1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2005 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #pragma ident "%Z%%M% %I% %E% SMI" 28 29 /* 30 * Various routines to handle identification 31 * and classification of x86 processors. 32 */ 33 34 #include <sys/types.h> 35 #include <sys/archsystm.h> 36 #include <sys/x86_archext.h> 37 #include <sys/kmem.h> 38 #include <sys/systm.h> 39 #include <sys/cmn_err.h> 40 #include <sys/sunddi.h> 41 #include <sys/sunndi.h> 42 #include <sys/cpuvar.h> 43 #include <sys/processor.h> 44 #include <sys/chip.h> 45 #include <sys/fp.h> 46 #include <sys/controlregs.h> 47 #include <sys/auxv_386.h> 48 #include <sys/bitmap.h> 49 #include <sys/controlregs.h> 50 #include <sys/memnode.h> 51 52 /* 53 * Pass 0 of cpuid feature analysis happens in locore. It contains special code 54 * to recognize Cyrix processors that are not cpuid-compliant, and to deal with 55 * them accordingly. For most modern processors, feature detection occurs here 56 * in pass 1. 57 * 58 * Pass 1 of cpuid feature analysis happens just at the beginning of mlsetup() 59 * for the boot CPU and does the basic analysis that the early kernel needs. 60 * x86_feature is set based on the return value of cpuid_pass1() of the boot 61 * CPU. 62 * 63 * Pass 1 includes: 64 * 65 * o Determining vendor/model/family/stepping and setting x86_type and 66 * x86_vendor accordingly. 67 * o Processing the feature flags returned by the cpuid instruction while 68 * applying any workarounds or tricks for the specific processor. 69 * o Mapping the feature flags into Solaris feature bits (X86_*). 70 * o Processing extended feature flags if supported by the processor, 71 * again while applying specific processor knowledge. 72 * o Determining the CMT characteristics of the system. 73 * 74 * Pass 1 is done on non-boot CPUs during their initialization and the results 75 * are used only as a meager attempt at ensuring that all processors within the 76 * system support the same features. 77 * 78 * Pass 2 of cpuid feature analysis happens just at the beginning 79 * of startup(). It just copies in and corrects the remainder 80 * of the cpuid data we depend on: standard cpuid functions that we didn't 81 * need for pass1 feature analysis, and extended cpuid functions beyond the 82 * simple feature processing done in pass1. 83 * 84 * Pass 3 of cpuid analysis is invoked after basic kernel services; in 85 * particular kernel memory allocation has been made available. It creates a 86 * readable brand string based on the data collected in the first two passes. 87 * 88 * Pass 4 of cpuid analysis is invoked after post_startup() when all 89 * the support infrastructure for various hardware features has been 90 * initialized. It determines which processor features will be reported 91 * to userland via the aux vector. 92 * 93 * All passes are executed on all CPUs, but only the boot CPU determines what 94 * features the kernel will use. 95 * 96 * Much of the worst junk in this file is for the support of processors 97 * that didn't really implement the cpuid instruction properly. 98 * 99 * NOTE: The accessor functions (cpuid_get*) are aware of, and ASSERT upon, 100 * the pass numbers. Accordingly, changes to the pass code may require changes 101 * to the accessor code. 102 */ 103 104 uint_t x86_feature = 0; 105 uint_t x86_vendor = X86_VENDOR_IntelClone; 106 uint_t x86_type = X86_TYPE_OTHER; 107 108 ulong_t cr4_value; 109 uint_t pentiumpro_bug4046376; 110 uint_t pentiumpro_bug4064495; 111 112 uint_t enable486; 113 114 /* 115 * This set of strings are for processors rumored to support the cpuid 116 * instruction, and is used by locore.s to figure out how to set x86_vendor 117 */ 118 const char CyrixInstead[] = "CyrixInstead"; 119 120 struct cpuidr { 121 uint32_t cp_eax; 122 uint32_t cp_ebx; 123 uint32_t cp_ecx; 124 uint32_t cp_edx; 125 }; 126 127 /* 128 * These constants determine how many of the elements of the 129 * cpuid we cache in the cpuid_info data structure; the 130 * remaining elements are accessible via the cpuid instruction. 131 */ 132 133 #define NMAX_CPI_STD 6 /* eax = 0 .. 5 */ 134 #define NMAX_CPI_EXTD 9 /* eax = 0x80000000 .. 0x80000008 */ 135 136 struct cpuid_info { 137 uint_t cpi_pass; /* last pass completed */ 138 /* 139 * standard function information 140 */ 141 uint_t cpi_maxeax; /* fn 0: %eax */ 142 char cpi_vendorstr[13]; /* fn 0: %ebx:%ecx:%edx */ 143 uint_t cpi_vendor; /* enum of cpi_vendorstr */ 144 145 uint_t cpi_family; /* fn 1: extended family */ 146 uint_t cpi_model; /* fn 1: extended model */ 147 uint_t cpi_step; /* fn 1: stepping */ 148 chipid_t cpi_chipid; /* fn 1: %ebx: chip # on ht cpus */ 149 uint_t cpi_brandid; /* fn 1: %ebx: brand ID */ 150 int cpi_clogid; /* fn 1: %ebx: thread # */ 151 uint_t cpi_ncpu_per_chip; 152 153 uint8_t cpi_cacheinfo[16]; /* fn 2: intel-style cache desc */ 154 uint_t cpi_ncache; /* fn 2: number of elements */ 155 struct cpuidr cpi_std[NMAX_CPI_STD]; /* 0 .. 5 */ 156 /* 157 * extended function information 158 */ 159 uint_t cpi_xmaxeax; /* fn 0x80000000: %eax */ 160 char cpi_brandstr[49]; /* fn 0x8000000[234] */ 161 uint8_t cpi_pabits; /* fn 0x80000006: %eax */ 162 uint8_t cpi_vabits; /* fn 0x80000006: %eax */ 163 struct cpuidr cpi_extd[NMAX_CPI_EXTD]; /* 0x80000000 .. 0x80000008 */ 164 /* 165 * supported feature information 166 */ 167 uint32_t cpi_support[4]; 168 #define STD_EDX_FEATURES 0 169 #define AMD_EDX_FEATURES 1 170 #define TM_EDX_FEATURES 2 171 #define STD_ECX_FEATURES 3 172 173 }; 174 175 176 static struct cpuid_info cpuid_info0; 177 178 /* 179 * These bit fields are defined by the Intel Application Note AP-485 180 * "Intel Processor Identification and the CPUID Instruction" 181 */ 182 #define CPI_FAMILY_XTD(cpi) BITX((cpi)->cpi_std[1].cp_eax, 27, 20) 183 #define CPI_MODEL_XTD(cpi) BITX((cpi)->cpi_std[1].cp_eax, 19, 16) 184 #define CPI_TYPE(cpi) BITX((cpi)->cpi_std[1].cp_eax, 13, 12) 185 #define CPI_FAMILY(cpi) BITX((cpi)->cpi_std[1].cp_eax, 11, 8) 186 #define CPI_STEP(cpi) BITX((cpi)->cpi_std[1].cp_eax, 3, 0) 187 #define CPI_MODEL(cpi) BITX((cpi)->cpi_std[1].cp_eax, 7, 4) 188 189 #define CPI_FEATURES_EDX(cpi) ((cpi)->cpi_std[1].cp_edx) 190 #define CPI_FEATURES_ECX(cpi) ((cpi)->cpi_std[1].cp_ecx) 191 #define CPI_FEATURES_XTD_EDX(cpi) ((cpi)->cpi_extd[1].cp_edx) 192 #define CPI_FEATURES_XTD_ECX(cpi) ((cpi)->cpi_extd[1].cp_ecx) 193 194 #define CPI_BRANDID(cpi) BITX((cpi)->cpi_std[1].cp_ebx, 7, 0) 195 #define CPI_CHUNKS(cpi) BITX((cpi)->cpi_std[1].cp_ebx, 15, 7) 196 #define CPI_CPU_COUNT(cpi) BITX((cpi)->cpi_std[1].cp_ebx, 23, 16) 197 #define CPI_APIC_ID(cpi) BITX((cpi)->cpi_std[1].cp_ebx, 31, 24) 198 199 #define CPI_MAXEAX_MAX 0x100 /* sanity control */ 200 #define CPI_XMAXEAX_MAX 0x80000100 201 202 /* 203 * Some undocumented ways of patching the results of the cpuid 204 * instruction to permit running Solaris 10 on future cpus that 205 * we don't currently support. Could be set to non-zero values 206 * via settings in eeprom. 207 */ 208 209 uint32_t cpuid_feature_ecx_include; 210 uint32_t cpuid_feature_ecx_exclude; 211 uint32_t cpuid_feature_edx_include; 212 uint32_t cpuid_feature_edx_exclude; 213 214 uint_t 215 cpuid_pass1(cpu_t *cpu) 216 { 217 uint32_t mask_ecx, mask_edx; 218 uint_t feature = X86_CPUID; 219 struct cpuid_info *cpi; 220 struct cpuidr *cp; 221 int xcpuid; 222 223 /* 224 * By convention, cpu0 is the boot cpu, which is called 225 * before memory allocation is available. Other cpus are 226 * initialized when memory becomes available. 227 */ 228 if (cpu->cpu_id == 0) 229 cpu->cpu_m.mcpu_cpi = cpi = &cpuid_info0; 230 else 231 cpu->cpu_m.mcpu_cpi = cpi = 232 kmem_zalloc(sizeof (*cpi), KM_SLEEP); 233 234 cp = &cpi->cpi_std[0]; 235 cp->cp_eax = __cpuid_insn(0, &cp->cp_ebx, &cp->cp_ecx, &cp->cp_edx); 236 cpi->cpi_maxeax = cp->cp_eax; 237 { 238 uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr; 239 *iptr++ = cp->cp_ebx; 240 *iptr++ = cp->cp_edx; 241 *iptr++ = cp->cp_ecx; 242 *(char *)&cpi->cpi_vendorstr[12] = '\0'; 243 } 244 245 /* 246 * Map the vendor string to a type code 247 */ 248 if (strcmp(cpi->cpi_vendorstr, "GenuineIntel") == 0) 249 cpi->cpi_vendor = X86_VENDOR_Intel; 250 else if (strcmp(cpi->cpi_vendorstr, "AuthenticAMD") == 0) 251 cpi->cpi_vendor = X86_VENDOR_AMD; 252 else if (strcmp(cpi->cpi_vendorstr, "GenuineTMx86") == 0) 253 cpi->cpi_vendor = X86_VENDOR_TM; 254 else if (strcmp(cpi->cpi_vendorstr, CyrixInstead) == 0) 255 /* 256 * CyrixInstead is a variable used by the Cyrix detection code 257 * in locore. 258 */ 259 cpi->cpi_vendor = X86_VENDOR_Cyrix; 260 else if (strcmp(cpi->cpi_vendorstr, "UMC UMC UMC ") == 0) 261 cpi->cpi_vendor = X86_VENDOR_UMC; 262 else if (strcmp(cpi->cpi_vendorstr, "NexGenDriven") == 0) 263 cpi->cpi_vendor = X86_VENDOR_NexGen; 264 else if (strcmp(cpi->cpi_vendorstr, "CentaurHauls") == 0) 265 cpi->cpi_vendor = X86_VENDOR_Centaur; 266 else if (strcmp(cpi->cpi_vendorstr, "RiseRiseRise") == 0) 267 cpi->cpi_vendor = X86_VENDOR_Rise; 268 else if (strcmp(cpi->cpi_vendorstr, "SiS SiS SiS ") == 0) 269 cpi->cpi_vendor = X86_VENDOR_SiS; 270 else if (strcmp(cpi->cpi_vendorstr, "Geode by NSC") == 0) 271 cpi->cpi_vendor = X86_VENDOR_NSC; 272 else 273 cpi->cpi_vendor = X86_VENDOR_IntelClone; 274 275 x86_vendor = cpi->cpi_vendor; /* for compatibility */ 276 277 /* 278 * Limit the range in case of weird hardware 279 */ 280 if (cpi->cpi_maxeax > CPI_MAXEAX_MAX) 281 cpi->cpi_maxeax = CPI_MAXEAX_MAX; 282 if (cpi->cpi_maxeax < 1) 283 goto pass1_done; 284 285 cp = &cpi->cpi_std[1]; 286 cp->cp_eax = __cpuid_insn(1, &cp->cp_ebx, &cp->cp_ecx, &cp->cp_edx); 287 288 /* 289 * Extract identifying constants for easy access. 