE FE00h--FEF7h FEF8h FEFCh FF00h FF02h
Note:
Contents Reserved SMBASE SMM-Revision Identifier I/O Instruction Restart Auto-Halt Restart
Size 248 Bytes Doubleword Doubleword Word Word
Allowable Access -- Read/Write Read-Only Read/Write
1. The offset for the SMM-revision identifier is compatible with previous implementations.
328
Chapter 10: System-Management Mode
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 10-2. Legacy SMM State-Save Area (continued)(Not used by AMD64 Architecture)
Offset (Hex) from SMBASE FF04h--FF87h FF88h FF8Ch--FF93h FF94h FF98h--FFA7h FFA8h FFACh FFB0h FFB4h FFB8h FFBCh FFC0h FFC4h FFC8h FFCCh FFD0h FFD4h FFD8h FFDCh FFE0h FFE4h FFE8h FFECh FFF0h
Note:
Contents Reserved GDT Base Reserved IDT Base Reserved ES CS SS DS FS GS LDT Base TR DR7 DR6 EAX ECX EDX EBX ESP EBP ESI EDI EIP
Size 132 Bytes Doubleword Quadword Doubleword 16 Bytes Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword Doubleword
Allowable Access -- Read-Only -- Read-Only --
Read-Only
Read-Only
Read-Only
Read/Write
Read/Write
1. The offset for the SMM-revision identifier is compatible with previous implementations.
Chapter 10: System-Management Mode
329
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 10-2. Legacy SMM State-Save Area (continued)(Not used by AMD64 Architecture)
Offset (Hex) from SMBASE FFF4h FFF8h FFFCh
Note:
Contents EFLAGS CR3 CR0
Size Doubleword Doubleword Doubleword
Allowable Access Read/Write Read-Only
1. The offset for the SMM-revision identifier is compatible with previous implementations.
10.1.4 SMM-Revision Identifier
The SMM-revision identifier specifies the SMM version and the available SMM extensions implemented by the processor. Software reads the SMM-revision identifier from offset FEFCh in the SMM state-save area of SMRAM. This offset location is compatible with earlier versions of SMM. Software must not write to this location. Doing so can produce undefined results. Figure 10-3 shows the format of the SMM-revision identifier.
Reserved Description SMM-Revision Level I/O Instruction Restart SMM Base Relocation
31
Bits 15:0 16 17
18 17 16 15 0
11
SMM-Revision Level
513-251eps
Figure 10-3.
SMM-Revision Identifier The fields within the SMM-revision identifier are: SMM-revision Level--Bits 15-0. Specifies the version of SMM supported by the processor. The SMM-revision level is of the form 0_xx64h, where xx starts with 00 and is incremented for later revisions to the SMM mechanism.
330
Chapter 10: System-Management Mode
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
I/O Instruction Restart--Bit 16. When set to 1, the processor supports restarting I/O instructions that are interrupted by an SMI. This bit is always set to 1 by implementations of the AMD64 architecture. See "I/O Instruction Restart" on page 336 for information on using this feature. SMM Base Relocation--Bit 17. When set to 1, the processor supports relocation of SMRAM. This bit is always set to 1 by implementations of the AMD64 architecture. See "SMBASE Register" on page 323 for information on using this feature. All remaining bits in the SMM-revision identifier are reserved.
10.2
Using SMM
SMM is entered using the system-management interrupt (SMI). SMI is an external non-maskable interrupt that operates differently from and independently of other interrupts. SMI has priority over all other external interrupts, including NMI (see "Priorities" on page 269 for a list of the interrupt priorities). SMIs are disabled when in SMM, which prevents reentrant calls to the SMM handler. When an SMI is received by the processor, the processor stops fetching instructions and waits for currently-executing instructions to complete and write their results. The SMI also waits for all buffered memory writes to update the caches or system memory. When these activities are complete, the processor uses implementation-dependent external signalling to acknowledge back to the system that it has received the SMI.
10.2.1 SystemManagement Interrupt (SMI)
10.2.2 SMM OperatingEnvironment
The SMM operating-environment is similar to real mode, except that the segment limits in SMM are 4 Gbytes rather than 64 Kbytes. This allows an SMM handler to address memory in the range from 0h to 0FFFF_FFFFh. As with real mode, segment-base addresses are restricted to 20 bits in SMM, and the default operand-size and address-size is 16 bits. To address memory locations above 1 Mbyte, the SMM handler must use the 32-bit operand-size-override and address-size-override prefixes. After saving the processor state in the SMRAM state-save area, a processor running in SMM sets the segment-selector registers and control registers into a state consistent with real mode.
Chapter 10: System-Management Mode
331
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Other registers are also initialized upon entering SMM, as shown in Table 10-3. Table 10-3. SMM Register Initialization
Register Selector CS Base Limit Selector DS, ES, FS, GS, SS Base Limit RIP RFLAGS CR0 CR4 DR7 EFER Initial SMM Contents SMBASE right-shifted 4 bits SMBASE FFFF_FFFFh 0000h 0000_0000_0000_0000h FFFF_FFFFh 0000_0000_0000_8000h 0000_0000_0000_0002h PE, EM, TS, PG bits cleared to 0. All other bits are unmodified. 0000_0000_0000_0000h 0000_0000_0000_0400h 0000_0000_0000_0000h
10.2.3 Exceptions and Interrupts
All hardware interrupts are disabled upon entering SMM, but exceptions and software interrupts are not disabled. If necessary, the SMM handler can re-enable hardware interrupts. Software that handles interrupts in SMM should consider the following: SMI--If an SMI occurs while the processor is in SMM, it is latched by the processor. The latched SMI occurs when the processor leaves SMM. NMI--If an NMI occurs while the processor is in SMM, it is latched by the processor, but the NMI handler is not invoked until the processor leaves SMM with the execution of an RSM instruction. A pending NMI causes the handler to be invoked immediately after the RSM completes and before the first instruction in the interrupted program is executed. An SMM handler can unmask NMI interrupts by simply executing an IRET. Upon completion of the IRET
332
Chapter 10: System-Management Mode
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
instruction, the processor recognizes the pending NMI, and transfers control to the NMI handler. Once an NMI is recognized within SMM using this technique, subsequent NMIs are recognized until SMM is exited. Later SMIs cause NMIs to be masked, until the SMM handler unmasks them. Exceptions--Exceptions (internal processor interrupts) are not disabled and can occur while in SMM. Therefore, the SMM-handler software should be written to avoid generating exceptions. Software Interrupts--The software-interrupt instructions (BOUND, INTn, INT3, and INTO) can be executed while in SMM. However, it is not recommended that the SMM handler use these instructions. Maskable Interrupts--RFLAGS.IF is cleared to 0 by the processor when SMM is entered. Software can re-enable maskable interrupts while in SMM, but it must follow the guidelines listed below for handling interrupts. Debug Interrupts--The processor disables the debug interrupts when SMM is entered by clearing DR7 to 0 and clearing RFLAGS.TF to 0. The SMM handler can re-enable the debug facilities while in SMM, but it must follow the guidelines listed below for handling interrupts. INIT--The processor does not recognize INIT while in SMM. Because the RFLAGS.IF bit is cleared when entering SMM, the HLT instruction should not be executed in SMM without first setting the RFLAGS.IF bit to 1. Setting this bit to 1 allows the processor to exit the halt state by using an external maskable interrupt. In the cases where an SMM handler must accept and handle interrupts and exceptions, several guidelines must be followed: Interrupt handlers must be loaded and accessible before enabling interrupts. A real-mode interrupt-vector table located at virtual (linear) address 0 is required. Segments accessed by the interrupt handler cannot have a base address greater than 20 bits because of the real-mode addressing used in SMM. In SMM, the 16-bit value stored in the segment-selector register is left-shifted four bits to form the 20-bit segment-base address, like real mode.
Chapter 10: System-Management Mode
333
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Only the IP (rIP[15:0]) is pushed onto the stack as a result of an interrupt in SMM, because of the real-mode addressing used in SMM. If the SMM handler is interrupted at a codesegment offset above 64 Kbytes, then the return address on the stack must be adjusted by the interrupt-handler, and a RET instruction with a 32-bit operand-size override must be used to return to the SMM handler. If the interrupt-handler is located below 1 Mbyte, and the SMM handler is located above 1 Mbyte, a RET instruction cannot be used to return to the SMM handler. In this case, the interrupt handler can adjust the return pointer on the stack, and use a far CALL to transfer control back to the SMM handler. 10.2.4 Invalidating the Caches The processor can cache SMRAM-memory locations. If the system implements physically separate SMRAM and system memory, it is possible for SMRAM and system memory locations to alias into identical cache locations. In some processor implementations, the cache contents must be written to memory and invalidated when SMM is entered and exited. This prevents the processor from using previously-cached mainmemory locations as aliases for SMRAM-memory locations when SMM is entered, and vice-versa when SMM is exited. Implementations of the AMD64 architecture do not require cache invalidation when entering and exiting SMM. Internally, the processor keeps track of SMRAM and system-memory accesses separately and properly handles situations where aliasing occurs. Cached system memory and SMRAM locations can persist across SMM mode changes. Removal of the requirement to writeback and invalidate the cache simplifies SMM entry and exit and allows SMM code to execute more rapidly. 10.2.5 Saving Additional Processor State Several registers are not saved or restored automatically by the SMM mechanism. These are: The 128-bit media instruction registers. The 64-bit media instruction registers. The x87 floating-point registers. The page-fault linear-address register (CR2). The task-priority register (CR8). The debug registers, DR0, DR1, DR2, and DR3.
334
Chapter 10: System-Management Mode
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The memory-type range registers (MTRRs). Model-specific registers (MSRs). These registers are not saved because SMM handlers do not normally use or modify them. If an SMI results in a processor reset (due to powering down the processor, for example) or the SMM handler modifies the contents of the unsaved registers, the SMM handler should save and restore the original contents of those registers. The unsaved registers, along with those stored in the SMRAM state-save area, need to be saved in a nonvolatile storage location if a processor reset occurs. The SMM handler should execute the CPUID instruction to determine the feature set available in the processor, and be able to save and restore the registers required by those features. The SMM handler can execute any of the 128-bit media, 64-bit media, or x87 instructions. A simple method for saving and restoring those registers is to use the FXSAVE and FXRSTOR instructions, respectively, if it is supported by the processor. See "Saving Media and x87 Processor State" on page 346 for information on saving and restoring those registers. Floating-point exceptions can occur when the SMM handler uses media or x87 floating-point instructions. If the SMM handler uses floating-point exception handlers, they must follow the usage guidelines established in "Exceptions and Interrupts" on page 332. A simple method for dealing with floating-point exceptions while in SMM is to simply mask all exception conditions using the appropriate floating-point control register. When the exceptions are masked, the processor handles floating-point exceptions internally in a default manner, and allows execution to continue uninterrupted. 10.2.6 Operating in Protected Mode and Long Mode Software can enable protected mode from SMM and it can also enable and activate long mode. An SMM handler can use this capability to enter 64-bit mode and save additional processor state that cannot be accessed from outside 64-bit mode (for example, the most-significant 32 bits of CR2). The auto-halt restart entry is located at offset FEC9h in the SMM state-save area. The size of this field is one byte, as compared with two bytes in previous versions of SMM.
10.2.7 Auto-Halt Restart
Chapter 10: System-Management Mode
335
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
When entering SMM, the processor loads the auto-halt restart entry to indicate whether SMM was entered from the halt state, as follows: Bit 0 indicates the processor state upon entering SMM: - When set to 1, the processor entered SMM from the halt state. - When cleared to 0, the processor did not enter SMM from the halt state. Bits 7-1 are cleared to 0. The SMM handler can write the auto-halt restart entry to specify whether the return from SMM should take the processor back to the halt state or to the instruction-execution state specified by the SMM state-save area. The values written are: Clear to 00h--The processor returns to the state specified by the SMM state-save area. Set to any non-zero value--The processor returns to the halt state. If the return from SMM takes the processor back to the halt state, the HLT instruction is not re-executed. However, the halt special bus-cycle is driven on the processor bus after the RSM instruction executes. The result of entering SMM from a non-halt state and returning to a halt state is not predictable. 10.2.8 I/O Instruction Restart The I/O-instruction restart entry is located at offset FEC8h in the SMM state-save area. The size of this field is one byte, as compared with two bytes in previous versions of SMM. The I/Oinstruction restart mechanism is supported when the I/Oinstruction restart bit (bit 16) in the SMM-revision identifier is set to 1. This bit is always set to 1 in the AMD64 architecture. When an I/O instruction is interrupted by an SMI, the I/Oinstruction restart entry specifies whether the interrupted I/O instruction should be re-executed following an RSM that returns from SMM. Re-executing a trapped I/O instruction is useful, for example, when an I/O write is performed to a powered-down disk drive. When this occurs, the system logic monitoring the access can issue an SMI to have the SMM handler power-up the disk drive and retry the I/O write. The SMM handler does this by querying system logic and detecting the failed I/O write, asking system logic to initiate the disk336 Chapter 10: System-Management Mode
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
drive power-up sequence, enabling the I/O instruction restart mechanism, and returning from SMM. Upon returning from SMM, the I/O write to the disk drive is restarted. When an SMI occurs, the processor always clears the I/Oinstruction restart entry to 0. If the SMI interrupted an I/O instruction, then the SMM handler can modify the I/Oinstruction restart entry as follows: Clear to 00h (default value)--The I/O instruction is not restarted, and the instruction following the interrupted I/Oinstruction is executed. When a REP (repeat) prefix is used with an I/O instruction, it is possible that the next instruction to be executed is the next I/O instruction in the repeat loop. Set to any non-zero value--The I/O instruction is restarted. While in SMM, the handler must determine the cause of the SMI and examine the processor state at the time the SMI occurred to determine whether or not an I/O instruction was interrupted. Writing a non-zero value into the I/O-instruction restart entry when the interrupted instruction is a non-I/O instruction produces undefined results. Some implementations provide model-specific facilities to assist in this determination. For more information, refer to the BIOS writer's guide for the implementation.
10.3
Leaving SMM
Software leaves SMM and returns to the interrupted program by executing the RSM instruction. RSM causes the processor to load the interrupted state from the SMRAM state-save area and then transfer control back to the interrupted program. RSM cannot be executed in any mode other than SMM, otherwise an invalid-opcode exception (#UD) occurs. An RSM causes a processor shutdown if an invalid-state condition is found in the SMRAM state-save area. Only an external reset, external processor-initialization, or nonmaskable external interrupt (NMI) can cause the processor to leave the shutdown state. The invalid SMRAM state-save-area conditions that can cause a processor shutdown during an RSM are: CR0.PE=0 and CR0.PG=1.
Chapter 10: System-Management Mode
337
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
CR0.CD=0 and CR0.NW=1. Certain reserved bits are set to 1, including: - Any CR0 bit in the range 63-32 is set to 1. - Any unsupported bit in CR3 is set to 1. - Any unsupported bit in CR4 is set to 1. - Any DR6 bit or DR7 bit in the range 63-32 is set to 1. - Any unsupported bit in EFER is set to 1. Invalid returns to long mode, including: - EFER.LME=1, CR0.PG=1, and CR4.PAE=0. - EFER.LME=1, CR0.PG=1, CR4.PAE=1, CS.L=1, and CS.D=1. The SSM revision identifier is modified. Some SMRAM state-save-area conditions are ignored, and the registers, or bits within the registers, are restored in a default manner by the processor. This avoids a processor shutdown when an invalid condition is stored in SMRAM. The default conditions restored by the processor are: The EFER.LMA register bit is set to the value obtained by logically ANDing the SMRAM values of EFER.LME, CR0.PG, and CR4.PAE. The rFLAGS.VM register bit is set to the value obtained by logically ANDing the SMRAM values of rFLAGS.VM, CR0.PE, and the inverse of EFER.LMA. The base values of FS, GS, GDTR, IDTR, LDTR, and TR are restored in canonical form. Those values are sign-extended to bit 63 using the most-significant implemented bit. Unimplemented segment-base bits in the CS, DS, ES, and SS registers are cleared to 0.
338
Chapter 10: System-Management Mode
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
11
128-Bit, 64-Bit, and x87 Programming
This chapter describes the system-software implications of supporting applications that use the 128-bit media, 64-bit media, and x87 instructions. Throughout this chapter, these instructions are collectively referred to as media and x87 (media/x87) instructions. A complete listing of the instructions that fall in this category--and the detailed operation of each instruction--can be found in volumes 4 and 5. Refer to Volume 1 for information on using these instructions in application software.
11.1
Overview of System-Software Considerations
Processor implementations can support different combinations of the 128-bit media, 64-bit media, and x87 instruction sets. Two sets of registers--independent of the general-purpose registers--support these instructions. The 128-bit media instructions operate on the XMM registers, and the 64-bit media and x87-instructions operate on the aliased MMX/x87 registers. The 128-bit media and x87 floating-point instruction sets have special status registers, control registers, exception vectors, and system-software control bits for managing the operating environment. System software that supports use of these instructions must be able to manage these resources properly including: Detecting support for the instruction set, and enabling any optional features, as necessary. Saving and restoring the processor media or x87 state. Execution of floating-point instructions (media or x87) can produce exceptions. System software must supply exception handlers for all unmasked floating-point exceptions.
11.2
Determining Media and x87 Feature Support
The support of 128-bit media, 64-bit media, and x87 instructions is implementation dependent. System software executes the CPUID instruction to determine whether a processor implements any of these features (see "Processor Feature Identification" on page 78 for more information on using the CPUID instruction). After CPUID is executed with an extended
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
339
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
function codes 1 and 8000_0001h, feature support can be determined by examining the contents of the EDX register. General guidelines for determining feature support are given in the list below. A few instructions belong to more than one instruction subset. Refer to "Instruction Subsets and CPUID Feature Sets" in Volume 3 for specific information. 128-bit media instructions are supported when: - EDX[25]=1 for SSE instructions. (Returned by CPUID function code 1.) - EDX[26]=1 for SSE2 instructions. (Returned by CPUID function code 1.) 64-bit media instructions are supported when: - EDX[23]=1 for MMX instructions. (Returned by CPUID function codes 1h and 8000_0001h.) - EDX[22]=1 for AMD extensions to MMXTM instructions. (Returned by CPUID function code 8000_0001h.) - EDX[31]=1 for AMD 3DNow!TM instructions. (Returned by CPUID function code 8000_0001h.) - EDX[30]=1 for AMD extensions to 3DNow! instructions. (Returned by CPUID function code 8000_0001h.) x87 floating-point instructions are supported when: - EDX[0]=1. (Returned by function codes 1h and 8000_0001h.) FXSAVE and FXRSTOR instructions are supported when: - EDX[24]=1. These instructions save and restore the entire media and x87 processor state. (Returned by function codes 1h and 8000_0001h.) If software attempts to execute an instruction belonging to an unsupported instruction subset, an invalid-opcode exception (#UD) occurs. For a summary of instruction subsets, see "Instruction Subsets and CPUID Feature Sets" in Volume 3.
11.3
Enabling 128-Bit Media Instructions
Use of the 128-bit media instructions requires system software to support SSE and/or SSE2 features, but also the FXSAVE and FXRSTOR instructions, which are used to save and restore the 128-bit media state (see "FXSAVE and FXRSTOR Instructions" on page 351). When these instructions are supported, system
340
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
software must set CR4.OSFXSR=1 to let the processor know that the software uses these instructions. When the processor detects CR4.OSFXSR=1, it allows execution of the 128-bit m e d i a i n s t r u c t i o n s . I f s y s t e m s o f t wa re d o e s n o t s e t CR4.OSFXSR to 1, attempts to execute 128-bit media instructions cause an invalid-opcode exception (#UD). System software must also clear the CR0.EM (emulate coprocessor) bit to 0, otherwise an attempt to execute a 128-bit media instruction causes a #UD exception. System software should also set the CR0.MP (monitor coprocessor) bit to 1. When CR0.EM=0 and CR0.MP=1, all media instructions, x87 instructions, and the FWAIT/WAIT instructions cause a device-not-available exception (#NM) when the CR0.TS bit is set. System software can use the #NM exception to perform lazy context switching, saving and restoring media and x87 state only when necessary after a task switch. See "CR0 Register" on page 55 for more information. System software must supply an exception handler if unmasked 128-bit media floating-point exceptions are allowed to occur. When an unmasked exception is detected, the processor transfers control to the SIMD floating-point exception (#XF) handler provided by the operating system. System software must let the processor know that the #XF handler is available by setting CR4.OSXMMEXCPT to 1. If this bit is set to 1, the processor transfers control to the #XF handler when it detects an unmasked exception, otherwise a #UD exception occurs. When the processor detects a masked exception, it handles it in a default manner regardless of the CR4.OSXMMEXCPT value.
11.4
Media and x87 Processor State
The media and x87 processor state includes the contents of the registers used by 128-bit media, 64-bit media, and x87 instructions. System software that supports such applications must be capable of saving and restoring these registers.
11.4.1 128-Bit Media State
Figure 11-1 on page 342 shows the registers whose contents are affected by execution of 128-bit media instructions. These include:
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
341
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
xmm0-xmm15--Sixteen 128-bit media registers. In legacy and compatibility modes, software access is limited to the first eight registers, XMM0-XMM7. MXCSR--The 32-bit control and status register. All of the above registers are visible to 128-bit media application software. Refer to "128-Bit Media and Scientific Programming" in Volume 1 for more information on these registers.
XMM Data Registers
127 0
xmm0 xmm1 xmm2 xmm3 xmm4 xmm5 xmm6 xmm7 xmm8 xmm9 xmm10 xmm11 xmm12 xmm13 xmm14 xmm15
128-Bit Media Control and Status Register
31
MXCSR
0
513-270.eps
Figure 11-1. 128-Bit Media-Instruction State
342
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
11.4.2 64-Bit Media State
Figure 11-2 shows the register contents that are affected by execution of 64-bit media instructions. These registers include: mmx0-mmx7--Eight 64-bit media registers. FSW--Two fields (TOP and ES) in the 16-bit x87 status word register. FTW--The 16-bit x87 tag word.
MMX Data Registers
79 64 63 0
mmx0 mmx1 mmx2 mmx3 mmx4 mmx5 mmx6 mmx7 TOP ES
Visible to application software Written by processor hardware
fpr0 fpr1 fpr2 fpr3 fpr4 fpr5 fpr6 fpr7
x87 Status Word x87 Tag Word
15 0
FSW FTW
513-272.eps
Figure 11-2.
64-Bit Media-Instruction State
Th e 6 4 - b i t m e d i a i n s t r u c t i o n s a n d x 8 7 f l o a t i n g - p o i n t instructions share the same physical data registers. Figure 11-2 shows how the 64-bit registers (MMX0-MMX7) are aliased onto the low 64 bits of the 80-bit x87 floating-point physical data registers (FPR0-FPR7). Refer to "64-Bit Media Programming" in Volume 1 for more information on these registers. Of the registers shown in Figure 11-2, only the eight 64-bit MMX registers are visible to 64-bit media application software. The processor maintains the contents of the two fields of the x87 status word--top-of-stack-pointer (TOP) and exception Chapter 11: 128-Bit, 64-Bit, and x87 Programming 343
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
summary (ES)--and the 16-bit x87 tag word during execution of 64-bit media instructions, as described in "Actions Taken on Executing 64-Bit Media Instructions" in Volume 1. 64-bit media instructions do not generate x87 floating-point exceptions, nor do they set any status flags. However, 64-bit media instructions can trigger an unmasked floating-point exception caused by a previously executed x87 instruction. 64bit media instructions do this by reading the x87 FSW.ES bit to determine whether such an exception is pending. 11.4.3 x87 State Figure 11-3 on page 345 shows the registers whose contents are affected by execution of x87 floating-point instructions. These registers include: fpr0-fpr7--Eight 80-bit floating-point physical registers. FCW--The 16-bit x87 control word register. FSW--The 16-bit x87 status word register. FTW--The 16-bit x87 tag word. Last x87 Instruction Pointer--This value is a pointer (32-bit, 48-bit, or 64-bit, depending on effective operand size and mode) to the last non-control x87 floating-point instruction executed. Last x87 Data Pointer--The pointer (32-bit, 48-bit, or 64-bit, depending on effective operand size and mode) to the data operand referenced by the last non-control x87 floatingpoint instruction executed, if that instruction referenced memory; if it did not, then this value is implementation dependent. Last x87 Opcode--An 11-bit permutation of the instruction opcode from the last non-control x87 floating-point instruction executed. Of the registers shown in Figure 11-3, only FPR0-FPR7, FCW, and FSW are directly updated by x87 application software. The processor maintains the contents of the FTW, instruction and data pointers, and opcode registers during execution of x87 instructions. Refer to "Registers" in Volume 1 for more information on these registers. The 11-bit instruction opcode register holds a permutation of the two-byte instruction opcode from the last non-control x87 instruction executed by the processor. (For a definition of non-
344
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
control x87 instruction, see "Control" in Volume 1.) The opcode field is formed as follows: Opcode Register Field[10:8] = First x87 opcode byte[2:0]. Opcode Register Field[7:0] = Second x87 opcode byte[7:0]. For example, the x87 opcode D9 F8h is stored in the opcode register as 001_1111_1000b. The low-order three bits of the first opcode byte, D9h (1101_1001b), are stored in opcode-register bits 10-8. The second opcode byte, F8h (1111_1000b), is stored in bits 7-0 of the opcode register. The high-order five bits of the first opcode byte (1101_1b) are not needed because they are identical for all x87 instructions.
x87 Data Registers
79 0
fpr0 fpr1 fpr2 fpr3 fpr4 fpr5 fpr6 fpr7
x87 Control Word x87 Status Word x87 Tag Word
15
FCW Control Word FSW Status Word FTW Tag Word
0
Last x87 Instruction Pointer Last x87 Data Pointer
63
Opcode
10 0
513-271.eps
Figure 11-3.
x87-Instruction State 345
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
11.4.4 Saving Media and x87 Processor State
In most cases, operating systems, exception handlers, and device drivers should save and restore the media and/or x87 processor state between task switches or other interventions in the execution of 128-bit, 64-bit, or x87 procedures. Application programs are also free to save and restore state at any time. In general, system software should use the FXSAVE and FXRSTOR instructions to save and restore the entire media and x87 processor state. The FSAVE/FNSAVE and FRSTOR instructions can be used for saving and restoring the x87 state. Because the 64-bit media registers are physically aliased onto t h e x 8 7 re g i s t e rs , t h e F S AV E / F N S AV E a n d F R S TO R instructions can also be used to save and restore the 64-bit media state. However, FSAVE/FNSAVE and FRSTOR do not save or restore the 128-bit media state. FSAVE/FNSAVE and FRSTOR Instructions. The FSAVE/FNSAVE and FRSTOR instructions save and restore the entire register state fo r 6 4 - b i t m e d i a i n s t r u c t i o n s a n d x 8 7 f l o a t i n g - p o i n t instructions. The FSAVE instruction stores the register state, but only after handling any pending unmasked-x87 floatingpoint exceptions. The FNSAVE instruction stores the register state but skips the reporting and handling of these exceptions. The state of all MMX/FPR registers is saved, as well as all other x87 state (the control word register, status word register, tag word, instruction pointer, data pointer, and last opcode). After saving this state, the tag state for all MMX/FPR registers is changed to empty and is thus available for a new procedure. Starting on page 348, Figure 11-4 through Figure 11-7 show the memory formats used by the FSAVE/FNSAVE and FRSTOR instructions when storing the x87 state in various processor modes and using various effective-operand sizes. This state includes: x87 Data Registers - FPR0-FPR7 80-bit physical data registers. x87 Environment - FCW: x87 control word register - FSW: x87 status word register - FTW: x87 tag word - Last x87 instruction pointer - Last x87 data pointer
346
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
-
Last x87 opcode
The eight data registers are stored in the 80 bytes following the environment information. Instead of storing these registers in their physical order (FPR0-FPR7), the processor stores the registers in the their stack order, ST(0)-ST(7), beginning with the top-of-stack, ST(0).
