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 (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #pragma ident "%Z%%M% %I% %E% SMI" 27 28 /* 29 * Various routines to handle identification 30 * and classification of x86 processors. 31 */ 32 33 #include <sys/types.h> 34 #include <sys/archsystm.h> 35 #include <sys/x86_archext.h> 36 #include <sys/kmem.h> 37 #include <sys/systm.h> 38 #include <sys/cmn_err.h> 39 #include <sys/sunddi.h> 40 #include <sys/sunndi.h> 41 #include <sys/cpuvar.h> 42 #include <sys/processor.h> 43 #include <sys/sysmacros.h> 44 #include <sys/pg.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/memnode.h> 50 51 /* 52 * Pass 0 of cpuid feature analysis happens in locore. It contains special code 53 * to recognize Cyrix processors that are not cpuid-compliant, and to deal with 54 * them accordingly. For most modern processors, feature detection occurs here 55 * in pass 1. 56 * 57 * Pass 1 of cpuid feature analysis happens just at the beginning of mlsetup() 58 * for the boot CPU and does the basic analysis that the early kernel needs. 59 * x86_feature is set based on the return value of cpuid_pass1() of the boot 60 * CPU. 61 * 62 * Pass 1 includes: 63 * 64 * o Determining vendor/model/family/stepping and setting x86_type and 65 * x86_vendor accordingly. 66 * o Processing the feature flags returned by the cpuid instruction while 67 * applying any workarounds or tricks for the specific processor. 68 * o Mapping the feature flags into Solaris feature bits (X86_*). 69 * o Processing extended feature flags if supported by the processor, 70 * again while applying specific processor knowledge. 71 * o Determining the CMT characteristics of the system. 72 * 73 * Pass 1 is done on non-boot CPUs during their initialization and the results 74 * are used only as a meager attempt at ensuring that all processors within the 75 * system support the same features. 76 * 77 * Pass 2 of cpuid feature analysis happens just at the beginning 78 * of startup(). It just copies in and corrects the remainder 79 * of the cpuid data we depend on: standard cpuid functions that we didn't 80 * need for pass1 feature analysis, and extended cpuid functions beyond the 81 * simple feature processing done in pass1. 82 * 83 * Pass 3 of cpuid analysis is invoked after basic kernel services; in 84 * particular kernel memory allocation has been made available. It creates a 85 * readable brand string based on the data collected in the first two passes. 86 * 87 * Pass 4 of cpuid analysis is invoked after post_startup() when all 88 * the support infrastructure for various hardware features has been 89 * initialized. It determines which processor features will be reported 90 * to userland via the aux vector. 91 * 92 * All passes are executed on all CPUs, but only the boot CPU determines what 93 * features the kernel will use. 94 * 95 * Much of the worst junk in this file is for the support of processors 96 * that didn't really implement the cpuid instruction properly. 97 * 98 * NOTE: The accessor functions (cpuid_get*) are aware of, and ASSERT upon, 99 * the pass numbers. Accordingly, changes to the pass code may require changes 100 * to the accessor code. 101 */ 102 103 uint_t x86_feature = 0; 104 uint_t x86_vendor = X86_VENDOR_IntelClone; 105 uint_t x86_type = X86_TYPE_OTHER; 106 107 uint_t pentiumpro_bug4046376; 108 uint_t pentiumpro_bug4064495; 109 110 uint_t enable486; 111 112 /* 113 * This set of strings are for processors rumored to support the cpuid 114 * instruction, and is used by locore.s to figure out how to set x86_vendor 115 */ 116 const char CyrixInstead[] = "CyrixInstead"; 117 118 /* 119 * monitor/mwait info. 120 * 121 * size_actual and buf_actual are the real address and size allocated to get 122 * proper mwait_buf alignement. buf_actual and size_actual should be passed 123 * to kmem_free(). Currently kmem_alloc() and mwait happen to both use 124 * processor cache-line alignment, but this is not guarantied in the furture. 125 */ 126 struct mwait_info { 127 size_t mon_min; /* min size to avoid missed wakeups */ 128 size_t mon_max; /* size to avoid false wakeups */ 129 size_t size_actual; /* size actually allocated */ 130 void *buf_actual; /* memory actually allocated */ 131 uint32_t support; /* processor support of monitor/mwait */ 132 }; 133 134 /* 135 * These constants determine how many of the elements of the 136 * cpuid we cache in the cpuid_info data structure; the 137 * remaining elements are accessible via the cpuid instruction. 138 */ 139 140 #define NMAX_CPI_STD 6 /* eax = 0 .. 5 */ 141 #define NMAX_CPI_EXTD 9 /* eax = 0x80000000 .. 0x80000008 */ 142 143 struct cpuid_info { 144 uint_t cpi_pass; /* last pass completed */ 145 /* 146 * standard function information 147 */ 148 uint_t cpi_maxeax; /* fn 0: %eax */ 149 char cpi_vendorstr[13]; /* fn 0: %ebx:%ecx:%edx */ 150 uint_t cpi_vendor; /* enum of cpi_vendorstr */ 151 152 uint_t cpi_family; /* fn 1: extended family */ 153 uint_t cpi_model; /* fn 1: extended model */ 154 uint_t cpi_step; /* fn 1: stepping */ 155 chipid_t cpi_chipid; /* fn 1: %ebx: chip # on ht cpus */ 156 uint_t cpi_brandid; /* fn 1: %ebx: brand ID */ 157 int cpi_clogid; /* fn 1: %ebx: thread # */ 158 uint_t cpi_ncpu_per_chip; /* fn 1: %ebx: logical cpu count */ 159 uint8_t cpi_cacheinfo[16]; /* fn 2: intel-style cache desc */ 160 uint_t cpi_ncache; /* fn 2: number of elements */ 161 uint_t cpi_ncpu_shr_last_cache; /* fn 4: %eax: ncpus sharing cache */ 162 id_t cpi_last_lvl_cacheid; /* fn 4: %eax: derived cache id */ 163 uint_t cpi_std_4_size; /* fn 4: number of fn 4 elements */ 164 struct cpuid_regs **cpi_std_4; /* fn 4: %ecx == 0 .. fn4_size */ 165 struct cpuid_regs cpi_std[NMAX_CPI_STD]; /* 0 .. 5 */ 166 /* 167 * extended function information 168 */ 169 uint_t cpi_xmaxeax; /* fn 0x80000000: %eax */ 170 char cpi_brandstr[49]; /* fn 0x8000000[234] */ 171 uint8_t cpi_pabits; /* fn 0x80000006: %eax */ 172 uint8_t cpi_vabits; /* fn 0x80000006: %eax */ 173 struct cpuid_regs cpi_extd[NMAX_CPI_EXTD]; /* 0x8000000[0-8] */ 174 id_t cpi_coreid; /* same coreid => strands share core */ 175 int cpi_pkgcoreid; /* core number within single package */ 176 uint_t cpi_ncore_per_chip; /* AMD: fn 0x80000008: %ecx[7-0] */ 177 /* Intel: fn 4: %eax[31-26] */ 178 /* 179 * supported feature information 180 */ 181 uint32_t cpi_support[5]; 182 #define STD_EDX_FEATURES 0 183 #define AMD_EDX_FEATURES 1 184 #define TM_EDX_FEATURES 2 185 #define STD_ECX_FEATURES 3 186 #define AMD_ECX_FEATURES 4 187 /* 188 * Synthesized information, where known. 