AMD-X5-133AFZ

PRELIMINARY Am5X86™ Microprocessor Family DISTINCTIVE CHARACTERISTICS s High-Performance Design Industry-standard write-back cache support Frequent instructions execute in one clock 105.6-million bytes/second burst bus at 33 MHz Flexible write-through and write-back address control — Advanced 0.35-µ CMOS-process technology — Dynamic bus sizing for 8-, 16-, and 32-bit buses — Supports “soft reset” capability s High On-Chip Integration — 16-Kbyte unified code and data cache — Floating-point unit — Paged, virtual memory management s Enhanced System and Power Management — Stop clock control for reduced power consumption — Industry-standard two-pin System Management Interrupt (SMI) for power management independent of processor operating mode and operating system — Static design with Auto Halt power-down support — Wide range of chipsets supporting SMM available to allow product differentiation — — — — s Complete 32-Bit Architecture — Address and data buses — All registers — 8-, 16-, and 32-bit data types s Standard Features — 3-V core with 5-V tolerant I/O — Available in a 133-MHz version Advanced Micro Devices — Binary compatible with all Am486®DX, Am486DX2, and Am486DX4 microprocessors — Wide range of chipsets and support available through the AMD FusionPCSM Program s 168-pin PGA package or 208-pin SQFP package s IEEE 1149.1 JTAG Boundary-Scan Compatibility s Supports Environmental Protection Agency's Energy Star program — 3-V operation reduces power consumption up to 40% — Energy management capability provides excellent base for energy-efficient design — Works with a variety of energy-efficient, powermanaged devices GENERAL DESCRIPTION The Am5X86™ microprocessor is an addition to the AMD microprocessor product family. The new processor enhances system performance by raising the microprocessor operating frequency to the highest levels allowed by current manufacturing technology, while maintaining complete compatibility with the standard Am486 processor architecture and Microsoft® Windows®. The CPUs incorporate write-back cache, flexible clock control, and enhanced SMM. Table 1 shows available processors in the Am5X86 microprocessor family. Operating Frequency 133 MHz 133 MHz Table 1. Clocking Options Input Clock 33 MHz 33 MHz Available Package 168-pin PGA 208-pin SQFP The Am5X86 microprocessor family allows write-back configuration through software and cacheable access control. On-chip cache lines are configurable as either write-through or write-back. The CPU clock control feature permits the CPU clock to be stopped under controlled conditions, allowing reduced power consumption during system inactivity. The SMM function is implemented with an industry standard two-pin interface. Publication # 19751 Rev: C Amendment/0 Issue Date: March 1996 This document contains information on a product under development at Advanced Micro Devices. The information is intended to help you evaluate this product. AMD reserves the right to change or discontinue work on this proposed product without notice. AMD PRELIMINARY BLOCK DIAGRAM 2 Am5X86 Microprocessor PRELIMINARY AMD ORDERING INFORMATION Standard Products AMD standard products are available in several packages and operating ranges. The order number (Valid Combination) is formed by a combination of the elements below. AMD-X5 – 133 A D W Case Temperature W= 55° C Z = 85° C Operating Voltage D = 3.45 V F = 3.3 V Package Type A =168-pin PGA S = 208-pin SQFP Clock Speed 133 = 133 MHz Family/Core AMD-X5 Valid Combinations OPN AMD-X5-133ADW AMD-X5-133ADZ AMD-X5-133SFZ AMD-X5-133SDZ Package Type PGA PGA SQFP SQFP Operating Voltage 3.45 V 3.45 V 3.3 V 3.45 V Case Temperature (Max) 55° C 85° C 85° C 85° C Valid Combinations Valid Combinations list configurations planned to be supported in volume for this device. Consult the local AMD sales office to confirm availability of specific valid combinations and to check on newly released combinations. Am5X86 Microprocessor 3 AMD PRELIMINARY Table of Contents 1 Connection Diagrams and Pin Designations ......................................................................................... 8 1.1 168-Pin PGA (Pin Grid Array) Package .......................................................................................... 8 1.2 168-Pin PGA Designations (Functional Grouping) ......................................................................... 9 1.3 208-Pin SQFP (Shrink Quad Flat Pack) Package ........................................................................ 10 1.4 208-Pin SQFP Designations (Functional Grouping) ..................................................................... 11 Logic Symbol ...................................................................................................................................... 12 Pin Description .................................................................................................................................... 13 Functional Description ........................................................................................................................ 18 4.1 Overview ....................................................................................................................................... 18 4.2 Memory ......................................................................................................................................... 18 4.3 Modes of Operation ...................................................................................................................... 18 4.3.1 Real mode ........................................................................................................................... 18 4.3.2 Virtual mode ........................................................................................................................ 18 4.3.3 Protected mode ................................................................................................................... 18 4.3.4 System Management mode ................................................................................................ 18 4.4 Cache Architecture ....................................................................................................................... 18 4.4.1 Write-Through Cache .......................................................................................................... 18 4.4.2 Write-Back Cache ............................................................................................................... 18 4.5 Write-Back Cache Protocol ........................................................................................................... 19 4.5.1 Cache Line Overview .......................................................................................................... 19 4.5.2 Line Status and Line State .................................................................................................. 19 4.5.2.1 Invalid ......................................................................................................................... 19 4.5.2.2 Exclusive .................................................................................................................... 19 4.5.2.3 Shared ....................................................................................................................... 19 4.5.2.4 Modified ..................................................................................................................... 19 4.6 Cache Replacement Description .................................................................................................. 20 4.7 Memory Configuration ................................................................................................................... 20 4.7.1 Cacheability ......................................................................................................................... 20 4.7.2 Write-Through/Write-Back ................................................................................................... 20 4.8 Cache Functionality in Write-Back mode ...................................................................................... 20 4.8.1 Processor-Initiated Cache Functions and State Transitions ............................................... 20 4.8.2 Snooping Actions and State Transitions ............................................................................. 21 4.8.2.1 Difference between Snooping Access Cases ............................................................ 21 4.8.2.2 HOLD Bus Arbitration Implementation ....................................................................... 22 4.8.2.2.1 Processor-Induced Bus Cycles ........................................................................ 22 4.8.2.2.2 External Read ................................................................................................... 22 4.8.2.2.3 External Write ................................................................................................... 22 4.8.2.2.4 HOLD/HLDA External Access TIming .............................................................. 22 4.8.3 External Bus Master Snooping Actions ............................................................................... 25 4.8.3.1 Snoop Miss ................................................................................................................. 25 4.8.3.2 Snoop Hit to a Non-Modified Line .............................................................................. 25 4.8.4 Write-Back Case ................................................................................................................. 25 4.8.5 Write-Back and Pending Access ......................................................................................... 26 4.8.5.1 HOLD/HLDA Write-Back Design Considerations ....................................................... 27 4.8.5.2 AHOLD Bus Arbitration Implementation .................................................................... 28 4.8.5.3 Normal Write-Back ..................................................................................................... 28 4.8.6 Reordering of Write-Backs (AHOLD) with BOFF ................................................................. 29 4.8.7 Special Scenarios For AHOLD Snooping ............................................................................ 30 4.8.7.1 Write Cycle Reordering due to Buffering ................................................................... 30 4.8.7.2 BOFF Write-Back Arbitration Implementation ............................................................ 32 4.8.8 BOFF Design Considerations .............................................................................................. 32 4.8.8.1 Cache Line Fills ......................................................................................................... 32 4.8.8.2 Cache Line Copy-Backs ............................................................................................ 32 4.8.8.3 Locked Accesses ....................................................................................................... 32 2 3 4 4 Am5X86 Microprocessor PRELIMINARY AMD 5 6 7 4.8.9 BOFF During Write-Back ..................................................................................................... 32 4.8.10 Snooping Characteristics During a Cache Line Fill ........................................................... 32 4.8.11 Snooping Characteristics During a Copy-Back ................................................................. 32 4.9 Cache Invalidation and Flushing in Write-Back mode .................................................................. 33 4.9.1 Cache Invalidation through Software .................................................................................. 33 4.9.2 Cache Invalidation through Hardware ................................................................................. 33 4.9.3 Snooping During Cache Flushing ........................................................................................ 34 4.10 Burst Write .................................................................................................................................. 34 4.10.1 Locked Accesses .............................................................................................................. 35 4.10.2 Serialization ....................................................................................................................... 35 4.10.3 PLOCK Operation in Write-Through mode ........................................................................ 36 Clock Control ...................................................................................................................................... 36 5.1 Clock Generation .......................................................................................................................... 36 5.2 Stop Clock ..................................................................................................................................... 36 5.2.1 External Interrupts in Order of Priority ................................................................................. 36 5.3 Stop Grant Bus Cycle ................................................................................................................... 36 5.4 Pin State during Stop Grant .......................................................................................................... 37 5.5 Clock Control State Diagram ........................................................................................................ 37 5.5.1 Normal State ........................................................................................................................ 37 5.5.2 Stop Grant State .................................................................................................................. 37 5.5.3 Stop Clock State .................................................................................................................. 39 5.5.4 Auto Halt Power Down State ............................................................................................... 39 5.5.5 Stop Clock Snoop State (Cache Invalidations) .................................................................... 39 5.5.6 Cache Flush State ............................................................................................................... 39 SRESET Function ............................................................................................................................... 39 System Management mode ................................................................................................................ 39 7.1 Overview ....................................................................................................................................... 39 7.2 Terminology .................................................................................................................................. 40 7.3 System Management Interrupt Processing ................................................................................... 40 7.3.1 System Management Interrupt Processing ......................................................................... 41 7.3.2 SMI Active (SMIACT) .......................................................................................................... 41 7.3.3 SMRAM ............................................................................................................................... 42 7.3.4 SMRAM State Save Map .................................................................................................... 43 7.4 Entering System Management mode ............................................................................................ 44 7.5 Exiting System Management mode .............................................................................................. 44 7.6 Processor Environment ................................................................................................................. 44 7.7 Executing System Management mode Handler ............................................................................ 45 7.7.1 Exceptions and Interrupts with System Management mode ............................................... 46 7.7.2 SMM Revisions Identifier ..................................................................................................... 46 7.7.3 Auto HALT Restart .............................................................................................................. 47 7.7.4 I/O Trap Restart ................................................................................................................... 47 7.7.5 I/O Trap Word ...................................................................................................................... 47 7.7.6 SMM Base Relocation ......................................................................................................... 48 7.8 SMM System Design Considerations ........................................................................................... 48 7.8.1 SMRAM Interface ................................................................................................................ 48 7.8.2 Cache Flushes .................................................................................................................... 49 7.8.3 A20M Pin ............................................................................................................................. 49 7.8.4 CPU Reset during SMM ...................................................................................................... 52 7.8.5 SMM and Second Level Write Buffers ................................................................................ 52 7.8.6 Nested SMI and I/O Restart ................................................................................................ 52 7.9 SMM Software Considerations ..................................................................................................... 52 7.9.1 SMM Code Considerations ................................................................................................. 52 7.9.2 Exception Handling ............................................................................................................. 52 7.9.3 Halt during SMM .................................................................................................................. 53 7.9.4 Relocating SMRAM to an Address above 1 Mbyte ............................................................. 53 Am5X86 Microprocessor 5 AMD PRELIMINARY 8 9 10 11 12 13 Test Registers 4 and 5 Modifications .................................................................................................. 53 8.1 TR4 Definition ................................................................................................................................ 53 8.2 TR5 Definition ................................................................................................................................ 54 8.3 Using TR4 and TR5 for Cache Testing.......................................................................................... 55 8.3.1 Example 1: Reading the Cache (Write-back mode only) ..................................................... 55 8.3.2 Example 2: Writing the Cache .............................................................................................. 55 8.3.3 Example 3: Flushing the Cache ........................................................................................... 55 Am5X86 CPU Functional Differences ................................................................................................. 55 9.1 Status after Reset ......................................................................................................................... 55 9.2 Cache Status ................................................................................................................................ 55 9.3 CLKMUL Pin ................................................................................................................................. 55 Am5X86 CPU Identification ................................................................................................................. 56 10.1 DX Register at RESET ................................................................................................................ 56 10.2 CPUID Instruction ....................................................................................................................... 56 10.2.1 CPUID Timing ................................................................................................................... 56 10.2.2 CPUID Operation .............................................................................................................. 56 Electrical Data ..................................................................................................................................... 57 11.1 Power and Grounding ................................................................................................................. 57 11.1.1 Power Connections ........................................................................................................... 57 11.1.2 Power Decoupling Recommendations .............................................................................. 57 11.1.3 Other Connection Recommendations ............................................................................... 57 Package Thermal Specifications ......................................................................................................... 65 Physical Dimensions ........................................................................................................................... 66 LIST OF FIGURES Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 Figure 19 Figure 20 Figure 21 Figure 22 Figure 23 Figure 24 Figure 25 Figure 26 Figure 27 Figure 28 Figure 29 Processor-Induced Line Transitions in Write-Back mode ....................................................... 20 Snooping State Transitions ..................................................................................................... 21 Typical System Block Diagram for HOLD/HLDA Bus Arbitration ............................................ 22 External Read ......................................................................................................................... 23 External Write .......................................................................................................................... 23 Snoop of On-Chip Cache That Does Not Hit a Line ................................................................ 24 Snoop of On-Chip Cache That Hits a Non-modified Line ........................................................ 24 Snoop That Hits a Modified Line (Write-Back) ........................................................................ 25 Write-Back and Pending Access ............................................................................................. 26 Valid HOLD Assertion During Write-Back ............................................................................... 27 Closely Coupled Cache Block Diagram .................................................................................. 28 Snoop Hit Cycle with Write-Back ............................................................................................. 29 Cycle Reordering with BOFF (Write-Back) .............................................................................. 30 Write Cycle Reordering Due to Buffering ................................................................................ 31 Latest Snooping of Copy-Back ................................................................................................ 33 Burst Write .............................................................................................................................. 34 Burst Read with BOFF Assertion ............................................................................................ 34 Burst Write with BOFF Assertion ............................................................................................. 