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
347
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Bit Offset 31 16 15 ST(7) (79-48) ... ST(1) (15-0) ST(0) (63-32) ST(0) (31-0) Reserved, IGN Data Offset (31-0) 00000b Instruction Opcode (10-0) Instruction Offset (31-0) Reserved, IGN x87 Tag Word (FTW) Instruction CS Selector (15-0) Data DS Selector (15-0) ST(0) (79-64)
Byte 0 Offset +68h
...
...
...
+1Ch
+18h
+14h
+10h
+0Ch
+08h
Reserved, IGN
x87 Status Word (FSW)
+04h
Reserved, IGN
x87 Control Word (FCW)
+00h
Figure 11-4.
FSAVE/FNSAVE Image (32-Bit, Protected Mode)
348
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Bit Offset 31 16 15 ST(7) (79-48) ... ST(1) (15-0) ST(0) (63-32) ST(0) (31-0) 0000b Data Offset (31-16) Reserved, IGN Instruction Offset (31-16) Reserved, IGN 0000 0000 0000b Data Offset (15-0) 0 Instruction Opcode (10-0) ST(0) (79-64)
Byte 0 Offset +68h
...
...
...
+1Ch
+18h
+14h
0000b
+10h
Instruction Offset (15-0) x87 Tag Word (FTW)
+0Ch
Reserved, IGN
+08h
Reserved, IGN
x87 Status Word (FSW)
+04h
Reserved, IGN
x87 Control Word (FCW)
+00h
Figure 11-5.
FSAVE/FNSAVE Image (32-Bit, Real/Virtual-8086 Modes)
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
349
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Bit Offset 31 Not Part of x87 State 16 15 ST(7) (79-64) ... ST(0) (79-48) ST(0) (47-16) ST(0) (15-0) Data Offset (15-0) Instruction Offset (15-0) x87 Status Word (FSW) Data DS Selector (15-0) Instruction CS Selector (15-0) x87 Tag Word (FTW)
Byte 0 Offset +5Ch
...
+14h
+10h
+0Ch
+08h
+04h
x87 Control Word (FCW)
+00h
Figure 11-6.
FSAVE/FNSAVE Image (16-Bit, Protected Mode)
350
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Bit Offset 31 Not Part of x87 State 16 15 ST(7) (79-64) ... ST(0) (79-48) ST(0) (47-16) ST(0) (15-0) Data Offset (15-0) Data (19-16) Instruc. 0 (19-16) 0000 0000 0000b Instruction Opcode (10-0)
Byte 0 Offset +5Ch
...
+14h
+10h
+0Ch
+08h
Instruction Offset (15-0)
x87 Tag Word (FTW)
+04h
x87 Status Word (FSW)
x87 Control Word (FCW)
+00h
Figure 11-7.
FSAVE/FNSAVE Image (16-Bit, Real/Virtual-8086 Modes)
FLDENV/FNLDENV and FSTENV Instructions. The FLDENV/FNLDENV and FSTENV instructions load and store only the x87 floatingp o i n t e n v i r o n m e n t . T h e s e i n s t r u c t i o n s , u n l i ke t h e FSAVE/FNSAVE and FRSTOR instructions, do not save or restore the x87 data registers. FXSAVE and FXRSTOR Instructions. Th e F X S AV E a n d F X R S TO R instructions save and restore the entire 128-bit media, 64-bit media, and x87 state. These instructions usually execute faster than FSAVE/FNSAVE and FRSTOR because they do not normally save and restore the x87 exception pointers (lastinstruction pointer, last data-operand pointer, and last opcode). Chapter 11: 128-Bit, 64-Bit, and x87 Programming 351
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
The only case in which they do save the exception pointers is the relatively rare case in which the exception-summary bit in the x87 status word (FSW.ES) is set to 1, indicating that an unmasked exception has occurred. The FXSAVE and FXRSTOR memory format contains fields for storing these values. Unlike FSAVE and FNSAVE, the FXSAVE instruction does not alter the x87 tag word. Therefore, the contents of the shared 64bit MMX and 80-bit FPR registers can remain valid after an FXSAVE instruction (or any other value the tag bits indicated before the save). Also, FXSAVE (like FNSAVE) does not check for pending unmasked-x87 floating-point exceptions. Figure 11-8 on page 353 shows the memory format of the media x87 state in long mode. When in long mode using a 64-bit operand size, the format shown in Figure 11-8 is used. If a 32-bit operand size is used (in long mode), the memory format is the same, except that RIP and RDS are stored as sel:offset pointers, as shown in Figure 11-9 on page 354.
352
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
F
E
D
C
B
A
9
8
7
6
5
4
3
2
1
0
Byte
Reserved, IGN ... Reserved, IGN XMM15 XMM14 XMM13 XMM12 XMM11 XMM10 XMM9 XMM8 XMM7 XMM6 XMM5 XMM4 XMM3 XMM2 XMM1 XMM0 Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN MXCSR_MASK RIP1 1. Stored as sel:offset if operand size is 32 bits. MXCSR FOP 0 MMX7/ST(7) MMX6/ST(6) MMX5/ST(5) MMX4/ST(4) MMX3/ST(3) MMX2/ST(2) MMX1/ST(1) MMX0/ST(0) RDP1 FTW FSW FCW
+1F0h ... +1A0h +190h +180h +170h +160h +150h +140h +130h +120h +110h +100h +F0h +E0h +D0h +C0h +B0h +A0h +90h +80h +70h +60h +50h +40h +30h +20h +10h +00h
Figure 11-8.
FXSAVE and FXRSTOR Image (Long Mode)
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
353
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
F
E
D
C
B
A
9
8 ...
7
6
5
4
3
2
1
0
Byte +1F0h ... +120h +110h +100h +F0h +E0h +D0h +C0h +B0h +A0h
Reserved, IGN Reserved, IGN XMM7 XMM6 XMM5 XMM4 XMM3 XMM2 XMM1 XMM0 Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN MXCSR_MASK rsrvd, IGN CS MXCSR EIP rsrvd, IGN FOP 0 MMX7/ST(7) MMX6/ST(6) MMX5/ST(5) MMX4/ST(4) MMX3/ST(3) MMX2/ST(2) MMX1/ST(1) MMX0/ST(0) DS FTW FSW DP FCW
+90h +80h +70h +60h +50h +40h +30h +20h +10h +00h
Figure 11-9.
FXSAVE and FXRSTOR Image (Legacy Mode) Software can read and write all fields within the FXSAVE and FXRSTOR memory image. These fields include: FCW--Bytes 01h-00h. x87 control word. FSW--Bytes 03h-02h. x87 status word. FTW--Byte 04h. x87 tag word. See "FXSAVE Format for x87 Tag Word" on page 356 for additional information on the FTW format saved by the FXSAVE instruction. (Byte 05h contains the value 00h.)
354
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
FOP--Bytes 07h-06h. last x87 opcode. Last x87 Instruction Pointer--A pointer to the last noncontrol x87 floating-point instruction executed by the processor: - RIP (64-bit format)--Bytes 0Fh-08h. 64-bit offset into the code segment (used without a CS selector). - EIP (32-bit format)--Bytes 0Bh-08h. 32-bit offset into the code segment. - CS (32-bit format)--Bytes 0Dh-0Ch. Segment selector portion of the pointer. Last x87 Data Pointer--If the last non-control x87 floating point instruction referenced memory, this value is a pointer to the data operand referenced by the last non-control x87 floating-point instruction executed by the processor: - RDP (64-bit format)--Bytes 17h-10h. 64-bit offset into the data segment (used without a DS selector). - DP (32-bit format)--Bytes 13h-10h. 32-bit offset into the data segment. - DS (32-bit format)--Bytes 15h-14h. Segment selector portion of the pointer. If the last non-control x87 instruction did not reference memory, then the value in the pointer is implementation dependent. MXCSR--Bytes 1Bh-18h. 128-bit media-instruction control and status register. This register is saved only if CR4.OSFXSR is set to 1. MXCSR_MASK--Bytes 1Fh-1Ch. Set bits in MXCSR_MASK indicate supported feature bits in MXCSR. For example, if bit 6 (the DAZ bit) in the returned MXCSR_MASK field is set to 1, the DAZ mode and the DAZ flag in MXCSR are supported. Cleared bits in MXCSR_MASK indicate reserved bits in MXCSR. If software attempts to set a reserved bit in the MXCSR register, a #GP exception will occur. To avoid this exception, after software clears the FXSAVE memory image and executes the FXSAVE instruction, software should use the value returned by the processor in the MXCSR_MASK field when writing a value to the MXCSR register, as follows: . MXCSR_MASK = 0: If the processor writes a zero value into the MXCSR_MASK field, the denormals-arezeros (DAZ) mode and the DAZ flag in MXCSR are not Chapter 11: 128-Bit, 64-Bit, and x87 Programming 355
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
supported. Software should use the default mask value, 0000_FFBFh (bit 6, the DAZ bit, and bits 31-16 cleared to 0), to mask any value it writes to the MXCSR register to ensure that all reserved bits in MXCSR are written with 0, thus avoiding a #GP exception. . MXCSR_MASK 0: If the processor writes a non-zero value into the MXCSR_MASK field, software should AND this value with any value it writes to the MXCSR register. MMXn/FPRn--Bytes 9Fh-20h. Shared 64-bit media and x87 floating-point registers. As in the case of the x87 FSAVE instruction, these registers are stored in stack order ST(0)/MMX0-ST(7)/MMX7. The upper six bytes in the memory image for each register are reserved. XMMn--Bytes 11Fh-A0h. 128-bit media registers. These registers are saved only if CR4.OSFXSR is set to 1. FXSAVE Format for x87 Tag Word. Rather than saving the entire x87 tag word, FXSAVE saves a single-byte encoded version. FXSAVE encodes each of the eight two-bit fields in the x87 tag word as follows: Two-bit values of 00, 01, and 10 are encoded as a 1, indicating the corresponding x87 FPRn register holds a value. A two-bit value of 11 is encoded as a 0, indicating the corresponding x87 FPRn is empty. For example, assume an FSAVE instruction saves an x87 tag word with the value 83F1h. This tag-word value describes the x87 FPRn contents as follows:
x87 Register Tag Word Value (hex) Tag Value (binary) Meaning
FPR7 8 10 Special
FPR6
FPR5 3
FPR4
FPR3 F
FPR2
FPR1 1
FPR0
00 Valid
00 Valid
11 Empty
11 Empty
11 Empty
00 Valid
01 Zero
When an FXSAVE is used to write the x87 tag word to memory, it encodes the value as E3h. This encoded version describes the x87 FPRn contents as follows: 356 Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
x87 Register Encoded Tag Byte (hex) Tag Value (binary) Meaning
FPR7
FPR6 E
FPR5
FPR4
FPR3
FPR2 3
FPR1
FPR0
1 Valid
1 Valid
1 Valid
0 Empty
0 Empty
0 Empty
1 Valid
1 Valid
If necessary, software can decode the single-bit FXSAVE tagword fields into the two-bit field FSAVE uses by examining the contents of the corresponding FPR registers saved by FXSAVE. Table 11-1 shows how the FPR contents are used to find the equivalent FSAVE tag-field value. The fraction column refers to fraction portion of the extended-precision significand (bits 62-0). The integer bit column refers to the integer-portion of the significand (bit 63). See "x87 Floating-Point Programming" in Volume 1 for more information on floating-point numbering formats. Table 11-1. Deriving FSAVE Tag Field from FXSAVE Tag Field
Encoded FXSAVE Tag Field Exponent Integer Bit2 0 All 0s 0 1 1 1 (Valid) Neither all 0s nor all 1s 0 1 0 All 1s 1 0 (Empty)
Notes:
Fraction1 All 0s Not all 0s All 0s Not all 0s
Type of Value Zero Denormal Pseudo Denormal Unnormal
Equivalent FSAVE Tag Field 01 (Zero)
10 (Special)
don't care
Normal Pseudo Infinity or Pseudo NaN
00 (Valid)
All 0s Not all 0s
Infinity NaN Empty
10 (Special)
don't care
11 (Empty)
1. Bits 62-0 of the significand. Bit 62, the most-significant bit of the fraction, is also called the M bit. 2. Bit 63 of the significand, also called the J bit.
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
357
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Performance Considerations. When system software supports multitasking, it must be able to save the processor state for one task and load the state for another. For performance reasons, the media and/or x87 processor state is usually saved and loaded only when necessary. System software can save and load this state at the time a task switch occurs. However, if the new task does not use the state, loading the state is unnecessary and reduces performance. The task-switch bit (CR0.TS) is provided as a lazy contextswitch mechanism that allows system software to save and load the processor state only when necessary. When CR0.TS=1, a device-not-available exception (#NM) occurs when an attempt is made to execute a 128-bit media, 64-bit media, or x87 instruction. System software can use the #NM exception handler to save the state of the previous task, and restore the state of the current task. Before returning from the exception handler to the media or x87 instruction, system software must clear CR0.TS to 0 to allow the instruction to be executed. Using this approach, the processor state is saved only when the registers are used. In legacy mode, the hardware task-switch mechanism sets CR0.TS=1 during a task switch (see "Task Switched (TS) Bit" on page 56 for more information). In long mode, the hardware taskswitching is not supported, and the CR0.TS bit is not set by the processor. Instead, the architecture assumes that system software handles all task-switching and state-saving functions. If CR0.TS is to be used in long mode for controlling the save and restore of media or x87 state, system software must set and clear it explicitly.
358
Chapter 11: 128-Bit, 64-Bit, and x87 Programming
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
12
Task Management
This chapter describes the hardware task-management features. All of the legacy x86 task-management features are supported by the AMD64 architecture in legacy mode, but most features are not available in long mode. Long mode, however, requires system software to initialize and maintain certain taskmanagement re sources . The details of these resourceinitialization requirements for long mode are discussed in "Task-Management Resources" on page 360.
12.1
Hardware Multitasking Overview
A task (also called a process) is a program that the processor can execute, suspend, and later resume executing at the point of suspension. During the time a task is suspended, other tasks are allowed to execute. Each task has its own execution space, consisting of: Code segment and instruction pointer. Data segments. Stack segments for each privilege level. General-purpose registers. rFLAGS register. Local-descriptor table. Task register, and a link to the previously-executed task. I/O-permission and interrupt-permission bitmaps. Pointer to the page-translation tables (CR3). The state information defining this execution space is stored in the task-state segment (TSS) maintained for each task. Support for hardware multitasking is provided in legacy mode. Hardware multitasking provides automated mechanisms for switching tasks, saving the execution state of the suspended task, and restoring the execution state of the resumed task. When hardware multitasking is used to switch tasks, the processor takes the following actions: Suspends execution of the task, allowing any executing instructions to complete and save their results. Saves the task execution state in the task TSS.
Chapter 12: Task Management
359
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Loads the execution state for the new task from its TSS. Begins executing the new task at the location specified in the new task TSS. Software can switch tasks by branching to a new task using the CALL or JMP instructions. Exceptions and interrupts can also switch tasks if the exception or interrupt handlers are themselves separate tasks. IRET can be used to return to an earlier task.
12.2
Task-Management Resources
The hardware-multitasking features are available when protected mode is enabled (CR0.PE=1). Protected-mode software execution, by definition, occurs as part of a task. While s y s t e m s o f t wa re i s n o t re q u i re d t o u s e t h e h a rdwa re multitasking features, it is required to initialize certain taskmanagement resources for at least one task (the current task) when running in protected mode. This single task is needed to establish the protected-mode execution environment. The resources that must be initialized are: Task-State Segment (TSS)--A segment that holds the processor state associated with a task. TSS Descriptor--A segment descriptor that defines the taskstate segment. TSS Selector--A segment selector that references the TSS descriptor located in the GDT. Task Register--A register that holds the TSS selector and TSS descriptor for the current task. Figure 12-1 on page 361 shows the relationship of these resources to each other in both 64-bit and 32-bit operating environments.
360
Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Global-Descriptor Table
TSS Descriptor
Task Register (Visible)
15 0
Task Register (Hidden From Software)
TSS Selector
64-Bit or 32-Bit Base Address
32-Bit Limit
Attributes
+
Task-State Segment
I/O-Permission Bitmap Interrupt-Redirection Bitmap
I/O-Bitmap Base Address
513-254.eps
Figure 12-1. Task-Management Resources A fifth resource is available in legacy mode for use by system software that uses the hardware-multitasking mechanism to manage more than one task:
Chapter 12: Task Management
361
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Task-Gate Descriptor--This form of gate descriptor holds a reference to a TSS descriptor and is used to control access between tasks. The task-management resources are described in the following sections. 12.2.1 TSS Selector TSS selectors are selectors that point to task-state segment descriptors in the GDT. Their format is identical to all other segment selectors, as shown in Figure 12-2.
15 Selector Index 3 2
T I
1
0
RPL
Bits 15-3 2 1-0
Mnemonic TI RPL
Description Selector Index Table Indicator Requestor Privilege Level
Figure 12-2. Task-Segment Selector The selector format consists of the following fields: Selector Index. Bits 15-3. The selector-index field locates the TSS descriptor in the global-descriptor table. Table Indicator (TI) Bit. Bit 2. The TI bit must be cleared to 0, which indicates that the GDT is used. TSS descriptors cannot be located in the LDT. If a reference is made to a TSS descriptor in the LDT, a general-protection exception (#GP) occurs. Requestor Privilege-Level (RPL) Field. Bits 1-0. RPL represents the privilege level (CPL) the processor is operating under at the time the TSS selector is loaded into the task register. 12.2.2 TSS Descriptor The TSS descriptor is a system-segment descriptor, and it can be located only in the GDT. The format for an 8-byte, legacymode and compatibility-mode TSS descriptor can be found in "System Descriptors" on page 104. The format for a 16-byte, 64bit mode TSS descriptor can be found in "System Descriptors" on page 111. The fields within a TSS descriptor (all modes) are described in "Descriptor Format" on page 97. The following additional information applies to TSS descriptors:
362
Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Segment Limit--A TSS descriptor must have a segment limit value of at least 67h, which defines a minimum TSS size of 68h (104 decimal) bytes. If the limit is less than 67h, an invalid-TSS exception (#TS) occurs during the task switch. When an I/O-permission bitmap, interrupt-redirection bitmap, or additional state information is included in the TSS, the limit must be set to a value large enough to enclose that information. In this case, if the TSS limit is not large enough to hold the additional information, a #GP exception occurs when an attempt is made to access beyond the TSS limit. No check for the larger limit is performed during the task switch. Type--Four system-descriptor types are defined as TSS types, as shown in Table 4-5 on page 105. Bit 9 is used as the descriptor busy bit (B). This bit indicates that the task is busy when set to 1, and available when cleared to 0. Busy tasks are the currently running task and any previous (outer) tasks in a nested-task hierarchy. Task recursion is not supported, and a #GP exception occurs if an attempt is made to transfer control to a busy task. See "Nesting Tasks" on page 379 for additional information. In long mode, the 32-bit TSS types (available and busy) are redefined as 64-bit TSS types, and only 64-bit TSS descriptors can be used. Loading the task register with an available 64-bit TSS causes the processor to change the TSS descriptor type to indicate a busy 64-bit TSS. Because long mode does not support task switching, the TSS-descriptor busy bit is never cleared by the processor to indicate an available 64-bit TSS. Sixteen-bit TSS types are illegal in long mode. A generalprotection exception (#GP) occurs if a reference is made to a 16-bit TSS. 12.2.3 Task Register The task register (TR) points to the TSS location in memory, defines its size, and specifies its attributes. As with the other descriptor-table registers, the TR has two portions. A visible portion holds the TSS selector, and a hidden portion holds the TSS descriptor. When the TSS selector is loaded into the TR, the processor automatically loads the TSS descriptor from the GDT into the hidden portion of the TR. The TR is loaded with a new selector using the LTR instruction. The TR is also loaded during a task switch, as described in "Switching Tasks" on page 375. Chapter 12: Task Management 363
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Figure 12-3 shows the format of the TR in legacy mode.
Selector Descriptor Attributes 32-Bit Descriptor-Table Limit 32-Bit Descriptor-Table Base Address
Hidden From Software
513-221.eps
Figure 12-3. TR Format, Legacy Mode Figure 12-4 shows the format of the TR in long mode (both compatibility mode and 64-bit mode).
Selector Descriptor Attributes 32-Bit Descriptor-Table Limit 64-Bit Descriptor-Table Base Address
Hidden From Software
513-267.eps
Figure 12-4. TR Format, Long Mode The AMD64 architecture expands the TSS-descriptor baseaddress field to 64 bits so that system software running in long mode can access a TSS located anywhere in the 64-bit virtualaddress space. The processor ignores the 32 high-order baseaddress bits when running in legacy mode. Because the TR is loaded from the GDT, the system-segment descriptor format has been expanded to 16 bytes by the AMD64 architecture in support of 64-bit mode. See "System Descriptors" on page 111 for more information on this expanded format. The high-order base-address bits are only loaded from 64-bit mode using the 364 Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
LTR instruction. Figure 12-5 shows the relationship between the TSS and GDT.
Global Descriptor Table
Task State Segment
Task Selector TSS Attributes GDT Limit GDT Base Address TSS Limit TSS Base Address
Global Descriptor Table Register
Task Register
513-210.eps
Figure 12-5. Relationship between the TSS and GDT Long mode requires the use of a 64-bit TSS type, and this type must be loaded into the TR by executing the LTR instruction in 64-bit mode. Executing the LTR instruction in 64-bit mode loads the TR with the full 64-bit TSS base address from the 16-byte TSS descriptor format (compatibility mode can only load 8-byte system descriptors). A processor running in either compatibility mode or 64-bit mode uses the full 64-bit TR.base address. 12.2.4 Legacy TaskState Segment The task-state segment (TSS) is a data structure in memory that the processor uses to save and restore the execution state for a task when a task switch occurs. Figure 12-6 on page 366 shows the format of a legacy 32-bit TSS.
Chapter 12: Task Management
365
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Bit Offset 31 16 15 0
Byte Offset IOPB Base
I/O-Permission Bitmap (IOPB) (Up to 8 Kbytes) Interrupt-Redirection Bitmap (IRB) (Eight 32-Bit Locations) Operating-System Data Structure I/O-Permission Bitmap Base Address Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN Reserved, IGN EDI ESI EBP ESP EBX EDX ECX EAX EFLAGS EIP CR3 Reserved, IGN ESP2 Reserved, IGN ESP1 Reserved, IGN ESP0 Reserved, IGN Link (Prior TSS Selector) SS0 SS1 SS2 Reserved, IGN LDT Selector GS FS DS SS CS ES T
+64h +60h +5Ch +58h +54h +50h +4Ch +48h +44h +40h +3Ch +38h +34h +30h +2Ch +28h +24h +20h +1Ch +18h +14h +10h +0Ch +08h +04h +00h
Figure 12-6. Legacy 32-bit TSS
366
Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The 32-bit TSS contains three types of fields: Static fields are read by the processor during a task switch when a new task is loaded, but are not written by the processor when a task is suspended. Dynamic fields are read by the processor during a task switch when a new task is loaded, and are written by the processor when a task is suspended. Software-defined fields are read and written by software, but are not read or written by the processor. All but the first 104 bytes of a TSS can be defined for software purposes, minus any additional space required for the optional I/Opermission bitmap and interrupt-redirection bitmap. TSS fields are not read or written by the processor when the LTR instruction is executed. The LTR instruction loads the TSS descriptor into the TR and marks the task as busy, but it does not cause a task switch. The TSS fields used by the processor in legacy mode are: Link--Bytes 01h-00h, dynamic field. Contains a copy of the task selector from the previously-executed task. See "Nesting Tasks" on page 379 for additional information. Stack Pointers--Bytes 1Bh-04h, static field. Contains the privilege 0, 1, and 2 stack pointers for the task. These consist of the stack-segment selector (SSn), and the stack-segment offset (ESPn). CR3--Bytes 1Fh-1Ch, static field. Contains the pagetranslation-table base-address (CR3) register for the task. EIP--Bytes 23h-20h, dynamic field. Contains the instruction pointer (EIP) for the next instruction to be executed when the task is restored. EFLAGS--Bytes 27h-24h, dynamic field. Contains a copy of the EFLAGS image at the point the task is suspended. General-Purpose Registers--Bytes 47h-28h, dynamic field. Contains a copy of the EAX, ECX, EDX, EBX, ESP, EBP, ESI, and EDI values at the point the task is suspended. Segment-Selector Registers--Bytes 59h-48h, dynamic field. Contains a copy of the ES, CS, SS, DS, FS, and GS, values at the point the task is suspended.