189 */ 190 uint32_t cpi_chiprev; /* See X86_CHIPREV_* in x86_archext.h */ 191 const char *cpi_chiprevstr; /* May be NULL if chiprev unknown */ 192 uint32_t cpi_socket; /* Chip package/socket type */ 193 194 struct mwait_info cpi_mwait; /* fn 5: monitor/mwait info */ 195 }; 196 197 198 static struct cpuid_info cpuid_info0; 199 200 /* 201 * These bit fields are defined by the Intel Application Note AP-485 202 * "Intel Processor Identification and the CPUID Instruction" 203 */ 204 #define CPI_FAMILY_XTD(cpi) BITX((cpi)->cpi_std[1].cp_eax, 27, 20) 205 #define CPI_MODEL_XTD(cpi) BITX((cpi)->cpi_std[1].cp_eax, 19, 16) 206 #define CPI_TYPE(cpi) BITX((cpi)->cpi_std[1].cp_eax, 13, 12) 207 #define CPI_FAMILY(cpi) BITX((cpi)->cpi_std[1].cp_eax, 11, 8) 208 #define CPI_STEP(cpi) BITX((cpi)->cpi_std[1].cp_eax, 3, 0) 209 #define CPI_MODEL(cpi) BITX((cpi)->cpi_std[1].cp_eax, 7, 4) 210 211 #define CPI_FEATURES_EDX(cpi) ((cpi)->cpi_std[1].cp_edx) 212 #define CPI_FEATURES_ECX(cpi) ((cpi)->cpi_std[1].cp_ecx) 213 #define CPI_FEATURES_XTD_EDX(cpi) ((cpi)->cpi_extd[1].cp_edx) 214 #define CPI_FEATURES_XTD_ECX(cpi) ((cpi)->cpi_extd[1].cp_ecx) 215 216 #define CPI_BRANDID(cpi) BITX((cpi)->cpi_std[1].cp_ebx, 7, 0) 217 #define CPI_CHUNKS(cpi) BITX((cpi)->cpi_std[1].cp_ebx, 15, 7) 218 #define CPI_CPU_COUNT(cpi) BITX((cpi)->cpi_std[1].cp_ebx, 23, 16) 219 #define CPI_APIC_ID(cpi) BITX((cpi)->cpi_std[1].cp_ebx, 31, 24) 220 221 #define CPI_MAXEAX_MAX 0x100 /* sanity control */ 222 #define CPI_XMAXEAX_MAX 0x80000100 223 #define CPI_FN4_ECX_MAX 0x20 /* sanity: max fn 4 levels */ 224 225 /* 226 * Function 4 (Deterministic Cache Parameters) macros 227 * Defined by Intel Application Note AP-485 228 */ 229 #define CPI_NUM_CORES(regs) BITX((regs)->cp_eax, 31, 26) 230 #define CPI_NTHR_SHR_CACHE(regs) BITX((regs)->cp_eax, 25, 14) 231 #define CPI_FULL_ASSOC_CACHE(regs) BITX((regs)->cp_eax, 9, 9) 232 #define CPI_SELF_INIT_CACHE(regs) BITX((regs)->cp_eax, 8, 8) 233 #define CPI_CACHE_LVL(regs) BITX((regs)->cp_eax, 7, 5) 234 #define CPI_CACHE_TYPE(regs) BITX((regs)->cp_eax, 4, 0) 235 236 #define CPI_CACHE_WAYS(regs) BITX((regs)->cp_ebx, 31, 22) 237 #define CPI_CACHE_PARTS(regs) BITX((regs)->cp_ebx, 21, 12) 238 #define CPI_CACHE_COH_LN_SZ(regs) BITX((regs)->cp_ebx, 11, 0) 239 240 #define CPI_CACHE_SETS(regs) BITX((regs)->cp_ecx, 31, 0) 241 242 #define CPI_PREFCH_STRIDE(regs) BITX((regs)->cp_edx, 9, 0) 243 244 245 /* 246 * A couple of shorthand macros to identify "later" P6-family chips 247 * like the Pentium M and Core. First, the "older" P6-based stuff 248 * (loosely defined as "pre-Pentium-4"): 249 * P6, PII, Mobile PII, PII Xeon, PIII, Mobile PIII, PIII Xeon 250 */ 251 252 #define IS_LEGACY_P6(cpi) ( \ 253 cpi->cpi_family == 6 && \ 254 (cpi->cpi_model == 1 || \ 255 cpi->cpi_model == 3 || \ 256 cpi->cpi_model == 5 || \ 257 cpi->cpi_model == 6 || \ 258 cpi->cpi_model == 7 || \ 259 cpi->cpi_model == 8 || \ 260 cpi->cpi_model == 0xA || \ 261 cpi->cpi_model == 0xB) \ 262 ) 263 264 /* A "new F6" is everything with family 6 that's not the above */ 265 #define IS_NEW_F6(cpi) ((cpi->cpi_family == 6) && !IS_LEGACY_P6(cpi)) 266 267 /* Extended family/model support */ 268 #define IS_EXTENDED_MODEL_INTEL(cpi) (cpi->cpi_family == 0x6 || \ 269 cpi->cpi_family >= 0xf) 270 271 /* 272 * AMD family 0xf and family 0x10 socket types. 273 * First index : 274 * 0 for family 0xf, revs B thru E 275 * 1 for family 0xf, revs F and G 276 * 2 for family 0x10, rev B 277 * Second index by (model & 0x3) 278 */ 279 static uint32_t amd_skts[3][4] = { 280 /* 281 * Family 0xf revisions B through E 282 */ 283 #define A_SKTS_0 0 284 { 285 X86_SOCKET_754, /* 0b00 */ 286 X86_SOCKET_940, /* 0b01 */ 287 X86_SOCKET_754, /* 0b10 */ 288 X86_SOCKET_939 /* 0b11 */ 289 }, 290 /* 291 * Family 0xf revisions F and G 292 */ 293 #define A_SKTS_1 1 294 { 295 X86_SOCKET_S1g1, /* 0b00 */ 296 X86_SOCKET_F1207, /* 0b01 */ 297 X86_SOCKET_UNKNOWN, /* 0b10 */ 298 X86_SOCKET_AM2 /* 0b11 */ 299 }, 300 /* 301 * Family 0x10 revisions A and B 302 * It is not clear whether, as new sockets release, that 303 * model & 0x3 will id socket for this family 304 */ 305 #define A_SKTS_2 2 306 { 307 X86_SOCKET_F1207, /* 0b00 */ 308 X86_SOCKET_F1207, /* 0b01 */ 309 X86_SOCKET_F1207, /* 0b10 */ 310 X86_SOCKET_F1207, /* 0b11 */ 311 } 312 }; 313 314 /* 315 * Table for mapping AMD Family 0xf and AMD Family 0x10 model/stepping 316 * combination to chip "revision" and socket type. 317 * 318 * The first member of this array that matches a given family, extended model 319 * plus model range, and stepping range will be considered a match. 320 */ 321 static const struct amd_rev_mapent { 322 uint_t rm_family; 323 uint_t rm_modello; 324 uint_t rm_modelhi; 325 uint_t rm_steplo; 326 uint_t rm_stephi; 327 uint32_t rm_chiprev; 328 const char *rm_chiprevstr; 329 int rm_sktidx; 330 } amd_revmap[] = { 331 /* 332 * =============== AuthenticAMD Family 0xf =============== 333 */ 334 335 /* 336 * Rev B includes model 0x4 stepping 0 and model 0x5 stepping 0 and 1. 337 */ 338 { 0xf, 0x04, 0x04, 0x0, 0x0, X86_CHIPREV_AMD_F_REV_B, "B", A_SKTS_0 }, 339 { 0xf, 0x05, 0x05, 0x0, 0x1, X86_CHIPREV_AMD_F_REV_B, "B", A_SKTS_0 }, 340 /* 341 * Rev C0 includes model 0x4 stepping 8 and model 0x5 stepping 8 342 */ 343 { 0xf, 0x04, 0x05, 0x8, 0x8, X86_CHIPREV_AMD_F_REV_C0, "C0", A_SKTS_0 }, 344 /* 345 * Rev CG is the rest of extended model 0x0 - i.e., everything 346 * but the rev B and C0 combinations covered above. 347 */ 348 { 0xf, 0x00, 0x0f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_CG, "CG", A_SKTS_0 }, 349 /* 350 * Rev D has extended model 0x1. 351 */ 352 { 0xf, 0x10, 0x1f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_D, "D", A_SKTS_0 }, 353 /* 354 * Rev E has extended model 0x2. 355 * Extended model 0x3 is unused but available to grow into. 356 */ 357 { 0xf, 0x20, 0x3f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_E, "E", A_SKTS_0 }, 358 /* 359 * Rev F has extended models 0x4 and 0x5. 