290 */ 291 cpi->cpi_model = CPI_MODEL(cpi); 292 cpi->cpi_family = CPI_FAMILY(cpi); 293 294 if (cpi->cpi_family == 0xf) { 295 cpi->cpi_family += CPI_FAMILY_XTD(cpi); 296 cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4; 297 } 298 299 cpi->cpi_step = CPI_STEP(cpi); 300 cpi->cpi_brandid = CPI_BRANDID(cpi); 301 302 /* 303 * *default* assumptions: 304 * - believe %edx feature word 305 * - ignore %ecx feature word 306 * - 32-bit virtual and physical addressing 307 */ 308 mask_edx = 0xffffffff; 309 mask_ecx = 0; 310 311 cpi->cpi_pabits = cpi->cpi_vabits = 32; 312 313 switch (cpi->cpi_vendor) { 314 case X86_VENDOR_Intel: 315 if (cpi->cpi_family == 5) 316 x86_type = X86_TYPE_P5; 317 else if (cpi->cpi_family == 6) { 318 x86_type = X86_TYPE_P6; 319 pentiumpro_bug4046376 = 1; 320 pentiumpro_bug4064495 = 1; 321 /* 322 * Clear the SEP bit when it was set erroneously 323 */ 324 if (cpi->cpi_model < 3 && cpi->cpi_step < 3) 325 cp->cp_edx &= ~CPUID_INTC_EDX_SEP; 326 } else if (cpi->cpi_family == 0xf) { 327 x86_type = X86_TYPE_P4; 328 /* 329 * We don't currently depend on any of the %ecx 330 * features until Prescott, so we'll only check 331 * this from P4 onwards. We might want to revisit 332 * that idea later. 333 */ 334 mask_ecx = 0xffffffff; 335 } else if (cpi->cpi_family > 0xf) 336 mask_ecx = 0xffffffff; 337 break; 338 case X86_VENDOR_IntelClone: 339 default: 340 break; 341 case X86_VENDOR_AMD: 342 #if defined(OPTERON_ERRATUM_108) 343 if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) { 344 cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0; 345 cpi->cpi_model = 0xc; 346 } else 347 #endif 348 if (cpi->cpi_family == 5) { 349 /* 350 * AMD K5 and K6 351 * 352 * These CPUs have an incomplete implementation 353 * of MCA/MCE which we mask away. 354 */ 355 mask_edx = 356 CPUID_INTC_EDX_DE | 357 CPUID_INTC_EDX_PSE | 358 CPUID_INTC_EDX_TSC | 359 CPUID_INTC_EDX_MSR | 360 CPUID_INTC_EDX_CX8 | 361 CPUID_INTC_EDX_PGE; 362 if (cpi->cpi_model == 0) { 363 /* 364 * Model 0 uses the wrong (APIC) bit 365 * to indicate PGE. Fix it here. 366 */ 367 if (cp->cp_edx & 0x200) { 368 cp->cp_edx &= ~0x200; 369 cp->cp_edx |= CPUID_INTC_EDX_PGE; 370 } 371 } else if (cpi->cpi_model >= 6) 372 mask_edx |= CPUID_INTC_EDX_MMX; 373 } else if (cpi->cpi_family >= 0xf) { 374 /* SSE3 and CX16, at least, are valid; enable all */ 375 mask_ecx = 0xffffffff; 376 } 377 break; 378 case X86_VENDOR_TM: 379 /* 380 * workaround the NT workaround in CMS 4.1 381 */ 382 if (cpi->cpi_family == 5 && cpi->cpi_model == 4 && 383 (cpi->cpi_step == 2 || cpi->cpi_step == 3)) 384 cp->cp_edx |= CPUID_INTC_EDX_CX8; 385 break; 386 case X86_VENDOR_Centaur: 387 /* 388 * workaround the NT workarounds again 389 */ 390 if (cpi->cpi_family == 6) 391 cp->cp_edx |= CPUID_INTC_EDX_CX8; 392 break; 393 case X86_VENDOR_Cyrix: 394 /* 395 * We rely heavily on the probing in locore 396 * to actually figure out what parts, if any, 397 * of the Cyrix cpuid instruction to believe. 398 */ 399 switch (x86_type) { 400 case X86_TYPE_CYRIX_486: 401 mask_edx = 0; 402 break; 403 case X86_TYPE_CYRIX_6x86: 404 mask_edx = 0; 405 break; 406 case X86_TYPE_CYRIX_6x86L: 407 mask_edx = 408 CPUID_INTC_EDX_DE | 409 CPUID_INTC_EDX_CX8; 410 break; 411 case X86_TYPE_CYRIX_6x86MX: 412 mask_edx = 413 CPUID_INTC_EDX_DE | 414 CPUID_INTC_EDX_MSR | 415 CPUID_INTC_EDX_CX8 | 416 CPUID_INTC_EDX_PGE | 417 CPUID_INTC_EDX_CMOV | 418 CPUID_INTC_EDX_MMX; 419 break; 420 case X86_TYPE_CYRIX_GXm: 421 mask_edx = 422 CPUID_INTC_EDX_MSR | 423 CPUID_INTC_EDX_CX8 | 424 CPUID_INTC_EDX_CMOV | 425 CPUID_INTC_EDX_MMX; 426 break; 427 case X86_TYPE_CYRIX_MediaGX: 428 break; 429 case X86_TYPE_CYRIX_MII: 430 case X86_TYPE_VIA_CYRIX_III: 431 mask_edx = 432 CPUID_INTC_EDX_DE | 433 CPUID_INTC_EDX_TSC | 434 CPUID_INTC_EDX_MSR | 435 CPUID_INTC_EDX_CX8 | 436 CPUID_INTC_EDX_PGE | 437 CPUID_INTC_EDX_CMOV | 438 CPUID_INTC_EDX_MMX; 439 break; 440 default: 441 break; 442 } 443 break; 444 } 445 446 /* 447 * Now we've figured out the masks that determine 448 * which bits we choose to believe, apply the masks 449 * to the feature words, then map the kernel's view 450 * of these feature words into its feature word. 451 */ 452 cp->cp_edx &= mask_edx; 453 cp->cp_ecx &= mask_ecx; 454 455 /* 456 * fold in fix ups 457 */ 458 459 cp->cp_edx |= cpuid_feature_edx_include; 460 cp->cp_edx &= ~cpuid_feature_edx_exclude; 461 462 463 cp->cp_ecx |= cpuid_feature_ecx_include; 464 cp->cp_ecx &= ~cpuid_feature_ecx_exclude; 465 466 if (cp->cp_edx & CPUID_INTC_EDX_PSE) 467 feature |= X86_LARGEPAGE; 468 if (cp->cp_edx & CPUID_INTC_EDX_TSC) 469 feature |= X86_TSC; 470 if (cp->cp_edx & CPUID_INTC_EDX_MSR) 471 feature |= X86_MSR; 472 if (cp->cp_edx & CPUID_INTC_EDX_MTRR) 473 feature |= X86_MTRR; 474 if (cp->cp_edx & CPUID_INTC_EDX_PGE) 475 feature |= X86_PGE; 476 if (cp->cp_edx & CPUID_INTC_EDX_CMOV) 477 feature |= X86_CMOV; 478 if (cp->cp_edx & CPUID_INTC_EDX_MMX) 479 feature |= X86_MMX; 480 if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 && 481 (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0) 482 feature |= X86_MCA; 483 if (cp->cp_edx & CPUID_INTC_EDX_PAE) 484 feature |= X86_PAE; 485 if (cp->cp_edx & CPUID_INTC_EDX_CX8) 486 feature |= X86_CX8; 487 /* 488 * Once this bit was thought questionable, but it looks like it's 489 * back, as of Application Note 485 March 2005 (24161829.pdf) 490 */ 491 if (cp->cp_ecx & CPUID_INTC_ECX_CX16) 492 feature |= X86_CX16; 493 if (cp->cp_edx & CPUID_INTC_EDX_PAT) 494 feature |= X86_PAT; 495 if (cp->cp_edx & CPUID_INTC_EDX_SEP) 496 feature |= X86_SEP; 497 if (cp->cp_edx & CPUID_INTC_EDX_FXSR) { 498 /* 499 * In our implementation, fxsave/fxrstor 500 * are prerequisites before we'll even 501 * try and do SSE things. 502 */ 503 if (cp->cp_edx & CPUID_INTC_EDX_SSE) 504 feature |= X86_SSE; 505 if (cp->cp_edx & CPUID_INTC_EDX_SSE2) 506 feature |= X86_SSE2; 507 if (cp->cp_ecx & CPUID_INTC_ECX_SSE3) 508 feature |= X86_SSE3; 509 } 510 if (cp->cp_edx & CPUID_INTC_EDX_DE) 511 cr4_value |= CR4_DE; 512 513 if (feature & X86_PAE) 514 cpi->cpi_pabits = 36; 515 516 /* 517 * Hyperthreading configuration is slightly tricky on Intel 518 * and pure clones, and even trickier on AMD. 519 * 520 * (AMD chose to set the HTT bit on their CMP processors, 521 * even though they're not actually hyperthreaded. Thus it 522 * takes a bit more work to figure out what's really going 523 * on ... see the handling of the HTvalid bit below) 524 */ 525 if (cp->cp_edx & CPUID_INTC_EDX_HTT) { 526 cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi); 527 if (cpi->cpi_ncpu_per_chip > 1) 528 feature |= X86_HTT; 529 } 530 531 /* 532 * Work on the "extended" feature information, doing 533 * some basic initialization for cpuid_pass2() 534 */ 535 xcpuid = 0; 536 switch (cpi->cpi_vendor) { 537 case X86_VENDOR_Intel: 538 if (cpi->cpi_family >= 0xf) 539 xcpuid++; 540 break; 541 case X86_VENDOR_AMD: 542 if (cpi->cpi_family > 5 || 543 (cpi->cpi_family == 5 && cpi->cpi_model >= 1)) 544 xcpuid++; 545 break; 546 case X86_VENDOR_Cyrix: 547 /* 548 * Only these Cyrix CPUs are -known- to support 549 * extended cpuid operations. 550 */ 551 if (x86_type == X86_TYPE_VIA_CYRIX_III || 552 x86_type == X86_TYPE_CYRIX_GXm) 553 xcpuid++; 554 break; 555 case X86_VENDOR_Centaur: 556 case X86_VENDOR_TM: 557 default: 558 xcpuid++; 559 break; 560 } 561 562 if (xcpuid) { 563 cp = &cpi->cpi_extd[0]; 564 cpi->cpi_xmaxeax = cp->cp_eax = __cpuid_insn(0x80000000, 565 &cp->cp_ebx, &cp->cp_ecx, &cp->cp_edx); 566 } 567 568 if (cpi->cpi_xmaxeax & 0x80000000) { 569 570 if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX) 571 cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX; 572 573 switch (cpi->cpi_vendor) { 574 case X86_VENDOR_Intel: 575 case X86_VENDOR_AMD: 576 if (cpi->cpi_xmaxeax < 0x80000001) 577 break; 578 cp = &cpi->cpi_extd[1]; 579 cp->cp_eax = __cpuid_insn(0x80000001, 580 &cp->cp_ebx, &cp->cp_ecx, &cp->cp_edx); 581 if (cpi->cpi_vendor == X86_VENDOR_AMD && 582 cpi->cpi_family == 5 && 583 cpi->cpi_model == 6 && 584 cpi->cpi_step == 6) { 585 /* 586 * K6 model 6 uses bit 10 to indicate SYSC 587 * Later models use bit 11. Fix it here. 588 */ 589 if (cp->cp_edx & 0x400) { 590 cp->cp_edx &= ~0x400; 591 cp->cp_edx |= CPUID_AMD_EDX_SYSC; 592 } 593 } 594 595 /* 596 * Compute the additions to the kernel's feature word. 597 */ 598 if (cp->cp_edx & CPUID_AMD_EDX_NX) 599 feature |= X86_NX; 600 601 /* 602 * Unless both the HTT bit is set, and the 603 * HTvalid bit is set, then we're not actually 604 * HyperThreaded at all.. 605 */ 606 if (cpi->cpi_vendor == X86_VENDOR_AMD && 607 (feature & X86_HTT) == X86_HTT && 608 (cp->cp_ecx & CPUID_AMD_ECX_HTvalid) == 0) 609 feature &= ~X86_HTT; 610 #if defined(_LP64) 611 /* 612 * It's really tricky to support syscall/sysret in 613 * the i386 kernel; we rely on sysenter/sysexit 614 * instead. In the amd64 kernel, things are -way- 615 * better. 