Chapter 12: Task Management
367
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
LDT Segment-Selector Register--Bytes 63h-60h, static field. Contains the local-descriptor-table segment selector for the task. T (Trap) Bit--Bit 0 of byte 64h, static field. This bit, when set to 1, causes a debug exception (#DB) to occur on a task switch. See "Task-Switch Breakpoints" on page 400 for additional information. I/O-Permission Bitmap Base Address--Bytes 67h-66h, static field. This field represents a 16-bit offset into the TSS. This offset points to the beginning of the I/O-permission bitmap, and the end of the interrupt-redirection bitmap. I/O-Permission Bitmap--Static field. This field specifies protection for I/O-port addresses (up to the 64K ports supported by the processor), as follows: - Whether the port can be accessed at any privilege level. - Whether the port can be accessed outside the privilege level established by EFLAGS.IOPL. - Whether the port can be accessed when the processor is running in virtual-8086 mode. Because one bit is used per 8-byte I/O-port, this bitmap can take up to 8 Kbytes of TSS space. The bitmap can be located anywhere within the first 64 Kbytes of the TSS, as long as it is above byte 103. The last byte of the bitmap must contain all ones (0FFh). See "I/O-Permission Bitmap" on page 369 for more information. Interrupt-Redirection Bitmap--Static field. This field defines how each of the 256-possible software interrupts is directed in a virtual-8086 environment. One bit is used for each interrupt, for a total bitmap size of 32 bytes. The bitmap can be located anywhere above byte 103 within the first 64 Kbytes of the TSS. See "Interrupt Redirection of Software Interrupts" on page 299 for information on using this field. The TSS can be paged by system software. System software that uses the hardware task-switch mechanism must guarantee that a page fault does not occur during a task switch. Because the processor only reads and writes the first 104 TSS bytes during a task switch, this restriction only applies to those bytes. The simplest approach is to align the TSS on a page boundary so that all critical bytes are either present or not present. Then, if a page fault occurs when the TSS is accessed, it occurs before 368 Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
the first byte is read. If the page fault occurs after a portion of the TSS is read, the fault is unrecoverable. I/O-Permission Bitmap. The I/O-permission bitmap (IOPB) allows system software to grant less-privileged programs access to individual I/O ports, overriding the effect of RFLAGS.IOPL for those devices. When an I/O instruction is executed, the processor checks the IOPB only if the processor is in virtual x86 mode or the CPL is greater than the RFLAGS.IOPL field. Each bit in the IOPB corresponds to a byte I/O port. A word I/O port corresponds to two consecutive IOPB bits, and a doubleword I/O port corresponds to four consecutive IOPB bits. Access is granted to an I/O port of a given size when all IOPB bits corresponding to that port are clear. If any bits are set, a #GP occurs. The IOPB is located in the TSS, as shown by the example in Figure 12-7 on page 370. Each TSS can have a different copy of the IOPB, so access to individual I/O devices can be granted on a task-by-task basis. The I/O-permission bitmap base-address field located at byte 66h in the TSS is an offset into the TSS locating the start of the IOPB. If all 64K IO ports are supported, the IOPB base address must not be greater than 0DFFFh, otherwise accesses to the bitmap cause a #GP to occur. An extra byte must be present after the last IOPB byte. This byte must have all bits set to 1 (0FFh). This allows the processor to read two IOPB bytes each time an I/O port is accessed. By reading two IOPB bytes, the processor can check all bits when unaligned, multi-byte I/O ports are accessed.
Chapter 12: Task Management
369
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Bit Offset 31 1111_1111 0000 16 15 0
Byte Offset IOPB+Ch IOPB+8h IOPB+4h IOPB
I/O-Permission Bitmap Base Address
+64h ... +00h
Figure 12-7.
I/O-Permission Bitmap Example Bits in the IOPB sequentially correspond to I/O port addresses. The example in Figure 12-7 shows bits 12 through 15 in the second doubleword of the IOPB cleared to 0. Those bit positions correspond to byte I/O ports 44h through 47h, or alternatively, doubleword I/O port 44h. Because the bits are cleared to zero, software running at any privilege level can access those I/O ports. By adjusting the TSS limit, it may happen that some ports in the I/O-address space have no corresponding IOPB entry. Ports not represented by the IOPB will cause a #GP exception. Referring again to Figure 12-7, the last IOPB entry is at bit 23 in the fourth IOPB doubleword, which corresponds to I/O port 77h. In this example, all ports from 78h and above will cause a #GP exception, as if their permission bit was set to 1.
12.2.5 64-Bit Task State Segment
Although the hardware task-switching mechanism is not supported in long mode, a 64-bit task state segment (TSS) must still exist. System software must create at least one 64-bit TSS for use after activating long mode, and it must execute the LTR instruction, in 64-bit mode, to load the TR register with a pointer to the 64-bit TSS that serves both 64-bit-mode programs and compatibility-mode programs. The legacy TSS contains several fields used for saving and restoring processor-state information. The legacy fields include general-purpose register, EFLAGS, CR3 and segment-selector
370
Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
register state, among others. Those legacy fields are not supported by the 64-bit TSS. System software must save and restore the necessary processor-state information required by the software-multitasking implementation (if multitasking is supported). Figure 12-8 on page 372 shows the format of a 64-bit TSS. The 64-bit TSS holds several pieces of information important to long mode that are not directly related to the task-switch mechanism: RSPn--Bytes 1Bh-04h. The full 64-bit canonical forms of the stack pointers (RSP) for privilege levels 0 through 2. ISTn--Bytes 5Bh-24h. The full 64-bit canonical forms of the interrupt-stack-table (IST) pointers. See "Interrupt-Stack Table" on page 292 for a description of the IST mechanism. I/O Map Base Address--Bytes 67h-66h. The 16-bit offset to the I/O-permission bit map from the 64-bit TSS base. The function of this field is identical to that in a legacy 32-bit TSS. See "I/O-Permission Bitmap" on page 369 for more information.
Chapter 12: Task Management
371
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Bit Offset 31 16 15
0
Byte Offset IOPB Base
I/O-Permission Bitmap (IOPB) (Up to 8 Kbytes)
I/O Map Base Address Reserved, IGN IST7[63:32] IST7[31:0] IST6[63:32] IST6[31:0] IST5[63:32] IST5[31:0] IST4[63:32] IST4[31:0] IST3[63:32] IST3[31:0] IST2[63:32] IST2[31:0] IST1[63:32] IST1[31:0] Reserved, IGN RSP2[63:32] RSP2[31:0] RSP1[63:32] RSP1[31:0] RSP0[63:32] RSP0[31:0] Reserved, IGN
Reserved, IGN
+64h +60h +5Ch +58h +54h +50h +4Ch +48h +44h +40h +3Ch +38h +34h +30h +2Ch +28h +24h +20h +1Ch +18h +14h +10h +0Ch +08h +04h +00h
Figure 12-8. Long Mode TSS Format 372 Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
12.2.6 Task Gate Descriptor (Legacy Mode Only)
Task-gate descriptors hold a selector reference to a TSS and are used to control access between tasks. Unlike a TSS descriptor or other gate descriptors, a task gate can be located in any of the three descriptor tables (GDT, LDT, and IDT). Figure 12-9 shows the format of a task-gate descriptor.
31 Reserved, IGN
16 15 14 13 12 11 P DPL S Type
87 Reserved, IGN
0 +4
TSS Selector
Reserved, IGN
+0
Figure 12-9. Task-Gate Descriptor, Legacy Mode Only The task-gate descriptor fields are: System (S) and Type--Bits 12 and 11-8 (respectively) of byte +4. These bits are encoded by software as 00101b to indicate a task-gate descriptor type. Present (P)--Bit 15 of byte +4. The segment-present bit indicates the segment referenced by the gate descriptor is loaded in memory. If a reference is made to a segment when P=0, a segment-not-present exception (#NP) occurs. This bit is set and cleared by system software and is never altered by the processor. Descriptor Privilege-Level (DPL)--Bits 14-13 of byte +4. The DPL field indicates the gate-descriptor privilege level. DPL can be set to any value from 0 to 3, with 0 specifying the most privilege and 3 the least privilege.
12.3
Hardware Task-Management in Legacy Mode
This section describes the operation of the task-switch mechanism when the processor is running in legacy mode. None o f th ese features are s upported in long mode (either compatibility mode or 64-bit mode).
12.3.1 Task MemoryMapping
The hardware task-switch mechanism gives system software a great deal of flexibility in managing the sharing and isolation of memory--both virtual (linear) and physical--between tasks. 373
Chapter 12: Task Management
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Segmented Memory. The segmented memory for a task consists of the segments that are loaded during a task switch and any segments that are later accessed by the task code. The hardware task-switch mechanism allows tasks to either share segments with other tasks, or to access segments in isolation from one another. Tasks that share segments actually share a virtual-address (linear-address) space, but they do not necessarily share a physical-address space. When paging is enabled, the virtual-to-physical mapping for each task can differ, as is described in the following section. Shared segments do share physical memory when paging is disabled, because virtual addresses are used as physical addresses. A number of options are available to system software that shares segments between tasks: Sharing segment descriptors using the GDT. All tasks have access to the GDT, so it is possible for segments loaded in the GDT to be shared among tasks. Sharing segment descriptors using a single LDT. Each task has its own LDT, and that LDT selector is automatically saved and restored in the TSS by the processor during task switches. Tasks, however, can share LDTs simply by storing the same LDT selector in multiple TSSs. Using the LDT to manage segment sharing and segment isolation provides more flexibility to system software than using the GDT for the same purpose. Copying shared segment descriptors into multiple LDTs. Segment descriptors can be copied by system software into multiple LDTs that are otherwise not shared between tasks. Allowing segment sharing at the segment-descriptor level, rather than the LDT level or GDT level, provides the greatest flexibility to system software. In all three cases listed above, the actual data and instructions are shared between tasks only when the tasks' virtual-tophysical address mappings are identical. Paged Memory. Each task has its own page-translation table baseaddress (CR3) register, and that register is automatically saved and restored in the TSS by the processor during task switches. This allows each task to point to its own set of page-translation tables, so that each task can translate virtual addresses to physical addresses independently. Page translation must be enabled for changes in CR3 values to have an effect on virtual374 Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
to-physical address mapping. When page translation is disabled, the tables referenced by CR3 are ignored, and virtual addresses are equivalent to physical addresses. 12.3.2 Switching Tasks The hardware task-switch mechanism transfers program control to a new task when any of the following occur: A CALL or JMP instruction with a selector operand that references a task gate is executed. The task gate can be located in either the LDT or GDT. A CALL or JMP instruction with a selector operand that references a TSS descriptor is executed. The TSS descriptor must be located in the GDT. A software-interrupt instruction (INTn) is executed that references a task gate located in the IDT. An exception or external interrupt occurs, and the vector references a task gate located in the IDT. An IRET is executed while the EFLAGS.NT bit is set to 1, indicating that a return is being performed from an innerlevel task to an outer-level task. The new task is referenced using the selector stored in the current-task link field. See "Nesting Tasks" on page 379 for additional information. The RET instruction cannot be used to switch tasks. When a task switch occurs, the following operations are performed automatically by the processor: The processor performs privilege-checking to determine whether the currently-executing program is allowed to access the target task. If this check fails, the task switch is aborted without modifying the processor state, and a general-protection exception (#GP) occurs. The privilege checks performed depend on the cause of the task switch: - If the task switch is initiated by a CALL or JMP instruction through a TSS descriptor, the processor checks that both the currently-executing program CPL and the TSS-selector RPL are numerically less-than or equal-to the TSS-descriptor DPL. - If the task switch takes place through a task gate, the CPL and task-gate RPL are compared with the task-gate DPL, and no comparison is made using the TSSdescriptor DPL. See "Task Switches Using Task Gates" on page 377.
Chapter 12: Task Management
375
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
-
Software interrupts, hardware interrupts, and exceptions all transfer control without checking the task-gate DPL. - The IRET instruction transfers control without checking the TSS-descriptor DPL. The processor performs limit-checking on the target TSS descriptor to verify that the TSS limit is greater than or equal to 67h (at least 104 bytes). If this check fails, the task switch is aborted without modifying the processor state, and an invalid-TSS exception (#TS) occurs. The current-task state is saved in the TSS. This includes the next-instruction pointer (EIP), EFLAGS, the generalpurpose registers, and the segment-selector registers. Up to this point, any exception that occurs aborts the task switch without changing the processor state. From this point forward, any exception that occurs does so in the context of the new task. If an exception occurs in the context of the new task during a task switch, the processor finishes loading the new-task state without performing additional checks. The processor transfers control to the #TS handler after this state is loaded, but before the first instruction is executed in the new task. When a #TS occurs, it is possible that some of the state loaded by the processor did not participate in segment access checks. The #TS handler must verify that all segments are accessible before returning to the interrupted task. The task register (TR) is loaded with the new-task TSS selector, and the hidden portion of the TR is loaded with the new-task descriptor. The TSS now referenced by the processor is that of the new task. The current task is marked as busy. The previous task is marked as available or remains busy, based on the type of linkage. See "Nesting Tasks" on page 379 for more information. CR0.TS is set to 1. This bit can be used to save other processor state only when it becomes necessary. For more information, see the next section, "Saving Other Processor State." The new-task state is loaded from the TSS. This includes the next-instruction pointer (EIP), EFLAGS, the generalpurpose registers, and the segment-selector registers. The processor clears the segment-descriptor present (P) bits (in the hidden portion of the segment registers) to prevent 376 Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
access into the new segments, until the task switch completes successfully. The LDTR and CR3 registers are loaded from the TSS, changing the virtual-to-physical mapping from that of the old task to the new task. Because this is done in the middle of accessing the new TSS, system software must guarantee that TSS addresses are translated identically in all tasks. The descriptors for all previously-loaded segment selectors are loaded into the hidden portion of the segment registers. This sets or clears the P bits for the segments as specified by the new descriptor values. If the above steps complete successfully, the processor begins executing instructions in the new task beginning with the instruction referenced by the CS:EIP far pointer loaded from the new TSS. The privilege level of the new task is taken from the new CS segment selector's RPL. Saving Other Processor State. The processor does not automatically save the registers used by the media or x87 instructions. Instead, the processor sets CR0.TS to 1 during a task switch. Later, when an attempt is made to execute any of the media or x87 instructions while TS=1, a device-not-available exception (#NM) occurs. System software can then save the previous state of the media and x87 registers and clear the CR0.TS bit to 0 before executing the next media/x87 instruction. As a result, the media and x87 registers are saved only when necessary after a task switch. 12.3.3 Task Switches Using Task Gates When a control transfer to a new task occurs through a task gate, the processor reads the task-gate DPL (DPLG) from the task-gate descriptor. Two privilege checks, both of which must pass, are performed on DPLG before the task switch can occur successfully: The processor compares the CPL with DPLG. The CPL must be numerically less than or equal to DPLG for this check to pass. In other words, the following expression must be true: CPL DPLG. The processor compares the RPL in the task-gate selector with DPLG. The RPL must be numerically less than or equal to DPLG for this check to pass. In other words, the following expression must be true: RPL DPLG.
Chapter 12: Task Management
377
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Unlike call-gate control transfers, the processor does not read the DPL from the target TSS descriptor (DPLS) and compare it with the CPL when a task gate is used. Figure 12-10 on page 379 shows two examples of task-gate privilege checks. In Example 1, the privilege checks pass: The task-gate DPL (DPLG) is at the lowest privilege (3), specifying that software running at any privilege level (CPL) can access the gate. The selector referencing the task gate passes its privilege check because the RPL is numerically less than or equal to DPLG. In Example 2, both privilege checks fail: The task-gate DPL (DPLG) specifies that only software at privilege-level 0 can access the gate. The current program does not have enough privilege to access the task gate, because its CPL is 2. The selector referencing the task-gate descriptor does not have a high enough privilege to complete the reference. Its RPL is numerically greater than DPLG. Although both privilege checks failed in the example, if only one check fails, access into the target task is denied. Because the legacy task-switch mechanism is not supported in long mode, software cannot use task gates in long mode. Any attempt to transfer control to another task using a task gate in long mode causes a general-protection exception (#GP) to occur.
378
Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
CS Task-Gate Selector
CPL=2
RPL=3
DPLG=3 Task-Gate Descriptor DPLS TSS Descriptor Example 1: Privilege Check Passes Task-State Segment
Access Allowed
CS Task-Gate Selector
CPL=2
RPL=3
DPLG=0 Task-Gate Descriptor DPLS TSS Descriptor Example 2: Privilege Check Fails
513-255.eps
Task-State Segment
Access Denied
Figure 12-10.
Privilege-Check Examples for Task Gates The hardware task-switch mechanism supports task nesting through the use of EFLAGS nested-task (NT) bit and the TSS link-field. The manner in which these fields are updated and used during a task switch depends on how the task switch is initiated:
12.3.4 Nesting Tasks
Chapter 12: Task Management
379
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
The JMP instruction does not update EFLAGS.NT or the TSS link-field. Task nesting is not supported by the JMP instruction. The CALL instruction, INTn instructions, interrupts, and exceptions can only be performed from outer-level tasks to inner-level tasks. All of these operations set the EFLAGS.NT bit for the new task to 1 during a task switch, and copy the selector for the previous task into the new-task link field. An IRET instruction which returns to another task only occurs when the EFLAGS.NT bit for the current task is set to 1, and only can be performed from an inner-level task to an outer-level task. When an IRET results in a task switch, the new task is referenced using the selector stored in the current-TSS link field. The EFLAGS.NT bit for the current task is cleared to 0 during the task switch. Table 12-1 summarizes the effect various task-switch initiators have on EFLAGS.NT, the TSS link-field, and the TSS-busy bit. (For more information on the busy bit, see the next section, "Preventing Recursion.") Table 12-1. Effects of Task Nesting
Task-Switch Initiator JMP CALL INTn Interrupt Exception IRET
Note:
Old Task EFLAGS.NT -- Link (Selector) -- Busy Clear to 0 (was 1) EFLAGS.NT --
New Task Link (Selector) -- Busy Set to 1
--
-- (Was 1)
Set to 1
Old Task
Set to 1
Clear to 0 (was 1)
--
Clear to 0 (was 1)
--
"--" indicates no change is made.
Programs running at any privilege level can set EFLAGS.NT to 1 and execute the IRET instruction to transfer control to another task. System software can keep control over improperly nested-task switches by initializing the link field of all TSSs 380 Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
that it creates. That way, improperly nested-task switches always transfer control to a known task. Preventing Recursion. Ta s k re c u rs i o n i s n o t a l l owe d by t h e hardware task-switch mechanism. If recursive-task switches were allowed, they would replace a previous task-state image with a newer image, discarding the previous information. To prevent recursion from occurring, the processor uses the busy bit located in the TSS-descriptor type field (bit 9 of byte +4). Use of this bit depends on how the task switch is initiated: The JMP instruction clears the busy bit in the old task to 0 and sets the busy bit in the new task to 1. A generalprotection exception (#GP) occurs if an attempt is made to JMP to a task with a set busy bit. The CALL instruction, INTn instructions, interrupts, and exceptions set the busy bit in the new task to 1. The busy bit in the old task remains set to 1, preventing recursion through task-nesting levels. A general-protection exception (#GP) occurs if an attempt is made to switch to a task with a set busy bit. An IRET to another task (EFLAGS.NT must be 1) clears the busy bit in the old task to 0. The busy bit in the new task is not altered, because it was already set to 1. Table 12-1 on page 380 summarizes the effect various taskswitch initiators have on the TSS-busy bit.
Chapter 12: Task Management
381
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
382
Chapter 12: Task Management
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
13
Debug and Performance Resources
Testing, debug, and performance optimization consume a significant portion of the time needed to develop a new computer or software product and move it successfully into production. To stay competitive, product developers need tools that allow them to rapidly detect, isolate, and correct problems before a product is shipped. The goal of the debug and p e r f o r m a n c e f e a t u re s i n c o r p o ra t e d i n t o p ro c e s s o r implementations of the AMD64 architecture is to support the tool chain solutions used in software-product and hardwareproduct development. The debug and performance resources that can be supported by AMD64 architecture implementations include: Software Debug--The AMD64 architecture supports the legacy software-debug facilities, including the debug registers (DR0-DR7), debug exception, and breakpoint exception. Additional features are provided using modelspecific registers (MSRs). These registers are used to set breakpoints on branches, interrupts, and exceptions and to single step from one branch to the next. The software-debug capability is described in "Software-Debug Resources" on page 384. Performance Monitoring--Model-specific registers (MSRs) are provided to monitor events within an implementation of the AMD64 architecture. A set of control registers allow the selection of events to be monitored and a corresponding set of counter registers track the frequency of monitored events. These features are used to support software-performance and hardware-performance tuning. Performance monitoring is described in "Performance Optimization" on page 403. A given processor implementation may include additional debug capabilities that allow monitoring program execution and manipulation of processor state to be manipulated. These resources are typically intended for BIOS and platform development and, if available, are described in the BIOS and K e r n e l D e v e l o p e r 's G u i d e f o r A M D A t h l o n TM 6 4 a n d AMD OpteronTM Processors (order# 26094) for information specific to a particular processor.
Chapter 13: Debug and Performance Resources
383
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
13.1
Software-Debug Resources
Software can program breakpoints into the debug registers, causing a #DB exception when matches occur on instructionmemory addresses, data-memory addresses, or I/O addresses. The breakpoint exception (#BP) is also supported to allow software to set breakpoints by placing INT3 instructions in the instruction memory for a program. Program control is transferred to the breakpoint exception (#BP) handler when an INT3 instruction is executed. In addition to the debug features supported by the debug registers (DR0-DR7), the processor also supports features supported by model-specific registers (MSRs). Together, these capabilities provide a rich set of breakpoint conditions, including: Breakpoint On Address Match--Breakpoints occur when the address stored in a address-breakpoint register matches the address of an instruction or data reference. Up to four address-match breakpoint conditions can be set by software. Single Step All Instructions--Breakpoints can be set to occur on every instruction, allowing a debugger to examine the contents of registers as a program executes. Single Step Control Transfers--Breakpoints can be set to occur on control transfers, such as calls, jumps, interrupts, and exceptions. This can allow a debugger to narrow a problem search to a specific section of code before enabling single stepping of all instructions. Breakpoint On Any Instruction--Breakpoints can be set on any specific instruction using either the address-match breakpoint condition or using the INT3 instruction to force a breakpoint when the instruction is executed. Breakpoint On Task Switch--Software forces a #DB exception to occur when a task switch is performed to a task with the T bit in the TSS set to 1. Debuggers can use this capability to enable or disable debug conditions for a specific task. Problem areas can be identified rapidly using the information supplied by the debug registers when breakpoint conditions occur: Special conditions that cause a #DB exception are recorded in the DR6 debug-status register, including breakpoints due to task switches and single stepping. The DR6 register also
384
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
identifies which address-breakpoint register (DR0-DR3) caused a #DB exception due to an address match. When combined with the DR7 debug-control register settings, the cause of a #DB exception can be identified. To assist in analyzing the instruction sequence a processor follows in reaching its current state, the source and destination addresses of control-transfer events are saved by the processor. These include branches (calls and jumps), interrupts, and exceptions. Debuggers can use this information to narrow a problem search to a specific section of code before single stepping all instructions. 13.1.1 Debug Registers The AMD64 architecture supports the legacy debug registers, DR0-DR7. These registers are expanded to 64 bits by the AMD64 architecture. In legacy mode and in compatibility mode, only the lower 32 bits are used. In these modes, writes to a debug register fill the upper 32 bits with zeros, and reads from a debug register return only the lower 32 bits. In 64-bit mode, all 64 bits of the debug registers are read and written. Operandsize prefixes are ignored. The debug registers can be read and written only when the current-protection level (CPL) is 0 (most privileged). Attempts to read or write the registers at a lower-privilege level (CPL>0) cause a general-protection exception (#GP). Several debug registers described below are model-specific registers (MSRs). See "Software-Debug MSRs" on page 444 for a listing of the debug-MSR numbers and their reset values. Some processor implementations include additional MSRs used to support implementation-specific software debug features. For more information on these registers and their capabilities, refer to the BIOS writer's guide for the implementation. Address-Breakpoint Registers (DR0-DR3). Figure 13-1 on page 386 shows the format of the four address-breakpoint registers, DR0DR3. Software can load a virtual (linear) address into any of the four registers, and enable breakpoints to occur when the address matches an instruction or data reference. The MOV DRn instructions do not check that the virtual addresses loaded into DR0-DR3 are in canonical form. Breakpoint conditions are enabled using the debug-control register, DR7 (see "DebugControl Register (DR7)" on page 388).
Chapter 13: Debug and Performance Resources
385
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
63 Breakpoint 0 64-bit Virtual (linear) Address
0
63 Breakpoint 1 64-bit Virtual (linear) Address
0
63 Breakpoint 2 64-bit Virtual (linear) Address
0
63 Breakpoint 3 64-bit Virtual (linear) Address
0
Figure 13-1. Address-Breakpoint Registers (DR0-DR3) Reserved Debug Registers (DR4, DR5). The DR4 and DR5 registers are reserved and should not be used by software. These registers are aliased to the DR6 and DR7 registers, respectively. When the debug extensions are enabled (CR4.DE=1) attempts to access these registers cause an invalid-opcode exception (#UD). Debug-Status Register (DR6). Figure 13-2 on page 387 shows the format of the debug-status register, DR6. Debug status is loaded into DR6 when an enabled debug condition is encountered that causes a #DB exception. This register is never cleared by the processor and must be cleared by software after the contents have been read.
386
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
63 MBZ
32
31 Read as 1s
16
15
B T
14
B S
13
12 Read as 1s
4
3
B 3
2
B 2
1
B 1
0
B 0
B RAZ D
Bits Mnemonic 63-16 Reserved 15 BT 14 BS 13 BD 12 Reserved 11-4 Reserved 3 B3 2 B2 1 B1 0 B0
Description MBZ Breakpoint Task Switch Breakpoint Single Step Breakpoint Debug Access Detected Read as Zero Read as 1s Breakpoint #3 Condition Detected Breakpoint #2 Condition Detected Breakpoint #1 Condition Detected Breakpoint #0 Condition Detected
R/W R/W R/W R/W
R/W R/W R/W R/W
Figure 13-2. Debug-Status Register (DR6) The fields within the DR6 register are set by the processor and must be cleared by software. These fields are: Breakpoint-Condition Detected (B3-B0)--Bits 3-0. The processor individually sets these bits to 1 when the corresponding address-breakpoint register detects a breakpoint condition, as specified by the DR7 R/Wn and LENn controls. For example, B1 (bit 1) is set to 1 if an address-breakpoint condition is detected by DR1. These bits are only set if the #DB exception is enabled by the Ln and Gn control bits in DR7. Debug-Register-Access Detected (BD)--Bit 13. The processor sets this bit to 1 if software accesses any debug register (DR0-DR7) while the general-detect condition is enabled (DR7.GD=1). Single Step (BS)--Bit 14. The processor sets this bit to 1 if the #DB exception occurs as a result of single-step mode (rFLAGS.TF=1). Single-step mode has the highest-priority among debug exceptions. Other status bits within the DR6 register can be set by the processor along with the BS bit. Task-Switch (BT)--Bit 15. The processor sets this bit to 1 if the #DB exception occurred as a result of task switch to a task with a TSS T-bit set to 1.