360 */ 361 { 0xf, 0x40, 0x5f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_F, "F", A_SKTS_1 }, 362 /* 363 * Rev G has extended model 0x6. 364 */ 365 { 0xf, 0x60, 0x6f, 0x0, 0xf, X86_CHIPREV_AMD_F_REV_G, "G", A_SKTS_1 }, 366 367 /* 368 * =============== AuthenticAMD Family 0x10 =============== 369 */ 370 371 /* 372 * Rev A has model 0 and stepping 0/1/2 for DR-{A0,A1,A2}. 373 * Give all of model 0 stepping range to rev A. 374 */ 375 { 0x10, 0x00, 0x00, 0x0, 0x2, X86_CHIPREV_AMD_10_REV_A, "A", A_SKTS_2 }, 376 377 /* 378 * Rev B has model 2 and steppings 0/1/0xa/2 for DR-{B0,B1,BA,B2}. 379 * Give all of model 2 stepping range to rev B. 380 */ 381 { 0x10, 0x02, 0x02, 0x0, 0xf, X86_CHIPREV_AMD_10_REV_B, "B", A_SKTS_2 }, 382 }; 383 384 /* 385 * Info for monitor/mwait idle loop. 386 * 387 * See cpuid section of "Intel 64 and IA-32 Architectures Software Developer's 388 * Manual Volume 2A: Instruction Set Reference, A-M" #25366-022US, November 389 * 2006. 390 * See MONITOR/MWAIT section of "AMD64 Architecture Programmer's Manual 391 * Documentation Updates" #33633, Rev 2.05, December 2006. 392 */ 393 #define MWAIT_SUPPORT (0x00000001) /* mwait supported */ 394 #define MWAIT_EXTENSIONS (0x00000002) /* extenstion supported */ 395 #define MWAIT_ECX_INT_ENABLE (0x00000004) /* ecx 1 extension supported */ 396 #define MWAIT_SUPPORTED(cpi) ((cpi)->cpi_std[1].cp_ecx & CPUID_INTC_ECX_MON) 397 #define MWAIT_INT_ENABLE(cpi) ((cpi)->cpi_std[5].cp_ecx & 0x2) 398 #define MWAIT_EXTENSION(cpi) ((cpi)->cpi_std[5].cp_ecx & 0x1) 399 #define MWAIT_SIZE_MIN(cpi) BITX((cpi)->cpi_std[5].cp_eax, 15, 0) 400 #define MWAIT_SIZE_MAX(cpi) BITX((cpi)->cpi_std[5].cp_ebx, 15, 0) 401 /* 402 * Number of sub-cstates for a given c-state. 403 */ 404 #define MWAIT_NUM_SUBC_STATES(cpi, c_state) \ 405 BITX((cpi)->cpi_std[5].cp_edx, c_state + 3, c_state) 406 407 static void 408 synth_amd_info(struct cpuid_info *cpi) 409 { 410 const struct amd_rev_mapent *rmp; 411 uint_t family, model, step; 412 int i; 413 414 /* 415 * Currently only AMD family 0xf and family 0x10 use these fields. 416 */ 417 if (cpi->cpi_family != 0xf && cpi->cpi_family != 0x10) 418 return; 419 420 family = cpi->cpi_family; 421 model = cpi->cpi_model; 422 step = cpi->cpi_step; 423 424 for (i = 0, rmp = amd_revmap; i < sizeof (amd_revmap) / sizeof (*rmp); 425 i++, rmp++) { 426 if (family == rmp->rm_family && 427 model >= rmp->rm_modello && model <= rmp->rm_modelhi && 428 step >= rmp->rm_steplo && step <= rmp->rm_stephi) { 429 cpi->cpi_chiprev = rmp->rm_chiprev; 430 cpi->cpi_chiprevstr = rmp->rm_chiprevstr; 431 cpi->cpi_socket = amd_skts[rmp->rm_sktidx][model & 0x3]; 432 return; 433 } 434 } 435 } 436 437 static void 438 synth_info(struct cpuid_info *cpi) 439 { 440 cpi->cpi_chiprev = X86_CHIPREV_UNKNOWN; 441 cpi->cpi_chiprevstr = "Unknown"; 442 cpi->cpi_socket = X86_SOCKET_UNKNOWN; 443 444 switch (cpi->cpi_vendor) { 445 case X86_VENDOR_AMD: 446 synth_amd_info(cpi); 447 break; 448 449 default: 450 break; 451 452 } 453 } 454 455 /* 456 * Apply up various platform-dependent restrictions where the 457 * underlying platform restrictions mean the CPU can be marked 458 * as less capable than its cpuid instruction would imply. 459 */ 460 #if defined(__xpv) 461 static void 462 platform_cpuid_mangle(uint_t vendor, uint32_t eax, struct cpuid_regs *cp) 463 { 464 switch (eax) { 465 case 1: 466 cp->cp_edx &= 467 ~(CPUID_INTC_EDX_PSE | 468 CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE | 469 CPUID_INTC_EDX_MCA | /* XXPV true on dom0? */ 470 CPUID_INTC_EDX_SEP | CPUID_INTC_EDX_MTRR | 471 CPUID_INTC_EDX_PGE | CPUID_INTC_EDX_PAT | 472 CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP | 473 CPUID_INTC_EDX_PSE36 | CPUID_INTC_EDX_HTT); 474 break; 475 476 case 0x80000001: 477 cp->cp_edx &= 478 ~(CPUID_AMD_EDX_PSE | 479 CPUID_INTC_EDX_VME | CPUID_INTC_EDX_DE | 480 CPUID_AMD_EDX_MTRR | CPUID_AMD_EDX_PGE | 481 CPUID_AMD_EDX_PAT | CPUID_AMD_EDX_PSE36 | 482 CPUID_AMD_EDX_SYSC | CPUID_INTC_EDX_SEP | 483 CPUID_AMD_EDX_TSCP); 484 cp->cp_ecx &= ~CPUID_AMD_ECX_CMP_LGCY; 485 break; 486 default: 487 break; 488 } 489 490 switch (vendor) { 491 case X86_VENDOR_Intel: 492 switch (eax) { 493 case 4: 494 /* 495 * Zero out the (ncores-per-chip - 1) field 496 */ 497 cp->cp_eax &= 0x03fffffff; 498 break; 499 default: 500 break; 501 } 502 break; 503 case X86_VENDOR_AMD: 504 switch (eax) { 505 case 0x80000008: 506 /* 507 * Zero out the (ncores-per-chip - 1) field 508 */ 509 cp->cp_ecx &= 0xffffff00; 510 break; 511 default: 512 break; 513 } 514 break; 515 default: 516 break; 517 } 518 } 519 #else 520 #define platform_cpuid_mangle(vendor, eax, cp) /* nothing */ 521 #endif 522 523 /* 524 * Some undocumented ways of patching the results of the cpuid 525 * instruction to permit running Solaris 10 on future cpus that 526 * we don't currently support. Could be set to non-zero values 527 * via settings in eeprom. 528 */ 529 530 uint32_t cpuid_feature_ecx_include; 531 uint32_t cpuid_feature_ecx_exclude; 532 uint32_t cpuid_feature_edx_include; 533 uint32_t cpuid_feature_edx_exclude; 534 535 void 536 cpuid_alloc_space(cpu_t *cpu) 537 { 538 /* 539 * By convention, cpu0 is the boot cpu, which is set up 540 * before memory allocation is available. All other cpus get 541 * their cpuid_info struct allocated here. 542 */ 543 ASSERT(cpu->cpu_id != 0); 544 cpu->cpu_m.mcpu_cpi = 545 kmem_zalloc(sizeof (*cpu->cpu_m.mcpu_cpi), KM_SLEEP); 546 } 547 548 void 549 cpuid_free_space(cpu_t *cpu) 550 { 551 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 552 int i; 553 554 ASSERT(cpu->cpu_id != 0); 555 556 /* 557 * Free up any function 4 related dynamic storage 558 */ 559 for (i = 1; i < cpi->cpi_std_4_size; i++) 560 kmem_free(cpi->cpi_std_4[i], sizeof (struct cpuid_regs)); 561 if (cpi->cpi_std_4_size > 0) 562 kmem_free(cpi->cpi_std_4, 563 cpi->cpi_std_4_size * sizeof (struct cpuid_regs *)); 564 565 kmem_free(cpu->cpu_m.mcpu_cpi, sizeof (*cpu->cpu_m.mcpu_cpi)); 566 } 567 568 #if !defined(__xpv) 569 570 static void 571 check_for_hvm() 572 { 573 struct cpuid_regs cp; 574 char *xen_str; 575 uint32_t xen_signature[4]; 576 extern int xpv_is_hvm; 577 578 /* 579 * In a fully virtualized domain, Xen's pseudo-cpuid function 580 * 0x40000000 returns a string representing the Xen signature in 581 * %ebx, %ecx, and %edx. %eax contains the maximum supported cpuid 582 * function. 