35 Entering Stop Grant State ....................................................................................................... 37 Stop Clock State Machine ....................................................................................................... 38 Recognition of Inputs when Exiting Stop Grant State ............................................................. 38 Basic SMI Interrupt Service ..................................................................................................... 40 Basic SMI Hardware Interface.................................................................................................. 41 SMI Timing for Servicing an I/O Trap ...................................................................................... 41 SMIACT Timing ....................................................................................................................... 42 Redirecting System Memory Address to SMRAM ................................................................... 42 Transition to and from SMM .................................................................................................... 44 Auto HALT Restart Register Offset .......................................................................................... 47 I/O Instruction Restart Register Offset .................................................................................... 47 6 Am5X86 Microprocessor PRELIMINARY AMD Figure 30 Figure 31 Figure 32 Figure 33 Figure 34 Figure 35 Figure 36 Figure 37 Figure 38 Figure 39 Figure 40 Figure 41 Figure 42 Figure 43 Figure 44 Figure 45 Figure 46 SMM Base Slot Offset ............................................................................................................. 48 SRAM Usage .......................................................................................................................... 48 SMRAM Location .................................................................................................................... 49 SMM Timing in Systems Using Non-Overlaid Memory Space and Write-Through Mode with Caching Enabled During SMM.......................................................................................... 50 SMM Timing in Systems Using Non-Overlaid Memory Spaces and Write-Back Mode with Caching Enabled During SMM ................................................................................................. 50 SMM Timing in Systems Using Non-Overlaid Memory Spaces and Write-Back Mode with Caching Disabled During SMM ................................................................................................ 50 SMM Timing in Systems Using Overlaid Memory Space and Write-Through Mode with Caching Enabled During SMM ................................................................................................. 51 SMM Timing in Systems Using Overlaid Memory Spaces and Write-Through Mode with Caching Disabled During SMM ................................................................................................ 51 SMM Timing in Systems Using Overlaid Memory Spaces and Configured in Write-Back Mode...................................................................................................................... 51 CLK Waveforms ...................................................................................................................... 61 Output Valid Delay Timing ...................................................................................................... 61 Maximum Float Delay Timing .................................................................................................. 62 PCHK Valid Delay Timing ....................................................................................................... 62 Input Setup and Hold Timing ................................................................................................... 63 RDY and BRDY Input Setup and Hold Timing ........................................................................ 63 TCK Waveforms ...................................................................................................................... 64 Test Signal Timing Diagram .................................................................................................... 64 LIST OF TABLES Table 1 Table 2 Table 3 Table 4 Table 5 Table 6 Table 7 Table 8 Table 9 Table 10 Table 11 Table 12 Table 13 Table 14 Table 15 Table 16 Table 17 Table 18 Table 19 Table 20 Table 21 Table 22 Table 23 Clocking Options ....................................................................................................................... 1 EADS Sample Time ................................................................................................................ 14 Cache Line Organization ......................................................................................................... 19 Legal Cache Line States ......................................................................................................... 19 MESI Cache Line Status ......................................................................................................... 20 Key to Switching Waveforms ................................................................................................... 22 WBINVD/INVD Special Bus Cycles ......................................................................................... 33 FLUSH Special Bus Cycles ..................................................................................................... 34 Pin State during Stop Grant Bus State .................................................................................... 37 SMRAM State Save Map ........................................................................................................ 43 SMM Initial CPU Core Register Settings ................................................................................. 45 Segment Register Initial States ............................................................................................... 45 SMM Revision Identifier .......................................................................................................... 46 SMM Revision Identifier Bit Definitions ................................................................................... 46 HALT Auto Restart Configuration ............................................................................................ 47 I/O Trap Word Configuration ................................................................................................... 47 Test Register TR4 Bit Descriptions ......................................................................................... 53 Test Register TR5 Bit Descriptions ......................................................................................... 53 CPU ID Codes ......................................................................................................................... 56 CPUID Instruction Description ................................................................................................. 56 Thermal Resistance (°C/W) θJC and θJA for the Am5X86 CPU in 168-Pin PGA Package ....... 65 Maximum TA at Various Airflows in °C .................................................................................... 65 Maximum TA for SQFP Package by Clock Frequency ............................................................. 65 Am5X86 Microprocessor 7 AMD PRELIMINARY 1 1.1 CONNECTION DIAGRAMS AND PIN DESIGNATIONS 168-pin PGA (Pin Grid Array) Package A D20 B D19 D21 C D11 D18 D D9 D13 E VSS VCC F DP1 D8 G VSS VCC H VSS D3 J INC D5 K VSS VCC L VSS D6 M VSS VCC N D2 D1 P D0 A29 Q A31 VSS R A28 A25 S A27 1 D22 A26 1 2 TCK VSS CLK D17 D10 D15 D12 DP2 D16 D14 D7 D4 DP0 A30 A17 VCC A23 2 3 D23 VSS VSS D25 VCC VCC VCC D27 D26 D13 D13 D13 VCC VCC VCC D8 D8 D8 VCC VCC VCC D3 D3 D3 D5 D5 D5 VCC VCC VCC D6 D6 D6 VCC VCC VCC D1 D1 D1 A29 A29 A29 A19 A21 A24 A22 VSS A18 VCC A15 VOLDET 3 4 DP3 A14 4 5 D24 VSS 5 6 VSS A12 6 7 D29 D31 VCC D28 D30 D13 D13 VCC VCC D8 D8 VCC VCC D3 D5 VCC D6 VCC VCC D1 D1 A29 A29 A20 A16 VCC VCC VSS 7 8 VSS D3 D5 VCC PIN SIDE VIEW D6 D3 D5 VCC D6 VSS 8 9 INV SMI SRESET VCC D8 VCC VCC D1 A29 A13 VCC VSS 9 10 VSS VCC UP D13 VCC D8 VCC D3 D5 VCC D6 VCC D1 A29 A9 VCC VSS 10 11 HITM CACHE SMIACT A5 A7 A2 A11 A8 VCC VSS 11 12 INC WB/WT INC TMS FERR A10 12 13 TDI VSS 13 14 IGNNE NMI FLUSH A20M HOLD KEN STPCLK BRDY BE2 BE0 VCC VSS PWT VCC VSS D/C VCC VSS LOCK HLDA BREQ M/IO W/R VCC A3 A6 14 15 INTR TDO RESET BS8 VCC VSS RDY BE3 VCC VSS VCC VSS BE1 PCD PLOCK BLAST A4 15 16 AHOLD EADS BS16 BOFF VSS PCHK CLKMUL ADS 16 17 A B C D E F G H J K L M N P Q R S 17 8 Am5X86 Microprocessor PRELIMINARY AMD 1.2 168-pin PGA Designations (Functional Grouping) Address Data Pin Name D0 D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 D11 D12 D13 D14 D15 D16 D17 D18 D19 D20 D21 D22 D23 D24 D25 D26 D27 D28 D29 D30 D31 Pin No. P-1 N-2 N-1 H-2 M-3 J-2 L-2 L-3 F-2 D-1 E-3 C-1 G-3 D-2 K-3 F-3 J-3 D-3 C-2 B-1 A-1 B-2 A-2 A-4 A-6 B-6 C-7 C-6 C-8 A-8 C-9 B-8 Control Pin Name A20M ADS AHOLD BE0 BE1 BE2 BE3 BLAST BOFF BRDY BREQ BS8 BS16 CACHE CLK CLKMUL D/C DP0 DP1 DP2 DP3 EADS FERR FLUSH HITM HLDA HOLD IGNNE INTR INV KEN LOCK M/IO NMI PCD PCHK PLOCK PWT RDY RESET SMI SMIACT SRESET STPCLK UP VOLDET WB/WT W/R Pin No. D-15 S-17 A-17 K-15 J-16 J-15 F-17 R-16 D-17 H-15 Q-15 D-16 C-17 B-12 C-3 R-17 M-15 N-3 F-1 H-3 A-5 B-17 C-14 C-15 A-12 P-15 E-15 A-15 A-16 A-10 F-15 N-15 N-16 B-15 J-17 Q-17 Q-16 L-15 F-16 C-16 B-10 C-12 C-10 G-15 C-11 S-4 B-13 N-17 Pin Name TCK TDI TDO TMS Test Pin No. A-3 A-14 B-16 B-14 INC Pin No. A-13 C-13 J-1 Vcc Pin No. B-7 B-9 B-11 C-4 C-5 E-2 E-16 G-2 G-16 H-16 K-2 K-16 L-16 M-2 M-16 P-16 R-3 R-6 R-8 R-9 R-10 R-11 R-14 Vss Pin No. A-7 A-9 A-11 B-3 B-4 B-5 E-1 E-17 G-1 G-17 H-1 H-17 K-1 K-17 L-1 L-17 M-1 M-17 P-17 Q-2 R-4 S-6 S-8 S-9 S-10 S-11 S-12 S-14 Pin No. Q-14 R-15 S-16 Q-12 S-15 Q-13 R-13 Q-11 S-13 R-12 S-7 Q-10 S-5 R-7 Q-9 Q-3 R-5 Q-4 Q-8 Q-5 Q-7 S-3 Q-6 R-2 S-2 S-1 R-1 P-2 P-3 Q-1 Pin Name A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A12 A13 A14 A15 A16 A17 A18 A19 A20 A21 A22 A23 A24 A25 A26 A27 A28 A29 A30 A31 Notes: 1. VOLDET is connected internally to VSS. 2. INC = Internal No Connect Am5X86 Microprocessor 9 AMD PRELIMINARY 1.3 208-pin SQFP (Shrink Quad Flat Pack) Package TOP VIEW 10 Am5X86 Microprocessor PRELIMINARY AMD 1.4 208-pin SQFP Designations (Functional Grouping) Address Data Pin Name D0 D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 D11 D12 D13 D14 D15 D16 D17 D18 D19 D20 D21 D22 D23 D24 D25 D26 D27 D28 D29 D30 D31 Pin No. 144 143 142 141 140 130 129 126 124 123 119 118 117 116 113 112 108 103 101 100 99 93 92 91 87 85 84 83 79 78 75 74 Control Pin Name A20M ADS AHOLD BE0 BE1 BE2 BE3 BLAST BOFF BRDY BREQ BS8 BS16 CACHE CLK CLKMUL D/C DP0 DP1 DP2 DP3 EADS FERR FLUSH HITM HLDA HOLD IGNNE INTR INV KEN LOCK M/IO NMI PCD PCHK PLOCK PWT RDY RESET SMI SRESET STPCLK SMIACT UP WB/WT W/R Pin No. 47 203 17 31 32 33 34 204 6 5 30 8 7 70 24 11 39 145 125 109 90 46 66 49 63 26 16 72 50 71 13 207 37 51 41 4 206 40 12 48 65 58 73 59 194 64 27 Pin Name TCK TDI TDO TMS Test Pin No. 18 168 68 167 INC Pin No. 3 67 96 127 Vcc Pin No. 2 9 14 19 20 22 23 25 29 35 38 42 44 45 54 56 60 62 69 77 80 82 86 89 95 98 102 106 111 114 121 128 131 133 134 136 137 139 150 155 162 163 169 172 176 179 183 185 188 191 198 200 205 Vss Pin No. 1 10 15 21 28 36 43 52 53 55 57 61 76 81 88 94 97 104 105 107 110 115 120 122 132 135 138 146 156 157 170 175 181 184 189 199 201 208 Pin Name A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A12 A13 A14 A15 A16 A17 A18 A19 A20 A21 A22 A23 A24 A25 A26 A27 A28 A29 A30 A31 Pin No. 202 197 196 195 193 192 190 187 186 182 180 178 177 174 173 171 166 165 164 161 160 159 158 154 153 152 151 149 148 147 Note: INC = Internal No Connect Am5X86 Microprocessor 11 AMD PRELIMINARY 2 LOGIC SYMBOL Clock Stop Clock Clock Multiplier Address Mask Upgrade Present Voltage Detect 28 2 CLK STPCLK CLKMUL A20M UP VOLDET A31–A4 A3–A2 BE3–BE0 BS8 D31–D0 32 Data Bus DP3–DP0 PCHK BRDY BLAST CACHE SMI SMIACT 4 Data Parity Address Bus Burst Control 4 SMM Bus Cycle Control BS16 ADS RDY M/IO D/C Am5X86 CPU PWT PCD WB/WT INV KEN FLUSH AHOLD EADS HITM Page Cacheability Bus Cycle Definition W/R LOCK PLOCK Cache Control/ Invalidation Interrupts INTR NMI RESET SRESET HOLD BOFF BREQ HLDA TCK TMS IGNNE FERR TDI TDO Bus Arbitration Numeric Error Reporting IEEE Test Port Access 12 Am5X86 Microprocessor PRELIMINARY AMD 3 PIN DESCRIPTIONS BE3–BE0 Byte Enable (Active Low; Outputs) The byte enable pins indicate which bytes are enabled and active during read or write cycles. During the first cache fill cycle, however, an external system should ignore these signals and assume that all bytes are active. s BE3 for D31–D24 s BE2 for D23–D16 s BE1 for D15–D8 s BE0 for D7–D0 The Am5X86 microprocessor provides the complete interface support offered by the Enhanced Am486 microprocessor family products. The CLKMUL pin settings have changed to accommodate the higher operating speed selection. A20M Address Bit 20 Mask (Active Low; Input) A Low signal on the A20M pin causes the microprocessor to mask address line A20 before performing a lookup to the internal cache, or driving a memory cycle on the bus. Asserting A20M causes the processor to wrap the address at 1 Mbyte, emulating Real mode operation. The signal is asynchronous, but must meet setup and hold times t20 and t21 for recognition during a specific clock. During normal operation, A20M should be sampled High at the falling edge of RESET. BE3–BE0 are active Low and are not driven during bus hold. BLAST Burst Last (Active Low; Output) Burst Last goes Low to tell the CPU that the next BRDY signal completes the burst bus cycle. BLAST is active for both burst and non-burst cycles. BLAST is active Low and is not driven during a bus hold. A31–A4/A3–A2 Address Lines (Inputs/Outputs)/(Outputs) Pins A31–A2 define a physical area in memory or indicate an input/output (I/O) device. Address lines A31–A4 drive addresses into the microprocessor to perform cache line invalidations. Input signals must meet setup and hold times t22 and t23. A31–A2 are not driven during bus or address hold. BOFF Back Off (Active Low; Input) This input signal forces the microprocessor to float all pins normally floated during hold, but HLDA is not asserted in response to BOFF. BOFF has higher priority than RDY or BRDY; if both are returned in the same clock, BOFF takes effect. The microprocessor remains in bus hold until BOFF goes High. If a bus cycle is in progress when BOFF is asserted, the cycle restarts. BOFF must meet setup and hold times t18 and t19 for proper operation. BOFF has an internal weak pull-up. ADS Address Status (Active Low; Output) A Low output from this pin indicates that a valid bus cycle definition and address are available on the cycle definition lines and address bus. ADS is driven active by the same clock as the addresses. ADS is active Low and is not driven during bus hold. BRDY Burst Ready Input (Active Low; Input) The BRDY signal performs the same function during a burst cycle that RDY performs during a non-burst cycle. BRDY indicates that the external system has presented valid data in response to a read, or that the external system has accepted data in response to a write. BRDY is ignored when the bus is idle and at the end of the first clock in a bus cycle. BRDY is sampled in the second and subsequent clocks of a burst cycle. The data presented on the data bus is strobed into the microprocessor when BRDY is sampled active. If RDY is returned simultaneously with BRDY, BRDY is ignored and the cycle is converted to a non-burst cycle. BRDY is active Low and has a small pull-up resistor, and must satisfy the setup and hold times t16 and t17. AHOLD Address Hold (Active High; Input) The external system may assert AHOLD to perform a cache snoop. In response to the assertion of AHOLD, the microprocessor stops driving the address bus A31– A2 in the next clock. The data bus remains active and data can be transferred for previously issued read or write bus cycles during address hold. AHOLD is recognized even during RESET and LOCK. The earliest that AHOLD can be deasserted is two clock cycles after EADS is asserted to start a cache snoop. If HITM is activated due to a cache snoop, the microprocessor completes the current bus activity and then asserts ADS and drives the address bus while AHOLD is active. This starts the write-back of the modified line that was the target of the snoop. BREQ Internal Cycle Pending (Active High; Output) BREQ indicates that the microprocessor has generated a bus request internally, whether or not the microprocessor is driving the bus. BREQ is active High and is floated only during Tri-state Test mode (see FLUSH). Am5X86 Microprocessor 13 AMD PRELIMINARY BS8/BS16 Bus Size 8 (Active Low; Input)/ Bus Size 16 (Active Low; Input) The BS8 and BS16 signals allow the processor to operate with 8-bit and 16-bit I/O devices by running multiple bus cycles to respond to data requests: four for 8-bit devices, and two for 16-bit devices. The bus sizing pins are sampled every clock. The microprocessor samples the pins every clock before RDY to determine the appropriate bus size for the requesting device. The signals are active Low input with internal pull-up resistors, and must satisfy setup and hold times t14 and t15 for correct operation. Bus sizing is not permitted during copy-back or write-back operation. BS8 and BS16 are ignored during copy-back or write-back cycles. D/C Data/Control (Output) This bus cycle definition pin distinguishes memory and I/O data cycles from control cycles. The control cycles are: s Interrupt Acknowledge s Halt/Special Cycle s Code Read (instruction fetching) DP3–DP0 Data Parity (Inputs/Outputs) Data parity is generated on all write data cycles with the same timing as the data driven by the microprocessor. Even parity information must be driven back into the microprocessor on the data parity pins with the same timing as read information to ensure that the processor uses the correct parity check. The signals read on these pins do not affect program execution. Input signals must meet setup and hold times t22 and t23. DP3–DP0 should be connected to VCC through a pull-up resistor in systems not using parity. DP3–DP0 are active High and are driven during the second and subsequent clocks of write cycles. CACHE Internal Cacheability (Active Low; Output) In Write-through mode, this signal always floats. In Write-back mode for processor-initiated cycles, a Low output on this pin indicates that the current read cycle is cacheable, or that the current cycle is a burst writeback or copy-back cycle. If the CACHE signal is driven High during a read, the processor will not cache the data even if the KEN pin signal is asserted. If the processor determines that the data is cacheable, CACHE goes active when ADS is asserted and remains in that state until the next RDY or BRDY is asserted. CACHE floats in response to a BOFF or HOLD request. EADS External Address Strobe (Active Low; Input) This signal indicates that a valid external address has been driven on the address pins A31–A4 of the microprocessor to be used for a cache snoop. This signal is recognized while the processor is in hold (HLDA is driven active), while forced off the bus with the BOFF input, or while AHOLD is asserted. The microprocessor ignores EADS at all other times. EADS is not recognized if HITM is active, nor during the clock after ADS, nor during the clock after a valid assertion of EADS. Snoops to the on-chip cache must be completed before another snoop cycle is initiated. Table 2 describes EADS when first sampled. EADS can be asserted every other clock cycle as long as the hold remains active and HITM remains inactive. INV is sampled in the same clock period that EADS is asserted. EADS has an internal weak pullup. Table 2. EADS Sample Time Trigger AHOLD HOLD BOFF EADS First Sampled Second clock after AHOLD asserted First clock after HLDA asserted Second clock after BOFF asserted CLK Clock (Input) The CLK input provides the basic microprocessor timing signal. The CLKMUL input selects the multiplier value used to generate the internal operating frequency for the Am5X86 microprocessor family. All external timing parameters are specified with respect to the rising edge of CLK. The clock signal passes through an internal Phase-Lock Loop (PLL). CLKMUL Clock Multiplier (Input) The microprocessor samples the CLKMUL input signal at RESET to determine the design operating frequency. An internal pull-up resistor connects to VCC, which selects Clock-tripled mode if the input is High or left floating. For Clock-quadrupled mode, the input must be pulled Low. For 133-MHz processors, this input must always be connected to VSS to ensure correct operation. D31–D0 Data Lines (Inputs/Outputs) Lines D31–D0 define the data bus. The signals must meet setup and hold times t22 and t23 for proper read operations. These pins are driven during the second and subsequent clocks of write cycles. Note: The triggering signal (AHOLD, HOLD, or BOFF ) must remain active for at least 1 clock after EADS to ensure proper operation. 