Chapter 13: Debug and Performance Resources
387
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
All remaining bits in the DR6 register are reserved. Reserved bits 31-16 and 11-4 must all be set to 1, while reserved bit 12 must be cleared to 0. In 64-bit mode, the upper 32 bits of DR6 are reserved and must be written with zeros. Writing a 1 to any of the upper 32 bits results in a general-protection exception, #GP(0). Debug-Control Register (DR7). Figure 13-3 shows the format of the debug-control register, DR7. DR7 is used to establish the breakpoint conditions for the address-breakpoint registers (DR0-DR3) and to enable debug exceptions for each addressbreakpoint register individually. DR7 is also used to enable the general-detect breakpoint condition.
63 MBZ 32
31
30
29
28
27
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12 Read as 0s
10 R A 1s
9
8
7
6
5
4
3
2
1
0
LEN3
R/W3
LEN2
R/W2
LEN1
R/W1
LEN0
R/W0
RAZ
G D
G E
L E
G 3
L 3
G 2
L 2
G 1
L 1
G 0
L 0
Bits Mnemonic 63-32 Reserved 31-30 LEN3 29-28 R/W3 27-26 LEN2 25-24 R/W2 23-22 LEN1 21-20 R/W1 19-18 LEN0 17-16 R/W0 15-14 Reserved 13 GD 12-11 Reserved 10 Reserved 9 GE 8 LE 7 G3 6 L3 5 G2 4 L2 3 G1 2 L1 1 G0 0 L0
Description MBZ Length of Breakpoint #3 Type of Transaction(s) to Trap Length of Breakpoint #2 Type of Transaction(s) to Trap Length of Breakpoint #1 Type of Transaction(s) to Trap Length of Breakpoint #0 Type of Transaction(s) to Trap Read as 0s General Detect Enabled Read as 0s Read as 1 Global Exact Breakpoint Enabled Local Exact Breakpoint Enabled Global Exact Breakpoint #3 Enabled Local Exact Breakpoint #3 Enabled Global Exact Breakpoint #2 Enabled Local Exact Breakpoint #2 Enabled Global Exact Breakpoint #1 Enabled Local Exact Breakpoint #1 Enabled Global Exact Breakpoint #0 Enabled Local Exact Breakpoint #0 Enabled
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Figure 13-3. Debug-Control Register (DR7)
388
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The fields within the DR7 register are all read/write. These fields are: Local-Breakpoint Enable (L3-L0)--Bits 6, 4, 2, and 0 (respectively). Software individually sets these bits to 1 to enable debug exceptions to occur when the corresponding address-breakpoint register (DRn) detects a breakpoint condition while executing the current task. For example, if L1 (bit 2) is set to 1 and an address-breakpoint condition is detected by DR1, a #DB exception occurs. These bits are cleared to 0 by the processor when a hardware task-switch occurs. Global-Breakpoint Enable (G3-G0)--Bits 7, 5, 3, and 1 (respectively). Software sets these bits to 1 to enable debug exceptions to occur when the corresponding addressbreakpoint register (DRn) detects a breakpoint condition while executing any task. For example, if G1 (bit 3) is set to 1 and an address-breakpoint condition is detected by DR1, a #DB exception occurs. These bits are never cleared to 0 by the processor. Local-Enable (LE)--Bit 8. Software sets this bit to 1 in legacy implementations to enable exact breakpoints while executing the current task. This bit is ignored by implementations of the AMD64 architecture. All breakpoint conditions, except certain string operations preceded by a repeat prefix, are exact. Global-Enable (GE)--Bit 9. Software sets this bit to 1 in legacy implementations to enable exact breakpoints while executing any task. This bit is ignored by implementations of the AMD64 architecture. All breakpoint conditions, except certain string operations preceded by a repeat prefix, are exact. General-Detect Enable (GD)--Bit 13. Software sets this bit to 1 to cause a debug exception to occur when an attempt is made to execute a MOV DRn instruction to any debug register (DR0-DR7). This bit is cleared to 0 by the processor when the #DB handler is entered, allowing the handler to read and write the DRn registers. The #DB exception occurs before executing the instruction, and DR6.BD is set by the processor. Software debuggers can use this bit to prevent the currently-executing program from interfering with the debug operation.
Chapter 13: Debug and Performance Resources
389
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Read/Write (R/W3-R/W0)--Bits 29-28, 25-24, 21-20, and 17-16 (respectively). Software sets these fields to control the breakpoint conditions used by the corresponding addressbreakpoint registers (DRn). For example, control-field R/W1 (bits 21-20) controls the breakpoint conditions for the DR1 register. The R/Wn control-field encodings specify the following conditions for an address-breakpoint to occur: - 00--Only on instruction execution. - 01--Only on data write. - 10--This encoding is further qualified by CR4.DE as follows: . CR4.DE=0--Condition is undefined. . CR4.DE=1--Only on I/O read or I/O write. - 11--Only on data read or data write. Length (LEN3-LEN0)--Bits 31-30, 27-26, 23-22, and 19-18 (respectively). Software sets these fields to control the range used in comparing a memory address with the corresponding address-breakpoint register (DRn). For example, controlfield LEN1 (bits 23-22) controls the breakpoint-comparison range for the DR1 register. The value in DRn defines the low-end of the address range used in the comparison. LENn is used to mask the low-order address bits in the corresponding DRn register so that they are not used in the address comparison. To work properly, breakpoint boundaries must be aligned on an address corresponding to the range size specified by LENn. The LENn control-field encodings specify the following addressbreakpoint-comparison ranges: - 00--1 byte. - 01--2 byte, must be aligned on a word boundary. - 10--8 byte, must be aligned on a quadword boundary. (Long mode only; otherwise undefined.) - 11--4 byte, must be aligned on a doubleword boundary. If the R/Wn field is used to specify instruction breakpoints (R/Wn=00), the corresponding LENn field must be set to 00. Setting LENn to any other value produces undefined results. All remaining bits in the DR7 register are reserved. Reserved bits 15-14 and 12-11 must all be cleared to 0, while reserved bit 10 must be set to 1. In 64-bit mode, the upper 32 bits of DR7 are reserved and must be written with zeros. Writing a 1 to any of 390 Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
the upper 32 bits results in a general-protection #GP(0) exception. 64-Bit-Mode Extended Debug Registers. In 64-bit mode, additional encodings for debug registers are available. The REX.R bit, in a REX prefix, is used to modify the ModRM reg field when that field encodes a control register, as shown in "REX Prefix-Byte Fields" in Volume 3. These additional encodings enable the processor to address DR8-DR15. Access to the DR8-DR15 registers is implementationdependent. The architecture does not require any of these extended debug registers to be implemented. Any attempt to access an unimplemented register results in an invalid-opcode exception (#UD). Debug-Control MSR (DebugCtlMSR). Figure 13-4 shows the format of the debug-control MSR, DebugCtlMSR. DebugCtlMSR provides additional debug controls over control-transfer recording and single stepping, and external-breakpoint reporting and trace messages. DebugCtlMSR is an MSR and is read and written using the RDMSR and WRMSR instructions.
63 Reserved 32
31 Reserved
6
5
P B 3
4
P B 2
3
P B 1
2
P B 0
1
B T F
0
L B R
Bits Mnemonic 63-6 Reserved 5 PB3 4 PB2 3 PB1 2 PB0 1 BTF 0 LBR
Description Performance Monitoring Pin Control Performance Monitoring Pin Control Performance Monitoring Pin Control Performance Monitoring Pin Control Branch Single Step Last-Branch Record
R/W R/W R/W R/W R/W R/W R/W
Figure 13-4. Debug-Control MSR (DebugCtlMSR) The fields within the DebugCtlMSR register are: Last-Branch Record (LBR)--Bit 0, read/write. Software sets this bit to 1 to cause the processor to record the source and target addresses of the last control transfer taken before a debug exception occurs. The recorded control transfers Chapter 13: Debug and Performance Resources 391
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
include branch instructions, interrupts, and exceptions. When a debug exception occurs, the processor clears this bit before transferring control to the debug-exception handler. This prevents control transfers into the debug-exception handler from recording the source and target addresses. Branch Single Step (BTF)--Bit 1, read/write. Software uses this bit to change the behavior of the rFLAGS.TF bit. When this bit is cleared to 0, the rFLAGS.TF bit controls instruction single stepping, (normal behavior). When this bit is set to 1, the rFLAGS.TF bit controls single stepping on control transfers. The single-stepped control transfers include branch instructions, interrupts, and exceptions. Control-transfer single stepping requires both BTF=1 and rFLAGS.TF=1. Performance-Monitoring/Breakpoint Pin-Control (PBi)--Bits 5-2, read/write. Software uses these bits to control the type of information reported by the four external performancemonitoring/breakpoint pins on the processor. When a PBi bit is cleared to 0, the corresponding external pin (BPi) reports performance-monitor information. When a PBi bit is set to 1, the corresponding external pin (BPi) reports breakpoint information. All remaining bits in the DebugCtlMSR register are reserved. Control-Transfer Recording MSRs. Figure 13-5 on page 393 shows the format of the 64-bit control-transfer recording MSRs: LastBranchToIP, LastBranchFromIP, LastExceptionToIP, and LastExceptionFromIP. These registers are loaded automatically by the processor when the DebugCtlMSR.LBR bit is set to 1. These MSRs are read-only.
392
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
63 LastBranchToIP - 64-bit Segment Offset (RIP)
0
63 LastBranchFromIP - 64-bit Segment Offset (RIP)
0
63 LastExceptionToIP - 64-bit Segment Offset (RIP)
0
63 LastExceptionFromIP - 64-bit Segment Offset (RIP)
0
Figure 13-5. Control-Transfer Recording MSRs When DebugCtlMSR.LBR=1, the control-transfer recording MSRs are updated as follows: LastBranchToIP Register--The processor loads this register with the segment offset of the first instruction to be executed after a control transfer occurs (also called the target rIP). The control transfers that cause this register to be written are branch instructions, interrupts, and exceptions. In the case of INTn and INTO instructions, interrupts, and exceptions, this register is loaded with the offset of the interrupt or exception handler. LastBranchFromIP Register--The processor loads this register during a control transfer (branch instructions, interrupt, or exception). In the case of a branch instruction, this register is loaded with the segment offset of the branch instruction. In the case of interrupts and exceptions, this register is loaded with the segment offset of the interrupted instruction. In both cases, the offset loaded is referred to as the source rIP. LastExceptionToIP Register--This register holds the target rIP of the last branch that occurred before the exception or interrupt. The processor loads this register with the contents of the LastBranchToIP register when an interrupt or exception occurs, before updating the LastBranchToIP register as described above.
Chapter 13: Debug and Performance Resources
393
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
LastExceptionFromIP Register--This register holds the source rIP of the last branch that occurred before the exception or interrupt. The processor loads this register with the contents of the LastBranchFromIP register when an interrupt or exception occurs, before updating the LastBranchFromIP register as described above.
13.2
Breakpoints
Breakpoints can be set to occur on either instruction addresses or data addresses using the breakpoint-address registers, DR0-DR3 (DRn). The values loaded into these registers represent the breakpoint-location virtual address. The debugcontrol register, DR7, is used to enable the breakpoint registers and to specify the type of access and the range of addresses that can trigger a breakpoint. Software enables the DRn registers using the corresponding local-breakpoint enable (Ln) or global-breakpoint enable (Gn) found in the DR7 register. Ln is used to enable breakpoints only while the current task is active, and it is cleared by the processor when a task switch occurs. Gn is used to enable breakpoints for all tasks, and it is never cleared by the processor. The R/Wn fields in DR7, along with the CR4.DE bit, specify the type of access required to trigger a breakpoint when an address match occurs on the corresponding DRn register. Breakpoints can be set to occur on instruction execution, data reads and writes, and I/O reads and writes. The R/Wn and CR4.DE encodings used to specify the access type are described on page 390 of "Debug-Control Register (DR7)." The LENn fields in DR7 specify the size of the address range used in comparison with data or instruction addresses. LENn is used to mask the low-order address bits in the corresponding DRn register so that they are not used in the address comparison. Breakpoint boundaries must be aligned on an address corresponding to the range size specified by LENn. Assuming the access type matches the type specified by R/Wn, a breakpoint occurs if any accessed byte falls within the range specified by LENn. For instruction breakpoints, LENn must specify a single-byte range. The LENn encodings used to
13.2.1 Setting Breakpoints
394
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
specify the address range are described on page 390 of "DebugControl Register (DR7)." Table 13-1 shows several examples of data accesses, and whether or not they cause a #DB exception to occur based on the breakpoint address in DRn and the breakpoint-address range specified by LENn. In this table, R/Wn always specifies read/write access. Table 13-1. Breakpoint-Setting Examples
Data-Access Address (hexadecimal) Access Size (bytes) Byte-Addresses in Data-Access (hexadecimal) DRn=F000, LENn=00 (1 Byte) EFFB EFFE F000 F001 F005 8 2 4 1 2 4 EFFB, EFFC, EFFD, EFFE, EFFF, F000, F001 EFFE, EFFF EFFE, EFFF, F000, F001 F000 F001, F002 F005, F006, F007, F008 DRn=F004, LENn=11 (4 Bytes) EFFB EFFE F000 F001 F005
Notes:
Breakpoint-Address Range (hexadecimal)
Result
#DB -- F000 #DB
--
8 2 4 1 2 4
EFFB, EFFC, EFFD, EFFE, EFFF, F000, F001 EFFE, EFFF EFFE, EFFF, F000, F001 F000 F001, F002 F005, F006, F007, F008 #DB F004-F007 --
"--" indicates no #DB occurs.
Chapter 13: Debug and Performance Resources
395
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 13-1. Breakpoint-Setting Examples (continued)
Data-Access Address (hexadecimal) Access Size (bytes) Byte-Addresses in Data-Access (hexadecimal) DRn=F005, LENn=10 (8 Bytes) EFFB EFFE F000 F001 F005
Notes:
Breakpoint-Address Range (hexadecimal)
Result
8 2 4 1 2 4
EFFB, EFFC, EFFD, EFFE, EFFF, F000, F001 EFFE, EFFF EFFE, EFFF, F000, F001 F000 F001, F002 F005, F006, F007, F008 F000-F007
#DB --
#DB
"--" indicates no #DB occurs.
13.2.2 Using Breakpoints
A debug exception (#DB) occurs when an enabled-breakpoint condition is encountered during program execution. The debughandler must check the debug-status register (DR6), the conditions enabled by the debug-control register (DR7), and the debug-control MSR (DebugCtlMSR), to determine the #DB cause. The #DB exception corresponds to interrupt-vector 1. See "#DB--Debug Exception (Vector 1)" on page 250. Instruction breakpoints and general-detect conditions cause the #DB exception to occur before the instruction is executed, while all other breakpoint and single-stepping conditions cause the #DB exception to occur after the instruction is executed. Table 13-2 on page 397 summarizes where the #DB exception occurs based on the breakpoint condition.
396
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 13-2.
Breakpoint Location by Condition
Breakpoint Condition Breakpoint Location Before Instruction is Executed
Instruction General Detect Data Write Only Data Read or Data Write I/O Read or I/O Write Single Step1 Task Switch
Note:
After Instruction is Executed1
After Instruction is Executed
1. Repeated operations (REP prefix) can breakpoint between iterations.
Instruction breakpoints and general-detect conditions have a lower interrupt-priority than the other breakpoint and singlestepping conditions (see "Priorities" on page 269). Databreakpoint conditions on the previous instruction occur before an instruction-breakpoint condition on the next instruction. However, if instruction and data breakpoints can occur as a result of executing a single instruction, the instruction breakpoint occurs first (before the instruction is executed), followed by the data breakpoint (after the instruction is executed). Instruction Breakpoints. Instruction breakpoints are set by loading a breakpoint-address register (DRn) with the desired instruction virtual-address, and then setting the corresponding DR7 fields as follows: Ln or Gn is set to 1 to enable the breakpoint for either the local task or all tasks, respectively. R/Wn is set to 00b to specify that the contents of DRn are to be compared only with the virtual address of the next instruction to be executed. LENn must be set to 00b. When a #DB exception occurs due to an instruction breakpointaddress in DRn, the corresponding Bn field in DR6 is set to 1 to indicate that a breakpoint condition occurred. The breakpoint occurs before the instruction is executed, and the breakpointChapter 13: Debug and Performance Resources 397
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
instruction address is pushed onto the debug-handler stack. If multiple instruction breakpoints are set, the debug handler can use the Bn field to identify which register caused the breakpoint. Returning from the debug handler causes the breakpoint instruction to be executed. Before returning from the debug handler, the rFLAGS.RF bit should be set to 1 to prevent a reoccurrence of the #DB exception due to the instructionbreakpoint condition. The processor ignores instructionbreakpoint conditions when rFLAGS.RF=1, until after the next instruction (in this case, the breakpoint instruction) is executed. After the next instruction is executed, the processor clears rFLAGS.RF to 0. Data Breakpoints. D a t a b re a k p o i n t s a re s e t by l o a d i n g a breakpoint-address register (DRn) with the desired data virtual-address, and then setting the corresponding DR7 fields as follows: Ln or Gn is set to 1 to enable the breakpoint for either the local task or all tasks, respectively. R/Wn is set to 01b to specify that the data virtual-address is compared with the contents of DRn only during a memorywrite. Setting this field to 11b specifies that the comparison takes place during both memory reads and memory writes. LENn is set to 00b, 01b, 11b, or 10b to specify an addressmatch range of one, two, four, or eight bytes, respectively. Long mode must be active to set LENn to 10b. When a #DB exception occurs due to a data breakpoint address in DRn, the corresponding Bn field in DR6 is set to 1 to indicate that a breakpoint condition occurred. The breakpoint occurs after the data-access instruction is executed, which means that the original data is overwritten by the data-access instruction. If the debug handler needs to report the previous data value, it must save that value before setting the breakpoint. Because the breakpoint occurs after the data-access instruction is executed, the address of the instruction following the dataaccess instruction is pushed onto the debug-handler stack. Repeated string instructions, however, can trigger a breakpoint before all iterations of the repeat loop have completed. When this happens, the address of the string instruction is pushed onto the stack during a #DB exception if the repeat loop is not 398 Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
complete. A subsequent IRET from the #DB handler returns to the string instruction, causing the remaining iterations to be executed. Most implementations cannot report breakpoints exactly for repeated string instructions, but instead report the breakpoint on an iteration later than the iteration where the breakpoint occurred. I/O Breakpoints. I/O breakpoints are set by loading a breakpointaddress register (DRn) with the I/O-port address to be trapped, and then setting the corresponding DR7 fields as follows: Ln or Gn is set to 1 to enable the breakpoint for either the local task or all tasks, respectively. R/Wn is set to 10b to specify that the I/O-port address is compared with the contents of DRn only during execution of an I/O instruction. This encoding of R/Wn is valid only when debug extensions are enabled (CR4.DE=1). LENn is set to 00b, 01b, or 11b to specify the breakpoint occurs on a byte, word, or doubleword I/O operation, respectively. The I/O-port address specified by the I/O instruction is zero extended by the processor to 64 bits before comparing it with the DRn registers. When a #DB exception occurs due to an I/O breakpoint in DRn, the corresponding Bn field in DR6 is set to 1 to indicate that a breakpoint condition occurred. The breakpoint occurs after the instruction is executed, which means that the original data is overwritten by the breakpoint instruction. If the debug handler needs to report the previous data value, it must save that value before setting the breakpoint. Because the breakpoint occurs after the instruction is executed, the address of the instruction following the I/O instruction is pushed onto the debug-handler stack, in most cases. In the case of INS and OUTS instructions that use the repeat prefix, however, the breakpoint occurs after the first iteration of the repeat loop. When this happens, the I/O-instruction address can be pushed onto the stack during a #DB exception if the repeat loop is not complete. A subsequent return from the debug handler causes the next I/O iteration to be executed. If the breakpoint condition is still set, the #DB exception reoccurs after that iteration is complete.
Chapter 13: Debug and Performance Resources
399
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Single Stepping. Single-step breakpoints are enabled by setting the rFLAGS.TF bit to 1. When single stepping is enabled, a #DB exception occurs after every instruction is executed until it is disabled by clearing rFLAGS.TF to 0. However, the instruction that sets the TF bit, and the instruction that follows it, is not single stepped. When a #DB exception occurs due to single stepping, the processor clears rFLAGS.TF to 0 before entering the debug handler, so that the debug handler itself is not single stepped. The processor also sets DR6.BS to 1, which indicates that the #DB exception occurred as a result of single stepping. The rFLAGS image pushed onto the debug-handler stack has the TF bit set, and single stepping resumes when a subsequent IRET pops the stack image into the rFLAGS register. Single-step breakpoints have a higher priority than external interrupts. If an external interrupt occurs during single stepping, control is transferred to the #DB handler first, causing the rFLAGS.TF bit to be cleared to 0. Next, before the first instruction in the debug handler is executed, the processor transfers control to the pending-interrupt handler. This allows external interrupts to be handled outside of single-step mode. The INTn, INT3, and INTO instructions clear the rFLAGS.TF bit to 0 when they are executed. If a debugger is used to single-step software that contains these instructions, it must emulate them instead of executing them. The single-step mechanism can also be set to single step only control transfers, rather than single step every instruction. See "Single Stepping Control Transfers" on page 402 for additional information. Task-Switch Breakpoints. Breakpoints can be set in a task TSS to raise a #DB exception after a task switch. Software enables a task breakpoint by setting the T bit in the TSS to 1. When a task switch occurs into a task with the T bit set, the processor completes loading the new task state. Before the first instruction is executed, the #DB exception occurs, and the processor sets DR6.BT to 1, indicating that the #DB exception occurred as a result of task breakpoint. The processor does not clear the T bit in the TSS to 0 when the #DB exception occurs. Software must explicitly clear this bit to disable the task breakpoint. Software should never set the T-bit 400 Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
in the debug-handler TSS if a separate task is used for #DB exception handling, otherwise the processor loops on the debug handler. General-Detect Condition. General-detect is a special debugexception condition that occurs when software running at any privilege level attempts to access any of the DRn registers while DR7.GD is set to 1. When a #DB exception occurs due to the general-detect condition, the processor clears DR7.GD to 0 and sets DR6.BD to 1. Clearing DR7.GD to 0 allows the debug handler to access the DRn registers without causing infinite #DB exceptions. A debugger enables general detection to prevent other software from accessing and interfering with the debug registers while they are in use by the debugger. The exception is taken before executing the MOV DRn instruction so that the DRn contents are not altered. 13.2.3 Breakpoint Instruction (INT3) The INT3 instruction, or the INTn instruction with an operand of 3, can be used to set breakpoints that transfer control to the breakpoint-exception (#BP) handler rather than the debugexception handler. When a debugger uses the breakpoint instructions to set breakpoints, it does so by replacing the first bytes of an instruction with the breakpoint instruction. The debugger replaces the breakpoint instructions with the original-instruction bytes to clear the breakpoint. INT3 is a single-byte instruction while INTn with an operand of 3 is a two-byte instruction. The instructions have slightly different effects on the breakpoint exception-handler stack. See "#BP--Breakpoint Exception (Vector 3)" on page 252 for additional information on this exception. 13.2.4 ControlTransfer Breakpoint Features A control transfers is accomplished by using one of following instructions: Jumps (JMP, Jcc, JCXZ, JECXZ). LOOPx. Software interrupts (INTn, INT 3, INTO). CALL. RET, IRET. SYSCALL, SYSRET. SYSENTER, SYSEXIT. Chapter 13: Debug and Performance Resources 401
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Recording Control Transfers. Software enables control-transfer recording by setting DebugCtlMSR.LBR to 1. Before setting this bit, software can initialize the control-transfer recording MSRs (LastBranchToIP, LastBranchFromIP, LastExceptionToIP, LastExceptionFromIP) to a known value (such as all zeros) using the WRMSR instruction. After the control-transfer recording mechanism is initialized, the processor updates the recording MSRs automatically when control transfers occur. Every branch instruction, interrupt, and exception causes the LastBranchToIP register to be loaded with the target rIP, and the LastBranchFromIP register to be loaded with the source rIP. Interrupts and exceptions cause the processor to copy the contents of LastBranchToIP into L a s t E x c e p t i o n To I P, a n d L a s t B ra n c h Fro m I P i n t o LastExceptionFromIP, before updating LastBranchToIP and LastBranchFromIP. The processor automatically disables control-transfer recording w h e n a d e b u g e x c e p t i o n ( # D B ) o c c u r s by c l e a r i n g DebugCtlMSR.LBR to 0. The contents of the control-transfer recording MSRs are not altered by the processor when the #DB occurs. Before exiting the debug-exception handler, software can set DebugCtlMSR.LBR to 1 to re-enable the recording mechanism. Debuggers can trace a control transfer backward from a bug to its source using the recording MSRs and the breakpoint-address registers. The debug handler does this by updating the breakpoint registers from the recording MSRs after a #DB exception occurs, and restarting the program. The program takes a #DB exception on the previous control transfer, and this process can be repeated. The debug handler cannot simply copy the contents of the recording MSR into the breakpoint-address register. The recording MSRs hold segment offsets, while the debug registers hold virtual (linear) addresses. The debug handler must calculate the virtual address by reading the codesegment selector (CS) from the interrupt-handler stack, then reading the segment-base address from the CS descriptor, and adding that base address to the offset in the recording MSR. The calculated virtual-address can then be used as a breakpoint address. Single Stepping Control Transfers. Software can enable controltransfer single stepping by setting DebugCtlMSR.BTF to 1 and 402 Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
rFLAGS.TF to 1. The processor automatically disables controltransfer single stepping when a debug exception (#DB) occurs by clearing DebugCtlMSR.BTF to 0. rFLAGS.TF is also cleared when a #DB exception occurs. Before exiting the debugexception handler, software must set both DebugCtlMSR.BTF and rFLAGS.TF to 1 to restart single stepping. When enabled, this single-step mechanism causes a #DB exception to occur on every branch instruction, interrupt, or exception. Debuggers can use this capability to perform a "coarse" single step across blocks of code (bound by control transfers), and then, as the problem search is narrowed, switch i n t o a " f i n e " s i n g l e - s t e p m o d e o n eve ry i n s t r u c t i o n (DebugCtlMSR.BTF=0, rFLAGS.TF=1). Debuggers can use both the single-step mechanism and recording mechanism to support full backward and forward tracing of control transfers.