616 */ 617 if (cp->cp_edx & CPUID_AMD_EDX_SYSC) 618 feature |= X86_ASYSC; 619 620 /* 621 * While we're thinking about system calls, note 622 * that AMD processors don't support sysenter 623 * in long mode at all, so don't try to program them. 624 */ 625 if (x86_vendor == X86_VENDOR_AMD) 626 feature &= ~X86_SEP; 627 #endif 628 break; 629 default: 630 break; 631 } 632 633 switch (cpi->cpi_vendor) { 634 case X86_VENDOR_Intel: 635 case X86_VENDOR_AMD: 636 if (cpi->cpi_xmaxeax < 0x80000008) 637 break; 638 cp = &cpi->cpi_extd[8]; 639 cp->cp_eax = __cpuid_insn(0x80000008, 640 &cp->cp_ebx, &cp->cp_ecx, &cp->cp_edx); 641 /* 642 * Virtual and physical address limits from 643 * cpuid override previously guessed values. 644 */ 645 cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0); 646 cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8); 647 648 /* 649 * This -might- be a CMP processor? 650 */ 651 if (cpi->cpi_vendor == X86_VENDOR_AMD) { 652 cpi->cpi_ncpu_per_chip = 653 1 + BITX(cp->cp_ecx, 7, 0); 654 if (cpi->cpi_ncpu_per_chip > 1) 655 feature |= X86_CMP; 656 } 657 break; 658 default: 659 break; 660 } 661 } 662 663 if ((feature & (X86_HTT | X86_CMP)) == 0) { 664 cpi->cpi_chipid = -1; 665 cpi->cpi_clogid = 0; 666 } else if (cpi->cpi_ncpu_per_chip > 1) { 667 uint_t i, cid_shift, apic_id; 668 669 for (i = 1, cid_shift = 0; 670 i < cpi->cpi_ncpu_per_chip; i <<= 1) 671 cid_shift++; 672 apic_id = CPI_APIC_ID(cpi); 673 cpi->cpi_chipid = apic_id >> cid_shift; 674 cpi->cpi_clogid = apic_id & ((1 << cid_shift) - 1); 675 } 676 677 pass1_done: 678 cpi->cpi_pass = 1; 679 return (feature); 680 } 681 682 /* 683 * Make copies of the cpuid table entries we depend on, in 684 * part for ease of parsing now, in part so that we have only 685 * one place to correct any of it, in part for ease of 686 * later export to userland, and in part so we can look at 687 * this stuff in a crash dump. 688 */ 689 690 /*ARGSUSED*/ 691 void 692 cpuid_pass2(cpu_t *cpu) 693 { 694 uint_t n, nmax; 695 int i; 696 struct cpuidr *cp; 697 uint8_t *dp; 698 uint32_t *iptr; 699 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 700 701 ASSERT(cpi->cpi_pass == 1); 702 703 if (cpi->cpi_maxeax < 1) 704 goto pass2_done; 705 706 if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD) 707 nmax = NMAX_CPI_STD; 708 /* 709 * (We already handled n == 0 and n == 1 in pass 1) 710 */ 711 for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) { 712 cp->cp_eax = __cpuid_insn(n, 713 &cp->cp_ebx, &cp->cp_ecx, &cp->cp_edx); 714 switch (n) { 715 case 2: 716 /* 717 * "the lower 8 bits of the %eax register 718 * contain a value that identifies the number 719 * of times the cpuid [instruction] has to be 720 * executed to obtain a complete image of the 721 * processor's caching systems." 722 * 723 * How *do* they make this stuff up? 724 */ 725 cpi->cpi_ncache = sizeof (*cp) * 726 BITX(cp->cp_eax, 7, 0); 727 if (cpi->cpi_ncache == 0) 728 break; 729 cpi->cpi_ncache--; /* skip count byte */ 730 731 /* 732 * Well, for now, rather than attempt to implement 733 * this slightly dubious algorithm, we just look 734 * at the first 15 .. 735 */ 736 if (cpi->cpi_ncache > (sizeof (*cp) - 1)) 737 cpi->cpi_ncache = sizeof (*cp) - 1; 738 739 dp = cpi->cpi_cacheinfo; 740 if (BITX(cp->cp_eax, 31, 31) == 0) { 741 uint8_t *p = (void *)&cp->cp_eax; 742 for (i = 1; i < 3; i++) 743 if (p[i] != 0) 744 *dp++ = p[i]; 745 } 746 if (BITX(cp->cp_ebx, 31, 31) == 0) { 747 uint8_t *p = (void *)&cp->cp_ebx; 748 for (i = 0; i < 4; i++) 749 if (p[i] != 0) 750 *dp++ = p[i]; 751 } 752 if (BITX(cp->cp_ecx, 31, 31) == 0) { 753 uint8_t *p = (void *)&cp->cp_ecx; 754 for (i = 0; i < 4; i++) 755 if (p[i] != 0) 756 *dp++ = p[i]; 757 } 758 if (BITX(cp->cp_edx, 31, 31) == 0) { 759 uint8_t *p = (void *)&cp->cp_edx; 760 for (i = 0; i < 4; i++) 761 if (p[i] != 0) 762 *dp++ = p[i]; 763 } 764 break; 765 case 3: /* Processor serial number, if PSN supported */ 766 case 4: /* Deterministic cache parameters */ 767 case 5: /* Monitor/Mwait parameters */ 768 default: 769 break; 770 } 771 } 772 773 if ((cpi->cpi_xmaxeax & 0x80000000) == 0) 774 goto pass2_done; 775 776 if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD) 777 nmax = NMAX_CPI_EXTD; 778 /* 779 * Copy the extended properties, fixing them as we go. 780 * (We already handled n == 0 and n == 1 in pass 1) 781 */ 782 iptr = (void *)cpi->cpi_brandstr; 783 for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) { 784 cp->cp_eax = __cpuid_insn(n + 0x80000000, 785 &cp->cp_ebx, &cp->cp_ecx, &cp->cp_edx); 786 switch (n) { 787 case 2: 788 case 3: 789 case 4: 790 /* 791 * Extract the brand string 792 */ 793 *iptr++ = cp->cp_eax; 794 *iptr++ = cp->cp_ebx; 795 *iptr++ = cp->cp_ecx; 796 *iptr++ = cp->cp_edx; 797 break; 798 case 5: 799 switch (cpi->cpi_vendor) { 800 case X86_VENDOR_AMD: 801 /* 802 * The Athlon and Duron were the first 803 * parts to report the sizes of the 804 * TLB for large pages. Before then, 805 * we don't trust the data. 806 */ 807 if (cpi->cpi_family < 6 || 808 (cpi->cpi_family == 6 && 809 cpi->cpi_model < 1)) 810 cp->cp_eax = 0; 811 break; 812 default: 813 break; 814 } 815 break; 816 case 6: 817 switch (cpi->cpi_vendor) { 818 case X86_VENDOR_AMD: 819 /* 820 * The Athlon and Duron were the first 821 * AMD parts with L2 TLB's. 822 * Before then, don't trust the data. 823 */ 824 if (cpi->cpi_family < 6 || 825 cpi->cpi_family == 6 && 826 cpi->cpi_model < 1) 827 cp->cp_eax = cp->cp_ebx = 0; 828 /* 829 * AMD Duron rev A0 reports L2 830 * cache size incorrectly as 1K 831 * when it is really 64K 832 */ 833 if (cpi->cpi_family == 6 && 834 cpi->cpi_model == 3 && 835 cpi->cpi_step == 0) { 836 cp->cp_ecx &= 0xffff; 837 cp->cp_ecx |= 0x400000; 838 } 839 break; 840 case X86_VENDOR_Cyrix: /* VIA C3 */ 841 /* 842 * VIA C3 processors are a bit messed 843 * up w.r.t. encoding cache sizes in %ecx 844 */ 845 if (cpi->cpi_family != 6) 846 break; 847 /* 848 * model 7 and 8 were incorrectly encoded 849 * 850 * xxx is model 8 really broken? 851 */ 852 if (cpi->cpi_model == 7 || 853 cpi->cpi_model == 8) 854 cp->cp_ecx = 855 BITX(cp->cp_ecx, 31, 24) << 16 | 856 BITX(cp->cp_ecx, 23, 16) << 12 | 857 BITX(cp->cp_ecx, 15, 8) << 8 | 858 BITX(cp->cp_ecx, 7, 0); 859 /* 860 * model 9 stepping 1 has wrong associativity 861 */ 862 if (cpi->cpi_model == 9 && cpi->cpi_step == 1) 863 cp->cp_ecx |= 8 << 12; 864 break; 865 case X86_VENDOR_Intel: 866 /* 867 * Extended L2 Cache features function. 868 * First appeared on Prescott. 869 */ 870 default: 871 break; 872 } 873 break; 874 default: 875 break; 876 } 877 } 878 879 pass2_done: 880 cpi->cpi_pass = 2; 881 } 882 883 static const char * 884 intel_cpubrand(const struct cpuid_info *cpi) 885 { 886 int i; 887 888 if ((x86_feature & X86_CPUID) == 0 || 889 cpi->cpi_maxeax < 1 || cpi->cpi_family < 5) 890 return ("i486"); 891 892 switch (cpi->cpi_family) { 893 case 5: 894 return ("Intel Pentium(r)"); 895 case 6: 896 switch (cpi->cpi_model) { 897 uint_t celeron, xeon; 898 const struct cpuidr *cp; 899 case 0: 900 case 1: 901 case 2: 902 return ("Intel Pentium(r) Pro"); 903 case 3: 904 case 4: 905 return ("Intel Pentium(r) II"); 906 case 6: 907 return ("Intel Celeron(r)"); 908 case 5: 909 case 7: 910 celeron = xeon = 0; 911 cp = &cpi->cpi_std[2]; /* cache info */ 912 913 for (i = 1; i < 3; i++) { 914 uint_t tmp; 915 916 tmp = (cp->cp_eax >> (8 * i)) & 0xff; 917 if (tmp == 0x40) 918 celeron++; 919 if (tmp >= 0x44 && tmp <= 0x45) 920 xeon++; 921 } 922 923 for (i = 0; i < 2; i++) { 924 uint_t tmp; 925 926 tmp = (cp->cp_ebx >> (8 * i)) & 0xff; 927 if (tmp == 0x40) 928 celeron++; 929 else if (tmp >= 0x44 && tmp <= 0x45) 930 xeon++; 931 } 932 933 for (i = 0; i < 4; i++) { 934 uint_t tmp; 935 936 tmp = (cp->cp_ecx >> (8 * i)) & 0xff; 937 if (tmp == 0x40) 938 celeron++; 939 else if (tmp >= 0x44 && tmp <= 0x45) 940 xeon++; 941 } 942 943 for (i = 0; i < 4; i++) { 944 uint_t tmp; 945 946 tmp = (cp->cp_edx >> (8 * i)) & 0xff; 947 if (tmp == 0x40) 948 celeron++; 949 else if (tmp >= 0x44 && tmp <= 0x45) 950 xeon++; 951 } 952 953 if (celeron) 954 return ("Intel Celeron(r)"); 955 if (xeon) 956 return (cpi->cpi_model == 5 ? 957 "Intel Pentium(r) II Xeon(tm)" : 958 "Intel Pentium(r) III Xeon(tm)"); 959 return (cpi->cpi_model == 5 ? 