14 Am5X86 Microprocessor PRELIMINARY AMD FERR Floating-Point Error (Active Low; Output) Driven active when a floating-point error occurs, FERR is similar to the ERROR pin on a 387 math coprocessor. FERR is included for compatibility with systems using DOS-type floating-point error reporting. FERR is active Low, and is not floated during bus hold, except during Tri-state Test mode (see FLUSH). pins. HLDA is asserted after completing the current bus cycle, burst cycle, or sequence of locked cycles. The microprocessor remains in this state until HOLD is deasserted. HOLD is active High and does not have an internal pull-down resistor. HOLD must satisfy setup and hold times t18 and t19 for proper operation. IGNNE Ignore Numeric Error (Active Low; Input) When this pin is asserted, the Am5X86 microprocessor will ignore a numeric error and continue executing non-control floating-point instructions. When IGNNE is deasserted, the Am5X86 microprocessor will freeze on a non-control floating-point instruction if a previous floating-point instruction caused an error. IGNNE has no effect when the NE bit in Control Register 0 is set. IGNNE is active Low and is provided with a small internal pullup resistor. IGNNE is asynchronous but must meet setup and hold times t20 and t21 to ensure recognition in any specific clock. FLUSH Cache Flush (Active Low; Input) In Write-back mode, FLUSH forces the microprocessor to write-back all modified cache lines and invalidate its internal cache. The microprocessor generates two flush acknowledge special bus cycles to indicate completion of the write-back and invalidation. In Write-through mode, FLUSH invalidates the cache without issuing a special bus cycle. FLUSH is an active Low input that needs to be asserted only for one clock. FLUSH is asynchronous, but setup and hold times t20 and t21 must be met for recognition in any specific clock. Sampling FLUSH Low in the clock before the falling edge of RESET causes the microprocessor to enter Tri-state Test mode. INTR Maskable Interrupt (Active High; Input) When asserted, this signal indicates that an external interrupt has been generated. If the internal interrupt flag is set in EFLAGS, active interrupt processing is initiated. The microprocessor generates two locked interrupt acknowledge bus cycles in response to the INTR pin going active. INTR must remain active until the interrupt acknowledges have been performed to ensure that the interrupt is recognized. INTR is active High and is not provided with an internal pull-down resistor. INTR is asynchronous, but must meet setup and hold times t20 and t21 for recognition in any specific clock. HITM Hit Modified Line (Active Low; Output) In Write-back mode (WB/WT=1 at RESET), HITM indicates that an external snoop cache tag comparison hit a modified line. When a snoop hits a modified line in the internal cache, the microprocessor asserts HITM two clocks after EADS is asserted. The HITM signal stays asserted (Low) until the last BRDY for the corresponding write-back cycle. At all other times, HITM is deasserted (High). During RESET, the HITM signal can be used to detect whether the CPU is operating in Write-back mode. In Write-back mode (WB/WT=1 at RESET), HITM is deasserted (driven High) until the first snoop that hits a modified line. In Write-through mode, HITM floats at all times. INV Invalidate (Active High; Input) The external system asserts INV to invalidate the cacheline state when an external bus master proposes a write. It is sampled together with A31–A4 during the clock in which EADS is active. INV has an internal weak pull-up. INV is ignored in Write-through mode. HLDA Hold Acknowledge (Active High; Output) The HLDA signal is activated in response to a hold request presented on the HOLD pin. HLDA indicates that the microprocessor has given the bus to another local bus master. HLDA is driven active in the same clock in which the microprocessor floats its bus. HLDA is driven inactive when leaving bus hold. HLDA is active High and remains driven during bus hold. HLDA is floated only during Tri-state Test mode (see FLUSH). KEN Cache Enable (Active Low; Input) KEN determines whether the current cycle is cacheable. When the microprocessor generates a cacheable cycle and KEN is active one clock before RDY or BRDY during the first transfer of the cycle, the cycle becomes a cache line fill cycle. Returning KEN active one clock before RDY during the last read in the cache line fill causes the line to be placed in the on-chip cache. KEN is active Low and is provided with a small internal pull-up resistor. KEN must satisfy setup and hold times t14 and t15 for proper operation. HOLD Bus Hold Request (Active High; Input) HOLD gives control of the microprocessor bus to another bus master. In response to HOLD going active, the microprocessor floats most of its output and input/output Am5X86 Microprocessor 15 AMD PRELIMINARY transaction requires more than one bus cycle. Examples of such operations are segment table descriptor reads (8 bytes) and cache line fills (16 bytes). The microprocessor drives PLOCK active until the addresses for the last bus cycle of the transaction have been driven, whether or not RDY or BRDY is returned. PLOCK is a function of the BS8, BS16, and KEN inputs. PLOCK should be sampled on the clock when RDY is returned. PLOCK is active Low and is not driven during bus hold. LOCK Bus Lock (Active Low; Output) A Low output on this pin indicates that the current bus cycle is locked. The microprocessor ignores HOLD when LOCK is asserted (although it does acknowledge AHOLD and BOFF). LOCK goes active in the first clock of the first locked bus cycle and goes inactive after the last clock of the last locked bus cycle. The last locked cycle ends when RDY is returned. LOCK is active Low and is not driven during bus hold. Locked read cycles are not transformed into cache fill cycles if KEN is active. PWT Page Write-Through (Active High; Output) This pin reflects the state of the PWT bit in the page table entry or page directory entry (programmable through the PWT bit in CR3). If paging is disabled, the CPU ignores the PWT bit and drives the PWT output Low. PWT has the same timing as the cycle definition pins (M/IO, D/C, and W/R). PWT is active High and is not driven during bus hold. M/IO Memory/Input-Output (Active High/Active Low; Output) A High output indicates a memory cycle. A Low output indicates an I/O cycle. NMI Non-Maskable Interrupt (Active High; Input) A High NMI input signal indicates that an external nonmaskable interrupt has occurred. NMI is rising-edge sensitive. NMI must be held Low for at least four CLK periods before this rising edge. The NMI input does not have an internal pull-down resistor. The NMI input is asynchronous, but must meet setup and hold times t20 and t21 for recognition in any specific clock. RESET Reset (Active High; Input) RESET forces the microprocessor to initialize. The microprocessor cannot begin execution of instructions until at least 1 ms after VCC and CLK have reached their proper DC and AC specifications. To ensure proper microprocessor operation, the RESET pin should remain active during this time. RESET is active High. RESET is asynchronous but must meet setup and hold times t20 and t21 to ensure recognition on any specific clock. PCD Page Cache Disable (Active High; Output) This pin reflects the state of the PCD bit in the page table entry or page directory entry (programmable through the PCD bit in CR3). If paging is disabled, the CPU ignores the PCD bit and drives the PCD output Low. PCD has the same timing as the cycle definition pins (M/IO, D/C, and W/R). PCD is active High and is not driven during bus hold. PCD is masked by the Cache Disable bit (CD) in Control Register 0 (CR0). RDY Non-Burst Ready (Active Low; Input) A Low input on this pin indicates that the current bus cycle is complete, that is, either the external system has presented valid data on the data pins in response to a read, or the external system has accepted data from the microprocessor in response to a write. RDY is ignored when the bus is idle and at the end of the bus cycle’s first clock. RDY is active during address hold. Data can be returned to the processor while AHOLD is active. RDY is active Low and does not have an internal pullup resistor. RDY must satisfy setup and hold times t16 and t17 for proper chip operation. PCHK Parity Status (Active Low; Output) Parity status is driven on the PCHK pin the clock after RDY for read operations. The parity status reflects data sampled at the end of the previous clock. A Low PCHK indicates a parity error. Parity status is checked only for enabled bytes as is indicated by the byte enable and bus size signals. PCHK is valid only in the clock immediately after read data is returned to the microprocessor; at all other times PCHK is inactive High. PCHK is floated only during Tri-state Test mode (see FLUSH). SMI SMM Interrupt (Active Low; Input) A Low signal on the SMI pin signals the processor to enter System Management mode (SMM). SMI is the highest level processor interrupt. The SMI signal is recognized on an instruction boundary, similar to the NMI and INTR signals. SMI is sampled on every rising clock edge. SMI is a falling-edge sensitive input. The SMI input has an internal pull-up resister. Recognition of SMI is guaranteed in a specific clock if it is asserted synchronously and meets the setup and hold times. If SMI is asserted asynchronously, it must go High for a minimum of two clocks before going Low, and it must remain Low PLOCK Pseudo-Lock (Active Low; Output) In Write-back mode, the processor forces the output High and the signal is always read as inactive. In Writethrough mode, PLOCK operates normally. When asserted, PLOCK indicates that the current bus 16 Am5X86 Microprocessor PRELIMINARY for at least two clocks to guarantee recognition. When the CPU recognizes SMI, it enters SMM before executing the next instruction and saves internal registers in SMM space. AMD rising edge of TCK during the SHIFT-IR and the SHIFT-DR TAP (Test Access Port) controller states. During all other TAP controller states, TDI is ignored. TDI uses an internal weak pull-up. SMIACT SMM Interrupt Active (Active Low; Output) SMIACT goes Low in response to SMI. It indicates that the processor is operating under SMM control. SMIACT remains Low until the processor receives a RESET signal or executes the Resume Instruction (RSM) to leave SMM. This signal is always driven. It does not float during bus HOLD or BOFF. TDO Test Data Output (Active High; Output) TDO is the serial output that shifts JTAG instructions and data out of the component. TDO is driven on the falling edge of TCK during the SHIFT-IR and SHIFT-DR TAP controller states. Otherwise, TDO is tri-stated. TMS Test Mode Select (Active High; Input) TMS is decoded by the JTAG TAP to select the operation of the test logic. TMS is sampled on the rising edge of TCK. To guarantee deterministic behavior of the TAP controller, the TMS pin has an internal pull-up resistor. Note: Do not use SRESET to exit from SMM. The system should block SRESET during SMM. SRESET Soft Reset (Active High; Input) The CPU samples SRESET on every rising clock edge. If SRESET is sampled active, the SRESET sequence begins on the next instruction boundary. SRESET resets the processor, but, unlike RESET, does not cause it to sample UP or WB/WT, or affect the FPU, cache, CD and NW bits in CR0, and SMBASE. SRESET is asynchronous and must meet the same timing as RESET. The SRESET input has an internal pull-down resistor. UP Write/Read (Input) The processor samples the Upgrade Present (UP) pin in the clock before the falling edge of RESET. If it is Low, the processor tri-states its outputs immediately. UP must remain asserted to keep the processor inactive. The pin uses an internal pull-up resistor. STPCLK Stop Clock (Active Low; Input) A Low input signal indicates a request has been made to turn off the CLK input. When the CPU recognizes a STPCLK, the processor: s s s VOLDET—(168-pin PGA package only) Voltage Detect (Output) VOLDET provides an external signal to allow the system to determine the CPU input power level (3 V or 5 V). For Am5X86 processors, the pin ties internally to VSS. Stops execution on the next instruction boundary (unless superseded by a higher priority interrupt) Empties all internal pipelines and write buffers Generates a Stop Grant acknowledge bus cycle WB/WT Write-Back/Write-Through (Input) If the processor samples WB/WT High at RESET, the processor is configured in Write-back mode and all subsequent cache line fills sample WB/WT on the same clock edge in which it finds either RDY or the first BRDY of a burst transfer to determine if the cache line is designated as Write-back mode or Write-through. If the signal is Low on the first BRDY or RDY, the cache line is write-through. If the signal is High, the cache line is writeback. If WB/WT is sampled Low at RESET, all cache line fills are write-through. WB/WT has an internal weak pull-down. STPCLK is active Low and has an internal pull-up resistor. STPCLK is asynchronous, but it must meet setup and hold times t20 and t21 to ensure recognition in any specific clock. STPCLK must remain active until the Stop Clock special bus cycle is issued and the system returns either RDY or BRDY. TCK Test Clock (Input) Test Clock provides the clocking function for the JTAG boundary scan feature. TCK clocks state information and data into the component on the rising edge of TCK on TMS and TDI, respectively. Data is clocked out of the component on the falling edge of TCK on TDO. W/R Write/Read (Output) A High output indicates a write cycle. A Low output indicates a read cycle. TDI Test Data Input (Input) TDI is the serial input that shifts JTAG instructions and data into the tested component. TDI is sampled on the Note: The Am5X86 microprocessor does not use the VCC5 pin used by some 3-V, clock-tripled, 486-based processors. The corresponding pin on the Am5X86 microprocessor is an Internal No Connect (INC). Am5X86 Microprocessor 17 AMD PRELIMINARY 4.3.3 Protected Mode 4 4.1 FUNCTIONAL DESCRIPTION Overview Am5X86 microprocessors use a 32-bit architecture with on-chip memory management and cache memory units. The instruction set includes the complete 486 microprocessor instruction set along with extensions to serve the new extended applications. All software written for the 486 microprocessor and previous members of the x86 architectural family can run on the Am5X86 microprocessor without modification. The on-chip Memory Management Unit (MMU) is completely compatible with the 486 MMU. The MMU includes a segmentation unit and a paging unit. Segmentation allows management of the logical address space by providing easy data and code relocatibility and efficient sharing of global resources. The paging mechanism operates beneath segmentation and is transparent to the segmentation process. Paging is optional and can be disabled by system software. Each segment can be divided into one or more 4-Kbyte segments. To implement a virtual memory system, the Am5X86 microprocessor supports full restartability for all page and segment faults. Protected mode provides access to the sophisticated memory management paging and privilege capabilities of the processor. 4.3.4 System Management Mode SMM is a special operating mode described in detail in Section 7. 4.4 Cache Architecture The Am5X86 microprocessor family supports a superset architecture of the standard 486 cache implementation. This architectural enhancement improves not only CPU performance, but total system performance. 4.4.1 Write-Through Cache The standard 486DX-type write-through cache architecture is characterized by the following: s External read accesses are placed in the cache if they meet proper caching requirements. s Subsequent reads to the data in the cache are made if the address is stored in the cache tag array. s Write operations to a valid address in the cache are 4.2 Memory updated in the cache and to external memory. This data writing technique is called write-through. The write-through cache implementation forces all writes to flow through to the external bus and back to main memory. Consequently, the write-through cache generates a large amount of bus traffic on the external data bus. 4.4.2 Write-Back Cache Memory is organized into one or more variable length segments, each up to 4 Gbytes (232 bytes). A segment can have attributes associated with it, including its location, size, type (i.e., stack, code, or data), and protection characteristics. Each task on a microprocessor can have a maximum of 16,381 segments, each up to 4 Gbytes. Thus, each task has a maximum of 64 Tbytes of virtual memory. The segmentation unit provides four levels of protection for isolating and protecting applications and the operating system from each other. The hardware-enforced protection allows high-integrity system designs. The microprocessor write-back cache architecture is characterized by the following: s External read accesses are placed in the cache if they meet proper caching requirements. s Subsequent reads to the data in the cache are made 4.3 Modes of Operation if the address is stored in the cache tag array. s Write operations to a valid address in the cache that The Am5X86 microprocessor has four modes of operation: Real Address mode (Real mode), Virtual 8086 Address mode (Virtual mode), Protected Address mode (Protected mode), and System Management mode (SMM). 4.3.1 Real Mode is in the write-through (shared) state is updated in the cache and to external memory. s Write operations to a valid address in the cache that is in the write-back (exclusive or modified) state is updated only in the cache. External memory is not updated at the time of the cache update. s Modified data is written back to external memory In Real mode, the Am5X86 microprocessor operates as a fast 8086. Real mode is required primarily to set up the processor for Protected mode operation. 4.3.2 Virtual Mode In Virtual mode, the processor appears to be in Real mode, but can use the extended memory accessing of Protected mode. when the modified cache line is being replaced with a new cache line (copy-back operation) or an external bus master has snooped a modified cache line (write-back). The write-back cache feature significantly reduces the amount of bus traffic on the external bus; however, it also adds complexity to the system design to maintain memory coherency. The write-back cache requires en- 18 Am5X86 Microprocessor PRELIMINARY hanced system support because the cache may contain data that is not identical to data in main memory at the same address location. 4.5.2 Line Status and Line State AMD 4.5 Write-Back Cache Protocol The Am5X86 microprocessor family write-back cache coherency protocol reduces bus activity while maintaining data coherency in a multimaster environment. The cache coherency protocol offers the following advantages: s No unnecessary bus traffic. The protocol dynamical- A cache line can occupy one of four legal states as indicated by bits S0 and S1. The line states are shown in Table 4. Each line in the cache is in one of these states. The state transition is induced either by the processor or during snooping from an external bus master. Table 4. Legal Cache Line States S1 0 0 1 1 S0 0 1 0 1 Line State Invalid Exclusive Modified Shared ly identifies shared data to the granularity of a cache line. This dynamic identification ensures that the traffic on the external bus is the minimum necessary to ensure coherency. s Software-transparent. Because the protocol gives the appearance of a single, unified memory, software does not have to maintain coherency or identify shared data. Application software developed for a system without a cache can run without modification. Software support is required only in the operating system to identify non-cacheable data regions. The Am5X86 microprocessor family implements a modified MESI protocol on systems with write-back cache support. MESI allows a cache line to exist in four states: modified, exclusive, shared, and invalid. The Am5X86 microprocessor family allocates memory in the cache due to a read miss. Write allocation is not implemented. To maintain coherency between cache and main memory, the MESI protocol has the following characteristics: s The system memory is always updated during a 4.5.2.1 Invalid An invalid cache line does not contain valid data for any external memory location. An invalid line does not participate in the cache coherency protocol. 4.5.2.2 Exclusive An exclusive line contains valid data for some external memory location. The data exactly matches the data in the external memory location. 4.5.2.3 Shared A shared line contains valid data for an external memory location, the data is shared by another cache, and the shared data matches the data in the external memory exactly; or the cache line is in Write-through mode. snoop when a modified line is hit. s If a modified line is hit by another master during 4.5.2.4 Modified snooping, the master is forced off the bus and the snooped cache writes back the modified line to the system memory. After the snooped cache completes the write, the forced-off bus master restarts the access and reads the modified data from memory. 4.5.1 Cache Line Overview To implement the Am5X86 microprocessor cache coherency protocol, each tag entry is expanded to 2 bits: S1 and S0. Each tag entry is associated with a cache line. Table 3 shows the cache line organization. Table 3. Cache Line Organization Data Words (32 Bits) D0 D1 D2 D3 A modified line contains valid data for an external memory location. However, the data does not match the data in the external location because the processor has modified the data since it was loaded from the external memory. A cache that contains a modified line is responsible for ensuring that the data is properly maintained. This means that in the case of an external access to that line from another external bus master, the modified line is first written back to the external memory before the other external bus master can complete its access. Table 5 shows the MESI cache line states and the corresponding availability of data. Address Tag and Status Address Tag, S1, S0 Am5X86 Microprocessor 19 AMD Table 5. MESI Cache Line Status Situation Modified Exclusive Line valid? Yes External memory is... A write to this cache line... out-ofdate Yes valid Shared Yes valid PRELIMINARY pin indicates whether an individual write access is executed as write-through or write-back. The Am5X86 microprocessor does this on an access-by-access basis. Once the cache line is in the cache, the STATUS bit is tested each time the processor writes to the cache line or a tag compare results in a hit during Bus-watching mode. If the WB⁄WT signal is Low during the first BRDY of the cache line read access, the cache line is considered a write-through access. Therefore, all writes to this location in the cache are reflected on the external bus, even if the cache line is write protected. Invalid No status unknown goes to does not goes does not go the bus go to the directly to to the bus and bus the bus updates 4.6 Cache Replacement Description 4.8 The cache line replacement algorithm uses the standard Am486 CPU pseudo LRU (Least-Recently Used) strategy. When a line must be placed in the internal cache, the microprocessor first checks to see if there is an invalid line available in the set. If no invalid line is available, the LRU algorithm replaces the least-recently used cache line in the four-way set with the new cache line. If the cache line for replacement is modified, the modified cache line is placed into the copy-back buffer for copying back to external memory, and the new cache line is placed into the cache. This copy-back ensures that the external memory is updated with the modified data upon replacement. Cache Functionality in Write-Back Mode The description of cache functionality in Write-back mode is divided into two sections: processor-initiated cache functions and snooping actions. 4.8.1 Processor-Initiated Cache Functions and State Transitions 4.7 Memory Configuration In computer systems, memory regions require specific caching and memory write methods. For example, some memory regions are non-cacheable while others are cacheable but are write-through. To allow maximum memory configuration, the microprocessor supports specific memory region requirements. All bus masters, such as DMA controllers, must reflect all data transfers on the microprocessor local bus so that the microprocessor can respond appropriately. 4.7.1 Cacheability The microprocessor contains two new buffers for use with the MESI protocol support: the copy-back buffer and the write-back buffer. The processor uses the copyback buffer for cache line replacement of modified lines. The write-back buffer is used when an external bus master hits a modified line in the cache during a snoop operation and the cache line is designated for write-back to main memory. Each buffer is four doublewords in size. Figure 1 shows a diagram of the state transitions induced by the local processor. When a read miss occurs, the line selected for replacement remains in the modified state until overwritten. A copy of the modified line is sent to the copy-back buffer to be written back after replacement. When reload has successfully completed, the line is set either to the exclusive or the shared state, depending on the state of PWT and WB/WT signals. The Am5X86 CPU caches data based on the state of the CD and NW bits in CR0, in conjunction with the KEN signal, at the time of a burst read access from memory. If the WB/WT signal is Low during the first BRDY, KEN meets the standard setup and hold requirements and the four 32-bit doublewords are still placed in the cache. However, all cacheable accesses in this mode are considered write-through. When the WB/WT is High during the first BRDY, the entire four 32-bit doubleword transfer is considered write-back. Invalid Read_Miss (WB/WT = 1) • (PWT = 0) Read_Miss [(WB/WT = 0) + (PWT = 1)] Read_Hit Exclusive Shared Write_Hit + Read_Hit Note: The CD bit in CR0 enables (0) or disables (1) the internal cache. The NW bit in CR0 enables (0) or disables (1) write-through and snooping cycles. RESET sets CD and NW to 1. Unlike RESET, however, SRESET does not invalidate the cache nor does it modify the values of CD and NW in CR0. 4.7.2 Write-Through/Write-Back If the CPU is operating in Write-back mode (i.e., the WB⁄WT pin was sampled High at RESET), the WB⁄WT 20 Write_Hit Modified Read_Hit + Write_Hit Note: Write_Hit generates external bus cycle. Figure 1. Processor-Induced Line Transitions in Write-Back Mode Am5X86 Microprocessor PRELIMINARY If the PWT signal is 0, the external WB/WT signal determines the new state of the line. If the WB/WT signal was asserted to 1 during reload, the line transits to the exclusive state. If the WB/WT signal was 0, the line transits to the shared state. If the PWT signal is 1, it overrides the WB/WT signal, forcing the line into the shared state. Therefore, if paging is enabled, the software programmed PWT bit can override the hardware signal WB/WT. Until the line is reallocated, a write is the only processor action that can change the state of the line. If the write occurs to a line in the exclusive state, the data is simply written into the cache and the line state is changed to modified. The modified state indicates that the contents of the line require copy-back to the main memory before the line is reallocated. If the write occurs to a line in the shared state, the cache performs a write of the data on the external bus to update the external memory. The line remains in the shared state until it is replaced with a new cache line or until it is flushed. In the modified state, the processor continues to write the line without any further external actions or state transitions. If the PWT or PCD bits are changed for a specified memory location, the tag bits in the cache are assumed to be correct. To avoid memory inconsistencies with respect to cacheability and write status, a cache copyback and invalidation should be invoked either by using the WBINVD instruction or asserting the FLUSH signal. 