13.3
Performance Optimization
The AMD64 architecture supports the performance-monitoring features introduced in earlier processor implementations. These features allow the selection of events to be monitored, and include a set of corresponding counter registers that track the frequency of monitored events. Software tools can use these features to identify performance bottlenecks, such as sections of code that have high cache-miss rates or frequently mispredicted branches. This information can then be used as a guide for improving or eliminating performance problems t h ro u g h s o f t wa re o p t i m i z a t i o n s o r h a rd wa re - d e s i g n improvements. The performance-monitoring features include: A set of performance-counter registers that count the frequency or duration of specific processor events. A set of performance-event-select registers used to specify the events that are be tracked by the performance-counter registers. A time-stamp counter that can be used to count processorclock cycles over time. Implementations are not required to support the performance counters and the event-select registers, or the time-stamp
Chapter 13: Debug and Performance Resources
403
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
counter. The presence of these features can be determined by executing the CPUID instruction, and checking the returned feature bits to determine their availability. See "Processor Feature Identification" on page 78 for more information on using the CPUID instruction. The registers used in support of performance monitoring are model-specific registers (MSRs). See "Performance-Monitoring MSRs" on page 445 for a listing of the performance-monitoring MSR numbers and their reset values. Each feature, and its use, is described in the following sections. 13.3.1 Performance Counters Performance counters are used to count specific processor events, such as data-cache misses, or the duration of events, such as the number of clocks it takes to return data from memory after a cache miss. During event counting, the processor increments the counter when it detects an occurrence of the event. During duration measurement, the processor counts the number of processor clocks it takes to complete an event. Each performance counter can be used to count one event, or measure the duration of one event, at a time. Implementations of the AMD64 architecture can support any number of performance counters, PerfCtrn. The length, in bits, of the performance counters is also implementation dependent, but the maximum length supported is 64 bits. Figure 13-6 shows the format of the PerfCtrn register.
63 PerfCtrn 0
Figure 13-6. Performance Counter (PerfCtrn) The PerfCtrn registers are model-specific registers that can be read using a special read performance-monitoring counter instruction, RDPMC. The RDPMC instruction loads the contents of the PerfCtrn register specified by the ECX register, into the EDX register and the EAX register. The high 32 bits are loaded into EDX, and the low 32 bits are loaded into EAX. RDPMC can be executed only at CPL=0, unless system software enables use of the instruction at all privilege levels. RDPMC can be enabled for use at all privilege levels by setting CR4.PCE (the performance-monitor counter-enable bit) to 1. 404 Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
When CR4.PCE = 0 and CPL > 0, attempts to execute RDPMC result in a general-protection exception (#GP). The performance counters can also be read and written by system software running at CPL=0 using the RDMSR and WRMSR instructions, respectively. Writing the performance counters can be useful if software wants to count a specific number of events, and then trigger an interrupt when that count is reached. An interrupt can be triggered when a performance counter overflows (see "Counter Overflow" on page 408 for additional information). Software should use the WRMSR instruction to load the count as a two's-complement negative number into the performance counter. This causes the counter to overflow after counting the appropriate number of times. The performance counters are not guaranteed to produce identical measurements each time they are used to measure a particular instruction sequence, and they should not be used to take measurements of very small instruction sequences. The RDPMC instruction is not serializing, and it can be executed out-of-order with respect to other instructions around it. Even w h e n b o u n d by s e r i a l i z i n g i n s t r u c t i o n s , t h e s y s t e m environment at the time the instruction is executed can cause events to be counted before the counter value is loaded into EDX:EAX. 13.3.2 Performance Event-Select Registers Performance event-select registers (PerfEvtSeln) are 32-bit registers used to specify the events counted by the performance counters, and to control other aspects of their operation. Each performance counter supported by the implementation has a corresponding event-select register that controls its operation. Figure 13-7 on page 406 shows the format of the PerfEvtSeln register.
Chapter 13: Debug and Performance Resources
405
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
31 Counter Mask
24
23
I N V
22
E N
21 r
20
I N T
19
P C
18
E
17
O S
16
U S R
15 Unit Mask
8
7 Event Mask
0
Bits 31-24 23 22 21 20 19 18 17 16 15-8 7-0
Mnemonic INV EN Reserved INT PC E OS USR
Description Counter Mask Invert Mask Counter Enable Interrupt Enable Pin Control Edge Detect Operating-System Mode User Mode Unit Mask Event Mask
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Figure 13-7. Performance Event-Select Register (PerfEvtSeln) The fields within the PerfEvtSeln register are: Event Mask--Bits 7-0, read/write. This field specifies both the event or event duration to be counted by the corresponding PerfCtrn register. The events that can be counted are implementation dependent. For more information, refer to the BIOS writer's guide for the implementation. Unit Mask--Bits 15-8, read/write. This field can be used to specify a particular processor unit to be monitored, if the event counted can be produced by multiple units within the processor. Implementations can also use this field to further specify or qualify a monitored event. Operating-System Mode (OS) and User Mode (USR)--Bits 17-16 (respectively), read/write. Software uses these bits to control the privilege level at which event counting is performed according to Table 13-3. Table 13-3. Operating-System Mode and User Mode Bits
USR Mode (Bit 16) 0 1 0 1 Event Counting No counting. Only at CPL > 0. Only at CPL = 0. At all privilege levels.
OS Mode (Bit 17) 0 0 1 1
406
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Edge Detect (E)--Bit 18, read/write. Software sets this bit to 1 to count the number of edge transitions from the negated to asserted state. This feature is useful when coupled with event-duration monitoring, as it can be used to calculate the average time spent in an event. Clearing this bit to 0 disables edge detection. Pin Control (PC)--Bit 19, read/write. Software sets this bit to 1 to cause the external PMi pins on the processor to toggle when the counter overflows. When this bit is cleared to 0, the processor toggles the PMi pins each time it increments the performance counter. Interrupt Enable (INT)--Bit 20, read/write. Software sets this bit to 1 to enable an interrupt to occur when the performance counter overflows (see "Counter Overflow" on page 408 for additional information). Clearing this bit to 0 disables the triggering of the interrupt. Counter Enable (EN)--Bit 22, read/write. Software sets this bit to 1 to enable the PerfEvtSeln register, and counting in the corresponding PerfCtrn register. Clearing this bit to 0 disables the register pair. Invert Mask (INV)--Bit 23, read/write. Software sets this bit to 1 to invert the comparison result performed on the counter-mask field, so that a less-than or equal-to comparison can be performed. Clearing this bit to 0 leaves the comparison result alone, so that a greater-than or equalto comparison is reported. Counter Mask--Bits 31-24, read/write. This field is used to set a threshold for counting multiple events that can occur in a single clock. If the number of events occurring in the single clock is greater than or equal to this field, the corresponding PerfCtrn register is incremented. PerfCtrn is not incremented if the number of events is less than the count mask. The INV bit, when set, causes the PerfCtrn register to be incremented when the comparison is less than or equal to the count mask. In this case, PerfCtrn is not incremented if the number of events is greater than the count mask. The performance event-select registers can be read and written only by system software running at CPL=0 using the RDMSR and WRMSR instructions, respectively. Any attempt to read or write these registers at CPL>0 causes a general-protection exception to occur. Chapter 13: Debug and Performance Resources 407
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
13.3.3 Using Performance Counters
Starting and Stopping. Performance counting in a PerfCtrn register is initiated by setting the corresponding PerfEvtSeln.EN bit to 1. Counting is stopped by clearing PerfEvtSeln.EN to 0. Software must initialize the remaining PerfEvtSeln fields with the appropriate setup information before or at the same time EN is set. Counting begins when the WRMSR instruction that sets PerfEvtSeln.EN to 1 completes execution. Counting stops when the WRMSR instruction that clears PerfEvtSeln.EN to 0 completes execution. Counter Overflow. Some processor implementations support an interrupt-on-overflow capability that allows an interrupt to occur when one of the PerfCtrn registers overflows. The source and type of interrupt is implementation dependent. Some implementations cause a debug interrupt to occur, while others make use of the local APIC to specify the interrupt vector and trigger the interrupt when an overflow occurs. Software controls the triggering of an interrupt by setting or clearing the PerfEvtSeln.INT bit. If system software makes use of the interrupt-on-overflow capability, an interrupt handler must be provided that can record information relevant to the counter overflow. Before returning from the interrupt handler, the performance counter can be re-initialized to its previous state so that another interrupt occurs when the appropriate number of events are counted.
13.3.4 Time-Stamp Counter
The time-stamp counter (TSC) is used to count processor-clock cycles. The TSC is cleared to 0 after a processor reset. After a reset, the TSC is incremented by one for every processor clock cycle. Each time the TSC is read, it returns a monotonicallylarger value than the previous value read from the TSC. When the TSC contains all ones, it wraps to zero. The TSC in a 1-GHz processor counts for almost 600 years before it wraps. Figure 13-8 shows the format of the 64-bit time-stamp counter (TSC).
0 TSC
63
Figure 13-8. Time-Stamp Counter (TSC)
408
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
The TSC is a model-specific register that can also be read using a special read time-stamp counter instruction, RDTSC. RDTSC loads the contents of the TSC into the EDX register and the EAX register. The high 32 bits are loaded into EDX, and the low 32 bits are loaded into EAX. RDTSC can be executed at any privilege level and from any processor mode. However, system software can disable the RDTSC instruction for programs that run at CPL > 0 by setting CR4.TSD (the time-stamp disable bit) to 1. When CR4.TSD = 1 and CPL > 0, attempts to execute RDSTC result in a general-protection exception (#GP). Some implementations allow the TSC register to be read and written using the RDMSR and WRMSR instructions, respectively. Support of this capability, however, is not required by the architecture, and software should avoid using these instructions to access the TSC. The programmer should use the CPUID instruction to determine whether these features are supported. If bit 4 returned by CPUID standard function 1 is set, then the processor supports TSC, the RDTSC instruction and CR4.TSD. The TSC register can be used by performance-analysis applications, along with the performance-monitoring registers, to help determine the relative frequency of an event or its duration. Software can also use the TSC to time software routines to help identify candidates for optimization. In general, the TSC should not be used to take very short time measurements, because the resulting measurement is not guaranteed to be identical each time it is made. The RDTSC instruction is not serializing, and it can be executed out-oforder with respect to other instructions around it. Even when bound by serializing instructions, the system environment at the time the instruction is executed can cause additional cycles to be counted before the TSC value is loaded into EDX:EAX.
Chapter 13: Debug and Performance Resources
409
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
410
Chapter 13: Debug and Performance Resources
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
14
Processor Initialization and Long-Mode Activation
This chapter describes the hardware actions taken following a processor reset and the steps that must be taken to initialize processor resources and activate long mode. In some cases the actions required are implementation-specific with references made to the appropriate implementation-specific documentation.
14.1
Reset and Initialization
System logic initializes the processor in one of two ways. One method is to assert an external reset signal (typically designated RESET#). The other method, referred to here as INIT, is performed using implementation-dependent external signalling. Both initialization techniques place the processor in real mode and initialize processor resources to a known, consistent state from which software can begin execution. The differences between the two methods are: RESET#--This method provides an optional built-in self test (BIST) that can be performed as part of the RESET# process. INIT--This method does not modify the following state: - Memory hierarchy, including internal and external caches. - 128-bit media, 64-bit media, or x87 resources. - Memory-type range registers (MTRRs). - Machine-check registers. Some, but not all, model-specific registers (MSRs) are modified by an INIT. BIST cannot be performed as part of the INIT process. The processor always performs RESET# after it is powered up, but RESET# can be performed at any time. An INIT can be performed at any time after the processor is powered up.
14.1.1 Built-In Self Test (BIST)
An optional built-in self-test can be performed after RESET# is a s s e r t e d . Th e m e ch a n i s m fo r t r i g g e r i n g t h e B I S T i s implementation-specific, and can be found in the hardware documentation for the implementation. The number of processor cycles BIST can consume before completing is also 411
Chapter 14: Processor Initialization and Long-Mode Activation
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
implementation-specific but typically consumes several million cycles. BIST can be used by system implementations to assist in verifying system integrity, thereby improving system reliability, availability, and serviceability. The internal BIST hardware generally tests all internal array structures for errors. These structures can include (but are not limited to): All internal caches, including the tag arrays as well as the data arrays. All TLBs. Internal ROMs, such as the microcode ROM and floatingpoint constant ROM. Branch-prediction structures. EAX is loaded with zero if BIST completes without detecting errors. If any hardware faults are detected during BIST, a nonzero value is loaded into EAX. 14.1.2 Clock Multiplier Selection The internal processor clock runs at some multiple of the system clock. The processor-to-system clock multiple does not have to be fixed by a processor implementation but instead can be programmable through hardware or software, or some combination of the two. For information on selecting the processor-clock multiplier, refer to the BIOS writer's guide for the implementation. Table 14-1 shows the initial processor state following either RESET# or INIT. Except as indicated, processor resources generally are set to the same value after either RESET# or INIT.
14.1.3 Processor Initialization State
Table 14-1. Initial Processor State
Processor Resource CR0 CR2, CR3, CR4 CR8 RFLAGS Reset (RESET#) Value 0000_0000_6000_0010h 0 0 0000_0000_0000_0002h Not modified Initialization (INIT) Value CD and NW are unchanged Bit 4 (reserved) = 1 All others = 0
412
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 14-1. Initial Processor State (continued)
Processor Resource EFER RIP 0 0000_0000_0000_FFF0h Selector = F000h Base = 0000_0000_FFFF_0000h Limit = FFFFh Attributes = See Table 14-2 Selector = 0000h Base = 0 Limit = FFFFh Attributes = See Table 14-2 Base = 0 Limit = FFFFh Selector = 0000h Base = 0 Limit = FFFFh Attributes = See Table 14-2 0 (non-zero if BIST is run and fails) 0 Reset (RESET#) Value Initialization (INIT) Value
CS
DS, ES, FS, GS, SS
GDTR, IDTR
LDTR, TR
RAX RDX RBX, RCX, RBP, RSP, RDI, RSI, R8, R9, R10, R11, R12, R13, R14, R15
Family/Model/Stepping, including extended family and extended model--see "Processor Implementation Information" on page 416 0 FPR0-FPR7 = 0 Control Word = 0040h Status Word = 0000h Tag Word = 5555h Instruction CS = 0000h Instruction Offset = 0 x87 Instruction Opcode = 0 Data-Operand DS = 0000h Data-Operand Offset = 0 MMX0-MMX7 = 0
x87 Floating-Point State
Not modified
64-Bit Media State
Not modified
Chapter 14: Processor Initialization and Long-Mode Activation
413
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 14-1. Initial Processor State (continued)
Processor Resource 128-Bit Media State Memory-Type Range Registers Machine-Check Registers DR0, DR1, DR2, DR3 DR6 DR7 Time-Stamp Counter Performance-Monitor Resources Other Model-Specific Registers Instruction and Data Caches Instruction and Data TLBs APIC SMRAM Base Address (SMBASE) Reset (RESET#) Value XMM0-XMM15 = 0 MXCSR = 1F80h See "Memory-Typing MSRs" on page 440 See "Machine-Check MSRs" on page 443 0 0000_0000_FFFF_0FF0h 0000_0000_0000_0400h 0 See "Performance-Monitoring MSRs" on page 445 See "MSR Cross-Reference" on page 435 Invalidated Enabled 0003_0000h Not modified Not modified Not modified Not modified Not modified Not modified Initialization (INIT) Value Not modified Not modified Not modified
Table 14-2 on page 415 shows the initial state of the segmentregister attributes (located in the hidden portion of the segment registers) following either RESET# or INIT.
414
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table 14-2. Initial State of Segment-Register Attributes
Attribute G D/B L (CS Only) P DPL Code Segment Data Segment LDTR TR Value (Binary) 0 0 0 1 00 S=1 Type = 1010 S=1 Type = 0010 S=0 Type = 0010 S=0 Type = 0011 Description Byte Granularity 16-Bit Segment Legacy-Mode Segment Segment is Present Privilege-Level 0 Executable/Readable Code Segment Read/Write Data Segment LDT Busy 16-Bit TSS
S and Type
14.1.4 Multiple Processor Initialization
Following reset in multiprocessor configurations, the processors use a multiple-processor initialization protocol to negotiate which processor becomes the bootstrap processor. This bootstrap processor then executes the system initialization code while the remaining processors wait for software initialization to complete. For further information, see the documentation for particular implementations of the architecture. After a RESET# or INIT, the processor is operating in 16-bit real mode. Normally within real mode, the code-segment baseaddress is formed by shifting the CS-selector value left four bits. The base address is then added to the value in EIP to form the physical address into memory. As a result, the processor can only address the first 1 Mbyte of memory when in real mode. However, immediately following RESET# or INIT, the CSselector register is loaded with F000h, but the CS base-address is not formed by left-shifting the selector. Instead, the CS baseaddress is initialized to FFFF_0000h. EIP is initialized to
14.1.5 Fetching the First Instruction
Chapter 14: Processor Initialization and Long-Mode Activation
415
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
FFF0h. Therefore, the first instruction fetched from memory is located at physical-address FFFF_FFF0h (FFFF_0000h + 0000_FFF0h). The CS base-address remains at this initial value until the CSselector register is loaded by software. This can occur as a result of executing a far jump instruction or call instruction, for example. When CS is loaded by software, the new base-address value is established as defined for real mode (by left shifting the selector value four bits).
14.2
Hardware Configuration
Software can read processor-identification information from the EDX register immediately following RESET# or INIT. This information can be used to initialize software to perform processor-specific functions. The information stored in EDX is defined as follows: Stepping ID (bits 3-0)--This field identifies the processorrevision level. Extended Model (bits 19-16) and Model (bits 7-4)--These fields combine to differentiate processor models within a instruction family. For example, two processors may share the same microarchitecture but differ in their feature set. Such processors are considered different models within the same instruction family. This is a split field, comprising an extended-model portion in bits 19-16 with a legacy portion in bits 7-4 Extended Family (bits 27-20) and Family (bits 11-8)--These fields combine to differentiate processors by their microarchitecture. The CPUID instruction can be used to obtain the same information. This is done by executing CPUID with either s t a n d a rd - f u n c t i o n c o d e 1 o r ex t e n d e d - f u n c t i o n c o d e 8000_0001h. Additional information about the processor and the features supported can be gathered using CPUID with other feature codes. See "Processor Feature Identification" on page 78 for additional information.
14.2.1 Processor Implementation Information
14.2.2 Enabling Internal Caches
Following a RESET# (but not an INIT), all instruction and data caches are disabled, and their contents are invalidated (the MOESI state is set to the invalid state). Software can enable Chapter 14: Processor Initialization and Long-Mode Activation
416
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
these caches by clearing the cache-disable bit (CR0.CD) to zero (RESET# sets this bit to 1). Software can further refine caching based on individual pages and memory regions. Refer to "Cache Control Mechanisms" on page 212 for more information on cache control. Memory-Type Range Registers (MTRRs). Following a RESET# (but not an INIT), the MTRRdefType register is cleared to 0, which disables the MTRR mechanism. The variable-range and fixedrange MTRR registers are not initialized and are therefore in an undefined state. Before enabling the MTRR mechanism, the initialization software (usually BIOS) must load these registers with a known value to prevent unexpected results. Clearing these registers, for example, sets memory to the uncacheable (UC) type. 14.2.3 Initializing Media and x87 Processor State Some resources used by x87 floating-point instructions and 128bit media instructions must be initialized by software before being used. Initialization software can use the CPUID instruction to determine whether the processor supports these instructions, and then initialize their resources as appropriate. x87 Floating-Point State Initialization. Table 14-3 on page 418 shows the differences between the initial x87 floating-point state fol l owi n g a R E S E T # a n d t h e state established by the FINIT/FNINIT instruction. An INIT does not modify the x87 floating-point state. The initialization software can execute an FINIT or FNINIT instruction to prepare the x87 floating-point unit for use by application software. The FINIT and FNINIT instructions have no effect on the 64-bit media state.
Chapter 14: Processor Initialization and Long-Mode Activation
417
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table 14-3. x87 Floating-Point State Initialization
x87 Floating-Point Resource FPR0-FPR7 0 0040h * Round to nearest * Single precision * Unmask all exceptions 0000h 5555h (FPRn contain zero) 0000h 0 0 0000h 0 FFFFh (FPRn are empty) RESET# FINIT/FNINIT Instructions Not modified 037Fh * Round to nearest * Extended precision * Mask all exceptions
Control Word
Status Word Tag Word Instruction CS Instruction Offset x87 Instruction Opcode Data-Operand DS Data-Operand Offset
Initialization software should also load the MP, EM, and NE bits in the CR0 register as appropriate for the operating system. The recommended settings for implementations of the AMD64 architecture are: MP=1--Setting MP to 1 causes a device-not-available exception (#NM) to occur when the FWAIT/WAIT instruction is executed and the task-switched bit (CR0.TS) is set to 1. This supports operating systems that perform lazy context-switching of x87 floating-point state. EM=0--Clearing EM to 0 allows the x87 floating-point unit to execute instructions rather than causing a #NM exception (CR0.EM=1). System software sets EM to 1 only when software emulation of x87 instructions is desired. NE=1--Setting NE to 1 causes x87 floating-point exceptions to be handled by the floating-point exception-pending exception (#MF) handler. Clearing this bit causes the processor to externally indicate the exception occurred, and an external device can then cause an external interrupt to occur in response.
418
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Refer to "CR0 Register" on page 55 for additional information on these control bits. 64-Bit Media State Initialization. There are no special requirements placed on software to initialize the processor state used by 64bit media instructions. This state is initialized completely by the processor following a RESET#. System software should leave CR0.EM cleared to 0 to allow execution of the 64-bit media instructions. If CR0.EM is set to 1, attempted execution of the 64-bit media instructions causes an invalid-opcode exception (#UD ). The 64-bit media state is not modified by an INIT. 128-Bit Media State Initialization. BIOS or system software must also prepare the processor to allow execution of 128-bit media instructions. The required preparations include: Leaving CR0.EM cleared to 0 to allow execution of the 128bit media instructions. If CR0.EM is set to 1, attempted execution of the 128-bit media instructions causes an invalid-opcode exception (#UD). Enabling the 128-bit media instructions by setting CR4.OSFXSR to 1. Software cannot execute the 128-bit media instructions unless this bit is set. Setting this bit also indicates that system software uses the FXSAVE and FXRSTOR instructions to save and restore, respectively, the 128-bit media state. These instructions also save and restore the 64-bit media state and x87 floating-point state. Indicating that system software uses the SIMD floatingpoint exception (#XF) for handling 128-bit media floatingpoint exceptions. This is done by setting CR4.OSXMMEXCPT to 1. Setting (optionally) the MXCSR mask bits to mask or unmask 128-bit media floating-point exceptions as desired. Because this register can be read and written by application software, it is not absolutely necessary for system software to initialize it. Refer to "CR4 Register" on page 60 for additional information on these CR4 control bits. 14.2.4 Model-Specific Initialization Implementations of the AMD64 architecture can contain model-specific features and registers that are not initialized by t h e p ro c e s s o r a n d t h e re fo re re q u i re s y s t e m - s o f t wa re 419
Chapter 14: Processor Initialization and Long-Mode Activation
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
initialization. System software must use the CPUID instruction to determine which features are supported. Model-specific features are generally configured using model-specific registers (MSRs), which can be read and written using the RDMSR and WRMSR instructions, respectively. Some of the model-specific features are pervasive across many processor implementations of the AMD64 architecture and are therefore described within this volume. These include: System-call extensions, which must be enabled in the EFER register before using the SYSCALL and SYSRET instructions. See "System-Call Extension (SCE) Bit" on page 70 for information on enabling these instructions. Memory-typing MSRs. See "Memory-Type Range Registers (MTRRs)" on page 417 for information on initializing and using these registers. The machine-check mechanism. See "Initializing the Machine-Check Mechanism" on page 315 for information on enabling and using this capability. Extensions to the debug mechanism. See "Software-Debug Resources" on page 384 for information on initializing and using these extensions. The performance-monitoring resources. See "Performance Optimization" on page 403 for information on initializing and using these resources. Initialization of other model-specific features used by the paget ra n s l a t i o n m e ch a n i s m a n d l o n g m o d e a re d e s c r i b e d throughout the remainder of this section. Some model-specific features are not pervasive across processor implementations and are therefore not described in this volume. For more information on these features and their initialization requirements, refer to the BIOS writer's guide for the implementation.
14.3
Initializing Real Mode
A basic real-mode (real-address-mode) operating environment must be initialized so that system software can initialize the protected-mode operating environment. This real-mode environment must include:
420
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
A real-mode IDT for vectoring interrupts and exceptions to the appropriate handlers while in real mode. The IDT baseaddress value in the IDTR initialized by the processor can be used, or system software can relocate the IDT by loading a new base-address into the IDTR. The real-mode interrupt and exception handlers. These must be loaded before enabling external interrupts. Because the processor can always accept a non-maskable interrupt (NMI), it is possible an NMI can occur before initializing the IDT or the NMI handler. System hardware must provide a mechanism for disabling NMIs to allow time for the IDT and NMI handler to be properly initialized. Alternatively, the IDT and NMI handler can be stored in non-volatile memory that is referenced by the initial values loaded into the IDTR. Maskable interrupts can be enabled by setting EFLAGS.IF after the real-mode IDT and interrupt handlers are initialized. A valid stack pointer (SS:SP) to be used by the interrupt mechanism should interrupts or exceptions occur. The values of SS:SP initialized by the processor can be used. One or more data-segment selectors for storing the protected-mode data structures that are created in real mode. Once the real-mode environment is established, software can begin initializing the protected-mode environment.
14.4
Initializing Protected Mode
Protected mode must be entered before activating long mode. A minimal protected-mode environment must be established to allow long-mode initialization to take place. This environment must include the following: A protected-mode IDT for vectoring interrupts and exceptions to the appropriate handlers while in protected mode. The protected-mode interrupt and exception handlers referenced by the IDT. Gate descriptors for each handler must be loaded in the IDT.