583 */ 584 cp.cp_eax = 0x40000000; 585 (void) __cpuid_insn(&cp); 586 xen_signature[0] = cp.cp_ebx; 587 xen_signature[1] = cp.cp_ecx; 588 xen_signature[2] = cp.cp_edx; 589 xen_signature[3] = 0; 590 xen_str = (char *)xen_signature; 591 if (strcmp("XenVMMXenVMM", xen_str) == 0 && cp.cp_eax <= 0x40000002) 592 xpv_is_hvm = 1; 593 } 594 #endif /* __xpv */ 595 596 uint_t 597 cpuid_pass1(cpu_t *cpu) 598 { 599 uint32_t mask_ecx, mask_edx; 600 uint_t feature = X86_CPUID; 601 struct cpuid_info *cpi; 602 struct cpuid_regs *cp; 603 int xcpuid; 604 #if !defined(__xpv) 605 extern int idle_cpu_prefer_mwait; 606 #endif 607 608 /* 609 * Space statically allocated for cpu0, ensure pointer is set 610 */ 611 if (cpu->cpu_id == 0) 612 cpu->cpu_m.mcpu_cpi = &cpuid_info0; 613 cpi = cpu->cpu_m.mcpu_cpi; 614 ASSERT(cpi != NULL); 615 cp = &cpi->cpi_std[0]; 616 cp->cp_eax = 0; 617 cpi->cpi_maxeax = __cpuid_insn(cp); 618 { 619 uint32_t *iptr = (uint32_t *)cpi->cpi_vendorstr; 620 *iptr++ = cp->cp_ebx; 621 *iptr++ = cp->cp_edx; 622 *iptr++ = cp->cp_ecx; 623 *(char *)&cpi->cpi_vendorstr[12] = '\0'; 624 } 625 626 /* 627 * Map the vendor string to a type code 628 */ 629 if (strcmp(cpi->cpi_vendorstr, "GenuineIntel") == 0) 630 cpi->cpi_vendor = X86_VENDOR_Intel; 631 else if (strcmp(cpi->cpi_vendorstr, "AuthenticAMD") == 0) 632 cpi->cpi_vendor = X86_VENDOR_AMD; 633 else if (strcmp(cpi->cpi_vendorstr, "GenuineTMx86") == 0) 634 cpi->cpi_vendor = X86_VENDOR_TM; 635 else if (strcmp(cpi->cpi_vendorstr, CyrixInstead) == 0) 636 /* 637 * CyrixInstead is a variable used by the Cyrix detection code 638 * in locore. 639 */ 640 cpi->cpi_vendor = X86_VENDOR_Cyrix; 641 else if (strcmp(cpi->cpi_vendorstr, "UMC UMC UMC ") == 0) 642 cpi->cpi_vendor = X86_VENDOR_UMC; 643 else if (strcmp(cpi->cpi_vendorstr, "NexGenDriven") == 0) 644 cpi->cpi_vendor = X86_VENDOR_NexGen; 645 else if (strcmp(cpi->cpi_vendorstr, "CentaurHauls") == 0) 646 cpi->cpi_vendor = X86_VENDOR_Centaur; 647 else if (strcmp(cpi->cpi_vendorstr, "RiseRiseRise") == 0) 648 cpi->cpi_vendor = X86_VENDOR_Rise; 649 else if (strcmp(cpi->cpi_vendorstr, "SiS SiS SiS ") == 0) 650 cpi->cpi_vendor = X86_VENDOR_SiS; 651 else if (strcmp(cpi->cpi_vendorstr, "Geode by NSC") == 0) 652 cpi->cpi_vendor = X86_VENDOR_NSC; 653 else 654 cpi->cpi_vendor = X86_VENDOR_IntelClone; 655 656 x86_vendor = cpi->cpi_vendor; /* for compatibility */ 657 658 /* 659 * Limit the range in case of weird hardware 660 */ 661 if (cpi->cpi_maxeax > CPI_MAXEAX_MAX) 662 cpi->cpi_maxeax = CPI_MAXEAX_MAX; 663 if (cpi->cpi_maxeax < 1) 664 goto pass1_done; 665 666 cp = &cpi->cpi_std[1]; 667 cp->cp_eax = 1; 668 (void) __cpuid_insn(cp); 669 670 /* 671 * Extract identifying constants for easy access. 672 */ 673 cpi->cpi_model = CPI_MODEL(cpi); 674 cpi->cpi_family = CPI_FAMILY(cpi); 675 676 if (cpi->cpi_family == 0xf) 677 cpi->cpi_family += CPI_FAMILY_XTD(cpi); 678 679 /* 680 * Beware: AMD uses "extended model" iff base *FAMILY* == 0xf. 681 * Intel, and presumably everyone else, uses model == 0xf, as 682 * one would expect (max value means possible overflow). Sigh. 683 */ 684 685 switch (cpi->cpi_vendor) { 686 case X86_VENDOR_Intel: 687 if (IS_EXTENDED_MODEL_INTEL(cpi)) 688 cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4; 689 break; 690 case X86_VENDOR_AMD: 691 if (CPI_FAMILY(cpi) == 0xf) 692 cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4; 693 break; 694 default: 695 if (cpi->cpi_model == 0xf) 696 cpi->cpi_model += CPI_MODEL_XTD(cpi) << 4; 697 break; 698 } 699 700 cpi->cpi_step = CPI_STEP(cpi); 701 cpi->cpi_brandid = CPI_BRANDID(cpi); 702 703 /* 704 * *default* assumptions: 705 * - believe %edx feature word 706 * - ignore %ecx feature word 707 * - 32-bit virtual and physical addressing 708 */ 709 mask_edx = 0xffffffff; 710 mask_ecx = 0; 711 712 cpi->cpi_pabits = cpi->cpi_vabits = 32; 713 714 switch (cpi->cpi_vendor) { 715 case X86_VENDOR_Intel: 716 if (cpi->cpi_family == 5) 717 x86_type = X86_TYPE_P5; 718 else if (IS_LEGACY_P6(cpi)) { 719 x86_type = X86_TYPE_P6; 720 pentiumpro_bug4046376 = 1; 721 pentiumpro_bug4064495 = 1; 722 /* 723 * Clear the SEP bit when it was set erroneously 724 */ 725 if (cpi->cpi_model < 3 && cpi->cpi_step < 3) 726 cp->cp_edx &= ~CPUID_INTC_EDX_SEP; 727 } else if (IS_NEW_F6(cpi) || cpi->cpi_family == 0xf) { 728 x86_type = X86_TYPE_P4; 729 /* 730 * We don't currently depend on any of the %ecx 731 * features until Prescott, so we'll only check 732 * this from P4 onwards. We might want to revisit 733 * that idea later. 734 */ 735 mask_ecx = 0xffffffff; 736 } else if (cpi->cpi_family > 0xf) 737 mask_ecx = 0xffffffff; 738 /* 739 * We don't support MONITOR/MWAIT if leaf 5 is not available 740 * to obtain the monitor linesize. 741 */ 742 if (cpi->cpi_maxeax < 5) 743 mask_ecx &= ~CPUID_INTC_ECX_MON; 744 break; 745 case X86_VENDOR_IntelClone: 746 default: 747 break; 748 case X86_VENDOR_AMD: 749 #if defined(OPTERON_ERRATUM_108) 750 if (cpi->cpi_family == 0xf && cpi->cpi_model == 0xe) { 751 cp->cp_eax = (0xf0f & cp->cp_eax) | 0xc0; 752 cpi->cpi_model = 0xc; 753 } else 754 #endif 755 if (cpi->cpi_family == 5) { 756 /* 757 * AMD K5 and K6 758 * 759 * These CPUs have an incomplete implementation 760 * of MCA/MCE which we mask away. 761 */ 762 mask_edx &= ~(CPUID_INTC_EDX_MCE | CPUID_INTC_EDX_MCA); 763 764 /* 765 * Model 0 uses the wrong (APIC) bit 766 * to indicate PGE. Fix it here. 767 */ 768 if (cpi->cpi_model == 0) { 769 if (cp->cp_edx & 0x200) { 770 cp->cp_edx &= ~0x200; 771 cp->cp_edx |= CPUID_INTC_EDX_PGE; 772 } 773 } 774 775 /* 776 * Early models had problems w/ MMX; disable. 777 */ 778 if (cpi->cpi_model < 6) 779 mask_edx &= ~CPUID_INTC_EDX_MMX; 780 } 781 782 /* 783 * For newer families, SSE3 and CX16, at least, are valid; 784 * enable all 785 */ 786 if (cpi->cpi_family >= 0xf) 787 mask_ecx = 0xffffffff; 788 /* 789 * We don't support MONITOR/MWAIT if leaf 5 is not available 790 * to obtain the monitor linesize. 791 */ 792 if (cpi->cpi_maxeax < 5) 793 mask_ecx &= ~CPUID_INTC_ECX_MON; 794 795 #if !defined(__xpv) 796 /* 797 * Do not use MONITOR/MWAIT to halt in the idle loop on any AMD 798 * processors. AMD does not intend MWAIT to be used in the cpu 799 * idle loop on current and future processors. 