4.8.2 Snooping Actions and State Transitions AMD ing the W/R signal. The microprocessor compares the address of the snoop request with addresses of lines in the cache and of any line in the copy-back buffer waiting to be transferred on the bus. It does not, however, compare with the address of write-miss data in the write buffers. Two clock cycles after sampling EADS, the microprocessor drives the results of the snoop on the HITM pin. If HITM is active, the line was found in the modified state; if inactive, the line was in the exclusive or shared state, or was not found. Figure 2 shows a diagram of the state transitions induced by snooping accesses. Invalid (EADS = 0 * INV = 1) + FLUSH = 0 (EADS = 0 * INV = 1) + FLUSH = 0 EADS = 0 * INV = 0 * FLUSH = 1 Exclusive EADS = 0 * INV = 0 * FLUSH = 1 (HITM asserted + write-back) Shared EADS = 0 * INV = 0 * FLUSH = 1 EADS = 0 * INV = 1 + FLUSH = 0 (HITM asserted + write-back) Modified To maintain cache coherency, the CPU must allow snooping by the current bus master. The bus master initiates a snoop cycle to check whether an address is cached in the internal cache of the microprocessor. A snoop cycle differs from any other cycle in that it is initiated externally to the microprocessor, and the signal for beginning the cycle is EADS instead of ADS. The address bus of the microprocessor is bidirectional to allow the address of the snoop to be driven by the system. A snoop access can begin during any hold state: s While HOLD and HLDA are asserted s While BOFF is asserted s While AHOLD is asserted Figure 2. Snooping State Transitions 4.8.2.1 Difference between Snooping Access Cases Snooping accesses are external accesses to the microprocessor. As described earlier, the snooping logic has a set of signals independent from the processor-related signals. Those signals are: s EADS s INV s HITM In the clock in which EADS is asserted, the microprocessor samples the INV input to qualify the type of inquiry. INV specifies whether the line (if found) must be invalidated (i.e., the MESI status changes to Invalid or I). A line is invalidated if the snoop access was generated due to a write of another bus master. This is indicated by INV set to 1. In the case of a read, the line does not have to be invalidated, which is indicated by INV set to 0. The core system logic can generate EADS by watching the ADS from the current bus master, and INV by watch- In addition to these signals, the address bus is required as an input. This is achieved by setting AHOLD, HOLD, or BOFF active. Snooping can occur in parallel with a processor-initiated access that has already been started. The two accesses depend on each other only when a modified line is written back. In this case, the snoop requires the use of the cycle control signals and the data bus. The following sections describe the scenarios for the HOLD, AHOLD, and BOFF implementations. Am5X86 Microprocessor 21 AMD PRELIMINARY . 4.8.2.2 HOLD Bus Arbitration Implementation Table 6. Key to Switching Waveforms Waveform Inputs Must be steady May change from H to L May change from L to H Don’t care; any change permitted Does not apply Outputs Will be steady Will change from H to L Will change from L to H Changing; state unknown Center line is High-impedance “Off” state The HOLD/HLDA bus arbitration scheme is used primarily in systems where all memory transfers are seen by the microprocessor. The HOLD/HLDA bus arbitration scheme permits simple write-back cache design while maintaining a relatively high performing system. Figure 3 shows a typical system block diagram for HOLD/HLDA bus arbitration. Note: To maintain proper system timing, the HOLD signal must remain active for one clock cycle after HITM transitions active. Deassertion of HOLD in the same clock cycle as HITM assertion may lead to unpredictable processor behavior. Local Bus Peripheral CPU Address Bus Data Bus 4.8.2.2.2 External Read I/O Bus Interface Address Bus Scenario: The data resides in external memory (see Figure 4). Step 1 The processor starts the external read access by asserting ADS = 0 and W/R = 0. Step 2 WB/WT is sampled in the same cycle as BRDY. If WB/WT = 1, the data resides in a write-back cacheable memory location. Step 3 The processor completes its burst read and asserts BLAST. L2 Cache DRAM Data Bus Slow Peripheral Figure 3. Typical System Block Diagram for HOLD/HLDA Bus Arbitration 4.8.2.2.3 External Write Scenario: The data is written to the external memory (see Figure 5). Step 1 The processor starts the external write access by asserting ADS = 0 and W/R = 1. Step 2 The processor completes its write to the core system logic. 4.8.2.2.1 Processor-Induced Bus Cycles In the following scenarios, read accesses are assumed to be cache line fills. The cases also assume that the core system logic does not return BRDY or RDY until HITM is sampled. The addition of wait states follows the standard 486 bus protocol. For demonstration purposes, only the zero wait state approach is shown. Table 6 explains the key to switching waveforms. 4.8.2.2.4 HOLD/HLDA External Access TIming In systems with two or more bus masters, each bus master is equipped with individual HOLD and HLDA control signals. These signals are then centralized to the core system logic that controls individual bus masters, depending on bus request signals and the HITM signal. 22 Am5X86 Microprocessor PRELIMINARY AMD CLK ADR M/IO W/R 1 n n+4 n+8 n+12 ADS BLAST 3 BRDY Data KEN WB/WT BOFF n 2 n+4 n+8 n+12 Note: The circled numbers in this figure represent the steps in section 4.8.2.2.2. Figure 4. External Read CLK ADR M/IO W/R ADS BLAST 2 BRDY Data WB/WT BOFF n 1 n Note: The circled numbers in this figure represent the steps in section 4.8.2.2.3. Figure 5. External Write Am5X86 Microprocessor 23 AMD PRELIMINARY CLK ADR valid INV valid  EADS Ž HITM HOLD Œ HLDA Note: The circled numbers in this figure represent the steps in section 4.8.3.1. Figure 6. Snoop of On-Chip Cache That Does Not Hit a Line CLK ADR INV EADS valid valid  Ž HITM HOLD Œ HLDA Note: The circled numbers in this figure represent the steps in section 4.8.3.2. Figure 7. Snoop of On-Chip Cache That Hits a Non-modified Line 24 Am5X86 Microprocessor PRELIMINARY 4.8.3 External Bus Master Snooping Actions The following scenarios describe the snooping actions of an external bus master. AMD sertion. In the fastest case, this means that HOLD was asserted one clock cycle before the HLDA response. 4.8.3.1 Snoop Miss Scenario: A snoop of the on-chip cache does not hit a line, as shown in Figure 6. Step 1 The microprocessor is placed in Snooping mode with HOLD. HLDA must be High for a minimum of one clock cycle before EADS assertion. In the fastest case, this means that HOLD was asserted one clock cycle before the HLDA response. Step 2 EADS and INV are applied to the microprocessor. If INV is 0, a read access caused the snooping cycle. If INV is 1, a write access caused the snooping cycle. Step 3 Two clock cycles after EADS is asserted, HITM becomes valid. Because the addressed line is not in the snooping cache, HITM is 1. Step 2 EADS and INV are applied to the microprocessor. If INV is 0, a read access caused the snooping cycle. If INV is 1, a write access caused the snooping cycle. Step 3 Two clock cycles after EADS is asserted, HITM becomes valid. In this case, HITM is 1. 4.8.4 Write-Back Case Scenario: Write-back accesses are always burst writes with a length of four 32-bit words. For burst writes, the burst always starts with the microprocessor line offset at 0. HOLD must be deasserted before the write-back can be performed (see Figure 8). Step 1 HOLD places the microprocessor in Snooping mode. HLDA must be High for a minimum of one clock cycle before EADS assertion. In the fastest case, this means that HOLD asserts one clock cycle before the HLDA response. Step 2 EADS and INV are asserted. If INV is 0, snooping is caused by a read access. If INV is 1, snooping is caused by a write access. EADS is not sampled again until after the modified line is written back to memory. It is detected again as early as in Step 11. 4.8.3.2 Snoop Hit to a Non-Modified Line Scenario: The snoop of the on-chip cache hits a line, and the line is not modified (see Figure 7). Step 1 The microprocessor is placed in Snooping mode with HOLD. HLDA must be High for a minimum of one clock cycle before EADS asCLK ADR M/IO CACHE W/R ADS BLAST n floating/tri-stated 5 n n+4 n+8 n+1 floating/tri-stated n 11 BRDY INV EADS valid 2 3 6 valid HITM 7 8 HOLD HLDA Data External bus master’s BOFF signal 1 4 9 n n+4 n+8 n+12 10 Note: The circled numbers in this figure represent the steps in section 4.8.4. Figure 8. Snoop That Hits a Modified Line (Write-Back) Am5X86 Microprocessor 25 AMD Step 3 Two clock cycles after EADS is asserted, HITM becomes valid, and is 0 because the line is modified. Step 4 In the next clock, the core system logic deasserts the HOLD signal in response to the HITM = 0 signal. The core system logic backs off the current bus master at the same time so that the microprocessor can access the bus. HOLD can be reasserted immediately after ADS is asserted for burst cycles. Step 5 The snooping cache starts it’s write-back of the modified line by asserting ADS = 0, CACHE = 0, and W/R = 1. The write access is a burst write. The number of clock cycles between deasserting HOLD to the snooping cache and first asserting ADS for the write-back cycles can vary. In this example, it is one clock cycle, which is the shortest possible time. Regardless of the number of clock cycles, the start of the writeback is seen by ADS going Low. Step 6 The write-back access is finished when BLAST and BRDY both are 0. Step 7 In the clock cycle after the final write-back access, the processor drives HITM back to 1. Step 8 HOLD is sampled by the microprocessor. Step 9 One cycle after sampling HOLD High, the microprocessor transitions HLDA transitions to 1, acknowledging the HOLD request. Step 10 The core system logic removes hold-off control to the external bus master. This allows the external bus master to immediately retry the aborted access. ADS is strobed Low, which generates EADS Low in the same clock cycle. Step 11 The bus master restarts the aborted access. EADS and INV are applied to the microprocessor as before. This starts another snoop cycle. The status of the addressed line is now either shared (INV = 0) or is changed to invalid (INV = 1). 4.8.5 Write-Back and Pending Access Scenario: The following occurs when, in addition to the write-back operation, other bus accesses initiated by the processor associated with the snooped cache are pending. The microprocessor gives the write-back access priority. This implies that if HOLD is deasserted, the microprocessor first writes back the modified line (see Figure 9). CLK ADR M/IO CACHE W/R ADS BLAST n floating/tri-stated n n+4 n+8 n+12 n 5 11 BRDY INV EADS valid 2 3 6 valid HITM 7 8 HOLD HLDA Data External bus master’s BOFF signal 1 4 9 n n+4 n+8 n+12 10 Note: The circled numbers in this figure represent the steps in section 4.8.5. Figure 9. Write-Back and Pending Access 26 Am5X86 Microprocessor PRELIMINARY Step 1 HOLD places the microprocessor in Snooping mode. HLDA must be High for a minimum of one clock cycle before EADS assertion. In the fastest case, this means that HOLD asserts one clock cycle before the HLDA response. Step 2 EADS and INV are asserted. If INV is 0, snooping is caused by a read access. If INV is 1, snooping is caused by a write access. EADS is not sampled again until after the modified line is written back to memory. It is detected again as early as in Step 11. Step 3 Two clock cycles after EADS is asserted, HITM becomes valid, and is 0 because the line is modified. Step 4 In the next clock the core system logic deasserts the HOLD signal in response to the HITM = 0. The core system logic backs off the current bus master at the same time so that the microprocessor can access the bus. HOLD can be reasserted immediately after ADS is asserted for burst cycles. Step 5 The snooping cache starts its write-back of the modified line by asserting ADS = 0, CACHE = 0, and W/R = 1. The write access is a burst write. The number of clock cycles between deasserting HOLD to the snooping cache and first asserting ADS for the write-back cycles can vary. In this example, it is one clock cycle, which is the shortest possible time. Regardless of the AMD number of clock cycles, the start of the writeback is seen by ADS going Low. Step 6 The write-back access is finished when BLAST and BRDY both are 0. Step 7 In the clock cycle after the final write-back access, the processor drives HITM back to 1. Step 8 HOLD is sampled by the microprocessor. Step 9 A minimum of 1 clock cycle after the completion of the pending access, HLDA transitions to 1, acknowledging the HOLD request. Step 10 The core system logic removes hold-off control to the external bus master. This allows the external bus master to immediately retry the aborted access. ADS is strobed Low, which generates EADS Low in the same clock cycle. Step 11 The bus master restarts the aborted access. EADS and INV are applied to the microprocessor as before. This starts another snoop cycle. The status of the addressed line is now either shared (INV = 0) or is changed to invalid (INV = 1). 4.8.5.1 HOLD/HLDA Write-Back Design Considerations When designing a write-back cache system that uses HOLD/HLDA as the bus arbitration method, the following considerations must be observed to ensure proper operation (see Figure 10). CLK ADS BLAST BRDY HOLD HITM HLDA Valid Hold Assertion Figure 10. Valid HOLD Assertion During Write-Back Am5X86 Microprocessor 27 AMD PRELIMINARY The following sections describe the snooping scenarios for the AHOLD implementation. Step 1 During a snoop to the on-chip cache that hits a modified cache line, the HOLD signal cannot be deasserted to the microprocessor until the next clock cycle after HITM transitions active. Step 2 After the write-back has commenced, the HOLD signal should be asserted no earlier than the next clock cycle after ADS goes active, and no later than in the final BRDY of the last write. Asserting HOLD later than the final BRDY may allow the microprocessor to permit a pending access to begin. Step 3 If RDY is returned instead of BRDY during a write-back, the HOLD signal can be reasserted at any time starting one clock after ADS goes active in the first transfer up to the final transfer when RDY is asserted. Asserting RDY instead of BRDY will not break the write-back cycle if HOLD is asserted. The processor ignores HOLD until the final write cycle of the write-back. 4.8.5.3 Normal Write-Back Scenario: This scenario assumes that a processor-initiated access has already started and that the external logic can finish that access even without the address being applied after the first clock cycle. Therefore, a snooping access with AHOLD can be done in parallel. In this case, the processor-initiated access is finished first, then the write-back is executed (see Figure 12). The sequence is as follows: Step 1 The processor initiates an external, simple, non-cacheable read access, strobing ADS = 0 and W/R = 0. The address is driven from the CPU. Step 2 In the same cycle, AHOLD is asserted to indicate the start of snooping. The address bus floats and becomes an input in the next clock cycle. Step 3 During the next clock cycles, the BRDY or RDY signal is not strobed Low. Therefore, the processor-initiated access is not finished. Step 4 Two clock cycles after AHOLD is asserted, the EADS signal is activated to start an actual snooping cycle, and INV is valid. If INV is 0, a read access caused the snooping cycle. If INV is 1, a write access caused the snooping cycle. Additional EADS are ignored due to the hit of a modified line. It is detected after HITM goes inactive. Step 5 Two clock cycles after EADS is asserted, the snooping signal HITM becomes valid. The line is modified; therefore, HITM is 0. Step 6 In this cycle, the processor-initiated access is finished. Step 7 Two clock cycles after the end of the processorinitiated access, the cache immediately starts writing back the modified line. This is indicated by ADS = 0 and W/R = 1. Note that AHOLD is still active and the address bus is still an input. However, the write-back access can be executed without any address. This is because the corresponding address must have been on the bus when EADS was strobed. Therefore, in the case of the core system logic, the address for the write-back must be latched with EADS to be available later. This is required only if AHOLD is not removed if HITM becomes 0. Otherwise, the address of the write-back is put onto the address bus by the microprocessor. 4.8.5.2 AHOLD Bus Arbitration Implementation The use of AHOLD as the control mechanism is often found in systems where an external second-level cache is closely coupled to the microprocessor. This tight coupling allows the microprocessor to operate with the least amount of stalling from external snooping of the on-chip cache. Additionally, snooping of the cache can be performed concurrently with an access by the microprocessor. This feature further improves the performance of the total system (see Figure 11). Note: To maintain proper system timing, the AHOLD signal must remain active for one clock cycle after HITM transitions active. Deassertion of AHOLD in the same clock cycle as HITM assertion may lead to unpredictable processor behavior. CPU Address Bus Data Bus L2 Cache Address Bus Data Bus DRAM I/O Bus Interface Address Bus Data Bus Slow Peripheral Figure 11. Closely Coupled Cache Block Diagram 28 Am5X86 Microprocessor PRELIMINARY CLK ADR M/IO CACHE W/R ADS BLAST 3 BRDY AHOLD 2 INV EADS HITM Data Read Wn W n+4 W n+8 W n+C from CPU to CPU AMD from CPU 1 7 9 6 8 4 10 5 Note: The circled numbers in this figure represent the steps in section 4.8.5.3. Figure 12. Snoop Hit Cycle with Write-Back Step 8 As an example, AHOLD is now removed. In the next clock cycle, the current address of the write-back access is driven onto the address bus. Step 9 The write-back access is finished when BLAST and BRDY both transition to 0. Step 10 In the clock cycle after the final write-back access, the snooping cache drives HITM back to 1. The status of the snooped and written-back line is now either shared (INV = 0) or is changed to invalid (INV = 1). 4.8.6 Reordering of Write-Backs (AHOLD) with BOFF Scenario: If there are outstanding processor-initiated cycles on the bus, asserting BOFF clears the bus pipeline. If a snoop causes HITM to be asserted, the first cycle issued by the microprocessor after deassertion of BOFF is the write-back cycle. After the write-back cycle, it reissues the aborted cycles. This translates into the following sequence: Step 1 The processor starts a cacheable burst read cycle. Step 2 One clock cycle later, AHOLD is asserted. This switches the address bus into an input one clock cycle after AHOLD is asserted. Step 3 Two clock cycles after AHOLD is asserted, the EADS and INV signals are asserted to start the snooping cycle. Step 4 Two clock cycles after EADS is asserted, HITM becomes valid. The line is modified, therefore HITM = 0. Step 5 Note that the processor-initiated access is not completed because BLAST = 1. Step 6 With HITM going Low, the core system logic asserts BOFF in the next clock cycle to the snooping processor to reorder the access. BOFF overrides BRDY. Therefore, the partial read is not used. It is reread later. As seen previously, the Bus Interface Unit (BIU) completes the processor-initiated access first if the snooping access occurs after the start of the processor-initiated access. If the HITM signal occurs one clock cycle before the ADS = 0 of the processor-initiated access, the writeback receives priority and is executed first. However, if the snooping access is executed after the start of the processor-initiated access, there is a methodology to reorder the access order. The BOFF signal delays outstanding processor-initiated cycles so that a snoop write-back can occur immediately (see Figure 13). Am5X86 Microprocessor 29 AMD CLK ADR M/IO CACHE W/R ADS BLAST R1 from CPU W1 to CPU PRELIMINARY don’t care W1 from CPU W2 W3 W4 R2 from CPU Œ  ” 11 • ’“ BRDY BOFF AHOLD INV EADS HITM Data R1 R2 W1 W2 W3 W4  ‘ Ž  12 Note: The circled numbers in this figure represent the steps in section 4.8.6. Figure 13. Cycle Reordering with BOFF (Write-Back) Step 7 One clock cycle later BOFF is deasserted. The write-back access starts one clock cycle later because the BOFF has cleared the bus pipeline. Step 8 AHOLD is deasserted. In the next clock cycle the address for the write-back is driven on the address bus. Step 9 One cycle after BOFF is deasserted, the cache immediately starts writing back the modified line. This is indicated by ADS = 0 and W/R = 1. Step 10 The write-back access is finished when BLAST and BRDY go active 0. Step 11 The BIU restarts the aborted cache line fill with the previous read. This is indicated by ADS = 0 and W/R = 0. Step 12 In the same clock cycle, the snooping cache drives HITM back to 1. Step 13 The previous read is now reread. 4.8.7 Special Scenarios for AHOLD Snooping 4.8.7.1 Write Cycle Reordering due to Buffering Scenario: The MESI cache protocol and the ability to perform and respond to snoop cycles guarantee that writes to the cache are logically equivalent to writes to memory. In particular, the order of read and write operations on cached data is the same as if the operations were on data in memory. Even non-cached memory read and write requests usually occur on the external bus in the same order that they were issued in the program. For example, when a write miss is followed by a read miss, the write data goes on the bus before the read request is put on the bus. However, the posting of writes in write buffers coupled with snooping cycles may cause the order of writes seen on the external bus to differ from the order they appear in the program. Consider the following example, which is illustrated in Figure 14. For simplicity, snooping signals that behave in their usual manner are not shown. Step 1 AHOLD is asserted. No further processor-initiated accesses to the external bus can be started. No other access is in progress. Step 2 The processor writes data A to the cache, resulting in a write miss. Therefore, the data is put into the write buffers, assuming they are not full. No external access can be started because AHOLD is still 1. In addition to the previously described scenarios, there are special scenarios regarding the time of the EADS and AHOLD assertion. The final result depends on the time EADS and AHOLD are asserted relative to other processor-initiated operations. 