Chapter 14: Processor Initialization and Long-Mode Activation
421
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
A GDT which contains: - A code descriptor for the code segment that is executed in protected mode. A read/write data segment that can be used as a protected-mode stack. This stack can be used by the interrupt mechanism if interrupts or exceptions occur.
Software can optionally load the GDT with one or more data segment descriptors, a TSS descriptor, and an LDT descriptor for use by long-mode initialization software. After the protected-mode data structures are initialized, system software must load the IDTR and GDTR (and optionally, the LDTR and TR) with pointers to those data structures. Once these registers are initialized, protected mode can be enabled by setting CR0.PE to 1. If legacy paging is used during the long-mode initialization process, the page-translation tables must be initialized before enabling paging. At a minimum, one page directory and one page table are required to support page translation. The CR3 register must be loaded with the starting physical address of the highest-level table supported in the page-translation hierarchy. After these structures are initialized and protected mode is enabled, paging can be enabled by setting CR0.PG to 1.
14.5
Initializing Long Mode
From protected mode, system software can initialize the data structures required by long mode and store them anywhere in the first 4 Gbytes of physical memory. These data structures can be relocated above 4 Gbytes once long mode is activated. The data structures required by long mode include the following: An IDT with 64-bit interrupt-gate descriptors. Long-mode interrupts are always taken in 64-bit mode, and the 64-bit gate descriptors are used to transfer control to interrupt handlers running in 64-bit mode. See "Long-Mode Interrupt Control Transfers" on page 287 for more information. The 64-bit mode interrupt and exception handlers to be used in 64-bit mode. Gate descriptors for each handler must be loaded in the 64-bit IDT.
422
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
A GDT containing segment descriptors for software running in 64-bit mode and compatibility mode, including: - Any LDT descriptors required by the operating system or application software. - A TSS descriptor for the single 64-bit TSS required by long mode. - Code descriptors for the code segments that are executed in long mode. The code-segment descriptors are used to specify whether the processor is operating in 64-bit mode or compatibility mode. See "Code-Segment Descriptors" on page 108, "Long (L) Attribute Bit" on page 109, and "CS Register" on page 87 for more information. - Data-segment descriptors for software running in compatibility mode. The DS, ES, and SS segments are ignored in 64-bit mode. See "Data-Segment Descriptors" on page 110 for more information. - FS and GS data-segment descriptors for 64-bit mode, if required by the operating system. If these segments are used in 64-bit mode, system software can also initialize the full 64-bit base addresses using the WRMSR instruction. See "FS and GS Registers in 64-Bit Mode" on page 88 for more information. The existing protected-mode GDT can be used to hold the long-mode descriptors described above. A single 64-bit TSS for holding the privilege-level 0, 1, and 2 stack pointers, the interrupt-stack-table pointers, and the I/O-redirection-bitmap base address (if required). This is the only TSS required, because hardware task-switching is not supported in long mode. See "64-Bit Task State Segment" on page 370 for more information. The 4-level page-translation tables required by long mode. Long mode also requires the use of physical-address extensions (PAE) to support physical-address sizes greater than 32 bits. See "Long-Mode Page Translation" on page 160 for more information. If paging is enabled during the initialization process, it must be disabled before enabling long mode. After the long-mode data structures are initialized, and paging is disabled, software can enable and activate long mode.
Chapter 14: Processor Initialization and Long-Mode Activation
423
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
14.6
Enabling and Activating Long Mode
Long mode is enabled by setting the long-mode enable control bit (EFER.LME) to 1. However, long mode is not activated until software also enables paging. When software enables paging while long mode is enabled, the processor activates long mode, which the processor indicates by setting the long-mode-active status bit (EFER.LMA) to 1. The processor behaves as a 32-bit x86 processor in all respects until long mode is activated, even if long mode is enabled. None of the new 64-bit data sizes, addressing, or system aspects available in long mode can be used until EFER.LMA=1. Table 14-4 shows the control-bit settings for enabling and activating the various operating modes of the AMD64 architecture. The default address and data sizes are shown for each mode. For the methods of overriding these default address and data sizes, see "Instruction Prefixes" in Volume 3. Table 14-4. Processor Operating Modes
Encoding Mode EFER.LMA1 Default Address Size (bits)2 Default Data Size (bits)2
Long Mode
64-Bit Mode 1 Compatibility Mode Legacy Mode 0
1
CS.D 0 1 0 1 0
CS.L
64 32 16 32 16
32 32 16 32 16
0
x
Note:
1. EFER.LMA is set by the processor when software sets EFER.LME and CR0.PG according to the sequence described in "Activating Long Mode" on page 425. 2. See "Instruction Prefixes" in Volume 1 for overrides to default sizes.
Long mode uses two code-segment-descriptor bits, CS.L and CS.D, to control the operating submodes. If long mode is active, CS.L = 1, and CS.D = 0, the processor is running in 64-bit mode, as shown in Table 14-4. With this encoding (CS.L=1, CS.D=0), 424 Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
default operand size is 32 bits and default address size is 64 bits. Using instruction prefixes, the default operand size can be overridden to 64 bits or 16 bits, and the default address size can be overridden to 32 bits. The final encoding of CS.L and CS.D in long mode (CS.L=1, CS.D=1) is reserved for future use. When long mode is active and CS.L is cleared to 0, the processor i s i n c o m p a t i b i l i t y m o d e , a s s h ow n i n Tab l e 1 4 -4 . I n compatibility mode, CS.D controls default operand and address sizes exactly as it does in the legacy x86 architecture. Setting CS.D to 1 specifies default operand and address sizes as 32 bits. Clearing CS.D to 0 specifies default operand and address sizes as 16 bits. 14.6.1 Activating Long Mode Switching the processor to long mode requires several steps. In general, the sequence involves disabling paging (CR0.PG=0), enabling physical-address extensions (CR4.PAE=1), loading CR3, enabling long mode (EFER.LME=1), and finally enabling paging (CR0.PG=1). Specifically, software must follow this sequence to activate long mode: 1. If starting from page-enabled protected mode, disable paging by clearing CR0.PG to 0. This requires that the MOV CR0 instruction used to disable paging be located in an identity-mapped page (virtual address equals physical address). 2. In any order: - Enable physical-address extensions by setting CR4.PAE to 1. Long mode requires the use of physical-address extensions (PAE) in order to support physical-address sizes greater than 32 bits. Physical-address extensions must be enabled before enabling paging. - Load CR3 with the physical base-address of the level-4 page-map-table (PML4). See "Long-Mode Page Translation" on page 160 for details on creating the 4level page translation tables required by long mode. - Enable long mode by setting EFER.LME to 1.
Chapter 14: Processor Initialization and Long-Mode Activation
425
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
3. Enable paging by setting CR0.PG to 1. This causes the processor to set the EFER.LMA bit to 1. The instruction following the MOV CR0 that enables paging must be a branch, and both the MOV CR0 and the following branch instruction must be located in an identity-mapped page. 14.6.2 Consistency Checks The processor performs long-mode consistency checks whenever software attempts to modify any of the control bits directly involved in activating long mode (EFER.LME, CR0.PG, and CR4.PAE). A general-protection exception (#GP) occurs when a consistency check fails. Long-mode consistency checks ensure that the processor does not enter an undefined mode or state that results in unpredictable behavior. Long-mode consistency checks cause a general-protection exception (#GP) to occur if: An attempt is made to enable or disable long mode while paging is enabled. Long mode is enabled, and an attempt is made to enable paging before enabling physical-address extensions (PAE). Long mode is enabled, and an attempt is made to enable paging while CS.L=1. Long mode is active and an attempt is made to disable physical-address extensions (PAE). Table 14-5 summarizes the long-mode consistency checks made during control-bit transitions. Table 14-5. Long-Mode Consistency Checks
Control Bit EFER.LME Transition 01 10 01 10 Check If (CR0.PG=1) then #GP(0) If (CR0.PG=1) then #GP(0) If ((EFER.LME=1) & (CR4.PAE=0) then #GP(0) If ((EFER.LME=1) & (CS.L=1)) then #GP(0) If (EFER.LMA=1) then #GP(0)
CR0.PG CR4.PAE
14.6.3 Updating System Descriptor Table References
Immediately after activating long mode, the system-descriptortable registers (GDTR, LDTR, IDTR, TR) continue to reference legacy descriptor tables. The tables referenced by these descriptors all reside in the lower 4 Gbytes of virtual-address Chapter 14: Processor Initialization and Long-Mode Activation
426
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
space. After activating long mode, 64-bit operating-system s o f t ware s ho u l d u s e t he L G D T, L L D T, L I D T, a n d LT R instructions to load the system descriptor-table registers with references to the 64-bit versions of the descriptor tables. See "Descriptor Tables" on page 89 for details on descriptor tables in long mode. Long mode requires 64-bit interrupt-gate descriptors to be stored in the interrupt-descriptor table (IDT). Software must not allow exceptions or interrupts to occur between the time long mode is activated and the subsequent update of the interrupt-descriptor-table register (IDTR) that establishes a reference to the 64-bit IDT. This is because the IDTR continues to reference a 32-bit IDT immediately after long mode is activated. If an interrupt or exception occurred before updating the IDTR, a legacy 32-bit interrupt gate would be referenced and interpreted as a 64-bit interrupt gate, with unpredictable results. External interrupts can be disabled using the CLI instruction. Non-maskable interrupts (NMI) and system-management interrupts (SMI) must be disabled using external hardware. See "Long-Mode Interrupt Control Transfers" on page 287 for more information on long mode interrupts. 14.6.4 Relocating Page-Translation Tables The long-mode page-translation tables must be located in the first 4 Gbytes of physical-address space before activating long mode. This is necessary because the MOV CR3 instruction used to initialize the page-map level-4 base address must be executed in legacy mode before activating long mode. Because the MOV CR3 is executed in legacy mode, only the low 32 bits of the register are written, which limits the location of the pagemap level-4 translation table to the low 4 Gbytes of memory. Software can relocate the page tables anywhere in physical memory, and re-initialize the CR3 register, after long mode is activated.
14.7
Leaving Long Mode
To return from long mode to legacy protected mode with paging enabled, software must deactivate and disable long mode using the following sequence: 1. Switch to compatibility mode and place the processor at the highest privilege level (CPL=0).
Chapter 14: Processor Initialization and Long-Mode Activation
427
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
2. Deactivate long mode by clearing CR0.PG to 0. This causes the processor to clear the LMA bit to 0. The MOV CR0 instruction used to disable paging must be located in an identity-mapped page. Once paging is disabled, the processor behaves as a standard 32-bit x86 processor. 3. Load CR3 with the physical base-address of the legacy page tables. 4. Disable long mode by clearing EFER.LME to 0. 5. Enable legacy page-translation by setting CR0.PG to 1. The instruction following the MOV CR0 that enables paging must be a branch, and both the MOV CR0 and the following branch instruction must be located in an identity-mapped page.
14.8
Long-Mode Initialization Example
Following is sample code that outlines the steps required to place the processor in long mode.
mydata segment para
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ; ; This generic data-segment holds pseudo-descriptors used ; by the LGDT and LIDT instructions. ; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; ; Establish a temporary 32-bit GDT and IDT. ; pGDT32 label fword dw dd fword dw dd ; Used by LGDT. ; GDT limit ... ; and 32-bit GDT base ; Used by LIDT. ; IDT limit ... ; and 32-bit IDT base
gdt32_limit gdt32_base
pIDT32
label
idt32_limit idt32_base
; ; Establish a 64-bit GDT and IDT (64-bit linear base; address) ;
428
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
pGDT64
label
tbyte dw dq tbyte dw dq
gdt64_limit gdt64_base
; Used by LGDT. ; GDT limit ... ; and 64-bit GDT base ; Used by LIDT. ; IDT limit ... ; and 64-bit IDT base
pIDT64
label
idt64_limit idt64_base
mydata
ends
; end of data segment
code16
segment para use16 ; 16-bit code segment
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ; ; 16-bit code, real mode ; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ; ; Initialize DS to point to the data segment containing ; pGDT32 and PIDT32. Set up a real-mode stack pointer, SS:SP, ; in case of interrupts and exceptions. ; cli mov ax, seg mydata mov ds, ax mov ax, seg mystack mov ss, ax mov sp, esp0
; ; Use CPUID to determine if the processor supports long mode. ; mov cpuid cmp jbe mov cpuid bt jnc eax, 80000000h ; ; eax, 80000000h ; no_long_mode ; eax, 80000001h ; ; edx, 29 ; no_long_mode ; Extended-function code 8000000h. Is largest extended function any function > 80000000h? If not, no long mode. Extended-function code 8000001h. Now EDX = extended-features flags. Test if long mode is supported. Exit if not supported.
; ; Load the 32-bit GDT before entering protected mode. ; This GDT must contain, at a minimum, the following ; descriptors:
Chapter 14: Processor Initialization and Long-Mode Activation
429
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
; ; ; ; ; ; ; ; ; ; ; ;
1) a CPL=0 16-bit code descriptor for this code segment. 2) a CPL=0 32/64-bit code descriptor for the 64-bit code. 3) a CPL=0 read/write data segment, usable as a stack (referenced by SS). Load the 32-bit IDT, in case any interrupts or exceptions occur after entering protected mode, but before enabling long mode). Initialize the GDTR and IDTR to point to the temporary 32-bit GDT and IDT, respectively.
lgdt lidt
ds:[pGDT32] ds:[pIDT32]
; ; Enable protected mode (CR0.PE=1). ; mov mov eax, 000000011h cr0, eax
; ; ; ; ; ;
Execute a far jump to turn protected mode on. code16_sel must point to the previously-established 16-bit code descriptor located in the GDT (for the code currently being executed).
db dw dw
0eah ;Far jump... offset now_in_prot;to offset... code16_sel ;in current code segment.
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ; ; At this point we are in 16-bit protected mode, but long ; mode is still disabled. ; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; now_in_prot:
; ; Set up the protected-mode stack pointer, SS:ESP. ; Stack_sel must point to the previously-established stack ; descriptor (read/write data segment), located in the GDT.
430
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
; Skip setting DS/ES/FS/GS, because we are jumping right to ; 64-bit code. ; mov mov mov ax, stack_sel ss, ax esp, esp0
; ; ; ; ; ;
Enable the 64-bit page-translation-table entries by setting CR4.PAE=1 (this is _required_ before activating long mode). Paging is not enabled until after long mode is enabled.
mov bts mov
eax, cr4 eax, 5 cr4, eax
; ; ; ; ; ; ;
Create the long-mode page tables, and initialize the 64-bit CR3 (page-table base address) to point to the base of the PML4 page table. The PML4 page table must be located below 4 Gbytes because only 32 bits of CR3 are loaded when the processor is not in 64-bit mode.
mov mov
eax, pml4_base ; Pointer to PML4 table (<4GB). cr3, eax ; Initialize CR3 with PML4 base.
; ; Enable long mode (set EFER.LME=1). ; mov ecx, 0c0000080h rdmsr bts eax, 8 wrmsr ; ; ; ; EFER MSR number. Read EFER. Set LME=1. Write EFER.
; ; Enable paging to activate long mode (set CR0.PG=1) ; mov bts mov eax, cr0 eax, 31 cr0, eax ; Read CR0. ; Set PE=1. ; Write CR0.
Chapter 14: Processor Initialization and Long-Mode Activation
431
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
; ; ; ; ; ; ; ; ;
At this point, we are in 16-bit compatibility mode ( LMA=1, CS.L=0, CS.D=0 ). Now, jump to the 64-bit code segment. The offset must be equal to the linear address of the 64-bit entry point, because 64-bit code is in an unsegmented address space. The selector points to the 32/64-bit code selector in the current GDT.
db db dd dw
066h 0eah start64_linear code64_sel ; End of the 16-bit code segment
code16ends
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; ; ;; ;;; Start of 64-bit code ;; ; ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; code64 para use64 ; At this point, we're using 64-bit code
start64:
; ; Point the 64-bit RSP register to the stack's _linear_ ; address. There is no need to set SS here, because the SS ; register is not used in 64-bit mode. ; mov rsp, stack0_linear
; ; ; ; ; ; ; ; ; ;
This LGDT is only needed if the long-mode GDT is to be located at a linear address above 4 Gbytes. If the long mode GDT is located at a 32-bit linear address, putting 64-bit descriptors in the GDT pointed to by [pGDT32] is just fine. pGDT64_linear is the _linear_ address of the 10-byte GDT pseudo-descriptor. The new GDT should have a valid CPL0 64-bit code segment descriptor at the entry-point corresponding to the current
432
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
; CS selector. Alternatively, a far transfer to a valid CPL0 ; 64-bit code segment descriptor in the new GDT must be done ; before enabling interrupts. ; lgdt [pGDT64_linear]
; ; ; ; ; ;
Load the 64-bit IDT. This is _required_, because the 64-bit IDT uses 64-bit interrupt descriptors, while the 32-bit IDT used 32-bit interrupt descriptors. pIDT64_linear is the _linear_ address of the 10-byte IDT pseudo-descriptor.
lidt
[pIDT64_linear]
; ; Set the current TSS. tss_sel should point to a 64-bit TSS ; descriptor in the current GDT. The TSS is used for ; inner-level stack pointers and the IO bit-map. ; mov ltr ax, tss_sel ax
; ; Set the current LDT. ldt_sel should point to a 64-bit LDT ; descriptor in the current GDT. ; mov lldt ax, ldt_sel ax
; ; ; ; ; ; ; ; ;
Using fs: and gs: prefixes on memory accesses still uses the 32-bit fs.base and gs.base. Reload these 2 registers before using the fs: and gs: prefixes. FS and GS can be loaded from the GDT using a normal "mov fs,foo" type instructions, which loads a 32-bit base into FS or GS. Alternatively, use WRMSR to assign 64-bit base values to MSR_FS_base or MSR_GS_base.
mov ecx, MSR_FS_base mov eax, FsbaseLow mov edx, FsbaseHi wrmsr
Chapter 14: Processor Initialization and Long-Mode Activation
433
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
; ; ; ; ; ; ; ; ;
Reload CR3 if long-mode page tables are to be located above 4 Gbytes. Because the original CR3 load was done in 32-bit legacy mode, it could only load 32 bits into CR3. Thus, the current page tables are located in the lower 4 Gbytes of physical memory. This MOV to CR3 is only needed if the actual long-mode page tables should be located at a linear address above 4 Gbytes.
mov mov
rax, final_pml4_base cr3, rax
; Point to PML4 ; Load 64-bit CR3
; ; Enable interrupts. ; sti ; Enabled INTR
434
Chapter 14: Processor Initialization and Long-Mode Activation
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Appendix A
MSR Cross-Reference
This appendix lists the MSRs that are defined in the AMD64 architecture. The AMD64 architecture supports some of the same MSRs as previous versions of the x86 architecture and implementations thereof. Where possible, the AMD64 architecture supports the same MSRs, for the same functions, as these previous architectures and implementations. The first section lists the MSRs according to their MSR address, and it gives a cross reference for additional information. The remaining sections list the MSRs by their functional group. Those sections also give a brief description of the register and specify the register reset value. Some MSRs are implementation-specific For information about these MSRs, see the documentation for specific implementations of the AMD64 architecture.
A.1
MSR Cross-Reference by MSR Address
Table A-1 lists the MSRs in the AMD64 architecture in order of MSR address.
Table A-1. MSRs of the AMD64 Architecture
MSR Address 0010h 00FEh 0174h 0175h 0176h 0179h 017Ah 017Bh 01D9h TSC MTRRcap SYSENTER_CS SYSENTER_ESP SYSENTER_EIP MCG_CAP MCG_STATUS MCG_CTL DebugCtlMSR Software Debug Machine Check "Machine-Check Global-Capabilities Register" on page 308 "Machine-Check Global-Status Register" on page 308 "Machine-Check Global-Control Register" on page 309 "Debug-Control MSR (DebugCtlMSR)" on page 391 System Software "SYSENTER and SYSEXIT MSRs" on page 184 MSR Name Functional Group Performance Memory Typing Cross-Reference "Time-Stamp Counter" on page 408 "Identifying MTRR Features" on page 226
Appendix A: MSR Cross-Reference
435
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table A-1. MSRs of the AMD64 Architecture (continued)
MSR Address 01DBh 01DCh 01DDh 01DEh 0200h 0201h 0202h 0203h 0204h 0205h 0206h 0207h 0208h 0209h 020Ah 020Bh 020Ch 020Dh 020Eh 020Fh MSR Name LastBranchFromIP LastBranchToIP LastExceptionFromIP LastExceptionToIP MTRRphysBase0 MTRRphysMask0 MTRRphysBase1 MTRRphysMask1 MTRRphysBase2 MTRRphysMask2 MTRRphysBase3 MTRRphysMask3 MTRRphysBase4 MTRRphysMask4 MTRRphysBase5 MTRRphysMask5 MTRRphysBase6 MTRRphysMask6 MTRRphysBase7 MTRRphysMask7 Memory Typing "Variable-Range MTRRs" on page 222 Software Debug "Control-Transfer Recording MSRs" on page 392 Functional Group Cross-Reference
436
Appendix A: MSR Cross-Reference
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table A-1. MSRs of the AMD64 Architecture (continued)
MSR Address 0250h 0258h 0259h 0268h 0269h 026Ah 026Bh 026Ch 026Dh 026Eh 026Fh 0277h 02FFh 0400h 0404h 0408h 040Ch 0410h 0401h 0405h 0409h 040Dh 0411h MSR Name MTRRfix64K_00000 MTRRfix16K_80000 MTRRfix16K_A0000 MTRRfix4K_C0000 MTRRfix4K_C8000 MTRRfix4K_D0000 MTRRfix4K_D8000 MTRRfix4K_E0000 MTRRfix4K_E8000 MTRRfix4K_F0000 MTRRfix4K_F8000 PAT MTRRdefType MC0_CTL MC1_CTL MC2_CTL MC3_CTL MC4_CTL MC0_STATUS MC1_STATUS MC2_STATUS MC3_STATUS MC4_STATUS Machine Check "Machine-Check Status Registers" on page 311 Machine Check See the documentation for particular implementations of the architecture. Memory Typing "PAT Register" on page 230 "Default-Range MTRRs" on page 225 Memory Typing "Fixed-Range MTRRs" on page 220 Functional Group Cross-Reference
Appendix A: MSR Cross-Reference
437
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table A-1. MSRs of the AMD64 Architecture (continued)
MSR Address 0402h 0406h 040Ah 040Eh 0412h 0403h 0407h 040Bh 040Fh 0413h C000_0080h C000_0081h C000_0082h C000_0083h C000_0084h C000_0100h C000_0101h C000_0102h C001_0000h C001_0001h C001_0002h C001_0003h C001_0004h C001_0005h C001_0006h C001_0007h C001_0010h MSR Name MC0_ADDR MC1_ADDR MC2_ADDR MC3_ADDR MC4_ADDR MC0_MISC MC1_MISC MC2_MISC MC3_MISC MC4_MISC EFER STAR LSTAR CSTAR SF_MASK FS.Base GS.Base KernelGSbase PerfEvtSel0 PerfEvtSel1 PerfEvtSel2 PerfEvtSel3 PerfCtr0 PerfCtr1 PerfCtr2 PerfCtr3 SYSCFG Memory Typing "System Configuration Register (SYSCFG)" on page 73 Performance "Performance Counters" on page 404 Performance "Performance Event-Select Registers" on page 405 System Software System Software "FS and GS Registers in 64-Bit Mode" on page 88 "SWAPGS Instruction" on page 185 System Software "SYSCALL and SYSRET MSRs" on page 183 System Software "Extended Feature Enable Register (EFER)" on page 69 Machine Check "Machine-Check Miscellaneous-Error Information Registers" on page 314 Machine Check "Machine-Check Address Registers" on page 314 Functional Group Cross-Reference
438
Appendix A: MSR Cross-Reference
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table A-1. MSRs of the AMD64 Architecture (continued)
MSR Address C001_0016h C001_0017h C001_0018h C001_0019h C001_001Ah C001_001Dh MSR Name IORRBase0 IORRMask0 IORRBase1 IORRMask1 TOP_MEM TOP_MEM2 Memory Typing "Top of Memory" on page 239 Memory Typing "IORRs" on page 237 Functional Group Cross-Reference
A.2
System-Software MSRs
Table A-2 lists the MSRs defined for general use by system software in controlling long mode and in allowing fast control transfers between applications and the operating system.
Table A-2.
MSR Address C000_0080h
System-Software MSR Cross-Reference
MSR Name EFER Description Contains control bits that enable extended features supported by the processor, including long mode. In legacy mode, used to specify the target address of a SYSCALL instruction, as well as the CS and SS selectors of the called and returned procedures. In 64-bit mode, used to specify the target RIP of a SYSCALL instruction. In compatibility mode, used to specify the target RIP of a SYSCALL instruction. SYSCALL Flags Mask Contains the 64-bit base address in the hidden portion of the FS register (the base address from the FS descriptor). Reset Value
0000_0000_0000_0000h
C000_0081h
STAR
undefined
C000_0082h C000_0083h C000_0084h C000_0100h
LSTAR CSTAR SF_MASK FS.Base
undefined undefined undefined undefined
Appendix A: MSR Cross-Reference
439
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table A-2.
MSR Address C000_0101h
System-Software MSR Cross-Reference (continued)
MSR Name GS.Base Description Contains the 64-bit base address in the hidden portion of the GS register (the base address from the GS descriptor). The SWAPGS instruction exchanges the value in KernelGSbase with the value in GS.base, providing a fast method for system software to load a pointer to system data-structures. In legacy mode, used to specify the CS selector of the procedure called by SYSENTER. In legacy mode, used to specify the stack pointer for the procedure called by SYSENTER. In legacy mode, used to specify the EIP of the procedure called by SYSENTER. Reset Value
undefined
C000_0102h
KernelGSbase
undefined
0174h 0175h 0176h
SYSENTER_CS SYSENTER_ESP SYSENTER_EIP
undefined undefined undefined
A.3
Memory-Typing MSRs
Table A-3 on page 441 lists the MSRs used to control memorytyping and the page-attribute-table mechanism.
440
Appendix A: MSR Cross-Reference
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Table A-3.