10h and future 800 * AMD processors use more power in MWAIT than HLT. 801 * Pre-family-10h Opterons do not have the MWAIT instruction. 802 */ 803 idle_cpu_prefer_mwait = 0; 804 #endif 805 806 break; 807 case X86_VENDOR_TM: 808 /* 809 * workaround the NT workaround in CMS 4.1 810 */ 811 if (cpi->cpi_family == 5 && cpi->cpi_model == 4 && 812 (cpi->cpi_step == 2 || cpi->cpi_step == 3)) 813 cp->cp_edx |= CPUID_INTC_EDX_CX8; 814 break; 815 case X86_VENDOR_Centaur: 816 /* 817 * workaround the NT workarounds again 818 */ 819 if (cpi->cpi_family == 6) 820 cp->cp_edx |= CPUID_INTC_EDX_CX8; 821 break; 822 case X86_VENDOR_Cyrix: 823 /* 824 * We rely heavily on the probing in locore 825 * to actually figure out what parts, if any, 826 * of the Cyrix cpuid instruction to believe. 827 */ 828 switch (x86_type) { 829 case X86_TYPE_CYRIX_486: 830 mask_edx = 0; 831 break; 832 case X86_TYPE_CYRIX_6x86: 833 mask_edx = 0; 834 break; 835 case X86_TYPE_CYRIX_6x86L: 836 mask_edx = 837 CPUID_INTC_EDX_DE | 838 CPUID_INTC_EDX_CX8; 839 break; 840 case X86_TYPE_CYRIX_6x86MX: 841 mask_edx = 842 CPUID_INTC_EDX_DE | 843 CPUID_INTC_EDX_MSR | 844 CPUID_INTC_EDX_CX8 | 845 CPUID_INTC_EDX_PGE | 846 CPUID_INTC_EDX_CMOV | 847 CPUID_INTC_EDX_MMX; 848 break; 849 case X86_TYPE_CYRIX_GXm: 850 mask_edx = 851 CPUID_INTC_EDX_MSR | 852 CPUID_INTC_EDX_CX8 | 853 CPUID_INTC_EDX_CMOV | 854 CPUID_INTC_EDX_MMX; 855 break; 856 case X86_TYPE_CYRIX_MediaGX: 857 break; 858 case X86_TYPE_CYRIX_MII: 859 case X86_TYPE_VIA_CYRIX_III: 860 mask_edx = 861 CPUID_INTC_EDX_DE | 862 CPUID_INTC_EDX_TSC | 863 CPUID_INTC_EDX_MSR | 864 CPUID_INTC_EDX_CX8 | 865 CPUID_INTC_EDX_PGE | 866 CPUID_INTC_EDX_CMOV | 867 CPUID_INTC_EDX_MMX; 868 break; 869 default: 870 break; 871 } 872 break; 873 } 874 875 #if defined(__xpv) 876 /* 877 * Do not support MONITOR/MWAIT under a hypervisor 878 */ 879 mask_ecx &= ~CPUID_INTC_ECX_MON; 880 #endif /* __xpv */ 881 882 /* 883 * Now we've figured out the masks that determine 884 * which bits we choose to believe, apply the masks 885 * to the feature words, then map the kernel's view 886 * of these feature words into its feature word. 887 */ 888 cp->cp_edx &= mask_edx; 889 cp->cp_ecx &= mask_ecx; 890 891 /* 892 * apply any platform restrictions (we don't call this 893 * immediately after __cpuid_insn here, because we need the 894 * workarounds applied above first) 895 */ 896 platform_cpuid_mangle(cpi->cpi_vendor, 1, cp); 897 898 /* 899 * fold in overrides from the "eeprom" mechanism 900 */ 901 cp->cp_edx |= cpuid_feature_edx_include; 902 cp->cp_edx &= ~cpuid_feature_edx_exclude; 903 904 cp->cp_ecx |= cpuid_feature_ecx_include; 905 cp->cp_ecx &= ~cpuid_feature_ecx_exclude; 906 907 if (cp->cp_edx & CPUID_INTC_EDX_PSE) 908 feature |= X86_LARGEPAGE; 909 if (cp->cp_edx & CPUID_INTC_EDX_TSC) 910 feature |= X86_TSC; 911 if (cp->cp_edx & CPUID_INTC_EDX_MSR) 912 feature |= X86_MSR; 913 if (cp->cp_edx & CPUID_INTC_EDX_MTRR) 914 feature |= X86_MTRR; 915 if (cp->cp_edx & CPUID_INTC_EDX_PGE) 916 feature |= X86_PGE; 917 if (cp->cp_edx & CPUID_INTC_EDX_CMOV) 918 feature |= X86_CMOV; 919 if (cp->cp_edx & CPUID_INTC_EDX_MMX) 920 feature |= X86_MMX; 921 if ((cp->cp_edx & CPUID_INTC_EDX_MCE) != 0 && 922 (cp->cp_edx & CPUID_INTC_EDX_MCA) != 0) 923 feature |= X86_MCA; 924 if (cp->cp_edx & CPUID_INTC_EDX_PAE) 925 feature |= X86_PAE; 926 if (cp->cp_edx & CPUID_INTC_EDX_CX8) 927 feature |= X86_CX8; 928 if (cp->cp_ecx & CPUID_INTC_ECX_CX16) 929 feature |= X86_CX16; 930 if (cp->cp_edx & CPUID_INTC_EDX_PAT) 931 feature |= X86_PAT; 932 if (cp->cp_edx & CPUID_INTC_EDX_SEP) 933 feature |= X86_SEP; 934 if (cp->cp_edx & CPUID_INTC_EDX_FXSR) { 935 /* 936 * In our implementation, fxsave/fxrstor 937 * are prerequisites before we'll even 938 * try and do SSE things. 939 */ 940 if (cp->cp_edx & CPUID_INTC_EDX_SSE) 941 feature |= X86_SSE; 942 if (cp->cp_edx & CPUID_INTC_EDX_SSE2) 943 feature |= X86_SSE2; 944 if (cp->cp_ecx & CPUID_INTC_ECX_SSE3) 945 feature |= X86_SSE3; 946 if (cpi->cpi_vendor == X86_VENDOR_Intel) { 947 if (cp->cp_ecx & CPUID_INTC_ECX_SSSE3) 948 feature |= X86_SSSE3; 949 if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_1) 950 feature |= X86_SSE4_1; 951 if (cp->cp_ecx & CPUID_INTC_ECX_SSE4_2) 952 feature |= X86_SSE4_2; 953 } 954 } 955 if (cp->cp_edx & CPUID_INTC_EDX_DE) 956 feature |= X86_DE; 957 if (cp->cp_ecx & CPUID_INTC_ECX_MON) { 958 cpi->cpi_mwait.support |= MWAIT_SUPPORT; 959 feature |= X86_MWAIT; 960 } 961 962 if (feature & X86_PAE) 963 cpi->cpi_pabits = 36; 964 965 /* 966 * Hyperthreading configuration is slightly tricky on Intel 967 * and pure clones, and even trickier on AMD. 968 * 969 * (AMD chose to set the HTT bit on their CMP processors, 970 * even though they're not actually hyperthreaded. Thus it 971 * takes a bit more work to figure out what's really going 972 * on ... see the handling of the CMP_LGCY bit below) 973 */ 974 if (cp->cp_edx & CPUID_INTC_EDX_HTT) { 975 cpi->cpi_ncpu_per_chip = CPI_CPU_COUNT(cpi); 976 if (cpi->cpi_ncpu_per_chip > 1) 977 feature |= X86_HTT; 978 } else { 979 cpi->cpi_ncpu_per_chip = 1; 980 } 981 982 /* 983 * Work on the "extended" feature information, doing 984 * some basic initialization for cpuid_pass2() 985 */ 986 xcpuid = 0; 987 switch (cpi->cpi_vendor) { 988 case X86_VENDOR_Intel: 989 if (IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf) 990 xcpuid++; 991 break; 992 case X86_VENDOR_AMD: 993 if (cpi->cpi_family > 5 || 994 (cpi->cpi_family == 5 && cpi->cpi_model >= 1)) 995 xcpuid++; 996 break; 997 case X86_VENDOR_Cyrix: 998 /* 999 * Only these Cyrix CPUs are -known- to support 1000 * extended cpuid operations. 