30 Am5X86 Microprocessor PRELIMINARY AMD CLK 2 Write Buffer Cached Data AHOLD EADS 5 HITM ADS BLAST 6 10 XXX B original A B modified 3 7 1 4 Ignored 9 11 BRDY Data B B+4 B+8 8 B+12 A Note: The circled numbers in this figure represent the steps in section 4.8.7.1. Figure 14. Write Cycle Reordering Due to Buffering Step 3 The next write of the processor hits the cache and the line is non-shared. Therefore, data B is written into the cache. The cache line transits to the modified state. Step 4 In the same clock cycle, a snoop request to the same address where data B resides is started because EADS = 0. The snoop hits a modified line. EADS is ignored due to the hit of a modified line, but is detected again as early as in step 10. Step 5 Two clock cycles after EADS asserts, HITM becomes valid. Step 6 Because the processor-initiated access cannot be finished (AHOLD is still 1), the BIU gives priority to a write-back access that does not require the use of the address bus. Therefore, in the clock cycle, the cache starts the write-back sequence indicated by ADS = 0 and W/R = 0. Step 7 During the write-back sequence, AHOLD is deasserted. Step 8 The write-back access is finished when BLAST and BRDY transition to 0. Step 9 After the last write-back access, the BIU starts writing data A from the write buffers. This is indicated by ADS = 0 and W/R = 0. Step 10 In the same clock cycle, the snooping cache drives HITM back to 1. Step 11 The write of data A is finished if BRDY transitions to 0 (BLAST = 0), because it is a single word. The software write sequence was first data A and then data B. But on the external bus the data appear first as data B and then data A. The order of writes is changed. In most cases, it is unnecessary to strictly maintain the ordering of writes. However, some cases (for example, writing to hardware control registers) require writes to be observed externally in the same order as programmed. There are two options to ensure serialization of writes, both of which drive the cache to Write-through mode: 1. Set the PWT bit in the page table entries. 2. Drive the WB/WT signal Low when accessing these memory locations. Option 1 is an operating-system-level solution not directly implemented by user-level code. Option 2, the hardware solution, is implemented at the system level. Am5X86 Microprocessor 31 AMD PRELIMINARY 4.8.10 Snooping Characteristics During a Cache Line Fill The microprocessor takes responsibility for responding to snoop cycles for a cache line only during the time that the line is actually in the cache or in a copy-back buffer. There are times during the cache line fill cycle and during the cache replacement cycle when the line is “in transit” and snooping responsibility must be taken by other system components. The following cases apply if snooping is invoked via AHOLD, and neither HOLD nor BOFF is asserted. s System designers should consider the possibility 4.8.7.2 BOFF Write-Back Arbitration Implementation The use of BOFF to perform snooping of the on-chip cache is used in systems where more than one cacheable bus master resides on the microprocessor bus. The BOFF signal forces the microprocessor to relinquish the bus in the following clock cycle, regardless of the type of bus cycle it was performing at the time. Consequently, the use of BOFF as a bus arbitrator should be implemented with care to avoid system problems. 4.8.8 BOFF Design Considerations The use of BOFF as a bus arbitration control mechanism is immediate. BOFF forces the microprocessor to abort an access in the following clock cycle after it is asserted. The following design issues must be considered. 4.8.8.1 Cache Line Fills The microprocessor aborts a cache line fill during a burst read if BOFF is asserted during the access. Upon regaining the bus, the read access commences where it left off when BOFF was recognized. External buffers should take this cycle continuation into consideration if BOFF is allowed to abort burst read cycles. that a snooping cycle may arrive at the same time as a cache line fill or replacement for the same address. If a snooping cycle arrives at the same time as a cache line fill with the same address, the CPU uses the cache line fill, but does not place it in the cache. s If a snooping cycle occurs at the same time as a 4.8.8.2 Cache Line Copy-Backs Similar to the burst read, the burst write also can be aborted at any time with the BOFF signal. Upon regaining access to the bus, the write continues from where it was aborted. External buffers and control logic should take into consideration the necessary control, if any, for burst write continuations. cache line fill with a different address, the cache line fill is placed into the cache unless EADS is recognized before the first BRDY but after ADS is asserted, or EADS is recognized on the last BRDY of the cache line fill. In these cases, the line is not placed into the cache. 4.8.11 Snooping Characteristics During a Copy-Back If a copy-back is occurring because of a cache line replacement, the address being replaced can be matched by a snoop until assertion of the last BRDY of the copyback. This is when the modified line resides in the copyback buffer. An EADS as late as two clocks before the last BRDY can cause HITM to be asserted. Figure 15 illustrates the microprocessor relinquishing responsibility of recognizing snoops for a line that is copied back. It shows the latest EADS assertion that can cause HITM assertion. HITM remains active for only one clock period in that example. HITM remains active through the last BRDY of the corresponding write-back; in that case, the write-back has already completed. This is the latest point where snooping can start, because two clock cycles later, the final BRDY of the write-back is applied. If a snoop cycle hits the copy-back address after the first BRDY of the copy-back and ADS has been issued, the microprocessor asserts HITM. Keep in mind that the write-back was initiated due to a read miss and not due to a snoop to a modified line. In the second case, no snooping is recognized if a modified line is detected. 4.8.8.3 Locked Accesses Locked bus cycles occur in various forms. Locked accesses occur during read-modify-write operations, interrupt acknowledges, and page table updates. Although asserting BOFF during a locked cycle is permitted, extreme care should be taken to ensure data coherency for semaphore updates and proper data ordering. 4.8.9 BOFF During Write-Back If BOFF is asserted during a write-back, the processor performing the write-back goes off the bus in the next clock cycle. If BOFF is released, the processor restarts that write-back access from the point at which it was aborted. The behavior is identical to the normal BOFF case that includes the abort and restart behavior. 32 Am5X86 Microprocessor PRELIMINARY AMD CLK ADR AHOLD EADS CACHE HITM ADS BRDY BLAST n S Address B Figure 15. Latest Snooping of Copy-Back 4.9 Cache Invalidation and Flushing in Write-Back Mode The Am5X86 microprocessor family supports cache invalidation and flushing, much like the Am486 microprocessor Write-through mode. However, the addition of the write-back cache adds some complexity. 4.9.1 Cache Invalidation through Software erated when the on-chip cache is invalidated. Table 7 specifies the special bus cycle states for the instructions WBINVD and INVD. Table 7. WBINVD/INVD Special Bus Cycles A32–A2 0000 0000 h 0000 0000 h M/IO D/C W/R BE3 BE2 BE1 BE0 Bus Cycle 0 0 0 0 1 1 0 1 1 1 1 0 1 1 Write-back1 Flush1, 2 To invalidate the on-chip cache, the Am5X86 microprocessor family uses the same instructions as the Am486 microprocessor family. The two invalidation instructions, INVD and WBINVD, while similar, are slightly different for use in the write-back environment. The WBINVD instruction first performs a write-back of the modified data in the cache to external memory. Then it invalidates the cache, followed by two special bus cycles. The INVD instruction only invalidates the cache, regardless of whether modified data exists, and follows with a special bus cycle. The utmost care should be taken when executing the INVD instruction to ensure memory coherency. Otherwise, modified data may be invalidated prior to writing back to main memory. In Write-back mode, WBINVD requires a minimum of 4100 internal clocks to search the cache for modified data. Writing back modified data adds to this minimum time. WBINVD can only be stopped by a RESET. Two special bus cycles follow the write-back of modified data upon execution of the WBINVD instruction: first the write-back, and then the flush special bus cycle. The INVD operates identically to the standard 486 microprocessor family in that the flush special bus cycle is gen- Notes: 1. WBINVD generates first write-back, then flush. 2. INVD generates only flush. 4.9.2 Cache Invalidation through Hardware The other mechanism for cache invalidation is the FLUSH pin. The FLUSH pin operates similarly to the WBINVD command, writing back modified cache lines to main memory. After the entire cache has copied back all the modified data, the microprocessor generates two special bus cycles. These special bus cycles signal to the external caches that the microprocessor on-chip cache has completed its copy-back and that the second level cache may begin its copy-back to memory, if so required. Two flush acknowledge cycles are generated after the FLUSH pin is asserted and the modified data in the cache is written back. As with the WBINVD instruction, in Write-back mode, a flush requires a minimum of 4100 internal clocks to test the cache for modified data. Writing back modified data adds to this minimum time. The flush operation can only be stopped by a RESET. Table 8 shows the special flush bus cycle configuration. Am5X86 Microprocessor 33 AMD Table 8. FLUSH Special Bus Cycles A32–A2 0000 0001h PRELIMINARY line write-backs and copy-backs. Standard write operations are still supported. Burst writes are always four 32-bit words and start at the beginning of a cache line address of 0 for the starting access. The timing of the BLAST and BRDY signals is identical to the burst read. Figure 16 shows a burst write access. (See Figure 17 and Figure 18 for burst read and burst write access with BOFF asserted.) In addition to using BLAST, the CACHE signal indicates burstable cycles. CACHE is a cycle definition pin used when in Write-back mode (CACHE floats in Write-through mode). For processor-initiated cycles, the signal indicates: s s M/IO D/C W/R BE3 BE2 BE1 BE0 Bus Cycle 0 0 0 0 1 1 0 1 1 1 1 0 1 1 First Flush Acknowledge Second Flush Acknowledge 0000 0001h 4.9.3 Snooping During Cache Flushing As with snooping during normal operation, snooping is permitted during a cache flush, whether initiated by the FLUSH pin or WBINVD instruction. After completion of the snoop, and write-back, if needed, the microprocessor completes the copy-back of modified cache lines. For a read cycle, the internal cacheability of the cycle For a write cycle, a burst write-back or copy-back, if KEN is asserted (for linefills). 4.10 Burst Write The Am5X86 microprocessor improves system performance by implementing a burst write feature for cache CLK ADR M/IO W/R CACHE ADS BLAST BRDY Data XX0 XX4 XX0 XX4 XX8 XXC XX8 XXC Figure 16. Burst Write CLK ADR M/IO W/R CACHE ADS BLAST XX0 XX4 XX4 XX8 XXC BRDY BOFF Data to CPU XX0 XX4 don’t care XX4 XX8 XXC Figure 17. Burst Read with BOFF Assertion 34 Am5X86 Microprocessor PRELIMINARY CLK ADR M/IO W/R CACHE ADS BLAST BRDY BOFF Data from CPU XX0 XX4 XX4 XX8 XXC XX0 XX4 XX4 XX8 XXC AMD Figure 18. Burst Write with BOFF Assertion CACHE is asserted for cacheable reads, cacheable code fetches, and write-backs/copy-backs. CACHE is deasserted for non-cacheable reads, translation lookaside buffer (TLB) replacements, locked cycles (except for write-back cycles generated by an external snoop operation that interrupts a locked read/modify/write sequence), I/O cycles, special cycles, and write-throughs. CACHE is driven to its valid level in the same clock as the assertion of ADS and remains valid until the next RDY or BRDY assertion. The CACHE output pin floats one clock after BOFF is asserted. Additionally, the signal floats when HLDA is asserted. The following steps describe the burst write sequence: 1. The access is started by asserting: ADS = 0, M/IO = 1, W/R = 1, CACHE = 0. The address offset always is 0, so the burst write always starts on a cache line boundary. CACHE transitions High (inactive) after the first BRDY. 2. In the second clock cycle, BLAST is 1 to indicate that the burst is not finished. 3. The burst write access is finished when BLAST is 0 and BRDY is 0. When the RDY signal is returned instead of the BRDY signal, the Am5X86 microprocessor halts the burst cycle and proceeds with the standard non-burst cycle. 4.10.1 Locked Accesses Locked accesses of an Am5X86 microprocessor occur for read-modify-write operations and interrupt acknowledge cycles. The timing is identical to the DX microprocessor, although the state transitions differ from the standard DX microprocessor. Unlike processor-initiated accesses, state transitions for locked accesses are seen by all processors in the system. Any locked read or write generates an external bus cycle, regardless of cache hit or miss. During locked cycles, the processor does not recognize a HOLD request, but it does recognize BOFF and AHOLD requests. Locked read operations always read data from the external memory, regardless of whether the data is in the cache. In the event that the data is in the cache and unmodified, the cache line is invalidated and an external read operation is performed. The data from the external memory is used instead of the data in the cache, thus ensuring that the locked read is seen by all other bus masters. If a locked read occurs, the data is in the cache, and it is modified. The microprocessor first copies back the data to external memory, invalidates the cache line, and then performs a read operation to the same location, thus ensuring that the locked read is seen by all other bus masters. At no time is the data in the cache used directly by the microprocessor or a locked read operation before reading the data from external memory. Since locked cycles always begin with a locked read access, and locked read cycles always invalidate a cache line, a locked write cycle to a valid cache line, either modified or unmodified, does not occur. 4.10.2 Serialization Locked accesses are totally serialized: s All reads and writes in the write buffer that precede the locked access are issued on the bus before the first locked access is executed. No read or write after the last locked access is issued internally or on the bus until the final RDY or BRDY for all locked accesses. It is possible to get a locked read, write-back, locked write cycle. s s Am5X86 Microprocessor 35 AMD PRELIMINARY In Write-back mode, the priority order of external interrupts is: 1. 2. 3. 4. 5. 6. 7. RESET FLUSH SRESET SMI NMI INTR STPCLK 4.10.3 PLOCK Operation in Write-Through Mode As described in Section 3, PLOCK is only used in Writethrough mode; the signal is driven inactive in Write-back mode. In Write-through mode, the processor drives PLOCK Low to indicate that the current bus transaction requires more than one bus cycle. The CPU continues to drive the signal Low until the transaction is completed, whether or not RDY or BRDY is returned. Refer to the pin description for additional information. 5 5.1 CLOCK CONTROL Clock Generation The Am5X86 CPU is driven by a 1x clock that relies on phased-lock loop (PLL) to generate the two internal clock phases: phase one and phase two. The rising edge of CLK corresponds to the start of phase one (ph1). All external timing parameters are specified relative to the rising edge of CLK. 5.2 Stop Clock The Am5X86 CPU also provides an interrupt mechanism, STPCLK, that allows system hardware to control the power consumption of the CPU by stopping the internal clock to the CPU core in a sequenced manner. The first low-power state is called the Stop Grant state. If the CLK input is completely stopped, the CPU enters into the Stop Clock state (the lowest power state). When the CPU recognizes a STPCLK interrupt, the processor: s Stops execution on the next instruction boundary STPCLK is active Low and has an internal pull-up resistor. STPCLK is asynchronous, but setup and hold times must be met to ensure recognition in any specific clock. STPCLK must remain active until the Stop Grant special bus cycle is asserted and the system responds with either RDY or BRDY. When the CPU enters the Stop Grant state, the internal pull-up resistor is disabled, reducing the CPU power consumption. The STPCLK input must be driven High (not floated) to exit the Stop Grant state. STPCLK must be deasserted for a minimum of five clocks after RDY or BRDY is returned active for the Stop Grant bus cycle before being asserted again. There are two regions for the Low-power mode supply current: 1. Low Power: Stop Grant state (fast wake-up, frequencyand voltage-dependent) 2. Lowest Power: Stop Clock state (slow wake-up, voltage-dependent) (unless superseded by a higher priority interrupt) s Waits for completion of cache flush s Stops the pre-fetch unit s Empties all internal pipelines and write buffers s Generates a Stop Grant bus cycle s Stops the internal clock 5.3 Stop Grant Bus Cycle The processor drives a special Stop Grant bus cycle to the bus after recognizing the STPCLK interrupt. This bus cycle is the same as the HALT cycle used by a standard Am486 microprocessor, with the exception that the Stop Grant bus cycle drives the value 0000 0010h on the address pins. s M/lO = 0 s D/C = 0 s W/R =1 s Address Bus = 0000 0010h (A4 = 1) s BE3–BE0 = 1011 s Data bus = undefined At this point the CPU is in the Stop Grant state. The CPU cannot respond to a STPCLK request from an HLDA state because it cannot empty the write buffers and, therefore, cannot generate a Stop Grant cycle. The rising edge of STPCLK signals the CPU to return to program execution at the instruction following the interrupted instruction. Unlike the normal interrupts (INTR and NMI), STPCLK does not initiate interrupt acknowledge cycles or interrupt table reads. 5.2.1 External Interrupts in Order of Priority In Write-through mode, the priority order of external interrupts is: 1. 2. 3. 4. 5. 6. 36 RESET/SRESET FLUSH SMI NMI INTR STPCLK The system hardware must acknowledge this cycle by returning RDY or BRDY, or the processor will not enter the Stop Grant state (see Figure 19). The latency between a STPCLK request and the Stop Grant bus cycle depends on the current instruction, the amount of data in the CPU write buffers, and the system memory performance. Am5X86 Microprocessor PRELIMINARY . AMD CLK STPCLK t20 ADDR t21 Stop Grant Bus cycle RDY Figure 19. Entering Stop Grant State 5.4 Pin State During Stop Grant Table 9 shows the pin states during Stop Grant Bus states. During the Stop Grant state, most output and input/output signals of the microprocessor maintain the level they held when entering the Stop Grant state. The data and data parity signals are tri-stated. In response to HOLD being driven active during the Stop Grant state (when the CLK input is running), the CPU generates HLDA and tri-states all output and input/output signals that are tri-stated during the HOLD/HLDA state. After HOLD is deasserted, all signals return to the same state they were before the HOLD/HLDA sequence. Table 9. Pin State During Stop Grant Bus State Signal A3–A2 A31–A4 D31–D0 BE3–BE0 DP3–DP0 W/R, D/C, M/IO, CACHE ADS LOCK, PLOCK BREQ HLDA BLAST FERR PCHK SMIACT HITM To achieve the lowest possible power consumption during the Stop Grant state, the system designer must ensure that the input signals with pull-up resistors are not driven Low, and the input signals with pull-down resistors are not driven High. All inputs except data bus pins must be driven to the power supply rails to ensure the lowest possible current consumption during Stop Grant or Stop Clock modes. For compatibility, data pins must be driven Low to achieve the lowest possible power consumption. 5.5 Clock Control State Diagram Figure 20 shows the state transitions during a Stop Clock cycle. 5.5.1 Normal State Type O I/O I/O O I/O O O O O O O O O O O State Previous State Previous State Floated Previous State Floated Previous State Inactive Inactive Previous State As per HOLD Previous State Previous State Previous State Previous State Previous State This is the normal operating state of the CPU. While in the normal state, the CLK input can be dynamically changed within the specified CLK period stability limits. 5.5.2 Stop Grant State The Stop Grant state provides a low-power state that can be entered by simply asserting the external STPCLK interrupt pin. When the Stop Grant bus cycle has been placed on the bus, and either RDY or BRDY is returned, the CPU is in this state. The CPU returns to the normal execution state 10–20 clock cycles after STPCLK has been deasserted. While in the Stop Grant state, the pull-up resistors on STPCLK and UP are disabled internally. The system must continue to drive these inputs to the state they were in immediately before the CPU entered the Stop Grant State. For minimum CPU power consumption, all other input pins should be driven to their inactive level while the CPU is in the Stop Grant state. Am5X86 Microprocessor 37 AMD PRELIMINARY (valid for Write-back mode only) Figure 20. Stop Clock State Machine CLK STPCLK Sampled STPCLK t20 t21 A NMI SMI Note: A = Earliest time at which NMI or SMI is recognized. Figure 21. Recognition of Inputs when Exiting Stop Grant State A RESET or SRESET brings the CPU from the Stop Grant state to the Normal state. The CPU recognizes the inputs required for cache invalidations (HOLD, AHOLD, BOFF, and EADS) as explained later. The CPU does not recognize any other inputs while in the Stop Grant state. Input signals to the CPU are not recognized until 1 clock after STPCLK is deasserted (see Figure 21). While in the Stop Grant state, the CPU does not recognize transitions on the interrupt signals (SMI, NMI, and INTR). Driving an active edge on either SMI or NMI does not guarantee recognition and service of the interrupt request following exit from the Stop Grant state. However, if one of the interrupt signals (SMI, NMI, or INTR) is driven active while the CPU is in the Stop Grant state, and held active for at least one CLK after STPCLK is deasserted, the corresponding interrupt will be serviced. 38 Am5X86 Microprocessor PRELIMINARY The Am5X86 CPU product family requires INTR to be held active until the CPU issues an interrupt acknowledge cycle to guarantee recognition. This condition also applies to the existing Am486 CPUs. In the Stop Grant state, the system can stop or change the CLK input. When the clock stops, the CPU enters the Stop Clock state. The CPU returns to the Stop Grant state immediately when the CLK input is restarted. You must hold the STPCLK input Low until a stabilized frequency has been maintained for at least 1 ms to ensure that the PLL has had sufficient time to stabilize. The CPU generates a Stop Grant bus cycle when entering the state from the Normal or the Auto HALT Power Down state. When the CPU enters the Stop Grant state from the Stop Clock state or the Stop Clock Snoop state, the CPU does not generate a Stop Grant bus cycle. 5.5.3 Stop Clock State AMD tioning active until the completion of the write-back. It then powers down and returns to the previous state. The CPU does not generate a bus cycle when it returns to the previous state. 5.5.6 Cache Flush State When configured in Write-back mode, the processor recognizes FLUSH for copying back modified cache lines to memory in the Auto Halt Power Down State or Normal State. Upon the completion of the cache flush, the processor returns to its prior state, and regenerates a special bus cycle, if necessary. 