960 "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" : 961 "Intel Pentium(r) III or Pentium(r) III Xeon(tm)"); 962 default: 963 break; 964 } 965 default: 966 break; 967 } 968 969 if (cpi->cpi_family <= 0xf && cpi->cpi_model <= 0xf && 970 cpi->cpi_brandid != 0) { 971 static const struct { 972 uint_t bt_bid; 973 const char *bt_str; 974 } brand_tbl[] = { 975 { 0x1, "Intel(r) Celeron(r)" }, 976 { 0x2, "Intel(r) Pentium(r) III" }, 977 { 0x3, "Intel(r) Pentium(r) III Xeon(tm)" }, 978 { 0x4, "Intel(r) Pentium(r) III" }, 979 { 0x6, "Mobile Intel(r) Pentium(r) III" }, 980 { 0x7, "Mobile Intel(r) Celeron(r)" }, 981 { 0x8, "Intel(r) Pentium(r) 4" }, 982 { 0x9, "Intel(r) Pentium(r) 4" }, 983 { 0xa, "Intel(r) Celeron(r)" }, 984 { 0xb, "Intel(r) Xeon(tm)" }, 985 { 0xc, "Intel(r) Xeon(tm) MP" }, 986 { 0xe, "Mobile Intel(r) Pentium(r) 4" }, 987 { 0xf, "Mobile Intel(r) Celeron(r)" } 988 }; 989 uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]); 990 uint_t sgn; 991 992 sgn = (cpi->cpi_family << 8) | 993 (cpi->cpi_model << 4) | cpi->cpi_step; 994 995 for (i = 0; i < btblmax; i++) 996 if (brand_tbl[i].bt_bid == cpi->cpi_brandid) 997 break; 998 if (i < btblmax) { 999 if (sgn == 0x6b1 && cpi->cpi_brandid == 3) 1000 return ("Intel(r) Celeron(r)"); 1001 if (sgn < 0xf13 && cpi->cpi_brandid == 0xb) 1002 return ("Intel(r) Xeon(tm) MP"); 1003 if (sgn < 0xf13 && cpi->cpi_brandid == 0xe) 1004 return ("Intel(r) Xeon(tm)"); 1005 return (brand_tbl[i].bt_str); 1006 } 1007 } 1008 1009 return (NULL); 1010 } 1011 1012 static const char * 1013 amd_cpubrand(const struct cpuid_info *cpi) 1014 { 1015 if ((x86_feature & X86_CPUID) == 0 || 1016 cpi->cpi_maxeax < 1 || cpi->cpi_family < 5) 1017 return ("i486 compatible"); 1018 1019 switch (cpi->cpi_family) { 1020 case 5: 1021 switch (cpi->cpi_model) { 1022 case 0: 1023 case 1: 1024 case 2: 1025 case 3: 1026 case 4: 1027 case 5: 1028 return ("AMD-K5(r)"); 1029 case 6: 1030 case 7: 1031 return ("AMD-K6(r)"); 1032 case 8: 1033 return ("AMD-K6(r)-2"); 1034 case 9: 1035 return ("AMD-K6(r)-III"); 1036 default: 1037 return ("AMD (family 5)"); 1038 } 1039 case 6: 1040 switch (cpi->cpi_model) { 1041 case 1: 1042 return ("AMD-K7(tm)"); 1043 case 0: 1044 case 2: 1045 case 4: 1046 return ("AMD Athlon(tm)"); 1047 case 3: 1048 case 7: 1049 return ("AMD Duron(tm)"); 1050 case 6: 1051 case 8: 1052 case 10: 1053 /* 1054 * Use the L2 cache size to distinguish 1055 */ 1056 return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ? 1057 "AMD Athlon(tm)" : "AMD Duron(tm)"); 1058 default: 1059 return ("AMD (family 6)"); 1060 } 1061 default: 1062 break; 1063 } 1064 1065 if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 && 1066 cpi->cpi_brandid != 0) { 1067 switch (BITX(cpi->cpi_brandid, 7, 5)) { 1068 case 3: 1069 return ("AMD Opteron(tm) UP 1xx"); 1070 case 4: 1071 return ("AMD Opteron(tm) DP 2xx"); 1072 case 5: 1073 return ("AMD Opteron(tm) MP 8xx"); 1074 default: 1075 return ("AMD Opteron(tm)"); 1076 } 1077 } 1078 1079 return (NULL); 1080 } 1081 1082 static const char * 1083 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type) 1084 { 1085 if ((x86_feature & X86_CPUID) == 0 || 1086 cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 || 1087 type == X86_TYPE_CYRIX_486) 1088 return ("i486 compatible"); 1089 1090 switch (type) { 1091 case X86_TYPE_CYRIX_6x86: 1092 return ("Cyrix 6x86"); 1093 case X86_TYPE_CYRIX_6x86L: 1094 return ("Cyrix 6x86L"); 1095 case X86_TYPE_CYRIX_6x86MX: 1096 return ("Cyrix 6x86MX"); 1097 case X86_TYPE_CYRIX_GXm: 1098 return ("Cyrix GXm"); 1099 case X86_TYPE_CYRIX_MediaGX: 1100 return ("Cyrix MediaGX"); 1101 case X86_TYPE_CYRIX_MII: 1102 return ("Cyrix M2"); 1103 case X86_TYPE_VIA_CYRIX_III: 1104 return ("VIA Cyrix M3"); 1105 default: 1106 /* 1107 * Have another wild guess .. 1108 */ 1109 if (cpi->cpi_family == 4 && cpi->cpi_model == 9) 1110 return ("Cyrix 5x86"); 1111 else if (cpi->cpi_family == 5) { 1112 switch (cpi->cpi_model) { 1113 case 2: 1114 return ("Cyrix 6x86"); /* Cyrix M1 */ 1115 case 4: 1116 return ("Cyrix MediaGX"); 1117 default: 1118 break; 1119 } 1120 } else if (cpi->cpi_family == 6) { 1121 switch (cpi->cpi_model) { 1122 case 0: 1123 return ("Cyrix 6x86MX"); /* Cyrix M2? */ 1124 case 5: 1125 case 6: 1126 case 7: 1127 case 8: 1128 case 9: 1129 return ("VIA C3"); 1130 default: 1131 break; 1132 } 1133 } 1134 break; 1135 } 1136 return (NULL); 1137 } 1138 1139 /* 1140 * This only gets called in the case that the CPU extended 1141 * feature brand string (0x80000002, 0x80000003, 0x80000004) 1142 * aren't available, or contain null bytes for some reason. 1143 */ 1144 static void 1145 fabricate_brandstr(struct cpuid_info *cpi) 1146 { 1147 const char *brand = NULL; 1148 1149 switch (cpi->cpi_vendor) { 1150 case X86_VENDOR_Intel: 1151 brand = intel_cpubrand(cpi); 1152 break; 1153 case X86_VENDOR_AMD: 1154 brand = amd_cpubrand(cpi); 1155 break; 1156 case X86_VENDOR_Cyrix: 1157 brand = cyrix_cpubrand(cpi, x86_type); 1158 break; 1159 case X86_VENDOR_NexGen: 1160 if (cpi->cpi_family == 5 && cpi->cpi_model == 0) 1161 brand = "NexGen Nx586"; 1162 break; 1163 case X86_VENDOR_Centaur: 1164 if (cpi->cpi_family == 5) 1165 switch (cpi->cpi_model) { 1166 case 4: 1167 brand = "Centaur C6"; 1168 break; 1169 case 8: 1170 brand = "Centaur C2"; 1171 break; 1172 case 9: 1173 brand = "Centaur C3"; 1174 break; 1175 default: 1176 break; 1177 } 1178 break; 1179 case X86_VENDOR_Rise: 1180 if (cpi->cpi_family == 5 && 1181 (cpi->cpi_model == 0 || cpi->cpi_model == 2)) 1182 brand = "Rise mP6"; 1183 break; 1184 case X86_VENDOR_SiS: 1185 if (cpi->cpi_family == 5 && cpi->cpi_model == 0) 1186 brand = "SiS 55x"; 1187 break; 1188 case X86_VENDOR_TM: 1189 if (cpi->cpi_family == 5 && cpi->cpi_model == 4) 1190 brand = "Transmeta Crusoe TM3x00 or TM5x00"; 1191 break; 1192 case X86_VENDOR_NSC: 1193 case X86_VENDOR_UMC: 1194 default: 1195 break; 1196 } 1197 if (brand) { 1198 (void) strcpy((char *)cpi->cpi_brandstr, brand); 1199 return; 1200 } 1201 1202 /* 1203 * If all else fails ... 1204 */ 1205 (void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr), 1206 "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family, 1207 cpi->cpi_model, cpi->cpi_step); 1208 } 1209 1210 /* 1211 * This routine is called just after kernel memory allocation 1212 * becomes available on cpu0, and as part of mp_startup() on 1213 * the other cpus. 1214 * 1215 * Fixup the brand string. 1216 */ 1217 /*ARGSUSED*/ 1218 void 1219 cpuid_pass3(cpu_t *cpu) 1220 { 1221 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 1222 1223 ASSERT(cpi->cpi_pass == 2); 1224 1225 if ((cpi->cpi_xmaxeax & 0x80000000) == 0) { 1226 fabricate_brandstr(cpi); 1227 goto pass3_done; 1228 } 1229 1230 /* 1231 * If we successfully extracted a brand string from the cpuid 1232 * instruction, clean it up by removing leading spaces and 1233 * similar junk. 1234 */ 1235 if (cpi->cpi_brandstr[0]) { 1236 size_t maxlen = sizeof (cpi->cpi_brandstr); 1237 char *src, *dst; 1238 1239 dst = src = (char *)cpi->cpi_brandstr; 1240 src[maxlen - 1] = '\0'; 1241 /* 1242 * strip leading spaces 1243 */ 1244 while (*src == ' ') 1245 src++; 1246 /* 1247 * Remove any 'Genuine' or "Authentic" prefixes 1248 */ 1249 if (strncmp(src, "Genuine ", 8) == 0) 1250 src += 8; 1251 if (strncmp(src, "Authentic ", 10) == 0) 1252 src += 10; 1253 1254 /* 1255 * Now do an in-place copy. 1256 * Map (R) to (r) and (TM) to (tm). 1257 * The era of teletypes is long gone, and there's 1258 * -really- no need to shout. 1259 */ 1260 while (*src != '\0') { 1261 if (src[0] == '(') { 1262 if (strncmp(src + 1, "R)", 2) == 0) { 1263 (void) strncpy(dst, "(r)", 3); 1264 src += 3; 1265 dst += 3; 1266 continue; 1267 } 1268 if (strncmp(src + 1, "TM)", 3) == 0) { 1269 (void) strncpy(dst, "(tm)", 4); 1270 src += 4; 1271 dst += 4; 1272 continue; 1273 } 1274 } 1275 *dst++ = *src++; 1276 } 1277 *dst = '\0'; 1278 1279 /* 1280 * Finally, remove any trailing spaces 1281 */ 1282 while (--dst > cpi->cpi_brandstr) 1283 if (*dst == ' ') 1284 *dst = '\0'; 1285 else 1286 break; 1287 } else 1288 fabricate_brandstr(cpi); 1289 1290 pass3_done: 1291 cpi->cpi_pass = 3; 1292 } 1293 1294 /* 1295 * This routine is called out of bind_hwcap() much later in the life 1296 * of the kernel (post_startup()). The job of this routine is to resolve 1297 * the hardware feature support and kernel support for those features into 1298 * what we're actually going to tell applications via the aux vector. 1299 */ 1300 uint_t 1301 cpuid_pass4(cpu_t *cpu) 1302 { 1303 struct cpuid_info *cpi; 1304 uint_t hwcap_flags = 0; 1305 1306 if (cpu == NULL) 1307 cpu = CPU; 1308 cpi = cpu->cpu_m.mcpu_cpi; 1309 1310 ASSERT(cpi->cpi_pass == 3); 1311 1312 if (cpi->cpi_maxeax >= 1) { 1313 uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES]; 1314 uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES]; 1315 1316 *edx = CPI_FEATURES_EDX(cpi); 1317 *ecx = CPI_FEATURES_ECX(cpi); 1318 1319 /* 1320 * [these require explicit kernel support] 1321 */ 1322 if ((x86_feature & X86_SEP) == 0) 1323 *edx &= ~CPUID_INTC_EDX_SEP; 1324 1325 if ((x86_feature & X86_SSE) == 0) 1326 *edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE); 1327 if ((x86_feature & X86_SSE2) == 0) 1328 *edx &= ~CPUID_INTC_EDX_SSE2; 1329 1330 if ((x86_feature & X86_HTT) == 0) 1331 *edx &= ~CPUID_INTC_EDX_HTT; 1332 1333 if ((x86_feature & X86_SSE3) == 0) 1334 *ecx &= ~CPUID_INTC_ECX_SSE3; 1335 1336 /* 1337 * [no explicit support required beyond x87 fp context] 1338 */ 1339 if (!fpu_exists) 1340 *edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX); 1341 1342 /* 1343 * Now map the supported feature vector to things that we 1344 * think userland will care about. 