1001 */ 1002 if (x86_type == X86_TYPE_VIA_CYRIX_III || 1003 x86_type == X86_TYPE_CYRIX_GXm) 1004 xcpuid++; 1005 break; 1006 case X86_VENDOR_Centaur: 1007 case X86_VENDOR_TM: 1008 default: 1009 xcpuid++; 1010 break; 1011 } 1012 1013 if (xcpuid) { 1014 cp = &cpi->cpi_extd[0]; 1015 cp->cp_eax = 0x80000000; 1016 cpi->cpi_xmaxeax = __cpuid_insn(cp); 1017 } 1018 1019 if (cpi->cpi_xmaxeax & 0x80000000) { 1020 1021 if (cpi->cpi_xmaxeax > CPI_XMAXEAX_MAX) 1022 cpi->cpi_xmaxeax = CPI_XMAXEAX_MAX; 1023 1024 switch (cpi->cpi_vendor) { 1025 case X86_VENDOR_Intel: 1026 case X86_VENDOR_AMD: 1027 if (cpi->cpi_xmaxeax < 0x80000001) 1028 break; 1029 cp = &cpi->cpi_extd[1]; 1030 cp->cp_eax = 0x80000001; 1031 (void) __cpuid_insn(cp); 1032 1033 if (cpi->cpi_vendor == X86_VENDOR_AMD && 1034 cpi->cpi_family == 5 && 1035 cpi->cpi_model == 6 && 1036 cpi->cpi_step == 6) { 1037 /* 1038 * K6 model 6 uses bit 10 to indicate SYSC 1039 * Later models use bit 11. Fix it here. 1040 */ 1041 if (cp->cp_edx & 0x400) { 1042 cp->cp_edx &= ~0x400; 1043 cp->cp_edx |= CPUID_AMD_EDX_SYSC; 1044 } 1045 } 1046 1047 platform_cpuid_mangle(cpi->cpi_vendor, 0x80000001, cp); 1048 1049 /* 1050 * Compute the additions to the kernel's feature word. 1051 */ 1052 if (cp->cp_edx & CPUID_AMD_EDX_NX) 1053 feature |= X86_NX; 1054 1055 #if defined(__amd64) 1056 /* 1 GB large page - enable only for 64 bit kernel */ 1057 if (cp->cp_edx & CPUID_AMD_EDX_1GPG) 1058 feature |= X86_1GPG; 1059 #endif 1060 1061 if ((cpi->cpi_vendor == X86_VENDOR_AMD) && 1062 (cpi->cpi_std[1].cp_edx & CPUID_INTC_EDX_FXSR) && 1063 (cp->cp_ecx & CPUID_AMD_ECX_SSE4A)) 1064 feature |= X86_SSE4A; 1065 1066 /* 1067 * If both the HTT and CMP_LGCY bits are set, 1068 * then we're not actually HyperThreaded. Read 1069 * "AMD CPUID Specification" for more details. 1070 */ 1071 if (cpi->cpi_vendor == X86_VENDOR_AMD && 1072 (feature & X86_HTT) && 1073 (cp->cp_ecx & CPUID_AMD_ECX_CMP_LGCY)) { 1074 feature &= ~X86_HTT; 1075 feature |= X86_CMP; 1076 } 1077 #if defined(__amd64) 1078 /* 1079 * It's really tricky to support syscall/sysret in 1080 * the i386 kernel; we rely on sysenter/sysexit 1081 * instead. In the amd64 kernel, things are -way- 1082 * better. 1083 */ 1084 if (cp->cp_edx & CPUID_AMD_EDX_SYSC) 1085 feature |= X86_ASYSC; 1086 1087 /* 1088 * While we're thinking about system calls, note 1089 * that AMD processors don't support sysenter 1090 * in long mode at all, so don't try to program them. 1091 */ 1092 if (x86_vendor == X86_VENDOR_AMD) 1093 feature &= ~X86_SEP; 1094 #endif 1095 if (x86_vendor == X86_VENDOR_AMD && 1096 cp->cp_edx & CPUID_AMD_EDX_TSCP) 1097 feature |= X86_TSCP; 1098 break; 1099 default: 1100 break; 1101 } 1102 1103 /* 1104 * Get CPUID data about processor cores and hyperthreads. 1105 */ 1106 switch (cpi->cpi_vendor) { 1107 case X86_VENDOR_Intel: 1108 if (cpi->cpi_maxeax >= 4) { 1109 cp = &cpi->cpi_std[4]; 1110 cp->cp_eax = 4; 1111 cp->cp_ecx = 0; 1112 (void) __cpuid_insn(cp); 1113 platform_cpuid_mangle(cpi->cpi_vendor, 4, cp); 1114 } 1115 /*FALLTHROUGH*/ 1116 case X86_VENDOR_AMD: 1117 if (cpi->cpi_xmaxeax < 0x80000008) 1118 break; 1119 cp = &cpi->cpi_extd[8]; 1120 cp->cp_eax = 0x80000008; 1121 (void) __cpuid_insn(cp); 1122 platform_cpuid_mangle(cpi->cpi_vendor, 0x80000008, cp); 1123 1124 /* 1125 * Virtual and physical address limits from 1126 * cpuid override previously guessed values. 1127 */ 1128 cpi->cpi_pabits = BITX(cp->cp_eax, 7, 0); 1129 cpi->cpi_vabits = BITX(cp->cp_eax, 15, 8); 1130 break; 1131 default: 1132 break; 1133 } 1134 1135 /* 1136 * Derive the number of cores per chip 1137 */ 1138 switch (cpi->cpi_vendor) { 1139 case X86_VENDOR_Intel: 1140 if (cpi->cpi_maxeax < 4) { 1141 cpi->cpi_ncore_per_chip = 1; 1142 break; 1143 } else { 1144 cpi->cpi_ncore_per_chip = 1145 BITX((cpi)->cpi_std[4].cp_eax, 31, 26) + 1; 1146 } 1147 break; 1148 case X86_VENDOR_AMD: 1149 if (cpi->cpi_xmaxeax < 0x80000008) { 1150 cpi->cpi_ncore_per_chip = 1; 1151 break; 1152 } else { 1153 /* 1154 * On family 0xf cpuid fn 2 ECX[7:0] "NC" is 1155 * 1 less than the number of physical cores on 1156 * the chip. In family 0x10 this value can 1157 * be affected by "downcoring" - it reflects 1158 * 1 less than the number of cores actually 1159 * enabled on this node. 1160 */ 1161 cpi->cpi_ncore_per_chip = 1162 BITX((cpi)->cpi_extd[8].cp_ecx, 7, 0) + 1; 1163 } 1164 break; 1165 default: 1166 cpi->cpi_ncore_per_chip = 1; 1167 break; 1168 } 1169 } else { 1170 cpi->cpi_ncore_per_chip = 1; 1171 } 1172 1173 /* 1174 * If more than one core, then this processor is CMP. 1175 */ 1176 if (cpi->cpi_ncore_per_chip > 1) 1177 feature |= X86_CMP; 1178 1179 /* 1180 * If the number of cores is the same as the number 1181 * of CPUs, then we cannot have HyperThreading. 1182 */ 1183 if (cpi->cpi_ncpu_per_chip == cpi->cpi_ncore_per_chip) 1184 feature &= ~X86_HTT; 1185 1186 if ((feature & (X86_HTT | X86_CMP)) == 0) { 1187 /* 1188 * Single-core single-threaded processors. 1189 */ 1190 cpi->cpi_chipid = -1; 1191 cpi->cpi_clogid = 0; 1192 cpi->cpi_coreid = cpu->cpu_id; 1193 cpi->cpi_pkgcoreid = 0; 1194 } else if (cpi->cpi_ncpu_per_chip > 1) { 1195 uint_t i; 1196 uint_t chipid_shift = 0; 1197 uint_t coreid_shift = 0; 1198 uint_t apic_id = CPI_APIC_ID(cpi); 1199 1200 for (i = 1; i < cpi->cpi_ncpu_per_chip; i <<= 1) 1201 chipid_shift++; 1202 cpi->cpi_chipid = apic_id >> chipid_shift; 1203 cpi->cpi_clogid = apic_id & ((1 << chipid_shift) - 1); 1204 1205 if (cpi->cpi_vendor == X86_VENDOR_Intel) { 1206 if (feature & X86_CMP) { 1207 /* 1208 * Multi-core (and possibly multi-threaded) 1209 * processors. 1210 */ 1211 uint_t ncpu_per_core; 1212 if (cpi->cpi_ncore_per_chip == 1) 1213 ncpu_per_core = cpi->cpi_ncpu_per_chip; 1214 else if (cpi->cpi_ncore_per_chip > 1) 1215 ncpu_per_core = cpi->cpi_ncpu_per_chip / 1216 cpi->cpi_ncore_per_chip; 1217 /* 1218 * 8bit APIC IDs on dual core Pentiums 1219 * look like this: 1220 * 1221 * +-----------------------+------+------+ 1222 * | Physical Package ID | MC | HT | 1223 * +-----------------------+------+------+ 1224 * <------- chipid --------> 1225 * <------- coreid ---------------> 1226 * <--- clogid --> 1227 * <------> 1228 * pkgcoreid 1229 * 1230 * Where the number of bits necessary to 1231 * represent MC and HT fields together equals 1232 * to the minimum number of bits necessary to 1233 * store the value of cpi->cpi_ncpu_per_chip. 1234 * Of those bits, the MC part uses the number 1235 * of bits necessary to store the value of 1236 * cpi->cpi_ncore_per_chip. 1237 */ 1238 for (i = 1; i < ncpu_per_core; i <<= 1) 1239 coreid_shift++; 1240 cpi->cpi_coreid = apic_id >> coreid_shift; 1241 cpi->cpi_pkgcoreid = cpi->cpi_clogid >> 1242 coreid_shift; 1243 } else if (feature & X86_HTT) { 1244 /* 1245 * Single-core multi-threaded processors. 