6 SRESET FUNCTION Stop Clock state is entered from the Stop Grant state by stopping the CLK input (either logic High or logic Low). None of the CPU input signals should change state while the CLK input is stopped. Any transition on an input signal (except INTR) before the CPU has returned to the Stop Grant state may result in unpredictable behavior. If INTR goes active while the CLK input is stopped, and stays active until the CPU issues an interrupt acknowledge bus cycle, it is serviced in the normal manner. System design must ensure the CPU is in the correct state prior to asserting cache invalidation or interrupt signals to the CPU. 5.5.4 Auto Halt Power Down State The Am5X86 microprocessor family supports a soft reset function through the SRESET pin. SRESET forces the processor to begin execution in a known state. The processor state after SRESET is the same as after RESET except that the internal caches, CD and NW in CR0, write buffers, SMBASE registers, and floating-point registers retain the values they had prior to SRESET, and cache snooping is allowed. The processor starts execution at physical address FFFFFFF0h. SRESET can be used to help performance for DOS extenders written for the 80286 processor. SRESET provides a method to switch from Protected to Real mode while maintaining the internal caches, CR0, and the FPU state. SRESET may not be used in place of RESET after power-up. In Write-back mode, once SRESET is sampled active, the SRESET sequence begins on the next instruction boundary (unless FLUSH or RESET occur before that boundary). When started, the SRESET sequence continues to completion and then normal processor execution resumes, independent of the deassertion of SRESET. If a snoop hits a modified line during SRESET, a normal write-back cycle occurs. ADS is asserted to drive the bus cycles even if SRESET is not deasserted. A HALT instruction causes the CPU to enter the Auto HALT Power Down state. The CPU issues a normal HALT bus cycle, and only transitions to the Normal state when INTR, NMI, SMI, RESET, or SRESET occurs. The system can generate a STPCLK while the CPU is in the Auto HALT Power Down state. The CPU generates a Stop Grant bus cycle when it enters the Stop Grant state from the HALT state. When the system deasserts the STPCLK interrupt, the CPU returns execution to the HALT state. The CPU generates a new HALT bus cycle when it reenters the HALT state from the Stop Grant state. 5.5.5 Stop Clock Snoop State (Cache Invalidations) 7 7.1 SYSTEM MANAGEMENT MODE Overview When the CPU is in the Stop Grant state or the Auto HALT Power Down state, the CPU recognizes HOLD, AHOLD, BOFF, and EADS for cache invalidation. When the system asserts HOLD, AHOLD, or BOFF, the CPU floats the bus accordingly. When the system asserts EADS, the CPU transparently enters Stop Clock Snoop state and powers up for one full clock to perform the required cache snoop cycle. If a modified line is snooped, a cache write-back occurs with HITM transi- The Am5X86 microprocessor supports four modes: Real, Virtual, Protected, and System Management mode (SMM). As an operating mode, SMM has a distinct processor environment, interface, and hardware/software features. SMM lets the system designer add new software-controlled features to the computer products that always operate transparent to the operating system (OS) and software applications. SMM is intended for use only by system firmware, not by applications software or general purpose systems software. The SMM architectural extension consists of the following elements: s System Management Interrupt (SMI) hardware in- terface Am5X86 Microprocessor 39 AMD s Dedicated and secure memory space (SMRAM) for SMI handler code and CPU state (context) data with a status signal for the system to decode access to that memory space, SMIACT s Resume (RSM) instruction, for exiting SMM s Special features, such as I/O Restart and I/O instruction information, for transparent power management of I/O peripherals, and Auto HALT Restart context normally consists of the CPU registers that fully represent the processor state. s Context Switch: A context switch is the process of either saving or restoring the context. The SMM discussion refers to the context switch as the process of saving/restoring the context while invoking/exiting SMM, respectively. s SMSAVE: A mechanism that saves and restores all 7.2 Terminology internal registers to and from SMRAM. The following terms are used throughout the discussion of System Management mode. s SMM: System Management mode. The operating 7.3 System Management Interrupt Processing environment that the processor (system) enters when servicing a System Management Interrupt. s SMI: System Management Interrupt. This is the trig- The system interrupts the normal program execution and invokes SMM by generating a System Management Interrupt (SMI) to the CPU. The CPU services the SMI by executing the following sequence (see Figure 22). 1. The CPU asserts the SMIACT signal, instructing the system to enable the SMRAM. 2. The CPU saves its state (internal register) to SMRAM. It starts at the SMBASE relative address location (see Section 7.3.3), and proceeds downward in a stack-like fashion. 3. The CPU switches to the SMM processor environment (an external pseudo-real mode). 4. The CPU then jumps to the absolute address of SMBASE + 8000h in SMRAM to execute the SMI handler. This SMI handler performs the system management activities. ger mechanism for the SMM interface. When SMI is asserted (SMI pin asserted Low) it causes the processor to invoke SMM. The SMI pin is the only means of entering SMM. s SMI handler: System Management mode handler. This is the code that is executed when the processor is in SMM. Example applications that this code might implement are a power management control or a system control function. s RSM: Resume instruction. This instruction is used by the SMI handler to exit the SMM and return to the interrupted OS or application process. s SMRAM: This is the physical memory dedicated to SMM. The SMI handler code and related data reside in this memory. The processor also uses this memory to store its context before executing the SMI handler. The operating system and applications should not have access to this memory space. s SMBASE: This is a control register that contains the Note: If the SMRAM shares the same physical address location with part of the system RAM, it is “overlaid” SMRAM. To preserve cache consistency and correct SMM operation in systems using overlaid SMRAM, the cache must be flushed via the FLUSH pin when entering SMM. 5. The SMI handler then executes the RSM instruction which restores the CPU’s context from SMRAM, deasserts the SMIACT signal, and then returns control to the previously interrupted program execution. base address that defines the SMRAM space. s Context: This term refers to the processor state. The SMM discussion refers to the context, or processor state, just before the processor invokes SMM. The SMI Instr #1 Instr #2 Instr Instr #3 #4 State Save SMI Handler RSM State Restore Instr #5 SMI SMIACT Figure 22. Basic SMI Interrupt Service 40 Am5X86 Microprocessor PRELIMINARY AMD SMIACT CPU SMI } SMI Interface Figure 23. Basic SMI Hardware Interface For uses such as fast enabling of external I/O devices, the SMSAVE mode permits the restarting of the I/O instructions and the HALT instruction. This is accomplished through I/O Trap Restart and Halt/Auto HALT Restart slots. Only I/O and HALT opcodes are restartable. Attempts to restart any other opcode may result in unpredictable behavior. The System Management Interrupt hardware interface consists of the SMI request input and the SMIACT output used by the system to decode the SMRAM (see Figure 23). 7.3.1 System Management Interrupt Processing Masking SMI prevents recursive calls. If another SMI occurs while SMI is masked, the pending SMI is recognized and executed on the next instruction boundary after the current SMI completes. This instruction boundary occurs before execution of the next instruction in the interrupted application code, resulting in back-to-back SMI handlers. Only one SMI signal can be pending while SMI is masked. The SMI signal is synchronized internally and must be asserted at least three clock cycles prior to asserting the RDY signal to guarantee recognition on a specific instruction boundary. This is important for servicing an I/O trap with an SMI handler. 7.3.2 SMI Active (SMIACT) SMI is a falling-edge-triggered, non-maskable interrupt request signal. SMI is an asynchronous signal, but setup and hold times must be met to guarantee recognition in a specific clock. The SMI input does not have to remain active until the interrupt is actually serviced. The SMI input needs to remain active for only a single clock if the required setup and hold times are met. SMI also works correctly if it is held active for an arbitrary number of clocks (see Figure 24). The SMI input must be held inactive for at least four clocks after it is asserted to reset the edge-triggered logic. A subsequent SMI may not be recognized if the SMI input is not held inactive for at least four clocks after being asserted. SMI, like NMI, is not affected by the IF bit in the EFLAGS register and is recognized on an instruction boundary. SMI does not break locked bus cycles. SMI has a higher priority than NMI and is not masked during an NMI. After SMI is recognized, the SMI signal is masked internally until the RSM instruction is executed and the interrupt service routine is complete. SMIACT indicates that the CPU is operating in SMM. The CPU asserts SMIACT in response to an SMI interrupt request on the SMI pin. SMIACT is driven active after the CPU has completed all pending write cycles (including emptying the write buffers), and before the first access to SMRAM when the CPU saves (writes) its state (or context) to SMRAM. SMIACT remains active until the last access to SMRAM when the CPU restores (reads) its state from SMRAM. The SMIACT signal does not float in response to HOLD. The SMIACT signal is used by the system logic to decode SMRAM. The number of clocks required to complete the SMM state save and restore is dependent on system memory performance. The values shown in Figure 25 assume 0 waitstate memory writes (2 clock cycles), 2–1–1–1 burst read cycles, and 0 wait-state non-burst reads (two clock cycles). Additionally, it is assumed that the data read during the SMM state restore sequence is not cacheable. The minimum time required to enter a SMSAVE SMI handler routine for the CPU (from the completion of the interrupted instruction) is given by: Latency to start of SMl handler = A + B + C = 161 clocks and the minimum time required to return to the interrupted application (following the final SMM instruction before RSM) is given by: Latency to continue application = E + F + G = 258 clocks CLK CLK2 SMI Sampled SMI tsu thd RDY Figure 24. SMI Timing for Servicing an I/O Trap Am5X86 Microprocessor 41 AMD T1 CLK CLK2 T2 PRELIMINARY SMI ADS B G RDY SMIACT C D E F A Normal State State Save SMM Handler State Restore Normal State Clock-Tripled CPU A: Last RDY from non-SMM transfer to SMIACT assertion B: SMIACT assertion to first ADS for SMM state save C: SMM state save (dependent on memory performance) D: SMI handler E: SMM state restore (dependent on memory performance) F: Last RDY from SMM transfer to deassertion of SMIACT G: SMIACT deassertion of first non-SMM ADS 2 CLKs minimum 15 CLKs minimum 100 CLKs User-determined 180 CLKs 2 CLKs minimum 20 CLKs minimum Clock-Quadrupled CPU 2 CLKs minimum 10 CLKs minimum 70 CLKs User-determined 120 CLKs 2 CLKs minimum 20 CLKs minimum Figure 25. SMIACT Timing 7.3.3 SMRAM from 3FFFFh downward to 3FE00h. After saving its state, the CPU jumps to the address location 38000h to begin executing the SMI handler. The system logic can choose to decode a larger area of SMRAM as needed. The size of this SMRAM can be between 32 Kbytes and 4 Gbytes.The system logic should provide a manual method for switching the SMRAM into system memory space when the CPU is not in SMM. This enables initialization of the SMRAM space (i.e., loading SMI handler) before executing the SMI handler during SMM (see Figure 26). System memory accesses redirected to SMRAM CPU accesses to system address space used for loading SMRAM The CPU uses the SMRAM space for state save and state restore operations during an SMI. The SMI handler, which also resides in SMRAM, uses the SMRAM space to store code, data, and stacks. In addition, the SMI handler can use the SMRAM for system management information such as the system configuration, configuration of a powered-down device, and system designer-specific information. Note: Access to SMRAM is through the CPU internal cache. To ensure cache consistency and correct operation, always assert the FLUSH pin in the same clock as SMI for systems using overlaid SMRAM. The CPU asserts SMIACT to indicate to the memory controller that it is operating in System Management mode. The system logic should ensure that only the CPU and SMI handler have access to this area. Alternate bus masters or DMA devices trying to access the SMRAM space when SMIACT is active should be directed to system RAM in the respective area. The system logic is minimally required to decode the physical memory address range 38000h–3FFFFh as SMRAM area. The CPU saves its state to the state save area 42 SMRAM System memory accesses not redirected to SMRAM Normal Memory Space Figure 26. Redirecting System Memory Address to SMRAM Am5X86 Microprocessor PRELIMINARY 7.3.4 SMRAM State Save Map AMD When SMI is recognized on an instruction boundary, the CPU core first sets the SMIACT signal Low, indicating to the system logic that accesses are now being made to the system-defined SMRAM areas. The CPU then writes its state to the state save area in the SMRAM. The state save area starts at SMBASE + [8000h + 7FFFh]. The default CS Base is 30000h; therefore, the default state save area is at 3FFFFh. In this case, the CS Base is also referred to as the SMBASE. Table 10. SMRAM State Save Map Register Offset* 7FFCh 7FF8h 7FF4h 7FF0h 7FECh 7FE8h 7FE4h 7FE0h 7FDCh 7FD8h 7FD4h 7FD0h 7FCCh 7FC8h 7FC4h 7FC0h 7FBCh 7FB8h 7FB4h 7FB0h 7FACh 7FA8h 7FA7h–7F98h 7F94h 7F93h–7F8Ch 7F88h 7F87h–7F08h 7F04h 7F02h 7F00h 7EFCh 7EF8h 7EF7h–7E00h Register CRO CR3 EFLAGS EIP EDI ESI EBP ESP EBX EDX ECX EAX DR6 DR7 TR* LDTR* GS* FS* DS* SS* CS* ES* Reserved IDT Base Reserved GDT Base Reserved I/O Trap Word Halt Auto Restart I/O Trap Restart SMM Revision Identifier State Dump Base Reserved Writable? No No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No No No No No No No No No No No No No No No No Yes Yes Yes Yes No If the SMBASE relocation feature is enabled, the SMRAM addresses can change. The following formula is used to determine the relocated addresses where the context is saved: SMBASE + [8000h + Register Offset], where the default initial SMBASE is 30000h and the Register Offset is listed in Table 10. Reserved spaces are for new registers in future CPUs. Some registers in the SMRAM state save area may be read and changed by the SMI handler, with the changed values restored to the processor register by the RSM instruction. Some register images are read-only, and must not be modified. (Modifying these registers results in unpredictable behavior.) The values stored in the “reserved” areas may change in future CPUs. An SMI handler should not rely on values stored in a reserved area. The following registers are written out during SMSAVE mode to the RESERVED memory locations (7FA7h– 7F98h, 7F93h–7F8Ch, and 7F87h–7F08h), but are not visible to the system software programmer: s DR3–DR0 s CR2 s CS, DS, ES, FS, GS, and SS hidden descriptor s s s s s registers EIP_Previous GDT Attributes and Limits IDT Attributes and Limits LDT Attributes, Base, and Limits TSS Attributes, Base, and Limits If an SMI request is issued to power down the CPU, the values of all reserved locations in the SMM state save area must be saved to non-volatile memory. The following registers are not automatically saved and restored by SMI and RSM: s TR7–TR3 s FPU registers: — — — — — — — STn FCS FSW Tag Word FP instruction pointer FP opcode Operand pointer Note: You can save the FPU state by using an FSAVE or FNSAVE instruction. For all SMI requests except for power down suspend/ resume, these registers do not have to be saved because their contents will not change. During a power down suspend/resume, however, a resume reset clears these registers back to their default values. In this case, the suspend SMI handler should read these registers directly to save them and restore them during the power up resume. Anytime the SMI handler changes these registers in the CPU, it must also save and restore them. Note: *Upper 2 bytes are not modified. Am5X86 Microprocessor 43 AMD PRELIMINARY processor enters the shutdown state. This occurs in the following situations: s The value in the State Dump base field is not a 7.4 Entering System Management Mode SMM is one of the major operating modes, along with Protected mode, Real mode, and Virtual mode. Figure 27 shows how the processor can enter SMM from any of the three modes and then return. 32-Kbyte aligned address s A combination of bits in CR0 is illegal: (PG=1 and PE=0) or (NW=1 and CD=0) Real mode Reset or PE=0 Reset Reset or RSM SMI RSM System Management mode SMI PE=1 In Shutdown mode, the processor stops executing instructions until an NMI interrupt is received or reset initialization is invoked. The processor generates a shutdown bus cycle. Three SMM features can be enabled by writing to control slots in the SMRAM state save area: 1. Auto HALT Restart. It is possible for the SMI request to interrupt the HALT state. The SMI handler can tell the RSM instruction to return control to the HALT instruction or to return control to the instruction following the HALT instruction by appropriately setting the Auto HALT Restart slot. The default operation is to restart the HALT instruction. 2. I/O Trap Restart. If the SMI was generated on an I/O access to a powered-down device, the SMI handler can instruct the RSM instruction to re-execute that I/O instruction by setting the I/O Trap Restart slot. 3. SMBASE Relocation. The system can relocate the SMRAM by setting the SMBASE Relocation slot in the state save area. The RSM instruction sets SMBASE in the processor based on the value in the SMBASE relocation slot. The SMBASE must be aligned on 32-Kbyte boundaries. A RESET also causes execution to exit from SMM. Protected mode VM=0 VM=1 RSM SMI Virtual mode Figure 27. Transition to and from SMM The external signal SMI causes the processor to switch to SMM. The RSM instruction exits SMM. SMM is transparent to applications, programs, and operating systems for the following reasons: s The only way to enter SMM is via a type of non- maskable interrupt triggered by an external signal s The processor begins executing SMM code from a separate address space, referred to earlier as system management RAM (SMRAM) s Upon entry into SMM, the processor saves the reg- 7.6 Processor Environment ister state of the interrupted program (depending on the save mode) in a part of SMRAM called the SMM context save space s All interrupts normally handled by the operating sys- tem or applications are disabled upon SMM entry s A special instruction, RSM, restores processor reg- isters from the SMM context save space and returns control to the interrupted program Similar to Real mode, SMM has no privilege levels or address mapping. SMM programs can execute all I/O and other system instructions and can address up to 4 Gbytes of memory. When an SMI signal is recognized on an instruction execution boundary, the processor waits for all stores to complete, including emptying the write buffers. The final write cycle is complete when the system returns RDY or BRDY. The processor then drives SMIACT active, saves its register state to SMRAM space, and begins to execute the SMI handler. SMI has greater priority than debug exceptions and external interrupts. This means that if more than one of these conditions occur at an instruction boundary, only the SMI processing occurs. Subsequent SMI requests are not acknowledged while the processor is in SMM. The first SMI request that occurs while the processor is in SMM is latched, and serviced when the processor exits SMM with the RSM instruction. Only one SMI signal is latched by the CPU while it is in SMM. When the CPU invokes SMM, the CPU core registers are initialized as indicated in Table 11. 7.5 Exiting System Management Mode The RSM instruction (opcode 0F AAh) leaves SMM and returns control to the interrupted program. The RSM instruction can be executed only in SMM. An attempt to execute the RSM instruction outside of SMM generates an invalid opcode exception. When the RSM instruction is executed and the processor detects invalid state information during the reloading of the save state, the 44 Am5X86 Microprocessor PRELIMINARY Table 11. SMM Initial CPU Core Register Settings Register General Purpose Registers EFLAGS CR0 DR6 DR7 GDTR, LDTR, IDTR, TSSR EIP SMM Initial State Unmodified 0000 0002h Bits 0, 2, 3, and 31 cleared (PE, EM, TS, and PG); rest unmodified Unpredictable state 0000 0400h Unmodified 0000 8000h AMD Table 12. Segment Register Initial States Segment Selector Register CS2 DS ES FS GS SS 3000h 0000h 0000h 0000h 0000h 0000h Base 30000h 00000000h 00000000h 00000000h 00000000h 00000000h Attributes 16-bit, expand up 16-bit, expand up 16-bit, expand up 16-bit, expand up 16-bit, expand up 16-bit, expand up Limit1 4 Gbytes 4 Gbytes 4 Gbytes 4 Gbytes 4 Gbytes 4 Gbytes Notes: Note: Interrupts from INT and NMI are disabled on SMM entry. 1. The segment limit check is 4 Gbytes instead of the usual 64 Kbytes. 2. The Selector value for CS remains at 3000h even if the SMBASE is changed. The following is a summary of the key features in the SMM environment: s Real mode style address calculation s 4-Gbyte limit checking s IF flag is cleared s NMI is disabled s TF flag in EFLAGS is cleared; single step traps are s s s s disabled DR7 is cleared; debug traps are disabled The RSM instruction no longer generates an invalid opcode error Default 16-bit opcode, register, and stack use All bus arbitration (HOLD, AHOLD, BOFF) inputs, and bus sizing (BS8, BS16) inputs operate normally while the CPU is in SMM The CS Base can be changed using the SMM Base relocation feature. When the SMI handler is invoked, the CPU’s PE and PG bits in CR0 are reset to 0. The processor is in an environment similar to Real mode, but without the 64-Kbyte limit checking. However, the default operand size and the default address size are set to 16 bits. The EM bit is cleared so that no exceptions are generated. (If the SMM was entered from Protected mode, the Real mode interrupt and exception support is not available.) The SMI handler should not use floating-point unit instructions until the FPU is properly detected (within the SMI handler) and the exception support is initialized. Because the segment bases (other than CS) are cleared to 0 and the segment limits are set to 4 Gbytes, the address space may be treated as a single flat 4-Gbyte linear space that is unsegmented. The CPU is still in Real mode and when a segment selector is loaded with a 16-bit value, that value is then shifted left by 4 bits and loaded into the segment base cache. In SMM, the CPU can access or jump anywhere within the 4-Gbyte logical address space. The CPU can also indirectly access or perform a near jump anywhere within the 4-Gbyte logical address space. 