1345 */ 1346 if (*edx & CPUID_INTC_EDX_SEP) 1347 hwcap_flags |= AV_386_SEP; 1348 if (*edx & CPUID_INTC_EDX_SSE) 1349 hwcap_flags |= AV_386_FXSR | AV_386_SSE; 1350 if (*edx & CPUID_INTC_EDX_SSE2) 1351 hwcap_flags |= AV_386_SSE2; 1352 if (*ecx & CPUID_INTC_ECX_SSE3) 1353 hwcap_flags |= AV_386_SSE3; 1354 1355 if (*edx & CPUID_INTC_EDX_FPU) 1356 hwcap_flags |= AV_386_FPU; 1357 if (*edx & CPUID_INTC_EDX_MMX) 1358 hwcap_flags |= AV_386_MMX; 1359 1360 if (*edx & CPUID_INTC_EDX_TSC) 1361 hwcap_flags |= AV_386_TSC; 1362 if (*edx & CPUID_INTC_EDX_CX8) 1363 hwcap_flags |= AV_386_CX8; 1364 if (*edx & CPUID_INTC_EDX_CMOV) 1365 hwcap_flags |= AV_386_CMOV; 1366 if (*ecx & CPUID_INTC_ECX_MON) 1367 hwcap_flags |= AV_386_MON; 1368 #if defined(CPUID_INTC_ECX_CX16) 1369 if (*ecx & CPUID_INTC_ECX_CX16) 1370 hwcap_flags |= AV_386_CX16; 1371 #endif 1372 } 1373 1374 if (cpuid_is_ht(cpu)) 1375 hwcap_flags |= AV_386_PAUSE; 1376 1377 if (cpi->cpi_xmaxeax < 0x80000001) 1378 goto pass4_done; 1379 1380 switch (cpi->cpi_vendor) { 1381 uint32_t junk, *edx; 1382 1383 case X86_VENDOR_Intel: /* sigh */ 1384 case X86_VENDOR_AMD: 1385 edx = &cpi->cpi_support[AMD_EDX_FEATURES]; 1386 1387 *edx = CPI_FEATURES_XTD_EDX(cpi); 1388 1389 /* 1390 * [no explicit support required beyond 1391 * x87 fp context and exception handlers] 1392 */ 1393 if (!fpu_exists) 1394 *edx &= ~(CPUID_AMD_EDX_MMXamd | 1395 CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx); 1396 1397 if ((x86_feature & X86_ASYSC) == 0) 1398 *edx &= ~CPUID_AMD_EDX_SYSC; 1399 if ((x86_feature & X86_NX) == 0) 1400 *edx &= ~CPUID_AMD_EDX_NX; 1401 #if !defined(_LP64) 1402 *edx &= ~CPUID_AMD_EDX_LM; 1403 #endif 1404 /* 1405 * Now map the supported feature vector to 1406 * things that we think userland will care about. 1407 */ 1408 if (*edx & CPUID_AMD_EDX_SYSC) 1409 hwcap_flags |= AV_386_AMD_SYSC; 1410 if (*edx & CPUID_AMD_EDX_MMXamd) 1411 hwcap_flags |= AV_386_AMD_MMX; 1412 if (*edx & CPUID_AMD_EDX_3DNow) 1413 hwcap_flags |= AV_386_AMD_3DNow; 1414 if (*edx & CPUID_AMD_EDX_3DNowx) 1415 hwcap_flags |= AV_386_AMD_3DNowx; 1416 break; 1417 1418 case X86_VENDOR_TM: 1419 edx = &cpi->cpi_support[TM_EDX_FEATURES]; 1420 (void) __cpuid_insn(0x80860001, &junk, &junk, edx); 1421 break; 1422 1423 default: 1424 break; 1425 } 1426 1427 pass4_done: 1428 cpi->cpi_pass = 4; 1429 return (hwcap_flags); 1430 } 1431 1432 1433 /* 1434 * Simulate the cpuid instruction using the data we previously 1435 * captured about this CPU. We try our best to return the truth 1436 * about the hardware, independently of kernel support. 1437 */ 1438 uint32_t 1439 cpuid_insn(cpu_t *cpu, 1440 uint32_t eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx) 1441 { 1442 struct cpuid_info *cpi; 1443 struct cpuidr *cp; 1444 1445 if (cpu == NULL) 1446 cpu = CPU; 1447 cpi = cpu->cpu_m.mcpu_cpi; 1448 1449 ASSERT(cpuid_checkpass(cpu, 3)); 1450 1451 /* 1452 * CPUID data is cached in two separate places: cpi_std for standard 1453 * CPUID functions, and cpi_extd for extended CPUID functions. 1454 */ 1455 if (eax <= cpi->cpi_maxeax && eax < NMAX_CPI_STD) 1456 cp = &cpi->cpi_std[eax]; 1457 else if (eax >= 0x80000000 && eax <= cpi->cpi_xmaxeax && 1458 eax < 0x80000000 + NMAX_CPI_EXTD) 1459 cp = &cpi->cpi_extd[eax - 0x80000000]; 1460 else 1461 /* 1462 * The caller is asking for data from an input parameter which 1463 * the kernel has not cached. In this case we go fetch from 1464 * the hardware and return the data directly to the user. 1465 */ 1466 return (__cpuid_insn(eax, ebx, ecx, edx)); 1467 *ebx = cp->cp_ebx; 1468 *ecx = cp->cp_ecx; 1469 *edx = cp->cp_edx; 1470 return (cp->cp_eax); 1471 } 1472 1473 int 1474 cpuid_checkpass(cpu_t *cpu, int pass) 1475 { 1476 return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL && 1477 cpu->cpu_m.mcpu_cpi->cpi_pass >= pass); 1478 } 1479 1480 int 1481 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n) 1482 { 1483 ASSERT(cpuid_checkpass(cpu, 3)); 1484 1485 return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr)); 1486 } 1487 1488 int 1489 cpuid_is_ht(cpu_t *cpu) 1490 { 1491 if (cpu == NULL) 1492 cpu = CPU; 1493 1494 ASSERT(cpuid_checkpass(cpu, 1)); 1495 1496 return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0); 1497 } 1498 1499 /* 1500 * AMD and Intel both implement the 64-bit variant of the syscall 1501 * instruction (syscallq), so if there's -any- support for syscall, 1502 * cpuid currently says "yes, we support this". 1503 * 1504 * However, Intel decided to -not- implement the 32-bit variant of the 1505 * syscall instruction, so we provide a predicate to allow our caller 1506 * to test that subtlety here. 1507 */ 1508 /*ARGSUSED*/ 1509 int 1510 cpuid_syscall32_insn(cpu_t *cpu) 1511 { 1512 ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1)); 1513 1514 if (x86_feature & X86_ASYSC) 1515 return (x86_vendor != X86_VENDOR_Intel); 1516 return (0); 1517 } 1518 1519 int 1520 cpuid_getidstr(cpu_t *cpu, char *s, size_t n) 1521 { 1522 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 1523 1524 static const char fmt[] = 1525 "x86 (%s family %d model %d step %d clock %d MHz)"; 1526 static const char fmt_ht[] = 1527 "x86 (chipid 0x%x %s family %d model %d step %d clock %d MHz)"; 1528 1529 ASSERT(cpuid_checkpass(cpu, 1)); 1530 1531 if (cpuid_is_ht(cpu)) 1532 return (snprintf(s, n, fmt_ht, cpi->cpi_chipid, 1533 cpi->cpi_vendorstr, cpi->cpi_family, cpi->cpi_model, 1534 cpi->cpi_step, cpu->cpu_type_info.pi_clock)); 1535 return (snprintf(s, n, fmt, 1536 cpi->cpi_vendorstr, cpi->cpi_family, cpi->cpi_model, 1537 cpi->cpi_step, cpu->cpu_type_info.pi_clock)); 1538 } 1539 1540 const char * 1541 cpuid_getvendorstr(cpu_t *cpu) 1542 { 1543 ASSERT(cpuid_checkpass(cpu, 1)); 1544 return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr); 1545 } 1546 1547 uint_t 1548 cpuid_getvendor(cpu_t *cpu) 1549 { 1550 ASSERT(cpuid_checkpass(cpu, 1)); 1551 return (cpu->cpu_m.mcpu_cpi->cpi_vendor); 1552 } 1553 1554 uint_t 1555 cpuid_getfamily(cpu_t *cpu) 1556 { 1557 ASSERT(cpuid_checkpass(cpu, 1)); 1558 return (cpu->cpu_m.mcpu_cpi->cpi_family); 1559 } 1560 1561 uint_t 1562 cpuid_getmodel(cpu_t *cpu) 1563 { 1564 ASSERT(cpuid_checkpass(cpu, 1)); 1565 return (cpu->cpu_m.mcpu_cpi->cpi_model); 1566 } 1567 1568 uint_t 1569 cpuid_get_ncpu_per_chip(cpu_t *cpu) 1570 { 1571 ASSERT(cpuid_checkpass(cpu, 1)); 1572 return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip); 1573 } 1574 1575 uint_t 1576 cpuid_getstep(cpu_t *cpu) 1577 { 1578 ASSERT(cpuid_checkpass(cpu, 1)); 1579 return (cpu->cpu_m.mcpu_cpi->cpi_step); 1580 } 1581 1582 chipid_t 1583 chip_plat_get_chipid(cpu_t *cpu) 1584 { 1585 ASSERT(cpuid_checkpass(cpu, 1)); 1586 1587 if (cpuid_is_ht(cpu)) 1588 return (cpu->cpu_m.mcpu_cpi->cpi_chipid); 1589 return (cpu->cpu_id); 1590 } 1591 1592 int 1593 chip_plat_get_clogid(cpu_t *cpu) 1594 { 1595 ASSERT(cpuid_checkpass(cpu, 1)); 1596 return (cpu->cpu_m.mcpu_cpi->cpi_clogid); 1597 } 1598 1599 void 1600 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits) 1601 { 1602 struct cpuid_info *cpi; 1603 1604 if (cpu == NULL) 1605 cpu = CPU; 1606 cpi = cpu->cpu_m.mcpu_cpi; 1607 1608 ASSERT(cpuid_checkpass(cpu, 1)); 1609 1610 if (pabits) 1611 *pabits = cpi->cpi_pabits; 1612 if (vabits) 1613 *vabits = cpi->cpi_vabits; 1614 } 1615 1616 /* 1617 * Returns the number of data TLB entries for a corresponding 1618 * pagesize. If it can't be computed, or isn't known, the 1619 * routine returns zero. If you ask about an architecturally 1620 * impossible pagesize, the routine will panic (so that the 1621 * hat implementor knows that things are inconsistent.) 1622 */ 1623 uint_t 1624 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize) 1625 { 1626 struct cpuid_info *cpi; 1627 uint_t dtlb_nent = 0; 1628 1629 if (cpu == NULL) 1630 cpu = CPU; 1631 cpi = cpu->cpu_m.mcpu_cpi; 1632 1633 ASSERT(cpuid_checkpass(cpu, 1)); 1634 1635 /* 1636 * Check the L2 TLB info 1637 */ 1638 if (cpi->cpi_xmaxeax >= 0x80000006) { 1639 struct cpuidr *cp = &cpi->cpi_extd[6]; 1640 1641 switch (pagesize) { 1642 1643 case 4 * 1024: 1644 /* 1645 * All zero in the top 16 bits of the register 1646 * indicates a unified TLB. Size is in low 16 bits. 1647 */ 1648 if ((cp->cp_ebx & 0xffff0000) == 0) 1649 dtlb_nent = cp->cp_ebx & 0x0000ffff; 1650 else 1651 dtlb_nent = BITX(cp->cp_ebx, 27, 16); 1652 break; 1653 1654 case 2 * 1024 * 1024: 1655 if ((cp->cp_eax & 0xffff0000) == 0) 1656 dtlb_nent = cp->cp_eax & 0x0000ffff; 1657 else 1658 dtlb_nent = BITX(cp->cp_eax, 27, 16); 1659 break; 1660 1661 default: 1662 panic("unknown L2 pagesize"); 1663 /*NOTREACHED*/ 1664 } 1665 } 1666 1667 if (dtlb_nent != 0) 1668 return (dtlb_nent); 1669 1670 /* 1671 * No L2 TLB support for this size, try L1. 1672 */ 1673 if (cpi->cpi_xmaxeax >= 0x80000005) { 1674 struct cpuidr *cp = &cpi->cpi_extd[5]; 1675 1676 switch (pagesize) { 1677 case 4 * 1024: 1678 dtlb_nent = BITX(cp->cp_ebx, 23, 16); 1679 break; 1680 case 2 * 1024 * 1024: 1681 dtlb_nent = BITX(cp->cp_eax, 23, 16); 1682 break; 1683 default: 1684 panic("unknown L1 d-TLB pagesize"); 1685 /*NOTREACHED*/ 1686 } 1687 } 1688 1689 return (dtlb_nent); 1690 } 1691 1692 /* 1693 * Return 0 if the erratum is not present or not applicable, positive 1694 * if it is, and negative if the status of the erratum is unknown. 