1246 */ 1247 cpi->cpi_coreid = cpi->cpi_chipid; 1248 cpi->cpi_pkgcoreid = 0; 1249 } 1250 } else if (cpi->cpi_vendor == X86_VENDOR_AMD) { 1251 /* 1252 * AMD CMP chips currently have a single thread per 1253 * core, with 2 cores on family 0xf and 2, 3 or 4 1254 * cores on family 0x10. 1255 * 1256 * Since no two cpus share a core we must assign a 1257 * distinct coreid per cpu, and we do this by using 1258 * the cpu_id. This scheme does not, however, 1259 * guarantee that sibling cores of a chip will have 1260 * sequential coreids starting at a multiple of the 1261 * number of cores per chip - that is usually the 1262 * case, but if the ACPI MADT table is presented 1263 * in a different order then we need to perform a 1264 * few more gymnastics for the pkgcoreid. 1265 * 1266 * In family 0xf CMPs there are 2 cores on all nodes 1267 * present - no mixing of single and dual core parts. 1268 * 1269 * In family 0x10 CMPs cpuid fn 2 ECX[15:12] 1270 * "ApicIdCoreIdSize[3:0]" tells us how 1271 * many least-significant bits in the ApicId 1272 * are used to represent the core number 1273 * within the node. Cores are always 1274 * numbered sequentially from 0 regardless 1275 * of how many or which are disabled, and 1276 * there seems to be no way to discover the 1277 * real core id when some are disabled. 1278 */ 1279 cpi->cpi_coreid = cpu->cpu_id; 1280 1281 if (cpi->cpi_family == 0x10 && 1282 cpi->cpi_xmaxeax >= 0x80000008) { 1283 int coreidsz = 1284 BITX((cpi)->cpi_extd[8].cp_ecx, 15, 12); 1285 1286 cpi->cpi_pkgcoreid = 1287 apic_id & ((1 << coreidsz) - 1); 1288 } else { 1289 cpi->cpi_pkgcoreid = cpi->cpi_clogid; 1290 } 1291 } else { 1292 /* 1293 * All other processors are currently 1294 * assumed to have single cores. 1295 */ 1296 cpi->cpi_coreid = cpi->cpi_chipid; 1297 cpi->cpi_pkgcoreid = 0; 1298 } 1299 } 1300 1301 /* 1302 * Synthesize chip "revision" and socket type 1303 */ 1304 synth_info(cpi); 1305 1306 pass1_done: 1307 #if !defined(__xpv) 1308 check_for_hvm(); 1309 #endif 1310 cpi->cpi_pass = 1; 1311 return (feature); 1312 } 1313 1314 /* 1315 * Make copies of the cpuid table entries we depend on, in 1316 * part for ease of parsing now, in part so that we have only 1317 * one place to correct any of it, in part for ease of 1318 * later export to userland, and in part so we can look at 1319 * this stuff in a crash dump. 1320 */ 1321 1322 /*ARGSUSED*/ 1323 void 1324 cpuid_pass2(cpu_t *cpu) 1325 { 1326 uint_t n, nmax; 1327 int i; 1328 struct cpuid_regs *cp; 1329 uint8_t *dp; 1330 uint32_t *iptr; 1331 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 1332 1333 ASSERT(cpi->cpi_pass == 1); 1334 1335 if (cpi->cpi_maxeax < 1) 1336 goto pass2_done; 1337 1338 if ((nmax = cpi->cpi_maxeax + 1) > NMAX_CPI_STD) 1339 nmax = NMAX_CPI_STD; 1340 /* 1341 * (We already handled n == 0 and n == 1 in pass 1) 1342 */ 1343 for (n = 2, cp = &cpi->cpi_std[2]; n < nmax; n++, cp++) { 1344 cp->cp_eax = n; 1345 1346 /* 1347 * CPUID function 4 expects %ecx to be initialized 1348 * with an index which indicates which cache to return 1349 * information about. The OS is expected to call function 4 1350 * with %ecx set to 0, 1, 2, ... until it returns with 1351 * EAX[4:0] set to 0, which indicates there are no more 1352 * caches. 1353 * 1354 * Here, populate cpi_std[4] with the information returned by 1355 * function 4 when %ecx == 0, and do the rest in cpuid_pass3() 1356 * when dynamic memory allocation becomes available. 1357 * 1358 * Note: we need to explicitly initialize %ecx here, since 1359 * function 4 may have been previously invoked. 1360 */ 1361 if (n == 4) 1362 cp->cp_ecx = 0; 1363 1364 (void) __cpuid_insn(cp); 1365 platform_cpuid_mangle(cpi->cpi_vendor, n, cp); 1366 switch (n) { 1367 case 2: 1368 /* 1369 * "the lower 8 bits of the %eax register 1370 * contain a value that identifies the number 1371 * of times the cpuid [instruction] has to be 1372 * executed to obtain a complete image of the 1373 * processor's caching systems." 1374 * 1375 * How *do* they make this stuff up? 1376 */ 1377 cpi->cpi_ncache = sizeof (*cp) * 1378 BITX(cp->cp_eax, 7, 0); 1379 if (cpi->cpi_ncache == 0) 1380 break; 1381 cpi->cpi_ncache--; /* skip count byte */ 1382 1383 /* 1384 * Well, for now, rather than attempt to implement 1385 * this slightly dubious algorithm, we just look 1386 * at the first 15 .. 1387 */ 1388 if (cpi->cpi_ncache > (sizeof (*cp) - 1)) 1389 cpi->cpi_ncache = sizeof (*cp) - 1; 1390 1391 dp = cpi->cpi_cacheinfo; 1392 if (BITX(cp->cp_eax, 31, 31) == 0) { 1393 uint8_t *p = (void *)&cp->cp_eax; 1394 for (i = 1; i < 4; i++) 1395 if (p[i] != 0) 1396 *dp++ = p[i]; 1397 } 1398 if (BITX(cp->cp_ebx, 31, 31) == 0) { 1399 uint8_t *p = (void *)&cp->cp_ebx; 1400 for (i = 0; i < 4; i++) 1401 if (p[i] != 0) 1402 *dp++ = p[i]; 1403 } 1404 if (BITX(cp->cp_ecx, 31, 31) == 0) { 1405 uint8_t *p = (void *)&cp->cp_ecx; 1406 for (i = 0; i < 4; i++) 1407 if (p[i] != 0) 1408 *dp++ = p[i]; 1409 } 1410 if (BITX(cp->cp_edx, 31, 31) == 0) { 1411 uint8_t *p = (void *)&cp->cp_edx; 1412 for (i = 0; i < 4; i++) 1413 if (p[i] != 0) 1414 *dp++ = p[i]; 1415 } 1416 break; 1417 1418 case 3: /* Processor serial number, if PSN supported */ 1419 break; 1420 1421 case 4: /* Deterministic cache parameters */ 1422 break; 1423 1424 case 5: /* Monitor/Mwait parameters */ 1425 { 1426 size_t mwait_size; 1427 1428 /* 1429 * check cpi_mwait.support which was set in cpuid_pass1 1430 */ 1431 if (!(cpi->cpi_mwait.support & MWAIT_SUPPORT)) 1432 break; 1433 1434 /* 1435 * Protect ourself from insane mwait line size. 1436 * Workaround for incomplete hardware emulator(s). 