MSR Address 00FEh 0200h 0202h 0204h 0206h 0208h 020Ah 020Ch 020Eh 0201h 0203h 0205h 0207h 0209h 020Bh 020Dh 020Fh
Memory-Typing MSR Cross-Reference
MSR Name MTRRcap MTRRphysBase0 MTRRphysBase1 MTRRphysBase2 MTRRphysBase3 MTRRphysBase4 MTRRphysBase5 MTRRphysBase6 MTRRphysBase7 MTRRphysMask0 MTRRphysMask1 MTRRphysMask2 MTRRphysMask3 MTRRphysMask4 MTRRphysMask5 MTRRphysMask6 MTRRphysMask7 Specifies the size of a variable-range memory region. Valid (bit 11) = 0 All Other Bits Undefined Specifies the memory-range base address in physical-address space of a variable-range memory region. These registers also specify the memory type used for the memory region. undefined Description A read-only register containing information describing the level of MTRR support provided by the processor. Reset Value
0000_0000_0000_0508h
Appendix A: MSR Cross-Reference
441
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table A-3.
MSR Address 0250h 0258h 0259h 0268h 0269h 026Ah 026Bh 026Ch 026Dh 026Eh 026Fh 0277h
Memory-Typing MSR Cross-Reference (continued)
MSR Name MTRRfix64K_00000 MTRRfix16K_80000 MTRRfix16K_A0000 MTRRfix4K_C0000 MTRRfix4K_C8000 MTRRfix4K_D0000 MTRRfix4K_D8000 MTRRfix4K_E0000 MTRRfix4K_E8000 MTRRfix4K_F0000 MTRRfix4K_F8000 PAT Used to extend the page-table entry format, allowing memory-type characterization on a physical-page basis. Sets the default memory-type for physical addresses not within ranges established by fixed-range and variable-range MTRRs. Contains control bits for enabling and configuring system bus features. Specifies the memory-range base address in physical-address space of a variable-range I/O region. Specifies the size of a variable-range I/O region. Sets the boundary between system memory and memory-mapped I/O for addresses below 4 Gbytes. Sets the boundary between system memory and memory-mapped I/O for addresses above 4 Gbytes. 0007_0406_0007_0406h Fixed-range MTRRs used to characterize the first 1 Mbyte of physical memory. Each 64-bit register contains eight type fields for characterizing a total of eight memory ranges. * MTRRfix64K_n characterizes 64 Kbyte ranges. * MTRRfix16K_n characterizes 16 Kbyte ranges. * MTRRfix4K_n characterizes 4 Kbyte ranges. Description Reset Value
undefined
02FFh
MTRRdefType
0000_0000_0000_0000h
C001_0010h C001_0016h C001_0018h C001_0017h C001_0019h C001_001Ah
SYSCFG IORRBase0 IORRBase1 IORRMask0 IORRMask1 TOP_MEM
0000_0000_0002_0601h
undefined Valid (bit 11) = 0 All Other Bits Undefined 0000_0000_0400_0000h
C001_001Dh
TOP_MEM2
undefined
442
Appendix A: MSR Cross-Reference
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
A.4
Machine-Check MSRs
Table A-4 lists the MSRs used in support of the machine-check mechanism.
Table A-4.
MSR Address 0179h
Machine-Check MSR Cross-Reference
MSR Name MCG_CAP Description A read-only register that specifies the machinecheck mechanism capabilities supported by the processor. Provides basic information about the processor state immediately after the occurrence of a machine-check error. Controls global reporting of machine-check errors from various sources. Controls error reporting for the data-cache-unit register bank. Controls error reporting for the instructioncache-unit register bank. Controls error reporting for the bus-unit register bank. Controls error reporting for the load/store-unit register bank. Reset Value
0000_0000_0000_0104h
017Ah
MCG_STATUS
undefined
017Bh 0400h 0404h 0408h 040Ch 0401h 0405h 0409h 040Dh 0402h 0406h 040Ah 040Eh
MCG_CTL MC0_CTL MC1_CTL MC2_CTL MC3_CTL MC0_STATUS MC1_STATUS MC2_STATUS MC3_STATUS MC0_ADDR MC1_ADDR MC2_ADDR MC3_ADDR
0000_0000_0000_0000h 0000_0000_0000_0000h 0000_0000_0000_0000h 0000_0000_0000_0000h 0000_0000_0000_0000h
Status registers for each error-reporting register bank, used to report machine-check error information for the specified register bank.
undefined
Reports the instruction memory-address or data memory-address responsible for the machinecheck error for the specified register bank.
undefined
Appendix A: MSR Cross-Reference
443
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
Table A-4.
MSR Address 0403h 0407h 040Bh 040Fh
Machine-Check MSR Cross-Reference (continued)
MSR Name MC0_MISC MC1_MISC MC2_MISC MC3_MISC Reports miscellaneous information about the machine-check error for the specified register bank. undefined Description Reset Value
A.5
Software-Debug MSRs
Table A-5 lists the MSRs used in support of the software-debug architecture.
Table A-5.
MSR Address
Software-Debug MSR Cross-Reference
MSR Name Description Provides debug controls for control-transfer recording and control-transfer single stepping, and external-breakpoint reporting and trace messages. During control-transfer recording, this register is loaded with the segment offset of the controltransfer source. During control-transfer recording, this register is loaded with the segment offset of the controltransfer target. When an interrupt occurs during control-transfer recording, this register is loaded with LastBranchFromIP before LastBranchFromIP is updated. When an interrupt occurs during control-transfer recording, this register is loaded with LastBranchToIP before LastBranchToIP is updated. Reset Value
01D9h
DebugCtlMSR
0000_0000_0000_0000h
01DBh
LastBranchFromIP
undefined
01DCh
LastBranchToIP
undefined
01DDh
LastExceptionFromIP
undefined
01DEh
LastExceptionToIP
undefined
444
Appendix A: MSR Cross-Reference
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
A.6
Performance-Monitoring MSRs
Table A-6 lists the MSRs used in support of performance monitoring, including the time-stamp counter.
Table A-6.
MSR Address 0010h C001_0000h C001_0001h C001_0002h C001_0003h C001_0004h C001_0005h C001_0006h C001_0007h
Performance-Monitoring MSR Cross-Reference
MSR Name TSC PerfEvtSel0 PerfEvtSel1 PerfEvtSel2 PerfEvtSel3 PerfCtr0 PerfCtr1 PerfCtr2 PerfCtr3 Used to count specific processor events, or the duration of events, as specified by the corresponding PerfEvtSeln register. undefined For the corresponding performance counter, this register specifies the events counted, and controls other aspects of counter operation. undefined Description Counts processor-clock cycles. It is incremented once for each processor-clock cycle. Reset Value 0000_0000_0000_0000h
Appendix A: MSR Cross-Reference
445
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
446
Appendix A: MSR Cross-Reference
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
Index
Symbols #AC............................................................. 263 #BP ............................................................. 252 #BR............................................................. 253 #D ....................................................... 262, 265 #DB............................................................. 250 #DE ............................................................ 250 #DF............................................................. 254 #GP............................................................. 259 #I ........................................................ 262, 265 #IA.............................................................. 262 #IS .............................................................. 262 #MC ............................................................ 264 #MF ............................................................ 262 #NM............................................................ 254 #NP............................................................. 257 #O ....................................................... 262, 265 #OF............................................................. 252 #P ....................................................... 262, 266 #PF ............................................................. 261 #SS ............................................................. 258 #TS ............................................................. 256 #U ....................................................... 262, 266 #UD ............................................................ 253 #XF............................................................. 265 #Z ....................................................... 262, 265 Numerics 128-bit media instructions enabling .................................................. 340 feature identification ............................ 340 MXCSR ................................................... 341 saving state............................................. 346 XMM registers.................................. 36, 341 16-bit mode............................................... xxvi 32-bit mode............................................... xxvi 64-bit media instructions causing #MF exception ......................... 344 feature identification ............................ 340 initializing .............................................. 419 MMX registers........................................ 343 saving state............................................. 346 64-bit mode......................................... xxvi, 15 A A bit ........................................... 101, 103, 170 abort........................................................... 245 AC bit ........................................................... 68 accessed (A) code segment ......................................... 101 data segment ......................................... 103 page-translation tables ......................... 170 address-breakpoint registers (DR0-DR3) 385 addressing RIP-relative.......................................... xxxii address-size prefix...................................... 39 ADDRV bit ................................................ 313 alignment check (rFLAGS.AC) ......... 68, 263 alignment mask (CR0.AM)................ 57, 263 alignment-check exception (#AC) 57, 68, 263 AM bit ......................................................... 57 architecture differences ............................ 29 ARPL instruction ..................................... 190 attributes .................................................... 96 available to software (AVL) descriptor................................................. 99 page-translation tables ......................... 171 AVL bit ................................................ 99, 171 B B3-B0 bits ................................................. 387 base address............ 92, 95, 99, 152, 161, 168 BD bit ........................................................ 387 benign exception...................................... 254 biased exponent ...................................... xxvi BIST ........................................................... 411 bootstrap processor.................................. 415 BOUND instruction.................................. 253 bound-range exception (#BR) ................. 253 branches ...................................................... 40 breakpoint ................................................ 394 determining cause................................. 396 on address match........................... 384, 398 on any instruction ................................. 384 on I/O ...................................................... 399 on instruction ........................................ 397 on task switch ................................ 384, 400 setting address....................................... 394 specifying address-match length ......... 394 breakpoint exception (#BP) .................... 252 BS bit ......................................................... 387 BT bit......................................................... 387 BTF bit ...................................................... 392 built-in self test (BIST) ............................ 411 C C bit ........................................................... 101 cache
Index
447
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
control mechanisms ............................... 212 control precedence ................................ 214 enabling .................................................. 416 index ....................................................... 211 invalidate................................................ 216 line .......................................................... 195 line writeback and invalidate............... 215 offset ....................................................... 211 organization............................................ 209 self-modifying code ............................... 211 set ............................................................ 210 tag............................................................ 211 way .......................................................... 210 writeback and invalidate ...................... 216 cache disable (CD) bit........................ 57, 212 cache disable (CD), memory type ........... 203 cache-coherency protocol......................... 199 losing coherency .................................... 201 CALL See call gate and control transfer. call gate ............................................. 106, 129 count field .............................................. 108 count field, long mode........................... 116 descriptor, long mode .............................. 41 jump through ......................................... 132 parameters.............................................. 135 privilege checks ..................................... 131 stack switch ............................................ 134 stack switch, long mode .................. 42, 136 canonical address form ........................ 5, 161 CD bit................................................... 57, 212 CD memory type ....................................... 203 CLFLUSH instruction .............................. 215 CLI instruction.......................................... 187 clock multiplier......................................... 412 CLTS instruction....................................... 187 code segment ................................ 33, 87, 100 64-bit mode............................................... 87 accessed (A) ........................................... 101 conforming (C) ....................................... 101 default-operand size (D) ....................... 102 ignored fields in 64-bit mode................ 109 long bit (L)........................................ 33, 109 long mode ............................................... 108 readable (R) ........................................... 101 type field ................................................ 101 coherency, cache ....................................... 196 commit ..................................................... xxvii commit, instruction results...................... 196 compatibility mode............................ xxvi, 15 conforming (C), code segment................. 101
consistency checks, long mode ............... 426 contributory exception ............................ 255 control registers ................................... 36, 53 control transfer......................................... 124 See also call gate and interrupt. call gate.................................................. 129 direct ...................................................... 125 far, conforming code segment .............. 128 far, nonconforming code segment........ 126 interrupt to higher privilege................ 278 interrupt to same privilege .................. 277 parameters ............................................. 135 stack switch............................................ 134 control-transfer recording MSRs ............ 392 coprocessor-segment-overrun exception 256 count field......................................... 116, 308 counter mask field ................................... 407 CPL ............................................................ 120 definition ............................................... 120 in call gate protection........................... 131 in data segment protection .................. 121 in interrupt to higher privilege ........... 280 in page protection ................................. 175 in protecting conforming CS ................ 128 in protecting nonconforming CS.......... 126 in stack segment protection ................. 123 privileged instructions.......................... 179 SYSCALL, SYSRET assumptions ........ 183 CPUID instruction ....................... 69, 78, 186 CR0 .............................................................. 55 alignment mask (AM) ..................... 57, 263 cache disable (CD) .......................... 57, 212 emulate coprocessor (EM).............. 56, 341 extension type (ET) ................................ 56 monitor coprocessor (MP) ...................... 55 not write-through (NW) .................. 57, 213 numeric error (NE) ......................... 57, 263 paging enable (PG) ......................... 58, 149 protection enable (PE) ............... 55, 80, 89 task switched (TS)........................... 56, 187 write protect (WP) .......................... 57, 176 CR1 .............................................................. 64 CR2 ...................................................... 58, 261 CR3 .......................... 32, 58, 59, 151, 161, 367 non-PAE paging ..................................... 151 PAE paging ...................................... 59, 151 PAE paging, long mode......................... 161 page-level cache disable (PCD) ... 152, 162 page-level write-through (PWT) .. 152, 161 table-base address ......................... 152, 161 CR4 .............................................................. 60
448
Index
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
debugging extensions (DE)..................... 61 machine-check enable (MCE)......... 62, 264 OS #XF support (OSXMMEXCPT) ..... 265, 341 OS FXSAVE/FXRSTOR support (OSFXSR) 340 page-global enable (PGE) ............... 63, 173 page-size extensions (PSE) ..... 62, 149, 154 performance counter enable (PCE) ...... 63, 188, 404 physical-address extensions (PAE) 62, 149, 160 protected-mode virtual interrupts (PVI)..... 61 time-stamp disable (TSD) ....... 61, 188, 409 virtual-8086 mode extensions (VME) ... 61, 296 CR5-CR7 ..................................................... 64 CR8 ...................................................... 64, 273 CR9-CR15 ................................................... 64 CS register................................................... 87 CSTAR register ......................... 183, 438, 439 D D bit ........................................... 102, 110, 170 D/B bit .................................................. 99, 104 data prefetch, cache ................................. 215 data segment ................................. 34, 87, 102 64-bit mode............................................... 88 accessed (A) ........................................... 103 default operand size (D) ....................... 104 expand down (E) .................................... 103 FS and GS ........................................... 34, 88 ignored fields in 64-bit mode................ 110 long mode ............................................... 110 privilege checks ..................................... 121 type field ................................................ 103 writable (W) ........................................... 103 DAZ bit ...................................................... 356 DE bit........................................................... 61 DE exception..................................... 262, 265 debug ........................................................... 26 See breakpoint and single-step. debug exception (#DB)............... 67, 250, 396 debug registers.................................... 36, 385 address-breakpoint registers (DR0-DR3) .... 385 control-transfer recording MSRs.......... 392 debug-control MSR (DebugCtlMSR) ... 391 debug-control register (DR7)................ 388 debug-status register (DR6).................. 386 reserved (DR4, DR5) ............................. 386
debug-control MSR (DebugCtlMSR)...... 391 debug-control register (DR7) .................. 388 DebugCtlMSR register ............ 391, 435, 444 debugging extensions (CR4.DE)............... 61 debug-status register (DR6) .................... 386 DEC instruction.......................................... 43 default operand size B bit, stack segment .............................. 104 D bit, code segment .............................. 102 D bit, data segment....................... 104, 140 D/B bit, descriptor................................... 99 with expand down ................................. 141 denormalized-operand exception (DE) . 262, 265 denormals-are-zeros (DAZ) mode ........... 355 descriptor.............................................. 82, 97 available to software (AVL).................... 99 code segment ........................................... 33 data segment ........................................... 34 default operand size (D/B)...................... 99 DPL........................................... 99, 120, 373 gate ........................................................... 34 granularity (G)....................................... 100 long mode............................................... 108 present (P) ....................................... 99, 373 S field ............................................... 99, 373 segment base ........................................... 99 segment limit........................................... 98 system segment ....................................... 34 TSS.......................................................... 362 type field.......................................... 99, 373 descriptor table .................................... 82, 89 global-descriptor table (GDT)................ 84 interrupt-descriptor table (IDT) ............ 46 local-descriptor table (LDT)................... 84 descriptor-table registers .................... 33, 84 64-bit mode ............................................ 116 GDTR........................................................ 91 IDTR ......................................................... 97 LDTR ........................................................ 93 device-not-available exception (#NM) .... 55, 56, 254 differences (architectural) ........................ 29 direct referencing................................... xxvii dirty (D), page-translation tables ........... 170 displacement .............................................. 39 displacements......................................... xxvii divide-by-zero-error exception (#DE)..... 250 double quadword.................................... xxvii double-fault exception (#DF).................. 254 doubleword ............................................. xxvii
Index
449
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
DP field...................................................... 355 DPL ............................................................ 120 data segment, 64-bit mode.................... 111 definition ................................................ 120 in call gate protection ........................... 131 in data segment protection................... 121 in interrupt stack switch ....................... 279 in interrupt to higher privilege ............ 280 in protecting conforming CS................. 128 in protecting nonconforming CS .......... 126 in stack segment protection.................. 123 in stack switching .................................. 135 DPL field ............................................. 99, 373 DR0-DR3 registers .................................... 385 DR4, DR5 registers ................................... 386 DR6 register .............................................. 386 DR7 register .............................................. 388 DS field ...................................................... 355 DS register............................................. 87, 88 E E bit ................................................... 103, 407 eAX-eSP register.................................. xxxiii EFER ..................................................... 37, 69 fast FXSAVE/FXRSTOR (FFXSR) ......... 71 long mode active (LMA).................. 70, 424 long mode enable (LME) ................ 70, 424 no-execute enable (NXE)........................ 71 system-call extension (SCE) ................... 70 EFER register ................................... 438, 439 effective address..................................... 2, 31 effective address size ........................... xxviii effective operand size .......................... xxviii EFLAGS See rFLAGS. eFLAGS register .................................... xxxiv EIP See rIP. eIP register............................................. xxxiv EIPV bit ..................................................... 309 element .................................................. xxviii EM bit .................................................. 56, 418 emulate coprocessor (CR0.EM)......... 56, 341 EN bit................................................. 313, 407 endian byte-ordering ............................. xxxvi environment .............................................. 346 error code page fault................................................ 268 selector ................................................... 267 ES register................................................... 88 ESP See rSP.
ET bit........................................................... 56 event handler, definition ......................... 243 event mask field ....................................... 406 exception handler, definition.................. 243 exceptions .............................................. xxviii abort ....................................................... 245 benign..................................................... 254 contributory ........................................... 255 definition of ........................................... 243 definition of vector ............................... 247 differences in long mode ........................ 46 error code, page fault ........................... 268 error code, selector ............................... 267 fault ........................................................ 245 floating-point priorities ........................ 271 imprecise................................................ 244 maskable 128-bit media floating point 246 maskable x87 floating point................. 246 masking during stack switches ............ 246 precise .................................................... 244 priorities ................................................ 269 trap ......................................................... 245 while in SMM......................................... 333 exclusive state, MOESI............................ 199 expand down (E) data segment ......................................... 103 stack segment ................................ 103, 141 exponent .................................................. xxvi extended family field............................... 416 extended model field............................... 416 extension type (CR0.ET) ........................... 56 F family field................................................ 416 far control transfer................................... 124 far return............................................. 42, 138 fast FXSAVE/FXRSTOR............................ 71 fault ........................................................... 245 FCW register ............................ 344, 346, 354 feature identification................................. 78 fence instructions..................................... 198 FFXSR bit ................................................... 71 fill, cache-line ........................................... 195 first instruction ........................................ 415 flat segmentation ............................. 7, 10, 82 FLDENV, FSTENV instructions ............. 351 floating-point exception pending (#MF) 262 caused by 64-bit media instructions.... 344 floating-point exception priorities ......... 271 flush........................................................ xxviii FOP register ............................................. 355 FPR registers .................................... 344, 346
450
Index
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
FS and GS .............................................. 34, 88 FS register ................................................... 88 FS.Base register ................................ 438, 439 FSAVE, FRSTOR instructions ................. 346 FSW register ..................... 343, 344, 346, 354 FTW register ..................... 343, 344, 346, 354 FXSAVE, FXRSTOR instructions 45, 63, 351 32-bit memory image............................. 354 64-bit memory image............................. 353 x87 tag word format............................... 356 G G bit ................................................... 100, 171 G3-G0 bits ................................................. 389 gate descriptors .......................................... 34 call gate .................................................. 106 DPL ......................................................... 120 ignored fields in long mode .................. 114 illegal types in long mode..................... 114 interrupt gate......................................... 106 long mode ....................................... 113, 116 redefined types in long mode ............... 114 target-segment offset ............................ 107 target-segment selector......................... 108 task gate ................................................. 106 trap gate ................................................. 106 GD bit ........................................................ 389 GDT .............................................................. 90 GDTR ........................................................... 91 GE bit......................................................... 389 general detect fault .......................... 250, 401 general-protection exception (#GP)........ 259 general-purpose registers (GPRs) ............. 35 global descriptor table (GDT) ............. 84, 90 base address, 64-bit mode ....................... 92 first entry.................................................. 90 limit check, long mode ............................ 92 global descriptor-table register (GDTR) .. 91 base address ............................................. 92 limit........................................................... 92 loading .................................................... 189 storing ..................................................... 189 global page (G), page-translation tables 171 global pages ........................................ 63, 173 granularity (G), descriptor .............. 100, 140 GS register................................................... 88 GS.Base register................................ 438, 440 H halt ............................................................. 190 HLT instruction......................................... 190
I I/O privilege level field (rFLAGS.IOPL) .. 66 I/O, memory-mapped................................ 234 I/O-permission bitmap in 32-bit TSS .......................................... 368 in 64-bit TSS .......................................... 371 I/O-permission bitmap (IOPB) ................ 369 ID bit ........................................................... 69 IDT............................................................... 96 IDTR ............................................................ 97 IE exception...................................... 262, 265 IF bit.................................................... 66, 301 IGN ........................................................... xxix immediate operand .................................... 39 imprecise exceptions and interrupts...... 244 INC instruction........................................... 43 indirect ..................................................... xxix inexact-result exception .................. 262, 266 INIT ........................................................... 411 initialization ............................................. 411 initialization (INIT).................................. 411 processor state....................................... 412 instructions (system-management) ........ 179 INT bit ....................................................... 407 INT3 instruction ............................... 252, 401 integer bit ................................................. 357 interrupt descriptor table (IDT) ............... 96 limit check, long mode............................ 97 interrupt descriptor-table register (IDTR) .... 97 loading.................................................... 189 storing .................................................... 189 interrupt flag (rFLAGS.IF) ............... 66, 187 interrupt gate ................................... 106, 287 IST field ................................................. 116 interrupt handler, definition................... 243 interrupt redirection........................ 284, 299 interrupt-descriptor table (IDT) index ....................................... 243, 276, 288 protected mode...................................... 276 real-address mode ................................. 274 interrupt-redirection bitmap .................. 368 interrupts definition of external............................ 243 definition of software ........................... 243 definition of vector ............................... 247 differences in long mode ........................ 46 external .................................................. 266 external maskable................................. 245 external nonmaskable .......................... 246 external-interrupt priorities ................ 272
Index
451
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
imprecise ................................................ 244 long mode summary .............................. 287 precise..................................................... 244 priorities ................................................. 269 returning from 64-bit mode .................. 294 returns .................................................... 283 software .................................................. 266 stack alignment, long mode.................. 291 stack pointer push, long mode ............. 289 stack switch, long mode .................. 47, 292 to higher privilege ................................. 278 to same privilege ................................... 277 while in SMM ......................................... 333 interrupt-stack table (IST)......... 47, 116, 292 in 64-bit TSS ........................................... 371 interrupt-vector table............................... 274 INTn instruction ............................... 266, 401 INTO instruction....................................... 252 INV bit ....................................................... 407 invalid arithmetic-operand exception .... 262 invalid state, MOESI ................................ 199 invalid-opcode exception (#UD)........ 44, 253 invalid-operation exception (IE) ..... 262, 265 invalid-TSS exception (#TS) .................... 256 INVD instruction .............................. 191, 216 INVLPG instruction.......................... 173, 191 IOPB ................................................... 368, 369 IOPL field ............................................ 66, 284 IOPL-sensitive instruction ....................... 296 IORRBasen registers ................ 238, 439, 442 IORRMaskn registers ....................... 238, 442 IORRs, variable-range.............................. 237 IRET instruction less privilege .......................................... 283 long mode ......................................... 47, 294 same privilege........................................ 283 IST field ..................................................... 116 J J bit ............................................................ 357 jump See call gate and control transfer. K KernelGSbase register ............. 185, 438, 440 L L bit............................................................ 109 L1 data cache ............................................ 195 L1 instruction cache ................................. 195 L2 cache..................................................... 195 L3-L0 bits .................................................. 389 LAR instruction ........................................ 189
LastBranchFromIP register ..... 393, 436, 444 LastBranchToIP register.......... 393, 436, 444 LastExceptionFromIP register................ 394 LastExceptionToIP register .................... 393 LastIntFromIP register .................... 436, 444 LastIntToIP register......................... 436, 444 LBR bit ...................................................... 391 LDT.............................................................. 92 LDTR ........................................................... 93 LDTselector field ..................................... 368 LE bit......................................................... 389 legacy mode ................................. xxix, 16, 30 legacy x86 ................................................ xxix LEN3-LEN0 bits....................................... 390 LFENCE instruction ................................ 215 LGDT instruction ..................................... 189 LIDT instruction....................................... 189 limit ......................................... 92, 95, 98, 363 linear address ............................................... 3 Link field .................................................. 367 LLDT instruction...................................... 189 LMA bit ....................................................... 70 LME bit ....................................................... 70 LMSW instruction .................................... 186 load ordering ............................................ 215 local descriptor table (LDT) ............... 84, 92 base address, 64-bit mode ...................... 95 limit check, long mode............................ 95 local descriptor-table register (LDTR)..... 93 attributes ................................................. 96 base address............................................. 95 hidden portion ......................................... 93 LDT selector ............................................ 95 limit .......................................................... 95 loading.................................................... 189 storing .................................................... 189 locality....................................................... 172 logical address .............................................. 2 long attribute (L) code segment ......................................... 109 effect on D bit........................................ 110 long mode..................................... xxix, 14, 29 activating ............................................... 425 consistency checks ................................ 426 differences from legacy mode................ 50 enabling ................................................. 424 enabling versus activating.................... 424 GDT requirements ................................ 423 IDT requirements .................................. 422 leaving .................................................... 427 page translation-table requirements ... 423
452
Index
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