1695 * 1696 * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm) 1697 * Processors" #25759, Rev 3.57, August 2005 1698 */ 1699 int 1700 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum) 1701 { 1702 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 1703 uint_t eax; 1704 1705 if (cpi->cpi_vendor != X86_VENDOR_AMD) 1706 return (0); 1707 1708 eax = cpi->cpi_std[1].cp_eax; 1709 1710 #define SH_B0(eax) (eax == 0xf40 || eax == 0xf50) 1711 #define SH_B3(eax) (eax == 0xf51) 1712 #define B(eax) (SH_B0(eax) | SH_B3(eax)) 1713 1714 #define SH_C0(eax) (eax == 0xf48 || eax == 0xf58) 1715 1716 #define SH_CG(eax) (eax == 0xf4a || eax == 0xf5a || eax == 0xf7a) 1717 #define DH_CG(eax) (eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0) 1718 #define CH_CG(eax) (eax == 0xf82 || eax == 0xfb2) 1719 #define CG(eax) (SH_CG(eax) | DH_CG(eax) | CH_CG(eax)) 1720 1721 #define SH_D0(eax) (eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70) 1722 #define DH_D0(eax) (eax == 0x10fc0 || eax == 0x10ff0) 1723 #define CH_D0(eax) (eax == 0x10f80 || eax == 0x10fb0) 1724 #define D0(eax) (SH_D0(eax) | DH_D0(eax) | CH_D0(eax)) 1725 1726 #define SH_E0(eax) (eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70) 1727 #define JH_E1(eax) (eax == 0x20f10) /* JH8_E0 had 0x20f30 */ 1728 #define DH_E3(eax) (eax == 0x20fc0 || eax == 0x20ff0) 1729 #define SH_E4(eax) (eax == 0x20f51 || eax == 0x20f71) 1730 #define BH_E4(eax) (eax == 0x20fb1) 1731 #define SH_E5(eax) (eax == 0x20f42) 1732 #define DH_E6(eax) (eax == 0x20ff2 || eax == 0x20fc2) 1733 #define JH_E6(eax) (eax == 0x20f12 || eax == 0x20f32) 1734 #define EX(eax) (SH_E0(eax) | JH_E1(eax) | DH_E3(eax) | SH_E4(eax) | \ 1735 BH_E4(eax) | SH_E5(eax) | DH_E6(eax) | JH_E6(eax)) 1736 1737 switch (erratum) { 1738 case 1: 1739 return (1); 1740 case 51: /* what does the asterisk mean? */ 1741 return (B(eax) || SH_C0(eax) || CG(eax)); 1742 case 52: 1743 return (B(eax)); 1744 case 57: 1745 return (1); 1746 case 58: 1747 return (B(eax)); 1748 case 60: 1749 return (1); 1750 case 61: 1751 case 62: 1752 case 63: 1753 case 64: 1754 case 65: 1755 case 66: 1756 case 68: 1757 case 69: 1758 case 70: 1759 case 71: 1760 return (B(eax)); 1761 case 72: 1762 return (SH_B0(eax)); 1763 case 74: 1764 return (B(eax)); 1765 case 75: 1766 return (1); 1767 case 76: 1768 return (B(eax)); 1769 case 77: 1770 return (1); 1771 case 78: 1772 return (B(eax) || SH_C0(eax)); 1773 case 79: 1774 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax)); 1775 case 80: 1776 case 81: 1777 case 82: 1778 return (B(eax)); 1779 case 83: 1780 return (B(eax) || SH_C0(eax) || CG(eax)); 1781 case 85: 1782 return (1); 1783 case 86: 1784 return (SH_C0(eax) || CG(eax)); 1785 case 88: 1786 #if !defined(__amd64) 1787 return (0); 1788 #else 1789 return (B(eax) || SH_C0(eax)); 1790 #endif 1791 case 89: 1792 return (1); 1793 case 90: 1794 return (B(eax) || SH_C0(eax) || CG(eax)); 1795 case 91: 1796 case 92: 1797 return (B(eax) || SH_C0(eax)); 1798 case 93: 1799 return (SH_C0(eax)); 1800 case 94: 1801 return (B(eax) || SH_C0(eax) || CG(eax)); 1802 case 95: 1803 #if !defined(__amd64) 1804 return (0); 1805 #else 1806 return (B(eax) || SH_C0(eax)); 1807 #endif 1808 case 96: 1809 return (B(eax) || SH_C0(eax) || CG(eax)); 1810 case 97: 1811 case 98: 1812 return (SH_C0(eax) || CG(eax)); 1813 case 99: 1814 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax)); 1815 case 100: 1816 return (B(eax) || SH_C0(eax)); 1817 case 101: 1818 case 103: 1819 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax)); 1820 case 104: 1821 return (SH_C0(eax) || CG(eax) || D0(eax)); 1822 case 105: 1823 case 106: 1824 case 107: 1825 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax)); 1826 case 108: 1827 return (DH_CG(eax)); 1828 case 109: 1829 return (SH_C0(eax) || CG(eax) || D0(eax)); 1830 case 110: 1831 return (D0(eax) || EX(eax)); 1832 case 111: 1833 return (CG(eax)); 1834 case 112: 1835 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax)); 1836 case 113: 1837 return (eax == 0x20fc0); 1838 case 114: 1839 return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax)); 1840 case 115: 1841 return (SH_E0(eax) || JH_E1(eax)); 1842 case 116: 1843 return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax)); 1844 case 117: 1845 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax)); 1846 case 118: 1847 return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) || 1848 JH_E6(eax)); 1849 case 121: 1850 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax)); 1851 case 122: 1852 return (SH_C0(eax) || CG(eax) || D0(eax) || EX(eax)); 1853 case 123: 1854 return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax)); 1855 case 131: 1856 return (1); 1857 1858 default: 1859 return (-1); 1860 } 1861 } 1862 1863 static const char assoc_str[] = "associativity"; 1864 static const char line_str[] = "line-size"; 1865 static const char size_str[] = "size"; 1866 1867 static void 1868 add_cache_prop(dev_info_t *devi, const char *label, const char *type, 1869 uint32_t val) 1870 { 1871 char buf[128]; 1872 1873 /* 1874 * ndi_prop_update_int() is used because it is desirable for 1875 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set. 1876 */ 1877 if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf)) 1878 (void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val); 1879 } 1880 1881 /* 1882 * Intel-style cache/tlb description 1883 * 1884 * Standard cpuid level 2 gives a randomly ordered 1885 * selection of tags that index into a table that describes 1886 * cache and tlb properties. 1887 */ 1888 1889 static const char l1_icache_str[] = "l1-icache"; 1890 static const char l1_dcache_str[] = "l1-dcache"; 1891 static const char l2_cache_str[] = "l2-cache"; 1892 static const char itlb4k_str[] = "itlb-4K"; 1893 static const char dtlb4k_str[] = "dtlb-4K"; 1894 static const char itlb4M_str[] = "itlb-4M"; 1895 static const char dtlb4M_str[] = "dtlb-4M"; 1896 static const char itlb424_str[] = "itlb-4K-2M-4M"; 1897 static const char dtlb44_str[] = "dtlb-4K-4M"; 1898 static const char sl1_dcache_str[] = "sectored-l1-dcache"; 1899 static const char sl2_cache_str[] = "sectored-l2-cache"; 1900 static const char itrace_str[] = "itrace-cache"; 1901 static const char sl3_cache_str[] = "sectored-l3-cache"; 1902 1903 static const struct cachetab { 1904 uint8_t ct_code; 1905 uint8_t ct_assoc; 1906 uint16_t ct_line_size; 1907 size_t ct_size; 1908 const char *ct_label; 1909 } intel_ctab[] = { 1910 /* maintain descending order! */ 1911 { 0xb3, 4, 0, 128, dtlb4k_str }, 1912 { 0xb0, 4, 0, 128, itlb4k_str }, 1913 { 0x87, 8, 64, 1024*1024, l2_cache_str}, 1914 { 0x86, 4, 64, 512*1024, l2_cache_str}, 1915 { 0x85, 8, 32, 2*1024*1024, l2_cache_str}, 1916 { 0x84, 8, 32, 1024*1024, l2_cache_str}, 1917 { 0x83, 8, 32, 512*1024, l2_cache_str}, 1918 { 0x82, 8, 32, 256*1024, l2_cache_str}, 1919 { 0x81, 8, 32, 128*1024, l2_cache_str}, /* suspect! */ 1920 { 0x7f, 2, 64, 512*1024, l2_cache_str}, 1921 { 0x7d, 8, 64, 2*1024*1024, sl2_cache_str}, 1922 { 0x7c, 8, 64, 1024*1024, sl2_cache_str}, 1923 { 0x7b, 8, 64, 512*1024, sl2_cache_str}, 1924 { 0x7a, 8, 64, 256*1024, sl2_cache_str}, 1925 { 0x79, 8, 64, 128*1024, sl2_cache_str}, 1926 { 0x78, 8, 64, 1024*1024, l2_cache_str}, 1927 { 0x72, 8, 0, 32*1024, itrace_str}, 1928 { 0x71, 8, 0, 16*1024, itrace_str}, 1929 { 0x70, 8, 0, 12*1024, itrace_str}, 1930 { 0x68, 4, 64, 32*1024, sl1_dcache_str}, 1931 { 0x67, 4, 64, 16*1024, sl1_dcache_str}, 1932 { 0x66, 4, 64, 8*1024, sl1_dcache_str}, 1933 { 0x60, 8, 64, 16*1024, sl1_dcache_str}, 1934 { 0x5d, 0, 0, 256, dtlb44_str}, 1935 { 0x5c, 0, 0, 128, dtlb44_str}, 1936 { 0x5b, 0, 0, 64, dtlb44_str}, 1937 { 0x52, 0, 0, 256, itlb424_str}, 1938 { 0x51, 0, 0, 128, itlb424_str}, 1939 { 0x50, 0, 0, 64, itlb424_str}, 1940 { 0x45, 4, 32, 2*1024*1024, l2_cache_str}, 1941 { 0x44, 4, 32, 1024*1024, l2_cache_str}, 1942 { 0x43, 4, 32, 512*1024, l2_cache_str}, 1943 { 0x42, 4, 32, 256*1024, l2_cache_str}, 1944 { 0x41, 4, 32, 128*1024, l2_cache_str}, 1945 { 0x3c, 4, 64, 256*1024, sl2_cache_str}, 1946 { 0x3b, 2, 64, 128*1024, sl2_cache_str}, 1947 { 0x39, 4, 64, 128*1024, sl2_cache_str}, 1948 { 0x30, 8, 64, 32*1024, l1_icache_str}, 1949 { 0x2c, 8, 64, 32*1024, l1_dcache_str}, 1950 { 0x29, 8, 64, 4096*1024, sl3_cache_str}, 1951 { 0x25, 8, 64, 2048*1024, sl3_cache_str}, 1952 { 0x23, 8, 64, 1024*1024, sl3_cache_str}, 1953 { 0x22, 4, 64, 512*1024, sl3_cache_str}, 1954 { 0x0c, 4, 32, 16*1024, l1_dcache_str}, 1955 { 0x0a, 2, 32, 8*1024, l1_dcache_str}, 1956 { 0x08, 4, 32, 16*1024, l1_icache_str}, 1957 { 0x06, 4, 32, 8*1024, l1_icache_str}, 1958 { 0x04, 4, 0, 8, dtlb4M_str}, 1959 { 0x03, 4, 0, 64, dtlb4k_str}, 1960 { 0x02, 4, 0, 2, itlb4M_str}, 1961 { 0x01, 4, 0, 32, itlb4k_str}, 1962 { 0 } 1963 }; 1964 1965 static const struct cachetab cyrix_ctab[] = { 1966 { 0x70, 4, 0, 32, "tlb-4K" }, 1967 { 0x80, 4, 16, 16*1024, "l1-cache" }, 1968 { 0 } 1969 }; 1970 1971 /* 1972 * Search a cache table for a matching entry 1973 */ 1974 static const struct cachetab * 1975 find_cacheent(const struct cachetab *ct, uint_t code) 1976 { 1977 if (code != 0) { 1978 for (; ct->ct_code != 0; ct++) 1979 if (ct->ct_code <= code) 1980 break; 1981 if (ct->ct_code == code) 1982 return (ct); 1983 } 1984 return (NULL); 1985 } 1986 1987 /* 1988 * Walk the cacheinfo descriptor, applying 'func' to every valid element 1989 * The walk is terminated if the walker returns non-zero. 1990 */ 1991 static void 1992 intel_walk_cacheinfo(struct cpuid_info *cpi, 1993 void *arg, int (*func)(void *, const struct cachetab *)) 1994 { 1995 const struct cachetab *ct; 1996 uint8_t *dp; 1997 int i; 1998 1999 if ((dp = cpi->cpi_cacheinfo) == NULL) 2000 return; 2001 for (i = 0; i < cpi->cpi_ncache; i++, dp++) 2002 if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) { 2003 if (func(arg, ct) != 0) 2004 break; 2005 } 2006 } 2007 2008 /* 2009 * (Like the Intel one, except for Cyrix CPUs) 2010 */ 2011 static void 2012 cyrix_walk_cacheinfo(struct cpuid_info *cpi, 2013 void *arg, int (*func)(void *, const struct cachetab *)) 2014 { 2015 const struct cachetab *ct; 2016 uint8_t *dp; 2017 int i; 2018 2019 if ((dp = cpi->cpi_cacheinfo) == NULL) 2020 return; 2021 for (i = 0; i < cpi->cpi_ncache; i++, dp++) { 2022 /* 2023 * Search Cyrix-specific descriptor table first .. 