7.7 Executing System Management Mode Handler The processor begins execution of the SMI handler at offset 8000h in the CS segment. The CS Base is initially 30000h, as shown in Table 12. Am5X86 Microprocessor 45 AMD 7.7.1 Exceptions and Interrupts with System Management Mode PRELIMINARY NMI interrupts are blocked on entry to the SMI handler. If an NMI request occurs during the SMI handler, it is latched and serviced after the processor exits SMM. Only one NMI request is latched during the SMI handler. If an NMI request is pending when the processor executes the RSM instruction, the NMI is serviced before the next instruction of the interrupted code sequence. Although NMI requests are blocked when the CPU enters SMM, they may be enabled through software by executing an IRET instruction. If the SMI handler requires the use of NMI interrupts, it should invoke a dummy interrupt service routine to execute an IRET instruction. When an IRET instruction is executed, NMI interrupt requests are serviced in the same Real mode manner in which they are handled outside of SMM. 7.7.2 SMM Revisions Identifier When the CPU enters SMM, it disables INTR interrupts, debug, and single step traps by clearing the EFLAGS, DR6, and DR7 registers. This prevents a debug application from accidentally breaking into an SMI handler. This is necessary because the SMI handler operates from a distinct address space (SMRAM) and the debug trap does not represent the normal system memory space. For an SMI handler to use the debug trap feature of the processor to debug SMI handler code, it must first ensure that an SMM-compliant debug handler is available. The SMI handler must also ensure DR3–DR0 is saved to be restored later. The debug registers DR3–DR0 and DR7 must then be initialized with the appropriate values. For the processor to use the single step feature of the processor, it must ensure that an SMM-compliant single step handler is available and then set the trap flag in the EFLAGS register. If the system design requires the processor to respond to hardware INTR requests while in SMM, it must ensure that an SMM-compliant interrupt handler is available, and then set the interrupt flag in the EFLAGS register (using the STI instruction). Software interrupts are not blocked on entry to SMM, and the system software designer must provide an SMM-compliant interrupt handler before attempting to execute any software interrupt instructions. Note that in SMM mode, the interrupt vector table has the same properties and location as the Real mode vector table. The 32-bit SMM Revision Identifier specifies the version of SMM and the extensions that are available on the processor. The fields of the SMM Revision Identifiers and bit definitions are shown in Table 13 and Table 14. Bit 17 or 16 indicates whether the feature is supported (1=supported, 0=not supported). The processor always reads the SMM Revision Identifier at the time of a restore. The I/O Trap Extension and SMM Base Relocation bits are fixed. The processor writes these bits out at the time it performs a save state. Note: Changing the state of the reserved bits may result in unpredictable processor behavior. Table 13. SMM Revision Identifier 31–18 Reserved 00000000000000 17 SMM Base Relocation 1 16 I/O Trap Extension 1 15–0 SMM Revision Level 0000h Table 14. SMM Revision Identifier Bit Definitions Bit Name SMM Base Relocation I/O Trap Extension Description 1=SMM Base Relocation Available 0=SMM Base Relocation Unavailable 1=I/O Trapping Available 0=I/O Trapping Unavailable Default State 1 State at SMM Entry 1 0 1 0 State at SMM Exit 1 0 1 0 Notes No Change in State No Change in State No Change in State No Change in State 1 46 Am5X86 Microprocessor PRELIMINARY 7.7.3 Auto HALT Restart . 15 0 AMD The Auto HALT Restart slot at register offset (word location) 7F02h in SMRAM indicates to the SMI handler that the SMI interrupted the CPU during a HALT state; bit 0 of slot 7F02h is set to 1 if the previous instruction was a HALT (see Figure 28). If the SMI did not interrupt the CPU in a HALT state, then the SMI microcode sets bit 0 of the Auto HALT Restart slot to 0. If the previous instruction was a HALT, the SMI handler can choose to either set or reset bit 0. If this bit is set to 1, the RSM microcode execution forces the processor to re-enter the HALT state. If this bit is set to 0 when the RSM instruction is executed, the processor continues execution with the instruction just after the interrupted HALT instruction. If the HALT instruction is restarted, the CPU will generate a memory access to fetch the HALT instruction (if it is not in the internal cache), and execute a HALT bus cycle. Register offset 7F00h I/O instruction restart slot Figure 29. I/O Instruction Restart Register Offset When the RSM instruction is executed — if the I/O instruction restart slot contains the value 0FFh — the CPU automatically re-executes the l/O instruction that the SMI signal trapped. If the I/O instruction restart slot contains the value 00h when the RSM instruction is executed, then the CPU does not re-execute the I/O instruction. The CPU automatically initializes the I/O instruction restart slot to 00h during SMM entry. The I/O instruction restart slot should be written only when the processor has generated an SMI on an I/O instruction boundary. Processor operation is unpredictable when the I/O instruction restart slot is set when the processor is servicing an SMI that originated on a non-I/O instruction boundary. If the system executes back-to-back SMI requests, the second SMI handler must not set the I/O instruction restart slot. The second back-to-back SMI signal will not have the I/O Trap Word set. 7.7.5 I/O Trap Word 15 Reserved 1 0 Register Offset 7F02h HALT Auto Restart Figure 28. Auto HALT Restart Register Offset Table 15 shows the possible restart configurations. If the interrupted instruction was not a HALT instruction (bit 0 is set to 0 in the Auto HALT Restart slot upon SMM entry), setting bit 0 to 1 will cause unpredictable behavior when the RSM instruction is executed. Table 15. HALT Auto Restart Configuration Value at Value Entry at Exit 0 0 1 1 0 1 0 1 The I/O Trap Word contains the address of the I/O access that forced the external chipset to assert SMI, whether it was a read or write access, and whether the instruction that caused the access to the I/O address was a valid I/O instruction. Table 16 shows the layout. Table 16. I/O Trap Word Configuration Processor Action on Exit 31–16 Return to next instruction in interrupted program Unpredictable Returns to instruction after HALT Returns to interrupted HALT instruction I/O Address 15–2 Reserved 1 Valid I/O Instruction 0 R/W Bits 31–16 contain the I/O address that was being accessed at the time SMI became active. Bits 15–2 are reserved. If the instruction that caused the I/O trap to occur was a valid I/O instruction (IN, OUT, INS, OUTS, REP INS, or REP OUTS), the Valid I/O Instruction bit is set. If it was not a valid I/O instruction, the bit is saved as a 0. For REP instructions, the external chip set should return a valid SMI within the first access. Bit 0 indicates whether the opcode that was accessing the I/O location was performing either a read (1) or a write (0) operation as indicated by the R/W bit. 7.7.4 I/O Trap Restart The I/O instruction restart slot (register offset 7F00h in SMRAM) gives the SMI handler the option of causing the RSM instruction to automatically re-execute the interrupted I/O instruction (see Figure 29). Am5X86 Microprocessor 47 AMD PRELIMINARY SMRAM SMBASE + 8000h + 7FFFh Start of State Save SMI Handler Entry Point SMBASE + 8000h If an SMI occurs and it does not trap an I/O instruction, the contents of the I/O address and R/W bit are unpredictable and should not be used. 7.7.6 SMM Base Relocation The Am5X86 CPU family provides a new control register, SMBASE. The SMRAM address space can be modified by changing the SMBASE register before exiting an SMI handler routine. SMBASE can be changed to any 32K-aligned value. (Values that are not 32K-aligned cause the CPU to enter the shutdown state when executing the RSM instruction.) SMBASE is set to the default value of 30000h on RESET. If SMBASE is changed by an SMI handler, all subsequent SMI requests initiate a state save at the new SMBASE. The SMBASE slot in the SMM state save area indicates and changes the SMI jump vector location and SMRAM save area. When bit 17 of the SMM Revision Identifier is set, then this feature exists and the SMRAM base and consequently, the jump vector, are as indicated by the SMM Base slot (see Figure 30). During the execution of the RSM instruction, the CPU reads this slot and initializes the CPU to use the new SMBASE during the next SMI. During an SMI, the CPU does its context save to the new SMRAM area pointed to by the SMBASE, stores the current SMBASE in the SMM Base slot (offset 7EF8h), and then starts execution of the new jump vector based on the current SMBASE (see Figure 31). 31 0 Register Offset 7EF8h SMBASE Figure 31. SRAM Usage To change the SMRAM base address and SMI jump vector location, SMI handler modifies the SMBASE slot. Upon executing an RSM instruction, the processor reads the SMBASE slot and stores it internally. Upon recognition of the next SMI request, the processor uses the new SMBASE slot for the SMRAM dump and SMI jump vector. If the modified SMBASE slot does not contain a 32-Kbyte aligned value, the RSM microcode causes the CPU to enter the shutdown state. 7.8 7.8.1 SMM System Design Considerations SMRAM Interface The hardware designed to control the SMRAM space must follow these guidelines: s Initialize SMRAM space during system boot up. Ini- SMM Base tialization must occur before the first SMI occurs. Initialization of SMRAM space must include installation of an SMI handler and may include installation of related data structures necessary for particular SMM applications. The memory controller interfacing SMRAM should provide a means for the initialization code to open the SMRAM space manually. s The memory controller must decode a minimum ini- Figure 30. SMM Base Slot Offset The SMBASE must be a 32-Kbyte aligned, 32-bit integer that indicates a base address for the SMRAM context save area and the SMI jump vector. For example, when the processor first powers up, the range for the SMRAM area is from 38000h–3FFFFh. The default value for SMBASE is 30000h. As illustrated in Figure 31, the starting address of the jump vector is calculated by: SMBASE + 8000h The starting address for the SMRAM state save area is calculated by: SMBASE + [8000h + 7FFFh] When this feature is enabled, the SMRAM register map is addressed according to the above formula. tial SMRAM address space of 38000h–3FFFFh. s Alternate bus masters (such as DMA controllers) must not be able to access SMRAM space. The system should allow only the CPU, either through SMI or during initialization, to access SMRAM. s To implement a 0-V suspend function, the system must have access to all normal system memory from within an SMI handler routine. If the SMRAM overlays normal system memory (see Figure 32), there must be a method to access overlaid system memory independently. 48 Am5X86 Microprocessor PRELIMINARY AMD SMRAM Normal memory SMRAM Normal memory Overlaid region Normal memory the FLUSH pin input or the WBINVD instruction, and therefore invokes a performance penalty. There is no flush penalty for systems configured in Write-through mode. If the flush at SMM entry is not done, the first SMM read could hit in a cache that contains normal memory space code/data instead of the required SMI handler, and the handler could not be executed. If the cache is not disabled and is not flushed at SMM exit, the normal read cycles after SMM may hit in a cache that may contain SMM code/data instead of the normal system memory contents. In Write-through mode, assert the FLUSH signal in response to the assertion of SMIACT at SMM entry, and, if required because the cache is enabled, assert FLUSH again in response to the deassertion of SMIACT at SMM exit (see Figure 36 and Figure 37). For systems configured in Write-back mode, assert FLUSH with SMI (see Figure 38). Reloading the state registers at the end of SMM restores cache functionality to its pre-SMM state. 7.8.3 A20M Pin Non-overlaid (no need to flush caches) Overlaid (caches must be flushed) Figure 32. SMRAM Location The recommended configuration is to use a separate (non-overlaid) physical address for SMRAM. This nonoverlaid scheme prevents the CPU from improperly accessing the SMRAM or system RAM directly or through the cache. Figure 33 shows the relative SMM timing for non-overlaid SMRAM for systems configured in Writethrough mode. For systems configured in Write-back mode, WB/WT must be driven Low (as shown in Figure 34) to force caching during SMM to be write-through. Alternately, caching can be disabled during SMM by deasserting KEN with SMI (as shown in Figure 35). When the default SMRAM location is used, however, SMRAM is overlaid with system main memory (at 38000h–3FFFFh). For simplicity, system designers may want to use this default address, or they may select another overlaid address range. However, in this case the system control circuitry must use SMIACT to distinguish between SMRAM and main system memory, and must restrict SMRAM space access to the CPU only. To maintain cache coherency and to ensure proper system operation in systems configured in Writethrough mode, the system must flush both the CPU internal cache and any second level caches in response to SMIACT going Low. A system that uses cache during SMM must flush the cache a second time in response to SMIACT going High (see Figure 36). If KEN is driven High when FLUSH is asserted, the cache is disabled and a second flush is not required (see Figure 37). If the system is configured in Write-back mode, the cache must be flushed when SMI is asserted and then disabled (see Figure 38). 7.8.2 Cache Flushes Systems based on the MS-DOS operating system contain a feature that enables the CPU address bit A20 to be forced to 0. This limits physical memory to a maximum of 1 Mbyte, and is provided to ensure compatibility with those programs that relied on the physical address wraparound functionality of the original IBM PC. The A20M pin on Am5X86 CPUs provides this function. When A20M is active, all external bus cycles drive A20 Low, and all internal cache accesses are performed with A20 Low. The A20M pin is recognized while the CPU is in SMM. The functionality of the A20M input must be recognized in two instances: 1. If the SMI handler needs to access system memory space above 1 Mbyte (for example, when saving memory to disk for a 0-V suspend), the A20M pin must be deasserted before the memory above 1 Mbyte is addressed. 2. If SMRAM has been relocated to address space above 1 Mbyte, and A20M is active upon entering SMM, the CPU attempts to access SMRAM at the relocated address, but with A20 Low. This could cause the system to crash, because there would be no valid SMM interrupt handler at the accessed location. To account for these two situations, the system designer must ensure that A20M is deasserted on entry to SMM. A20M must be driven inactive before the first cycle of the SMM state save, and must be returned to its original level after the last cycle of the SMM state restore. This can be done by blocking the assertion of A20M when SMIACT is active. 49 The CPU does not unconditionally flush its cache before entering SMM. Therefore, the designer must ensure that, for systems using overlaid SMRAM, the cache is flushed upon SMM entry and SMM exit if caching is enabled. Note: A cache flush in a system configured in Writeback mode requires a minimum of 4100 internal clocks to test the cache for modified data, whether invoked by Am5X86 Microprocessor AMD PRELIMINARY State Save SMI Handler State Resume Normal Cycle SMI RSM SMIACT Figure 33. SMM Timing in Systems Using Non-Overlaid Memory Space and Write-Through Mode with Caching Enabled During SMM State Save SMI Handler State Resume Normal Cycle SMI RSM SMIACT WB/WT Note: For proper operation of systems configured in Write-back mode when caching during SMM is allowed, force WB/WT Low to force all caching to be write-through during SMM. Figure 34. SMM Timing in Systems Using Non-Overlaid Memory Spaces and Write-Back Mode with Caching Enabled During SMM State Save SMI Handler State Resume Normal Cycle SMI RSM SMIACT KEN Figure 35. SMM Timing in Systems Using Non-Overlaid Memory Spaces and Write-Back Mode with Caching Disabled During SMM 50 Am5X86 Microprocessor PRELIMINARY AMD SMI State Save State Resume Instruction x+1 Normal Cycle Instruction x SMI Handler SMI RSM SMIACT FLUSH Cache contents invalidated Cache contents invalidated Figure 36. SMM Timing in Systems Using Overlaid Memory Space and Write-Through Mode with Caching Enabled During SMM SMI Instruction x+1 State Save State Resume Normal Cycle Instruction x SMI Handler SMI RSM SMIACT FLUSH KEN Cache contents invalidated Figure 37. SMM Timing in Systems Using Overlaid Memory Spaces and Write-Through Mode with Caching Disabled During SMM Cache Flush State State Save SMI Handler State Resume Normal Cycle SMI RSM SMIACT KEN FLUSH Cache must be empty Figure 38. SMM Timing in Systems Using Overlaid Memory Spaces and Configured in Write-Back Mode Am5X86 Microprocessor 51 AMD 7.8.4 CPU Reset During SMM PRELIMINARY quest, so that the l/O instruction can be re-executed. If a new SMI request is received while the CPU is executing an SMI handler, the CPU services this SMI request before restarting the original I/O instruction. If the I/O restart slot is set when the CPU executes the RSM instruction for the second SMI handler, the RSM microcode decrements the restored EIP again. EIP then points to an address different from the originally interrupted instruction, and the CPU begins execution at an incorrect entry point. To prevent this from occurring, the SMI handler routine must not set the I/O restart slot during the second of two consecutive SMI handlers. The system designer should take into account the following restrictions while implementing the CPU Reset logic: 1. When running software written for the 80286 CPU, a CPU RESET switches the CPU from Protected mode to Real mode. RESET and SRESET have a higher priority than SMI. When the CPU is in SMM, the SRESET to the CPU during SMM should be blocked until the CPU exits SMM. SRESET must be blocked beginning from the time when SMI is driven active. Care should be taken not to block the global system RESET, which may be necessary to recover from a system crash. 2. During execution of the RSM instruction to exit SMM, there is a small time window between the deassertion of SMIACT and the completion of the RSM microcode. If a Protected mode to Real mode SRESET is asserted during this window, it is possible that the SMRAM space will be violated. The system designer must guarantee that SRESET is blocked until at least 20 CPU clock cycles after SMIACT has been driven inactive or until the start of a bus cycle. 3. Any request for a CPU RESET for the purpose of switching the CPU from Protected mode to Real mode must be acknowledged after the CPU has exited SMM. To maintain software transparency, the system logic must latch any SRESET signals that are blocked during SMM. For these reasons, the SRESET signal should be used for any soft resets, and the RESET signal should be used for all hard resets. 7.8.5 SMM and Second Level Write Buffers 7.9 7.9.1 SMM Software Considerations SMM Code Considerations The default operand size and the default address size are 16 bits; however, operand-size override and address-size override prefixes can be used as needed to directly access data anywhere within the 4-Gbyte logical address space. With operand-size override prefixes, the SMI handler can use jumps, calls, and returns to transfer a control to any location within the 4-Gbyte space. Note, however, the following restrictions: 1. Any control transfer that does not have an operandsize override prefix truncates EIP to 16 Low-order bits. 2. Due to the Real mode style of base-address formation, a long jump or call cannot transfer control segment with a base address of more than 20 bits (1 Mbyte). 7.9.2 Exception Handling Upon entry into SMM, external interrupts that require handlers are disabled (the IF in EFLAGS is cleared). This is necessary because, while the processor is in SMM, it is running in a separate memory space. Consequently, the vectors stored in the interrupt descriptor table (IDT) for the prior mode are not applicable. Before allowing exception handling (or software interrupts), the SMM program must initialize new interrupt and exception vectors. The interrupt vector table for SMM has the same format as for Real mode. Until the interrupt vector table is correctly initialized, the SMI handler must not generate an exception (or software interrupt). Even though hardware interrupts are disabled, exceptions and software interrupts can still occur. Only a correctly written SMI handler can prevent internal exceptions. When new exception vectors are initialized, internal exceptions can be serviced. Restrictions are as follows: 1. Due to the Real mode style of base address formation, an interrupt or exception cannot transfer control to a segment with a base address of more than 20 bits. 2. An interrupt or exception cannot transfer control to a segment offset of more than 16 bits. 3. If exceptions or interrupts are allowed to occur, only the Low order 16 bits of the return address are Before the processor enters SMM, it empties its internal write buffers. This is to ensure that the data in the write buffers is written to normal memory space, not SMM space. When the CPU is ready to begin writing an SMM state save to SMRAM, it asserts SMIACT. SMIACT may be driven active by the CPU before the system memory controller has had an opportunity to empty the second level write buffers. To prevent the data from these second level write buffers from being written to the wrong location, the system memory controller needs to direct the memory write cycles to either SMM space or normal memory space. This can be accomplished by saving the status of SMIACT with the address for each word in the write buffers. 7.8.6 Nested SMI and I/O Restart Special care must be taken when executing an SMI handler for the purpose of restarting an l/O instruction. When the CPU executes a Resume (RSM) instruction with the l/O restart slot set, the restored EIP is modified to point to the instruction immediately preceding the SMI re52 Am5X86 Microprocessor PRELIMINARY pushed onto the stack. If the offset of the interrupted procedure is greater than 64 Kbytes, it is not possible for the interrupt/exception handler to return control to that procedure. (One work-around is to perform software adjustment of the return address on the stack.) 4. The SMBASE Relocation feature affects the way the CPU returns from an interrupt or exception during an SMI handler. AMD 8 TEST REGISTERS 4 AND 5 MODIFICATIONS The Cache Test Registers for the Am5X86 microprocessor are the same test registers (TR3, TR4, and TR5) provided in Am486 microprocessors. TR3 is the cache test data register. TR4, the cache test status register, and TR5, the cache test control register, operate together with TR3. If WB/WT meets the necessary setup timing and is sampled Low on the falling edge of RESET, the processor is placed in Write-through mode and the test register function is identical to the Am486 microprocessors. If WB/WT meets the necessary setup timing and is sampled High on the falling edge of RESET, the processor is placed in Write-back mode and the test registers TR4 and TR5 are modified to support the added write-back cache functionality. Tables 17 and 18 show the individual bit functions of these registers. Sections 8.1 and 8.2 provide a detailed description of the field functions. Note: The execution of an IRET instruction enables Non-Maskable Interrupt (NMI) processing. 7.9.3 Halt during SMM HALT should not be executed during SMM, unless interrupts have been enabled. Interrupts are disabled on entry to SMM. INTR and NMI are the only events that take the CPU out of HALT within SMM. 7.9.4 Relocating SMRAM to an Address above 1 Mbyte Within SMM (or Real mode), the segment base registers can be updated only by changing the segment register. The segment registers contain only 16 bits, which allows only 20 bits to be used for a segment base address (the segment register is shifted left 4 bits to determine the segment base address). If SMRAM is relocated to an address above 1 Mbyte, the segment registers can no longer be initialized to point to SMRAM. These areas can still be accessed by using address override prefixes to generate an offset to the correct address. For example, if the SMBASE has been relocated immediately below 16 Mbytes, the DS and ES registers are still initialized to 0000 0000h. Data in SMRAM can still be accessed by using 32-bit displacement registers move esi,OOFFxxxxh move ax,ds:[esi] ;64K segment immediately below 16M Note: TR3 has the same functions in both Write-through and Write-back modes.These functions are identical to the TR3 register functions provided by Am486 microprocessors. 8.1 TR4 Definition This section includes a detailed description of the bit fields defined for TR4. Note: Bits listed in Table 17 as Reserved or Not used are not included in these descriptions. s Tag (bits 31–12): Read/Write, always available in Write-through mode. Available only when EXT=0 in TR5 in Write-back mode. For a cache write, this is the tag that specifies the address in memory. On a cache look-up, this is tag for the selected entry in the cache. Table 17. Test Register TR4 Bit Descriptions 31 EXT = 0 30–29 28 27–26 25–24 23–22 Tag 21–20 19–16 15–12 11 0 10 Valid Valid 9–7 LRU LRU 6–3 Valid (rd) Valid (rd) 2–0 Not used Not used EXT = 1 Not used STn Rsvd. ST3 ST2 ST1 ST0 Reserved Not used Notes: 1. The values of STn and ST3–ST0 are: 00 = Invalid; 01 = Exclusive; 10 = Modified; 11 = Shared. 