relocating descriptor tables.................. 426 relocating page tables ........................... 427 TSS requirements .................................. 423 use of CS.L and CS.D............................. 424 long mode active (EFER.LMA) ......... 70, 424 long mode enable (EFER.LME) ........ 70, 424 LSB ............................................................ xxix lsb .............................................................. xxix LSTAR register ......................... 183, 438, 439 LTR instruction......................................... 189 M M bit........................................................... 357 machine check error codes.............................................. 312 error sources........................................... 306 error-reporting address register (MCi_ADDR) .................................... 314 error-reporting control register (MCi_CTL) 310 error-reporting miscellaneous register (MCi_MISC)...................................... 314 error-reporting register banks .............. 310 error-reporting status register (MCi_STATUS)................................. 311 global-capabilities register (MCG_CAP) ..... 308 global-control register (MCG_CTL) ..... 309 global-status register (MCG_STATUS) 308 initialization ........................................... 315 machine check registers........................... 307 machine-check enable (CR4.MCE) ... 62, 264 machine-check exception (#MC) ............. 264 fatal ......................................................... 306 recoverable ............................................. 306 mask ........................................................... xxx masking definition of interrupt........................... 243 MBZ............................................................ xxx MCA error code field ............................... 312 MCE bit ....................................................... 62 MCG_CAP register ................... 308, 435, 443 MCG_CTL register.................... 309, 435, 443 MCG_CTL_P bit ........................................ 308 MCG_STATUS register............. 308, 435, 443 MCi_ADDR registers ................ 314, 438, 443 MCi_CTL registers............ 310, 311, 437, 443 MCi_MISC registers.................. 314, 438, 444 MCi_STATUS registers............. 311, 437, 443 MCIP bit .................................................... 309 memory ...................................................... 193 memory addressing
canonical address form ............................. 5 effective address ....................................... 2 linear address ............................................ 3 logical address ........................................... 2 near pointers.............................................. 3 physical address ........................................ 3 real address.............................................. 11 RIP-relative address................................ 39 segment offset ........................................... 2 virtual address........................................... 3 memory management .................................. 6 memory serialization ............................... 215 memory system......................................... 193 memory type ............................................. 202 determining effective ........................... 227 memory-access ordering description ............................................. 196 read ordering ......................................... 197 write ordering........................................ 197 memory-mapped I/O directing reads and writes to ....... 235, 239 memory-type range register (MTRR)....... 37 combined with PAT ............................... 233 effect of paging cache controls ............ 227 effects with large page sizes ................ 228 fixed range ............................................. 220 identifying features .............................. 226 initial value............................................ 417 IORRBase............................................... 238 IORRMask ............................................. 238 MTRRcap ............................................... 226 MTRRdefType ....................................... 226 MTRRfix16K.......................................... 222 MTRRfix4K............................................ 222 MTRRfix64K.......................................... 222 MTRRphysBase ..................................... 223 MTRRphysMask .................................... 223 overlapping ranges................................ 229 type field, default ................................. 218 type field, extended .............................. 235 variable range........................................ 222 variable range size and alignment ...... 224 MFENCE instruction ............................... 215 MISCV bit ................................................. 313 MMX registers .......................................... 343 model field................................................ 416 model-specific error code field ............... 313 model-specific registers (MSRs) . 37, 71, 188 control-transfer recording .................... 392 debug extensions..................................... 76 debug-control MSR (DebugCtlMSR) ... 391
Index
453
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
FS.base...................................................... 89 GS.base ..................................................... 89 initializing .............................................. 419 machine check.................................. 77, 307 memory typing ................................. 75, 219 PAT .......................................................... 230 performance monitoring ................. 76, 404 SYSCFG .................................................... 73 system linkage ................................. 74, 183 time-stamp counter.......................... 76, 408 TOP_MEM .............................................. 239 TOP_MEM2 ............................................ 239 modes ........................................................... 12 16-bit ...................................................... xxvi 32-bit ...................................................... xxvi 64-bit ................................................ xxvi, 15 compatibility ................................... xxvi, 15 legacy ............................................... xxix, 16 long................................................... xxix, 14 protected ......................................... xxxi, 16 real ............................................... xxxi, 4, 16 virtual-8086 .................................. xxxiii, 17 modified state, MOESI............................. 199 MOESI ....................................................... 199 moffset ....................................................... xxx monitor coprocessor (CR0.MP) ................. 55 MOV CRn instruction............................... 186 MOV DRn instruction............................... 187 MOVSXD instruction.................................. 43 MP bit .................................................. 55, 418 MSB ............................................................ xxx msb ............................................................. xxx MSR ........................................................ xxxiv MSRs ............................................................ 71 MTRRcap register .................... 226, 435, 441 MTRRdefType register ............ 226, 437, 442 MTRRfix16K_n registers ......................... 222 MTRRfix4K_n registers ........................... 222 MTRRfix64K_n registers ......... 222, 437, 442 MtrrFixDramEn bit ............................ 74, 236 MtrrFixDramModEn bit..................... 74, 236 MTRRphysBasen registers....... 223, 436, 441 MTRRphysMaskn registers ............. 223, 441 MTRRs ....................................................... 219 MtrrTom2En bit .................................. 74, 241 MtrrVarDramEn bit............................ 74, 241 MXCSR field ............................................. 355 MXCSR register........................................ 342 MXCSR_MASK field ................................ 355 N NE bit................................................... 57, 418
near branch operand size, 64-bit mode....................... 40 near control transfer................................ 124 near pointers ................................................ 3 near return................................................ 137 nested task (rFLAGS.NT).................. 67, 379 NMI............................................................ 252 no-execute (NX) page protection...................................... 175 page-translation tables, bit in .............. 171 nonmaskable interrupt exception (NMI) 252 while in SMM......................................... 332 non-PAE paging ........................................ 150 CR3 format............................................. 151 NOP instruction.......................................... 43 not write-through (CR0.NW)............. 57, 213 NT bit .......................................................... 67 null selector ................................................ 86 64-bit mode far return .......................... 139 interrupt return from 64-bit mode ...... 295 long mode interrupts .................... 292, 293 long mode stack switch ......................... 136 numeric error (CR0.NE) .................... 57, 263 NW bit ......................................................... 57 NX bit ................................................ 171, 175 NXE bit ....................................................... 71 O octword...................................................... xxx OE exception .................................... 262, 265 offset.................................................. xxx, 107 operand-size prefix..................................... 38 operating modes ......................................... 12 OS bit......................................................... 406 OS FXSAVE/FXRSTOR support (CR4.OSFXSR) ................................ 340 OS unmasked exception support (CR4.OSXMMEXCPT) ............ 265, 341 OSFXSR bit ................................................ 63 OSXMMEXCPT bit .................................... 63 OVER bit................................................... 314 overflow.................................................... xxxi overflow exception (#OF) ........................ 252 overflow exception (OE).................. 262, 265 owned state, MOESI ................................ 199 P P bit ............................................. 99, 169, 373 packed ...................................................... xxxi PAE bit ................................................ 62, 149 PAE paging ......................................... 31, 150 CR3 format....................................... 59, 151 CR3 format, long mode......................... 161
454
Index
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
legacy mode............................................ 156 long mode ............................................... 162 page directory ........................................... 151 page size (PS) ......................... 150, 154, 156 page directory pointer...................... 151, 156 page size (PS), page-translation tables... 170 page table .................................................. 151 page translation ........................................ 145 page-attribute table (PAT)....................... 230 combined with MTRR ........................... 233 effect on memory access ....................... 232 identifying support................................ 232 indexing .................................................. 231 page-translation tables, bit in............... 171 page-fault exception (#PF) ..... 169, 175, 176, 261 page-fault virtual address ........................ 261 page-global enable (CR4.PGE).......... 63, 173 page-level cache disable (PCD) ............... 213 CR3, bit in .............................................. 152 page-translation tables, bit in............... 170 page-level write-through (PWT).............. 214 CR3, bit in .............................................. 152 page-translation tables, bit in............... 170 page-map level-4 ....................................... 161 page-size extensions (CR4.PSE). 32, 62, 149, 154 40-bit physical address support.... 150, 155 unsupported in long mode .................... 149 page-translation cache ............................. 172 page-translation tables............................... 32 accessed (A) ........................................... 170 available to software (AVL) .................. 171 dirty (D) .................................................. 170 global page (G)....................................... 171 hierarchy................................................. 147 no-execute .............................................. 171 page directory entry (PDE) .................. 151 page size (PS) ......................................... 170 page table entry (PTE).......................... 151 page-attribute table (PAT).................... 171 page-directory pointer entry (PDPE) ... 32, 151, 156 page-level cache disable (PCD) ............ 170 page-level write-through (PWT)........... 170 page-map level-4 entry (PML4E) ... 32, 161 physical-page base address................... 168 present (P).............................................. 169 read/write (R/W) .................................... 169 translation-table base address.............. 168 user/supervisor (U/S)............................. 169
paging.............................................. 8, 31, 145 See also PAE paging and non-PAE paging. effect of segment protection ................ 178 protection across translation hierarchy ...... 176 protection checks .................................. 174 supported translations.......................... 147 paging enable (CR0.PG).................... 58, 149 activating long mode..................... 149, 426 parameter count field .............................. 108 PAT See page-attribute table (PAT). PAT bit....................................................... 171 PAT register .............................. 230, 437, 442 PBi bits ...................................................... 392 PC bit......................................................... 407 PCC bit ...................................................... 313 PCD bit...................................... 152, 162, 170 PCE bit ........................................................ 63 PDE............................................................ 151 PDPE ......................................................... 151 PE bit........................................................... 55 PE exception..................................... 262, 266 PerfCtrn registers..................... 404, 438, 445 PerfEvtSeln registers............... 405, 438, 445 performance counter ............................... 188 performance counter enable (CR4.PCE) 63, 188, 404 performance optimization ................. 26, 403 performance-monitoring counter overflow.................................................. 408 PerfCtrn ................................................. 404 PerfEvtSeln............................................ 405 starting and stopping............................ 408 PG bit .................................................. 58, 149 PGE bit........................................................ 63 physical address ..................................... 3, 31 as index into cache................................ 211 physical memory .......................................... 4 physical-address extensions (CR4.PAE) . 31, 32, 62, 149, 160 activating long mode..................... 149, 425 See also PAE paging. POP instruction ........................................ 187 precise exceptions and interrupts .......... 244 precision exception (PE) ................. 262, 266 PREFETCH instruction........................... 215 present (P) descriptor......................................... 99, 373 page-translation tables ......................... 169 principle of locality.................................. 172
Index
455
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
priorities, interrupt .................................. 269 privilege level ........................................... 119 probe, cache ...................................... 196, 201 during cache disable.............................. 213 processor feature identification (rFLAGS.ID) ....................................... 69 processor halt............................................ 190 processor state .......................................... 412 protected mode .................................. xxxi, 16 initial operating environment .............. 421 protected-mode virtual interrupts (CR4.PVI) 61 protection checks adjusting RPL ........................................ 190 call gate .................................................. 131 checking access rights ........................... 189 data segment .......................................... 121 direct call, conforming .......................... 128 direct call, nonconforming.................... 126 enabling .................................................... 80 far return ................................................ 138 interrupt return ..................................... 283 interrupt to higher privilege ................ 280 limit check, 64-bit mode........................ 140 long mode changes .................................. 35 long mode interrupt .............................. 292 long mode interrupt return .................. 294 stack segment......................................... 123 type check............................................... 141 verifying read/write access ................... 190 protection enable (CR0.PE) .......... 55, 80, 89 PS bit ................................................. 150, 170 PSE bit ......................................................... 62 PSE paging .................................................. 32 PTE ............................................................ 151 PUSH instruction...................................... 187 PVI bit ......................................................... 61 PWT bit...................................... 152, 161, 170 Q quadword .................................................. xxxi R R bit ........................................................... 101 R/W bit............................................... 169, 176 R/W3-R/W0 bits ........................................ 390 r8-r15 ...................................................... xxxiv rAX-rSP................................................... xxxv RAZ ........................................................... xxxi RdMem, MTRR type field ................. 74, 235 RDMSR instruction ............................ 71, 188 RDP field ................................................... 355 RDPMC instruction .................... 63, 188, 404
RDTSC instruction..................... 61, 188, 409 read hit ...................................................... 195 read miss ................................................... 195 read ordering ............................................ 215 read/write (R/W) page protection...................................... 176 page-translation tables, bit in .............. 169 readable (R), code segment .................... 101 real address................................................. 11 real address mode. See real mode real mode ....................................... xxxi, 4, 16 initial operating environment.............. 420 registers See also entries for individual registers. 128-bit media registers (XMM) .............. 36 address-breakpoint registers (DR0-DR3).... 385 control registers................................. 36, 53 control-transfer recording MSRs ......... 392 CR0 ........................................................... 55 CR2 ......................................................... 261 CR3 ..................................... 32, 59, 151, 161 CR4 ........................................................... 60 CSTAR .................................................... 183 debug registers ................................ 36, 385 debug-control MSR (DebugCtlMSR) ... 391 debug-control register (DR7) ............... 388 debug-extension MSRs ........................... 76 debug-status register (DR6) ................. 386 descriptor-table registers.................. 33, 84 eAX-eSP ............................................. xxxiii EFER .................................................. 37, 69 eFLAGS ............................................... xxxiv eIP........................................................ xxxiv FPR................................................. 344, 346 FS and GS ................................................ 88 FS.base ..................................................... 89 GDTR........................................................ 91 GPRs......................................................... 35 GS.base..................................................... 89 IDTR ......................................................... 97 IORRBase............................................... 238 IORRMask ............................................. 238 last x87 data pointer ............. 344, 346, 355 last x87 instruction pointer .. 344, 346, 355 LDTR ........................................................ 93 LSTAR .................................................... 183 machine-check MSRs .............................. 77 MCG_CAP .............................................. 308 MCG_CTL .............................................. 309 MCG_STATUS ....................................... 308
456
Index
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
MCi_ADDR............................................. 314 MCi_CTL................................................. 310 MCi_MISC .............................................. 314 MCi_STATUS ......................................... 311 memory-type range register (MTRR) ... 37, 75, 219 MMX ....................................................... 343 model-specific registers (MSRs)............. 37 MTRR, fixed range ................................ 220 MTRR, variable range........................... 222 MTRRcap................................................ 226 MTRRdefType........................................ 226 MTRRfix16K .......................................... 222 MTRRfix4K ............................................ 222 MTRRfix64K .......................................... 222 MTRRphysBase...................................... 223 MTRRphysMask .................................... 223 MXCSR ................................................... 342 PAT .......................................................... 230 PerfCtrn .................................................. 404 PerfEvtSeln ............................................ 405 performance-monitoring MSRs .............. 76 r8-r15 ................................................... xxxiv rAX-rSP................................................ xxxv rFLAGS ..................................... xxxv, 36, 64 rIP.......................................................... xxxv rSP............................................................. 36 segment registers..................................... 86 STAR ....................................................... 183 SYSCFG .................................................... 73 SYSENTER_CS ...................................... 185 SYSENTER_EIP .................................... 185 SYSENTER_ESP.................................... 185 system-linkage MSRs .............................. 74 task-priority register (CR8) ...... 49, 64, 272 time-stamp counter.......................... 76, 408 TOP_MEM ........................................ 74, 239 TOP_MEM2 ...................................... 74, 239 x87 FCW ................................. 344, 346, 354 x87 floating-point processor state........ 344 x87 FSW.......................... 343, 344, 346, 354 x87 FTW.......................... 343, 344, 346, 354 x87 opcode.............................. 344, 347, 355 XMM registers........................................ 342 relative...................................................... xxxi replacement, cache-line ........................... 196 reset ........................................................... 411 processor state ....................................... 412 RESET# signal .......................................... 411 resume flag (rFLAGS.RF).......... 67, 251, 400 RET instruction ........................................ 137
from 64-bit mode ................................... 139 long mode......................................... 42, 138 popping null selector, 64-bit mode ...... 139 stack switch............................................ 138 retire, instruction ..................................... 196 revision history .......................................... xxi REX prefix .................................................. 37 RF bit .......................................................... 67 rFLAGS ................................................. 36, 64 alignment check (AC) ..................... 68, 263 I/O privilege level field (IOPL) .............. 66 interrupt flag (IF) ........................... 66, 187 nested task (NT).............................. 67, 379 processor feature identification (ID) .... 69 resume flag (RF) ..................... 67, 251, 400 trap flag (TF) ........................................... 65 virtual interrupt (VIF).................... 68, 298 virtual interrupt pending (VIP)..... 68, 298 virtual-8086 mode (VM) ......................... 67 rFLAGS register ..................................... xxxv rIP ................................................................ 36 rIP register.............................................. xxxv RIP-relative address................................... 39 RIP-relative addressing ......................... xxxii RIPV bit .................................................... 309 RPL...................................................... 85, 120 adjusting ................................................ 190 definition ............................................... 120 in call gate protection........................... 132 in data segment protection .................. 121 in far return ........................................... 138 in IRET instruction ............................... 283 in protecting conforming CS ................ 128 in protecting nonconforming CS.......... 126 in stack segment protection ................. 123 RPL field................................................... 362 RSM instruction ............................... 321, 337 rSP ............................................................... 36 call gate stack switch ............................ 134 implicit reference.................................... 39 S S bit ..................................................... 99, 373 SCE bit ........................................................ 70 segment base .............................................. 99 segment limit.............................................. 98 segment offset .............................................. 2 segment registers ................................. 84, 86 64-bit mode .............................................. 87 accessing ................................................ 188 hidden portion ......................................... 86 initializing unused registers................... 86
Index
457
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
segmentation........................................... 6, 33 64-bit mode............................................... 82 combining with paging............................ 10 flat segmentation........................... 7, 10, 82 multi-segmented model........................... 81 segment-not-present exception (#NP) .... 257 segment-override prefix............................. 38 selector ...................... 83, 84, 85, 95, 108, 362 selector index.............................................. 85 self-modifying code .................................. 211 serializing instructions............................. 216 set ............................................................. xxxii SF exception.............................................. 262 SFENCE instruction ................................. 215 SGDT instruction ...................................... 189 shared state, MOESI................................. 199 shut down .................................................. 255 SIDT instruction ....................................... 189 SIMD floating-point exception (#XF) ...... 63, 265, 341 single-step all instructions ............................... 384, 400 control-transfers............................. 384, 402 SLDT instruction ...................................... 189 SMBASE register ...................................... 323 SMI ............................................................. 321 SMM ........................................................... 321 SMM interrupts......................................... 332 SMM revision identifier ........................... 330 SMM state-save area................................. 324 SMRAM ..................................................... 322 SMRAM state-save area ........................... 324 SMSW instruction..................................... 186 speculative execution............................... 196 SS register ................................................... 88 SSE ........................................................... xxxii SSE-2 ........................................................ xxxii stack exception (#SS) ............................... 258 stack pointers in 32-bit TSS ........................................... 367 in 64-bit TSS ........................................... 371 stack segment...................................... 87, 102 64-bit mode............................................... 88 default operand size (D) ....................... 104 expand down (E) .................................... 103 privilege checks ..................................... 123 stack switch call gate .................................................. 134 call gate, long mode......................... 42, 136 far return ................................................ 138 interrupt ................................................. 279
interrupt return..................................... 283 interrupt, long mode............................... 47 stack-fault exception (SF) ....................... 262 STAR register ........................... 183, 438, 439 status word................................................ 186 stepping ID field ...................................... 416 STI instruction.......................................... 187 sticky bits ................................................ xxxii store ordering ........................................... 215 STR instruction ........................................ 189 supervisor page ........................................ 175 SWAPGS instruction ................................ 185 SYSCALL Flag Mask register ................. 183 SYSCALL, SYSRET instructions...... 70, 182 SYSCFG MtrrFixDramEn............................... 74, 236 MtrrFixDramModEn ....................... 74, 236 MtrrTom2En .......................................... 241 MtrrVarDramEn .............................. 74, 241 SYSCFG register ........................ 73, 438, 442 SYSENTER, SYSEXIT instructions ....... 184 illegal in long mode .............................. 184 SYSENTER_CS register .......... 185, 435, 440 SYSENTER_EIP register......... 185, 435, 440 SYSENTER_ESP register........ 185, 435, 440 system call and return ............................. 181 system data structures............................... 20 system management interrupt (SMI).... 321, 331 while in SMM......................................... 332 system management mode (SMM) ..... 17, 30 leaving .................................................... 337 long mode differences .......................... 322 operating environment ......................... 331 revision identifier ................................. 330 saving processor state........................... 334 SMBASE register................................... 323 SMRAM.................................................. 322 state-save area, AMD64 architecture .. 324 state-save area, legacy .......................... 328 system registers.......................................... 17 system segment ............................ 34, 97, 104 ignored fields in 64-bit mode ............... 112 illegal types in long mode .................... 111 long mode............................................... 111 type field................................................ 105 system-call extension (EFER.SCE) .......... 70 system-linkage MSRs......................... 74, 183 T T bit ........................................................... 368 table indicator, selector............................. 85
458
Index
24593--Rev. 3.09--September 2003
AMD 64-Bit Technology
task gate .................................................... 106 in task switching .................................... 377 long mode ............................................... 116 Task Register (TR) ..................................... 84 task register (TR) ..................................... 363 loading .................................................... 189 selector ................................................... 362 storing ..................................................... 189 task switch ......................................... 359, 375 disabled in long mode ............................. 49 lazy context switch........................... 56, 358 nesting tasks........................................... 379 preventing recursion ............................. 381 task switched (CR0.TS) ...................... 56, 187 task, execution space ............................... 359 task-priority register (CR8) ......... 49, 64, 272 task-state segment (TSS) descriptor ............................................... 362 dynamic fields ........................................ 367 I/O-permission bitmap................... 368, 371 interrupt-redirection bitmap ................ 368 interrupt-stack table.............................. 371 legacy 32-bit ........................................... 365 link field ................................................. 379 software-defined fields.......................... 367 stack pointers ................................. 367, 371 static fields ............................................. 367 TF bit ........................................................... 65 TI bit .................................................... 85, 362 time-stamp counter........................... 188, 408 time-stamp disable (CR4.TSD).. 61, 188, 409 TLB..................................................... 171, 172 explicit invalidation ...................... 174, 191 implicit invalidation .............................. 174 top of memory ........................................... 239 TOP_MEM register ............. 74, 239, 439, 442 TOP_MEM2 register ........... 74, 239, 439, 442 TPR register .................................. 49, 64, 273 TR register ........................................ 360, 363 translation lookaside buffer (TLB) ......... 172 trap............................................................. 245 trap flag (rFLAGS.TF)................................ 65 trap gate ............................................ 106, 287 TS bit ........................................................... 56 TSC register .............................. 408, 435, 445 TSD bit......................................................... 61 TSS ........................................... xxxii, 360, 365 TSS descriptor .......................................... 360 TSS selector ...................................... 108, 360 type check ................................................. 141 Type field..................................... 99, 363, 373
U U/S bit ............................................... 169, 175 UC bit ........................................................ 314 UC memory type ...................................... 202 UD2 instruction........................................ 253 UE exception .................................... 262, 266 uncacheable (UC), memory type ............ 202 uncacheable (UC-), memory type ........... 230 underflow ................................................ xxxii underflow exception (UE)............... 262, 266 unit mask field ......................................... 406 user page................................................... 175 user segment............................................... 97 user/supervisor (U/S) page protection...................................... 175 page-translation tables, bit in .............. 169 USR bit...................................................... 406 V VAL bit ...................................................... 314 Variable-range IORRs.............................. 237 vector....................................................... xxxii vector, interrupt ....................................... 247 VERR instruction..................................... 190 VERW instruction .................................... 190 VIF bit ......................................................... 68 VIP bit ......................................................... 68 virtual address........................................ 3, 31 virtual interrupt (rFLAGS.VIF)........ 68, 298 virtual interrupt pending (rFLAGS.VIP) 68, 298 virtual interrupts ................. 66, 68, 295, 298 virtual interrupts, protected mode......... 300 virtual memory ............................................. 4 virtual-8086 mode ........................... xxxiii, 17 interrupt to protected mode ................ 285 interrupts ............................................... 284 virtual-8086 mode (rFLAGS.VM) ............. 67 virtual-8086 mode extensions (CR4.VME) .... 61, 296 VM bit ......................................................... 67 VME........................................................... 296 VME bit............................................... 61, 284 W W bit .......................................................... 103 WAIT/FWAIT instruction .......................... 55 WB memory type ...................................... 204 WBINVD instruction........................ 191, 216 WC memory type...................................... 203 WP bit.................................................. 57, 176 WP memory type ...................................... 203
Index
459
AMD 64-Bit Technology
24593--Rev. 3.09--September 2003
writable (W), data segment ..................... 103 write buffer ....................................... 196, 205 emptying................................................. 206 write hit ..................................................... 195 write miss .................................................. 195 write ordering ................................... 197, 215 write protect (CR0.WP) ............................. 57 page protection ...................................... 176 write-back (WB), memory type................ 204 writeback, cache line................................ 196 write-combining (WC), memory type ..... 203 write-combining buffer .................... 196, 207 emptying................................................. 208 write-protect (WP), memory type ........... 203 write-though (WT), memory type............ 204 WrMem, MTRR type field ................. 74, 235 WRMSR instruction ........................... 71, 188 WT memory type ...................................... 204 X x87 control word ....................... 344, 346, 354 x87 data pointer register ......... 344, 346, 355 x87 environment ....................................... 346 x87 floating-point instructions feature identification ............................ 340 initializing .............................................. 417 processor state ....................................... 344 saving state............................................. 346 x87 floating-point state, initialization .... 413 x87 instruction pointer register ..... 344, 346, 355 x87 opcode register .................. 344, 347, 355 x87 status word ................. 343, 344, 346, 354 x87 tag word ...................... 343, 344, 346, 354 FXSAVE format ..................................... 356 XMM registers .......................................... 342 Z ZE exception ..................................... 262, 265 zero extension ....................................... 38, 39 zero-divide exception (ZE) .............. 262, 265
460
Index


▲Up To Search▲   

 
Price & Availability of AMD64

All Rights Reserved © IC-ON-LINE 2003 - 2022  
[Add Bookmark] [Contact Us] [Link exchange] [Privacy policy]
Mirror Sites :  [www.datasheet.hk]   [www.maxim4u.com]  [www.ic-on-line.cn] [www.ic-on-line.com] [www.ic-on-line.net] [www.alldatasheet.com.cn] [www.gdcy.com]  [www.gdcy.net]
 . . . . .
  We use cookies to deliver the best possible web experience and assist with our advertising efforts. By continuing to use this site, you consent to the use of cookies. For more information on cookies, please take a look at our Privacy Policy. X