2024 */ 2025 if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) { 2026 if (func(arg, ct) != 0) 2027 break; 2028 continue; 2029 } 2030 /* 2031 * .. else fall back to the Intel one 2032 */ 2033 if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) { 2034 if (func(arg, ct) != 0) 2035 break; 2036 continue; 2037 } 2038 } 2039 } 2040 2041 /* 2042 * A cacheinfo walker that adds associativity, line-size, and size properties 2043 * to the devinfo node it is passed as an argument. 2044 */ 2045 static int 2046 add_cacheent_props(void *arg, const struct cachetab *ct) 2047 { 2048 dev_info_t *devi = arg; 2049 2050 add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc); 2051 if (ct->ct_line_size != 0) 2052 add_cache_prop(devi, ct->ct_label, line_str, 2053 ct->ct_line_size); 2054 add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size); 2055 return (0); 2056 } 2057 2058 static const char fully_assoc[] = "fully-associative?"; 2059 2060 /* 2061 * AMD style cache/tlb description 2062 * 2063 * Extended functions 5 and 6 directly describe properties of 2064 * tlbs and various cache levels. 2065 */ 2066 static void 2067 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc) 2068 { 2069 switch (assoc) { 2070 case 0: /* reserved; ignore */ 2071 break; 2072 default: 2073 add_cache_prop(devi, label, assoc_str, assoc); 2074 break; 2075 case 0xff: 2076 add_cache_prop(devi, label, fully_assoc, 1); 2077 break; 2078 } 2079 } 2080 2081 static void 2082 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size) 2083 { 2084 if (size == 0) 2085 return; 2086 add_cache_prop(devi, label, size_str, size); 2087 add_amd_assoc(devi, label, assoc); 2088 } 2089 2090 static void 2091 add_amd_cache(dev_info_t *devi, const char *label, 2092 uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size) 2093 { 2094 if (size == 0 || line_size == 0) 2095 return; 2096 add_amd_assoc(devi, label, assoc); 2097 /* 2098 * Most AMD parts have a sectored cache. Multiple cache lines are 2099 * associated with each tag. A sector consists of all cache lines 2100 * associated with a tag. For example, the AMD K6-III has a sector 2101 * size of 2 cache lines per tag. 2102 */ 2103 if (lines_per_tag != 0) 2104 add_cache_prop(devi, label, "lines-per-tag", lines_per_tag); 2105 add_cache_prop(devi, label, line_str, line_size); 2106 add_cache_prop(devi, label, size_str, size * 1024); 2107 } 2108 2109 static void 2110 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc) 2111 { 2112 switch (assoc) { 2113 case 0: /* off */ 2114 break; 2115 case 1: 2116 case 2: 2117 case 4: 2118 add_cache_prop(devi, label, assoc_str, assoc); 2119 break; 2120 case 6: 2121 add_cache_prop(devi, label, assoc_str, 8); 2122 break; 2123 case 8: 2124 add_cache_prop(devi, label, assoc_str, 16); 2125 break; 2126 case 0xf: 2127 add_cache_prop(devi, label, fully_assoc, 1); 2128 break; 2129 default: /* reserved; ignore */ 2130 break; 2131 } 2132 } 2133 2134 static void 2135 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size) 2136 { 2137 if (size == 0 || assoc == 0) 2138 return; 2139 add_amd_l2_assoc(devi, label, assoc); 2140 add_cache_prop(devi, label, size_str, size); 2141 } 2142 2143 static void 2144 add_amd_l2_cache(dev_info_t *devi, const char *label, 2145 uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size) 2146 { 2147 if (size == 0 || assoc == 0 || line_size == 0) 2148 return; 2149 add_amd_l2_assoc(devi, label, assoc); 2150 if (lines_per_tag != 0) 2151 add_cache_prop(devi, label, "lines-per-tag", lines_per_tag); 2152 add_cache_prop(devi, label, line_str, line_size); 2153 add_cache_prop(devi, label, size_str, size * 1024); 2154 } 2155 2156 static void 2157 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi) 2158 { 2159 struct cpuidr *cp; 2160 2161 if (cpi->cpi_xmaxeax < 0x80000005) 2162 return; 2163 cp = &cpi->cpi_extd[5]; 2164 2165 /* 2166 * 4M/2M L1 TLB configuration 2167 * 2168 * We report the size for 2M pages because AMD uses two 2169 * TLB entries for one 4M page. 2170 */ 2171 add_amd_tlb(devi, "dtlb-2M", 2172 BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16)); 2173 add_amd_tlb(devi, "itlb-2M", 2174 BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0)); 2175 2176 /* 2177 * 4K L1 TLB configuration 2178 */ 2179 2180 switch (cpi->cpi_vendor) { 2181 uint_t nentries; 2182 case X86_VENDOR_TM: 2183 if (cpi->cpi_family >= 5) { 2184 /* 2185 * Crusoe processors have 256 TLB entries, but 2186 * cpuid data format constrains them to only 2187 * reporting 255 of them. 2188 */ 2189 if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255) 2190 nentries = 256; 2191 /* 2192 * Crusoe processors also have a unified TLB 2193 */ 2194 add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24), 2195 nentries); 2196 break; 2197 } 2198 /*FALLTHROUGH*/ 2199 default: 2200 add_amd_tlb(devi, itlb4k_str, 2201 BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16)); 2202 add_amd_tlb(devi, dtlb4k_str, 2203 BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0)); 2204 break; 2205 } 2206 2207 /* 2208 * data L1 cache configuration 2209 */ 2210 2211 add_amd_cache(devi, l1_dcache_str, 2212 BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16), 2213 BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0)); 2214 2215 /* 2216 * code L1 cache configuration 2217 */ 2218 2219 add_amd_cache(devi, l1_icache_str, 2220 BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16), 2221 BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0)); 2222 2223 if (cpi->cpi_xmaxeax < 0x80000006) 2224 return; 2225 cp = &cpi->cpi_extd[6]; 2226 2227 /* Check for a unified L2 TLB for large pages */ 2228 2229 if (BITX(cp->cp_eax, 31, 16) == 0) 2230 add_amd_l2_tlb(devi, "l2-tlb-2M", 2231 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0)); 2232 else { 2233 add_amd_l2_tlb(devi, "l2-dtlb-2M", 2234 BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16)); 2235 add_amd_l2_tlb(devi, "l2-itlb-2M", 2236 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0)); 2237 } 2238 2239 /* Check for a unified L2 TLB for 4K pages */ 2240 2241 if (BITX(cp->cp_ebx, 31, 16) == 0) { 2242 add_amd_l2_tlb(devi, "l2-tlb-4K", 2243 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0)); 2244 } else { 2245 add_amd_l2_tlb(devi, "l2-dtlb-4K", 2246 BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16)); 2247 add_amd_l2_tlb(devi, "l2-itlb-4K", 2248 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0)); 2249 } 2250 2251 add_amd_l2_cache(devi, l2_cache_str, 2252 BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12), 2253 BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0)); 2254 } 2255 2256 /* 2257 * There are two basic ways that the x86 world describes it cache 2258 * and tlb architecture - Intel's way and AMD's way. 2259 * 2260 * Return which flavor of cache architecture we should use 2261 */ 2262 static int 2263 x86_which_cacheinfo(struct cpuid_info *cpi) 2264 { 2265 switch (cpi->cpi_vendor) { 2266 case X86_VENDOR_Intel: 2267 if (cpi->cpi_maxeax >= 2) 2268 return (X86_VENDOR_Intel); 2269 break; 2270 case X86_VENDOR_AMD: 2271 /* 2272 * The K5 model 1 was the first part from AMD that reported 2273 * cache sizes via extended cpuid functions. 2274 */ 2275 if (cpi->cpi_family > 5 || 2276 (cpi->cpi_family == 5 && cpi->cpi_model >= 1)) 2277 return (X86_VENDOR_AMD); 2278 break; 2279 case X86_VENDOR_TM: 2280 if (cpi->cpi_family >= 5) 2281 return (X86_VENDOR_AMD); 2282 /*FALLTHROUGH*/ 2283 default: 2284 /* 2285 * If they have extended CPU data for 0x80000005 2286 * then we assume they have AMD-format cache 2287 * information. 2288 * 2289 * If not, and the vendor happens to be Cyrix, 2290 * then try our-Cyrix specific handler. 2291 * 2292 * If we're not Cyrix, then assume we're using Intel's 2293 * table-driven format instead. 