2. During a cache look-up, bit 11 is read only and always 0. The bit is read/write otherwise. Table 18. Test Register TR5 Bit Descriptions 31–20 Write-Back Not used 19 Ext 18–17 Set State 16 Reserved 15–12 Not used 11–4 Index 3–2 Entry 1–0 Control Control Write-Through Not used Index Entry Notes: 1. Bit 19 in TR5 is EXT. If EXT = 0, TR4 has the standard 486 processor definition for write-through cache. 2. The values of Set State are: 00 = Invalid; 01 = Exclusive; 10 = Modified; 11 = Shared. Am5X86 Microprocessor 53 AMD PRELIMINARY Set State bit locations in TR5 are used to set the Valid bit for the selected entry and set. s LRU (bits 9–7): Read Only, independent of the Ext s STn (bits 30–29): Read Only, available only in Write- back mode when Ext=1 in TR5. STn returns the status of the set (ST3, ST2, ST1, or ST0) specified by the TR5 Set State field (bits 18–17) during cache look-ups. Returned values are: — — — — 00 = invalid 01 = exclusive 10 = modified 11 = shared bit in TR5. On a cache look-up, these are the three LRU bits of the accessed set. On a cache write, these bits are ignored; the LRU bits in the cache are updated by the pseudo-LRU cache replacement algorithm. Write operations to these locations have no effect on the device. s Valid (bits 6–3): Read Only, independent of the Ext s ST3 (bits 27–26): Read Only, available only in Write- back mode when Ext=1 in TR5. ST3 returns the status of Set 3 during cache look-ups. Returned values are: — — — — 00 = invalid 01 = exclusive 10 = modified 11 = shared bit in TR5. On a cache look-up, these are the four Valid bits of the accessed set. In Write-back mode, these valid bits are set if a cache set is in the exclusive, modified, or shared state. Write operations to these locations have no effect on the device. 8.2 TR5 Definition This section includes a detailed description of the bit fields in the TR5. s ST2 (bits 25–24): Read Only, available only in Write- back mode when Ext=1 in TR5. ST2 returns the status of Set 2 during cache look-ups. Returned values are: — — — — 00 = invalid 01 = exclusive 10 = modified 11 = shared Note: Bits listed in Table 18 as Reserved or Not Used are not included in the descriptions. s Ext (bit 19): Read/Write, available only in Write-back mode. Ext, or extension, determines which bit fields are defined for TR4: the address TAG field, or the STn and ST3–ST0 status bit fields. In Write-through mode, the Ext bit is not accessible. The following describes the two states of Ext: — Ext = 0, bits 31–11 of TR4 contain the TAG address — Ext = 1, bits 30–29 of TR4 contain STn, bits 27– 20 contain ST3–ST0 s Set State (bits 18–17): Read/Write, available only in s ST1 (bits 23–22): Read Only, available only in Write- back mode when Ext=1 in TR5. ST1 returns the status of Set 1 during cache look-ups. Returned values are: — — — — 00 = invalid 01 = exclusive 10 = modified 11 = shared Write-back mode. The Set State field is used to change the MESI state of the set specified by the Index and Entry bits. The state is set by writing one of the following combinations to this field: — — — — 00 = invalid 01 = exclusive 10 = modified 11 = shared s ST0 (bits 21–20): Read Only, available only in Write- back mode when Ext=1 in TR5. ST0 returns the status of Set 0 during cache look-ups. Returned values are: — — — — 00 = invalid 01 = exclusive 10 = modified 11 = shared s Index (bits 11–4): Read/Write, independent of write- through or Write-back mode. Index selects one of the 256 cache lines. s Entry (bits 3–2): Read/Write, independent of write- s Valid (bit 10): Read/Write, independent of the Ext bit in TR5. This is the Valid bit for the accessed entry. On a cache look-up, Valid is a copy of one of the bits reported in bits 6–3. On a cache write in Writethrough mode, Valid becomes the new Valid bit for the selected entry and set. In Write-back mode, writing to the Valid bit has no effect and is ignored; the 54 through or Write-back mode. Entry selects between one of the four entries in the set addressed by the Set Select during a cache read or write. During cache fill buffer writes or cache read buffer reads, the value in the Entry field selects one of the four doublewords in a cache line. Am5X86 Microprocessor PRELIMINARY s Control (bits 1–0): Read/Write, independent of Write- AMD through or Write-back mode. The control bits determine which operation to perform. The following is a definition of the control operations: — 00 = Write to cache fill buffer, or read from cache read buffer — 01 = Perform cache write — 10 = Perform cache read — 11 = Flush the cache (mark all entries invalid) 6. In TR5, load 0 into the Ext field (bit 19), the required value into the Set State field (bits 18–17) (Writeback mode only), the required index into the Index field (bits 10–4), the required entry value into the Entry field (bits 3–2), and 01 into the Control field (bits 1–0). Loading the values into TR5 triggers the cache write. In Write-through mode, the Set State field is ignored, and the Valid bit (bit 10) in TR4 is used instead to define the state of the specified set. 8.3.3 Example 3: Flushing The Cache The cache flush mechanism functions in the same way in Write-back and Write-through modes. Load 11 into the Control field (bits 1–0) of TR5. All other fields are ignored, except for Ext in Write-back mode. The cache flush is triggered by loading the value into TR5. All of the LRU bits, Valid bits, and Set State bits are cleared. 8.3 Using TR4 and TR5 for Cache Testing The following paragraphs provide examples of testing the cache using TR4 and TR5. 8.3.1 Example 1: Reading The Cache (Write-back Mode Only) 1. Disable caching by setting the CD bit in the CR0 register. 2. In TR5, load 0 into the Ext field (bit 19), the required index into the Index field (bits 10–4), the required entry value into the Entry field (bits 3–2), and 10 into the Control field (bits 1–0). Loading the values into TR5 triggers the cache read. The cache read loads the TR4 register with the TAG for the read entry, and the LRU and Valid bits for the entire set that was read. The cache read loads 128 data bits into the cache read buffer. The entire buffer can be read by placing each of the four binary combinations in the Entry field and setting the Control field in TR5 to 00 (binary). Read each doubleword from the cache read buffer through TR3. 3. Reading the Set State fields in TR4 during Writeback mode is accomplished by setting the Ext field in TR5 to 1 and rereading TR4. 8.3.2 Example 2: Writing The Cache 1) Disable the cache by setting the CD bit in the CR0 register. 2. In TR5, load 0 into the Ext field (bit 19), the required entry value into the Entry field (bits 3–2), and 00 into the Control field (bits 1–0). 3. Load the TR3 register with the data to write to the cache fill buffer. The cache fill buffer write is triggered by loading TR3. 4. Repeat steps 2 and 3 for the remaining three doublewords in the cache fill buffer. 5. In TR4, load the required values into TAG field (bits 31–11) and the Valid field (bit 10). In Write-back mode, the Valid bit is ignored since the Set State field in TR5 is used in place of the TR4 Valid bit. The other bits in TR4 (9:0) have no effect on the cache write. 9 Am5X86 CPU Functional Differences Several important differences exist between Am5X86 microprocessors and standard Am486DX microprocessors: s s s The ID register contains a different version signature. The EADS function performs cache line write-backs of modified lines to memory in Write-back mode. A burst write feature is available for copy-backs. The FLUSH pin and WBINVD instruction copy back all modified data to external memory prior to issuing the special bus cycle or reset. The Am5X86 processor is functionally identical to the Enhanced Am486 processor except for the function of the CLKMUL pin (see Section 9.3) and the redefinition of TR4 and TR5 to access the 16-Kbyte cache (see Section 8). 9.1 Status after Reset The RESET state is invoked either after power up or after the RESET signal is applied according to the standard Am486DX microprocessor specification. 9.2 Cache Status After reset, the STATUS bits of all lines are set to 0. The LRU bits of each set are placed in a starting state. 9.3 CLKMUL Pin For the standard Am486 processor, the Enhanced Am486 processor, and the Am5X86 processor, if the CLKMUL pin is driven High at RESET, the processor uses a Clock-tripled mode. To ensure correct operation of the 133-MHz Am5X86 processor, always connect the CLKMUL input to VSS. Am5X86 Microprocessor 55 AMD PRELIMINARY When the parameter passed in EAX is zero, the register values returned upon instruction execution are: EAX[31:0] EBX[31:0] ECX[31:0] EDX[31:0] 00000001h 68747541h 444D4163h 69746E65h 10 Am5X86 CPU IDENTIFICATION The Am5X86 microprocessor supports two standard methods for identifying the CPU in a system. The reported values are assigned based on the RESET status of the WB/WT pin input (Low = write-through; High = write-back). 10.1 DX Register at RESET The DX register always contains a component identifier at the conclusion of RESET. The upper byte of DX (DH) contains 04 and the lower byte of DX (DL) contains a CPU type/stepping identifier (see Table 19). Table 19. CPU ID Codes CLKMUL Setting/Cache mode Write-through mode Write-back mode Component Revision ID (DH) ID (DL) 04 04 Ex Fx The values in EBX, ECX, and EDX indicate an AMD microprocessor. When taken in the proper order: s EBX (least significant bit to most significant bit) s EDX (least significant bit to most significant bit) s ECX (least significant bit to most significant bit) they decode to AuthenticAMD When the parameter passed in EAX is 1, the register values returned are: EAX[3:0] EAX[7:4] Stepping ID* model: Am5X86 CPU: Write-through mode = Eh Write-back mode = Fh Family 486 Instruction Set = 4h 0000 RESERVED 00000000h 00000000h 00000001h = all versions The 1 in bit 0 indicates that the FPU is present 10.2 CPUID Instruction The Am5X86 microprocessor family implements the CPUID instruction that makes information available to software about the family, model and stepping of the microprocessor on which it is executing. Support of this instruction is indicated by the presence of a user-modifiable bit in position EFLAGS.21, referred to as the EFLAGS.ID bit. This bit is reset to zero at device reset (RESET or SRESET) for compatibility with existing processor designs. 10.2.1 CPUID Timing CPUID execution timing depends on the selected EAX parameter values (see Table 20). Table 20. CPUID Instruction Description OP Instruction Code 0F A2 CPUID EAX Input Value 0 1 >1 CPU Core Description Clocks 41 AMD string 14 CPU ID Register 9 null registers EAX[11:8] EAX[15:12] EAX[31:16] EBX[31:0] ECX[31:0] EDX[31:0] Note: *Please contact AMD at (800) 222-9323 for stepping ID details. The value returned in EAX after CPUID instruction execution is identical to the value loaded into EDX upon device reset. Software must avoid any dependency upon the state of reserved processor bits. When the parameter passed in EAX is greater than one, register values returned upon instruction execution are: EAX[31:0] EBX[31:0] ECX[31:0] EDX[31:0] 00000000h 00000000h 00000000h 00000000h 10.2.2 CPUID Operation The CPUID instruction requires the user to pass an input parameter to the CPU in the EAX register. The CPU response is returned to the user in registers EAX, EBX, ECX, and EDX. Flags affected: No flags are affected. Exceptions: None 56 Am5X86 Microprocessor PRELIMINARY AMD 11 Electrical Data 11.1.2 Power Decoupling Recommendations Liberal decoupling capacitance should be placed near the microprocessor. The microprocessor, driving its 32bit parallel address and data buses at high frequencies, can cause transient power surges, particularly when driving large capacitive loads. Low inductance capacitors and interconnects are recommended for best high-frequency electrical performance. Inductance can be reduced by shortening circuit board traces between the microprocessor and the decoupling capacitors. Capacitors designed specifically for use with PGA packages are commercially available. 11.1.3 Other Connection Recommendations For reliable operation, always connect unused inputs to an appropriate signal level. Active Low inputs should be connected to VCC through a pull-up resistor. Pull-ups in the range of 20 KΩ are recommended. Active High inputs should be connected to GND. The following sections describe recommended electrical connections for the Am5X86 microprocessors and electrical specifications. 11.1 Power and Grounding 11.1.1 Power Connections Am5X86 microprocessors with 16 Kbytes of cache have modest power requirements. However, the high clock frequency output buffers can cause power surges as multiple output buffers drive new signal levels simultaneously. For clean, on-chip power distribution at high frequency, 23 VCC pins and 28 VSS pins feed the microprocessor in the 168-pin PGA package. The 208-pin SQFP package includes 53 VCC pins and 38 VSS pins. Power and ground connections must be made to all external VCC and VSS pins of the microprocessors. On a circuit board, all VCC pins must connect to a VCC plane. Likewise, all VSS pins must connect to a common GND plane. The Am5X86 microprocessor family requires only 3.3 V as input power. Unlike other 3-V processors, the Am5X86 microprocessor family does not require a VCC5 input of 5 V to indicate the presence of 5-V I/O devices on the system motherboard. For socket compatibility, this pin is INC, allowing the Am5X86 CPU to operate in 3-V sockets in systems that use 5-V I/O. Am5X86 Microprocessor 57 AMD PRELIMINARY ABSOLUTE MAXIMUM RATINGS Case Temperature under Bias . . . – 65°C to +110°C Storage Temperature . . . . . . . . . . – 65°C to +150°C Voltage on any pin with respect to ground . . . . . . – 0.5 V to Vcc +2.6 V Supply voltage with respect to VSS . . . . . . . . . . . . . . – 0.5 V to +4.6 V Stresses above those listed under Absolute Maximum Ratings may cause permanent device failure. Functionality at or above these limits is not implied. Exposure to Absolute Maximum Ratings for extended periods may affect device reliability. OPERATING RANGES Commercial (C) Devices TCASE . . . . . . . . . . . . . . . . . . . . . . . . . . . 0°C to 85°C VCC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 V ±0.3 V Operating Ranges define those limits between which the functionality of the device is guaranteed. DC CHARACTERISTICS over COMMERCIAL operating ranges VCC = 3.3 V ± 0.3 V; TCASE = 0°C to + 85°C Preliminary Info Min Max – 0.3 V 2.0 V +0.8 V VCC + 2.4 V 0.45 V 2.4 V 133 MHz 931 mA Note 1 Note 2 Typical supply current: 825 mA @ 133 MHz. Inputs at rails, outputs unloaded. Typical supply current for Stop Grant or Auto Halt mode: 50 mA @ 133 MHz. Typical supply current in Stop Clock mode is 600 µA. Note 3 Note 4 Note 5 Symbol VIL VIH VOL VOH ICC Parameter Input Low Voltage Input High Voltage Output Low Voltage Output High Voltage Power Supply Current: Notes ICCSTOPGRANT Input Current in Stop Grant or Auto Halt mode: or ICCAUTOHALT 133 MHz ICCSTPCLK ILI IIH IIL ILO CIN CO CCLK Input Current in Stop Clock mode Input Leakage Current Input Leakage Current Input Leakage Current Output Leakage Current Input Capacitance I/O or Output Capacitance CLK Capacitance 93 mA 5 mA ±15 µA 200 µA – 400 µA ±15 µA 10 pF 14 pF 12 pF FC = 1 MHz (Note 6) FC = 1 MHz (Note 6) FC = 1 MHz (Note 6) Notes: 1. This parameter is measured at: Address, Data, BE3 –BE0 = 4.0 mA; Definition, Control = 5.0 mA 2. This parameter is measured at: Address, Data, BE3 –BE0 = –1.0 mA; Definition, Control = –0.9 mA 3. This parameter is for inputs without internal pull-ups or pull-downs and 0 ≤ VIN ≤ VCC. 4. This parameter is for inputs with internal pull-downs and VIH = 2.4 V. 5. This parameter is for inputs with internal pull-ups and VIL = 0.45 V. 6. Not 100% tested. 58 Am5X86 Microprocessor PRELIMINARY AMD SWITCHING CHARACTERISTICS over COMMERCIAL operating ranges The AC specifications, provided in the AC characteristics table, consist of output delays, input setup requirements, and input hold requirements. All AC specifications are relative to the rising edge of the CLK signal. AC specifications measurement is defined by Figure 39. All timings are referenced to 1.5 V unless otherwise specified. Am5X86 microprocessor output delays are specified with minimum and maximum limits, measured as shown. The minimum microprocessor delay times are hold times provided to external circuitry. Input setup and hold times are specified as minimums, defining the smallest acceptable sampling window. Within the sampling window, a synchronous input signal must be stable for correct microprocessor operation. 33-MHz bus (133-MHz operating frequency) VCC = 3.3 V ±0.3 V; TCASE = 0°C to + 85°C; CL = 50 pF unless otherwise specified Symbol Frequency t1 t1a t2 t3 t4 t5 t6 t7 t8 t8a t9 t10 t11 t12 t13 t14 t15 t16 t17 t18 t18a t19 t20 t21 t22 CLK Period CLK Period Stability CLK High Time at 2 V CLK Low Time at 0.8 V CLK Fall Time (2 V–0.8 V) CLK Rise Time (0.8 V–2 V) A31–A2, PWT, PCD, BE3–BE0, M/IO, D/C, CACHE, W/R, ADS, LOCK, FERR, BREQ, HLDA, SMIACT, HITM Valid Delay A31–A2, PWT, PCD, BE3–BE0, M/IO, D/C, CACHE, W/R, ADS, LOCK Float Delay PCHK Valid Delay BLAST, PLOCK, Valid Delay BLAST, PLOCK, Float Delay D31–D0, DP3–DP0 Write Data Valid Delay D31–D0, DP3–DP0 Write Data Float Delay EADS, INV, WB/WT Setup Time EADS, INV, WB/WT Hold Time KEN, BS16, BS8 Setup Time KEN, BS16, BS8 Hold Time RDY, BRDY Setup Time RDY, BRDY Hold Time HOLD, AHOLD Setup Time BOFF Setup Time HOLD, AHOLD, BOFF Hold Time RESET, FLUSH, A20M, NMI, INTR, IGNNE, STPCLK, SRESET, SMI Setup Time RESET, FLUSH, A20M, NMI, INTR, IGNNE, STPCLK, SRESET, SMI Hold Time D31–D0, DP3–DP0, A31–A4 Read Setup Time 3 3 3 3 3 3 3 5 3 5 3 5 3 6 7 3 5 3 5 11 11 3 3 Parameter Preliminary Info Min Max 8 33 30 125 0.1% Unit MHz ns ∆ ns ns ns ns 39 39 39 39 Figure 39 Adjacent Clocks Notes 3 and 4 Note 3 Note 3 Note 3 Note 3 Note 5 14 20 14 14 20 14 20 ns ns ns ns ns ns ns ns ns ns ns ns ns ns ns ns ns ns ns 40 41 42 40 41 40 41 43 43 43 43 44 44 43 43 43 43 43 43, 44 Note 5 Note 5 Note 3 Note 3 Note 3 Notes Note 2 t23 3 ns 43, 44 D32–D0, DP3–DP0, A31–A4 Read Hold Time Notes: 1. Specifications assume CL = 50 pF. I/O Buffer model must be used to determine delays due to loading (trace and component). First Order I/O buffer models for the processor are available. 2. 0-MHz operation guaranteed during stop clock operation. 3. Not 100% tested. Guaranteed by design characterization. 4. For faster transitions (>0.1% between adjacent clocks), use the Stop Clock protocol to switch operating frequency. 5. All timings are referenced at 1.5 V (as illustrated in the listed figures) unless otherwise noted. Am5X86 Microprocessor 59 AMD PRELIMINARY Am5X86 Microprocessor AC Characteristics for Boundary Scan Test Signals at 25 MHz VCC = 3.3 V ±0.3 V; TCASE = 0°C to +85°C; CL = 50 pF unless otherwise specified Preliminary Info Symbol t24 Parameter TCK Frequency TCK Period TCK High Time at 2 V TCK Low Time at 0.8 V TCK Rise Time (0.8 V–2 V) TCK Fall Time (2 V–0.8 V) TDI, TMS Setup Time TDI, TMS Hold Time TDO Valid Delay TDO Float Delay All Outputs (Non-Test) Valid Delay All Outputs (Non-Test) Float Delay All Inputs (Non-Test) Setup Delay All Inputs (Non-Test) Hold Time Min Max 25 Unit MHz ns ns ns Figure 1x Clock 45, 46 45 45 45 45 46 46 46 46 46 46 46 46 Note 2 Note 2 Note 3 Note 3 Note 3 Note 3 Note 3 Note 3 Note 3 Note 3 Note 1 Notes t25 t26 t27 t28 t29 t30 t31 t32 t33 t34 t35 t36 t37 40 10 10 4 4 8 7 3 25 36 3 25 30 8 7 ns ns ns ns ns ns ns ns ns ns Notes: 1. TCK period ≥ CLK period. 2. Rise/Fall times can be relaxed by 1 ns per 10-ns increase in TCK period. 3. Parameter measured from TCK. 60 Am5X86 Microprocessor PRELIMINARY AMD Key to Switching Waveforms Waveform Inputs Must be steady May change from H to L May change from L to H Don’t care; any change permitted Does not apply Outputs Will be steady Will change from H to L Will change from L to H Changing; state unknown Center line is High-impedance “Off” state Figure 39. CLK Waveforms Figure 40. Output Valid Delay Timing Am5X86 Microprocessor 61 AMD PRELIMINARY Figure 41. Maximum Float Delay Timing Figure 42. PCHK Valid Delay Timing 62 Am5X86 Microprocessor PRELIMINARY AMD Figure 43. Input Setup and Hold Timing Figure 44. RDY and BRDY Input Setup and Hold Timing Am5X86 Microprocessor 63 AMD PRELIMINARY Figure 45. TCK Waveforms Figure 46. Test Signal Timing Diagram 64 Am5X86 Microprocessor PRELIMINARY AMD 12 PACKAGE THERMAL SPECIFICATIONS Where: TJ, TA, TCASE = Junction, Ambient, and Case Temperature = Junction-to-Case and Junction-to-Ambient θJC, θJA Thermal Resistance, respectively P = Maximum Power Consumption The Am5X86 microprocessor is specified for operation when TCASE (the case temperature) is within the range of 0°C to +55°C or +85°C. TCASE can be measured in any environment to determine whether the Am5X86 microprocessor is within specified operating range. The case temperature should be measured at the center of the top surface opposite the pins. The ambient temperature (TA) is guaranteed if TCASE is not violated. The ambient temperature can be calculated from θJC and θJA and from the following equations: TJ = TCASE + (P • θJC) TA = TJ – (P • θJA) TCASE = TA + (P • [θJA – θJC]) The values for θJA and θJC are given in Table 21 for the 1.75 sq. in., 168-pin, ceramic PGA. For the 208-pin SQFP plastic package, θJA = 14.0 and θJC = 1.5. Table 22 shows the TA allowable (without exceeding TCASE) at various airflows and TCASE values for the PGA package. Note that TA is greatly improved by attaching a heat sink to the package. P (the maximum power consumption) is calculated by using a maximum ICC value of 931 mA at 3.3 V. Table 23 shows the TA allowable (without exceeding TCASE) for the SQFP package using a maximum ICC value of 931 mA at 3.3 V. Table 21. Thermal Resistance (°C/W) θJC and θJA for the Am5X86 CPU in 168-Pin PGA Package Cooling Mechanism No Heat Sink Heat Sink* Heat Sink* and fan θJA vs. Airflow-Linear ft/min. (m/s) θJC 1.5 2.0 2.0 0 (0) 16.5 12.0 5.0 200 (1.01) 14.0 7.0 4.6 400 (2.03) 12.0 5.0 4.2 600 (3.04) 10.5 4.0 3.8 800 (4.06) 9.5 3.5 3.5 1000 (5.07) 9.0 3.25 3.25 Note: *0.350″ high unidirectional heat sink (Al alloy 6063-T5, 40 mil fin width, 155 mil center-to-center fin spacing) Table 22. Maximum TA at Various Airflows in °C TA by Cooling Type TA without Heat Sink TA with Heat Sink TA with Heat Sink and fan TA without Heat Sink TA with Heat Sink TA with Heat Sink and fan TCASE 55°C 55°C 55°C 85°C 85°C 85°C Clock 133 MHz 133 MHz 133 MHz 133 MHz 133 MHz 133 MHz Airflow-Linear ft/min. (m/sec) 0 200 400 (0) (1.01) (2.03) 8.9°C 16.6°C 22.7°C 24.3°C 39.6°C 45.8°C 45.8°C 47.0°C 48.2°C 38.9°C 46.6°C 52.7°C 54.3°C 69.6°C 75.8°C 75.8°C 77.0°C 78.2°C 600 (3.04) 27.3°C 48.9°C 49.5°C 57.3°C 78.9°C 79.5°C 800 (4.06) 30.4°C 50.4°C 50.4°C 60.4°C 80.4°C 80.4°C 1000 (5.07) 32.0°C 51.2°C 51.2°C 62.0°C 81.2°C 81.2°C Table 23. Maximum TA for SQFP Package by Clock Frequency TCASE 85°C Clock 133 MHz TA 46.6°C Am5X86 Microprocessor 65 AMD PRELIMINARY 13 PHYSICAL DIMENSIONS 168-Pin PGA 1.735 1.765 1.595 1.605 Index Corner Base Plane Seating Plane 1.595 1.605 0.017 0.020 1.735 1.765 0.090 0.110 0.105 0.125 0.140 0.180 0.110 0.140 Bottom View (Pins Facing Up) 0.025 0.045 Side View Notes: 1. All measurements are in inches. 2. Not to scale. For reference only. 3. BSC is an ANSI standard for Basic Space Centering. 66 Am5X86 Microprocessor PRELIMINARY AMD 208-Pin SQFP 30.40 30.80 27.90 28.10 Pin 208 25.50 REF Pin 156 Pin 1 I.D. 25.50 REF 27.90 28.10 30.40 30.80 Pin 52 Pin 104 Top View 0.50 BASIC 3.25 3.45 0.05 Min 0.50 0.75 3.70 Max. Seating Plane Side View Notes: 1. All measurements are in millimeters unless otherwise noted. 2. Not to scale. For reference only. Trademarks AMD, Am386, and Am486 are registered trademarks and Am5X86 is a trademark of Advanced Micro Devices, Inc. FusionPC is a service mark of Advanced Micro Devices, Inc. Microsoft and Windows are registered trademarks of Microsoft Corp. Product names used in this publication are for identification purposes only and may be trademarks of their respective companies. Am5X86 Microprocessor 67