2294 */ 2295 if (cpi->cpi_xmaxeax >= 0x80000005) 2296 return (X86_VENDOR_AMD); 2297 else if (cpi->cpi_vendor == X86_VENDOR_Cyrix) 2298 return (X86_VENDOR_Cyrix); 2299 else if (cpi->cpi_maxeax >= 2) 2300 return (X86_VENDOR_Intel); 2301 break; 2302 } 2303 return (-1); 2304 } 2305 2306 /* 2307 * create a node for the given cpu under the prom root node. 2308 * Also, create a cpu node in the device tree. 2309 */ 2310 static dev_info_t *cpu_nex_devi = NULL; 2311 static kmutex_t cpu_node_lock; 2312 2313 /* 2314 * Called from post_startup() and mp_startup() 2315 */ 2316 void 2317 add_cpunode2devtree(processorid_t cpu_id, struct cpuid_info *cpi) 2318 { 2319 dev_info_t *cpu_devi; 2320 int create; 2321 2322 mutex_enter(&cpu_node_lock); 2323 2324 /* 2325 * create a nexus node for all cpus identified as 'cpu_id' under 2326 * the root node. 2327 */ 2328 if (cpu_nex_devi == NULL) { 2329 if (ndi_devi_alloc(ddi_root_node(), "cpus", 2330 (pnode_t)DEVI_SID_NODEID, &cpu_nex_devi) != NDI_SUCCESS) { 2331 mutex_exit(&cpu_node_lock); 2332 return; 2333 } 2334 (void) ndi_devi_online(cpu_nex_devi, 0); 2335 } 2336 2337 /* 2338 * create a child node for cpu identified as 'cpu_id' 2339 */ 2340 cpu_devi = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID, 2341 cpu_id); 2342 if (cpu_devi == NULL) { 2343 mutex_exit(&cpu_node_lock); 2344 return; 2345 } 2346 2347 /* device_type */ 2348 2349 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi, 2350 "device_type", "cpu"); 2351 2352 /* reg */ 2353 2354 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2355 "reg", cpu_id); 2356 2357 /* cpu-mhz, and clock-frequency */ 2358 2359 if (cpu_freq > 0) { 2360 long long mul; 2361 2362 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2363 "cpu-mhz", cpu_freq); 2364 2365 if ((mul = cpu_freq * 1000000LL) <= INT_MAX) 2366 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2367 "clock-frequency", (int)mul); 2368 } 2369 2370 (void) ndi_devi_online(cpu_devi, 0); 2371 2372 if ((x86_feature & X86_CPUID) == 0) { 2373 mutex_exit(&cpu_node_lock); 2374 return; 2375 } 2376 2377 /* vendor-id */ 2378 2379 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi, 2380 "vendor-id", cpi->cpi_vendorstr); 2381 2382 if (cpi->cpi_maxeax == 0) { 2383 mutex_exit(&cpu_node_lock); 2384 return; 2385 } 2386 2387 /* 2388 * family, model, and step 2389 */ 2390 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2391 "family", CPI_FAMILY(cpi)); 2392 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2393 "cpu-model", CPI_MODEL(cpi)); 2394 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2395 "stepping-id", CPI_STEP(cpi)); 2396 2397 /* type */ 2398 2399 switch (cpi->cpi_vendor) { 2400 case X86_VENDOR_Intel: 2401 create = 1; 2402 break; 2403 default: 2404 create = 0; 2405 break; 2406 } 2407 if (create) 2408 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2409 "type", CPI_TYPE(cpi)); 2410 2411 /* ext-family */ 2412 2413 switch (cpi->cpi_vendor) { 2414 case X86_VENDOR_Intel: 2415 case X86_VENDOR_AMD: 2416 create = cpi->cpi_family >= 0xf; 2417 break; 2418 default: 2419 create = 0; 2420 break; 2421 } 2422 if (create) 2423 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2424 "ext-family", CPI_FAMILY_XTD(cpi)); 2425 2426 /* ext-model */ 2427 2428 switch (cpi->cpi_vendor) { 2429 case X86_VENDOR_Intel: 2430 case X86_VENDOR_AMD: 2431 create = CPI_MODEL(cpi) == 0xf; 2432 break; 2433 default: 2434 create = 0; 2435 break; 2436 } 2437 if (create) 2438 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2439 "ext-model", CPI_MODEL_XTD(cpi)); 2440 2441 /* generation */ 2442 2443 switch (cpi->cpi_vendor) { 2444 case X86_VENDOR_AMD: 2445 /* 2446 * AMD K5 model 1 was the first part to support this 2447 */ 2448 create = cpi->cpi_xmaxeax >= 0x80000001; 2449 break; 2450 default: 2451 create = 0; 2452 break; 2453 } 2454 if (create) 2455 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2456 "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8)); 2457 2458 /* brand-id */ 2459 2460 switch (cpi->cpi_vendor) { 2461 case X86_VENDOR_Intel: 2462 /* 2463 * brand id first appeared on Pentium III Xeon model 8, 2464 * and Celeron model 8 processors and Opteron 2465 */ 2466 create = cpi->cpi_family > 6 || 2467 (cpi->cpi_family == 6 && cpi->cpi_model >= 8); 2468 break; 2469 case X86_VENDOR_AMD: 2470 create = cpi->cpi_family >= 0xf; 2471 break; 2472 default: 2473 create = 0; 2474 break; 2475 } 2476 if (create && cpi->cpi_brandid != 0) { 2477 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2478 "brand-id", cpi->cpi_brandid); 2479 } 2480 2481 /* chunks, and apic-id */ 2482 2483 switch (cpi->cpi_vendor) { 2484 case X86_VENDOR_Intel: 2485 case X86_VENDOR_AMD: 2486 /* 2487 * first available on Pentium IV and Opteron (K8) 2488 */ 2489 create = cpi->cpi_family >= 0xf; 2490 break; 2491 default: 2492 create = 0; 2493 break; 2494 } 2495 if (create) { 2496 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2497 "chunks", CPI_CHUNKS(cpi)); 2498 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2499 "apic-id", CPI_APIC_ID(cpi)); 2500 if (cpi->cpi_chipid >= 0) 2501 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2502 "chip#", cpi->cpi_chipid); 2503 } 2504 2505 /* cpuid-features */ 2506 2507 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2508 "cpuid-features", CPI_FEATURES_EDX(cpi)); 2509 2510 2511 /* cpuid-features-ecx */ 2512 2513 switch (cpi->cpi_vendor) { 2514 case X86_VENDOR_Intel: 2515 create = cpi->cpi_family >= 0xf; 2516 break; 2517 default: 2518 create = 0; 2519 break; 2520 } 2521 if (create) 2522 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2523 "cpuid-features-ecx", CPI_FEATURES_ECX(cpi)); 2524 2525 /* ext-cpuid-features */ 2526 2527 switch (cpi->cpi_vendor) { 2528 case X86_VENDOR_AMD: 2529 case X86_VENDOR_Cyrix: 2530 case X86_VENDOR_TM: 2531 case X86_VENDOR_Centaur: 2532 /* 2533 * The extended cpuid features are not relevant on 2534 * Intel but are available from the AMD K5 model 1 2535 * and most Cyrix GXm and later. 2536 */ 2537 create = cpi->cpi_xmaxeax >= 0x80000001; 2538 break; 2539 default: 2540 create = 0; 2541 break; 2542 } 2543 if (create) 2544 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 2545 "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi)); 2546 2547 /* 2548 * Brand String first appeared in Intel Pentium IV, AMD K5 2549 * model 1, and Cyrix GXm. On earlier models we try and 2550 * simulate something similar .. so this string should always 2551 * same -something- about the processor, however lame. 2552 */ 2553 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi, 2554 "brand-string", cpi->cpi_brandstr); 2555 2556 /* 2557 * Finally, cache and tlb information 2558 */ 2559 switch (x86_which_cacheinfo(cpi)) { 2560 case X86_VENDOR_Intel: 2561 intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props); 2562 break; 2563 case X86_VENDOR_Cyrix: 2564 cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props); 2565 break; 2566 case X86_VENDOR_AMD: 2567 amd_cache_info(cpi, cpu_devi); 2568 break; 2569 default: 2570 break; 2571 } 2572 2573 mutex_exit(&cpu_node_lock); 2574 } 2575 2576 struct l2info { 2577 int *l2i_csz; 2578 int *l2i_lsz; 2579 int *l2i_assoc; 2580 int l2i_ret; 2581 }; 2582 2583 /* 2584 * A cacheinfo walker that fetches the size, line-size and associativity 2585 * of the L2 cache 2586 */ 2587 static int 2588 intel_l2cinfo(void *arg, const struct cachetab *ct) 2589 { 2590 struct l2info *l2i = arg; 2591 int *ip; 2592 2593 if (ct->ct_label != l2_cache_str && 2594 ct->ct_label != sl2_cache_str) 2595 return (0); /* not an L2 -- keep walking */ 2596 2597 if ((ip = l2i->l2i_csz) != NULL) 2598 *ip = ct->ct_size; 2599 if ((ip = l2i->l2i_lsz) != NULL) 2600 *ip = ct->ct_line_size; 2601 if ((ip = l2i->l2i_assoc) != NULL) 2602 *ip = ct->ct_assoc; 2603 l2i->l2i_ret = ct->ct_size; 2604 return (1); /* was an L2 -- terminate walk */ 2605 } 2606 2607 static void 2608 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i) 2609 { 2610 struct cpuidr *cp; 2611 uint_t size, assoc; 2612 int *ip; 2613 2614 if (cpi->cpi_xmaxeax < 0x80000006) 2615 return; 2616 cp = &cpi->cpi_extd[6]; 2617 2618 if ((assoc = BITX(cp->cp_ecx, 15, 12)) != 0 && 2619 (size = BITX(cp->cp_ecx, 31, 16)) != 0) { 2620 uint_t cachesz = size * 1024; 2621 2622 2623 if ((ip = l2i->l2i_csz) != NULL) 2624 *ip = cachesz; 2625 if ((ip = l2i->l2i_lsz) != NULL) 2626 *ip = BITX(cp->cp_ecx, 7, 0); 2627 if ((ip = l2i->l2i_assoc) != NULL) 2628 *ip = assoc; 2629 l2i->l2i_ret = cachesz; 2630 } 2631 } 2632 2633 int 2634 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc) 2635 { 2636 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 2637 struct l2info __l2info, *l2i = &__l2info; 2638 2639 l2i->l2i_csz = csz; 2640 l2i->l2i_lsz = lsz; 2641 l2i->l2i_assoc = assoc; 2642 l2i->l2i_ret = -1; 2643 2644 switch (x86_which_cacheinfo(cpi)) { 2645 case X86_VENDOR_Intel: 2646 intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo); 2647 break; 2648 case X86_VENDOR_Cyrix: 2649 cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo); 2650 break; 2651 case X86_VENDOR_AMD: 2652 amd_l2cacheinfo(cpi, l2i); 2653 break; 2654 default: 2655 break; 2656 } 2657 return (l2i->l2i_ret); 2658 } 2659