1437 */ 1438 mwait_size = (size_t)MWAIT_SIZE_MAX(cpi); 1439 if (mwait_size < sizeof (uint32_t) || 1440 !ISP2(mwait_size)) { 1441 #if DEBUG 1442 cmn_err(CE_NOTE, "Cannot handle cpu %d mwait " 1443 "size %ld", 1444 cpu->cpu_id, (long)mwait_size); 1445 #endif 1446 break; 1447 } 1448 1449 cpi->cpi_mwait.mon_min = (size_t)MWAIT_SIZE_MIN(cpi); 1450 cpi->cpi_mwait.mon_max = mwait_size; 1451 if (MWAIT_EXTENSION(cpi)) { 1452 cpi->cpi_mwait.support |= MWAIT_EXTENSIONS; 1453 if (MWAIT_INT_ENABLE(cpi)) 1454 cpi->cpi_mwait.support |= 1455 MWAIT_ECX_INT_ENABLE; 1456 } 1457 break; 1458 } 1459 default: 1460 break; 1461 } 1462 } 1463 1464 if ((cpi->cpi_xmaxeax & 0x80000000) == 0) 1465 goto pass2_done; 1466 1467 if ((nmax = cpi->cpi_xmaxeax - 0x80000000 + 1) > NMAX_CPI_EXTD) 1468 nmax = NMAX_CPI_EXTD; 1469 /* 1470 * Copy the extended properties, fixing them as we go. 1471 * (We already handled n == 0 and n == 1 in pass 1) 1472 */ 1473 iptr = (void *)cpi->cpi_brandstr; 1474 for (n = 2, cp = &cpi->cpi_extd[2]; n < nmax; cp++, n++) { 1475 cp->cp_eax = 0x80000000 + n; 1476 (void) __cpuid_insn(cp); 1477 platform_cpuid_mangle(cpi->cpi_vendor, 0x80000000 + n, cp); 1478 switch (n) { 1479 case 2: 1480 case 3: 1481 case 4: 1482 /* 1483 * Extract the brand string 1484 */ 1485 *iptr++ = cp->cp_eax; 1486 *iptr++ = cp->cp_ebx; 1487 *iptr++ = cp->cp_ecx; 1488 *iptr++ = cp->cp_edx; 1489 break; 1490 case 5: 1491 switch (cpi->cpi_vendor) { 1492 case X86_VENDOR_AMD: 1493 /* 1494 * The Athlon and Duron were the first 1495 * parts to report the sizes of the 1496 * TLB for large pages. Before then, 1497 * we don't trust the data. 1498 */ 1499 if (cpi->cpi_family < 6 || 1500 (cpi->cpi_family == 6 && 1501 cpi->cpi_model < 1)) 1502 cp->cp_eax = 0; 1503 break; 1504 default: 1505 break; 1506 } 1507 break; 1508 case 6: 1509 switch (cpi->cpi_vendor) { 1510 case X86_VENDOR_AMD: 1511 /* 1512 * The Athlon and Duron were the first 1513 * AMD parts with L2 TLB's. 1514 * Before then, don't trust the data. 1515 */ 1516 if (cpi->cpi_family < 6 || 1517 cpi->cpi_family == 6 && 1518 cpi->cpi_model < 1) 1519 cp->cp_eax = cp->cp_ebx = 0; 1520 /* 1521 * AMD Duron rev A0 reports L2 1522 * cache size incorrectly as 1K 1523 * when it is really 64K 1524 */ 1525 if (cpi->cpi_family == 6 && 1526 cpi->cpi_model == 3 && 1527 cpi->cpi_step == 0) { 1528 cp->cp_ecx &= 0xffff; 1529 cp->cp_ecx |= 0x400000; 1530 } 1531 break; 1532 case X86_VENDOR_Cyrix: /* VIA C3 */ 1533 /* 1534 * VIA C3 processors are a bit messed 1535 * up w.r.t. encoding cache sizes in %ecx 1536 */ 1537 if (cpi->cpi_family != 6) 1538 break; 1539 /* 1540 * model 7 and 8 were incorrectly encoded 1541 * 1542 * xxx is model 8 really broken? 1543 */ 1544 if (cpi->cpi_model == 7 || 1545 cpi->cpi_model == 8) 1546 cp->cp_ecx = 1547 BITX(cp->cp_ecx, 31, 24) << 16 | 1548 BITX(cp->cp_ecx, 23, 16) << 12 | 1549 BITX(cp->cp_ecx, 15, 8) << 8 | 1550 BITX(cp->cp_ecx, 7, 0); 1551 /* 1552 * model 9 stepping 1 has wrong associativity 1553 */ 1554 if (cpi->cpi_model == 9 && cpi->cpi_step == 1) 1555 cp->cp_ecx |= 8 << 12; 1556 break; 1557 case X86_VENDOR_Intel: 1558 /* 1559 * Extended L2 Cache features function. 1560 * First appeared on Prescott. 1561 */ 1562 default: 1563 break; 1564 } 1565 break; 1566 default: 1567 break; 1568 } 1569 } 1570 1571 pass2_done: 1572 cpi->cpi_pass = 2; 1573 } 1574 1575 static const char * 1576 intel_cpubrand(const struct cpuid_info *cpi) 1577 { 1578 int i; 1579 1580 if ((x86_feature & X86_CPUID) == 0 || 1581 cpi->cpi_maxeax < 1 || cpi->cpi_family < 5) 1582 return ("i486"); 1583 1584 switch (cpi->cpi_family) { 1585 case 5: 1586 return ("Intel Pentium(r)"); 1587 case 6: 1588 switch (cpi->cpi_model) { 1589 uint_t celeron, xeon; 1590 const struct cpuid_regs *cp; 1591 case 0: 1592 case 1: 1593 case 2: 1594 return ("Intel Pentium(r) Pro"); 1595 case 3: 1596 case 4: 1597 return ("Intel Pentium(r) II"); 1598 case 6: 1599 return ("Intel Celeron(r)"); 1600 case 5: 1601 case 7: 1602 celeron = xeon = 0; 1603 cp = &cpi->cpi_std[2]; /* cache info */ 1604 1605 for (i = 1; i < 4; i++) { 1606 uint_t tmp; 1607 1608 tmp = (cp->cp_eax >> (8 * i)) & 0xff; 1609 if (tmp == 0x40) 1610 celeron++; 1611 if (tmp >= 0x44 && tmp <= 0x45) 1612 xeon++; 1613 } 1614 1615 for (i = 0; i < 2; i++) { 1616 uint_t tmp; 1617 1618 tmp = (cp->cp_ebx >> (8 * i)) & 0xff; 1619 if (tmp == 0x40) 1620 celeron++; 1621 else if (tmp >= 0x44 && tmp <= 0x45) 1622 xeon++; 1623 } 1624 1625 for (i = 0; i < 4; i++) { 1626 uint_t tmp; 1627 1628 tmp = (cp->cp_ecx >> (8 * i)) & 0xff; 1629 if (tmp == 0x40) 1630 celeron++; 1631 else if (tmp >= 0x44 && tmp <= 0x45) 1632 xeon++; 1633 } 1634 1635 for (i = 0; i < 4; i++) { 1636 uint_t tmp; 1637 1638 tmp = (cp->cp_edx >> (8 * i)) & 0xff; 1639 if (tmp == 0x40) 1640 celeron++; 1641 else if (tmp >= 0x44 && tmp <= 0x45) 1642 xeon++; 1643 } 1644 1645 if (celeron) 1646 return ("Intel Celeron(r)"); 1647 if (xeon) 1648 return (cpi->cpi_model == 5 ? 1649 "Intel Pentium(r) II Xeon(tm)" : 1650 "Intel Pentium(r) III Xeon(tm)"); 1651 return (cpi->cpi_model == 5 ? 1652 "Intel Pentium(r) II or Pentium(r) II Xeon(tm)" : 1653 "Intel Pentium(r) III or Pentium(r) III Xeon(tm)"); 1654 default: 1655 break; 1656 } 1657 default: 1658 break; 1659 } 1660 1661 /* BrandID is present if the field is nonzero */ 1662 if (cpi->cpi_brandid != 0) { 1663 static const struct { 1664 uint_t bt_bid; 1665 const char *bt_str; 1666 } brand_tbl[] = { 1667 { 0x1, "Intel(r) Celeron(r)" }, 1668 { 0x2, "Intel(r) Pentium(r) III" }, 1669 { 0x3, "Intel(r) Pentium(r) III Xeon(tm)" }, 1670 { 0x4, "Intel(r) Pentium(r) III" }, 1671 { 0x6, "Mobile Intel(r) Pentium(r) III" }, 1672 { 0x7, "Mobile Intel(r) Celeron(r)" }, 1673 { 0x8, "Intel(r) Pentium(r) 4" }, 1674 { 0x9, "Intel(r) Pentium(r) 4" }, 1675 { 0xa, "Intel(r) Celeron(r)" }, 1676 { 0xb, "Intel(r) Xeon(tm)" }, 1677 { 0xc, "Intel(r) Xeon(tm) MP" }, 1678 { 0xe, "Mobile Intel(r) Pentium(r) 4" }, 1679 { 0xf, "Mobile Intel(r) Celeron(r)" }, 1680 { 0x11, "Mobile Genuine Intel(r)" }, 1681 { 0x12, "Intel(r) Celeron(r) M" }, 1682 { 0x13, "Mobile Intel(r) Celeron(r)" }, 1683 { 0x14, "Intel(r) Celeron(r)" }, 1684 { 0x15, "Mobile Genuine Intel(r)" }, 1685 { 0x16, "Intel(r) Pentium(r) M" }, 1686 { 0x17, "Mobile Intel(r) Celeron(r)" } 1687 }; 1688 uint_t btblmax = sizeof (brand_tbl) / sizeof (brand_tbl[0]); 1689 uint_t sgn; 1690 1691 sgn = (cpi->cpi_family << 8) | 1692 (cpi->cpi_model << 4) | cpi->cpi_step; 1693 1694 for (i = 0; i < btblmax; i++) 1695 if (brand_tbl[i].bt_bid == cpi->cpi_brandid) 1696 break; 1697 if (i < btblmax) { 1698 if (sgn == 0x6b1 && cpi->cpi_brandid == 3) 1699 return ("Intel(r) Celeron(r)"); 1700 if (sgn < 0xf13 && cpi->cpi_brandid == 0xb) 1701 return ("Intel(r) Xeon(tm) MP"); 1702 if (sgn < 0xf13 && cpi->cpi_brandid == 0xe) 1703 return ("Intel(r) Xeon(tm)"); 1704 return (brand_tbl[i].bt_str); 1705 } 1706 } 1707 1708 return (NULL); 1709 } 1710 1711 static const char * 1712 amd_cpubrand(const struct cpuid_info *cpi) 1713 { 1714 if ((x86_feature & X86_CPUID) == 0 || 1715 cpi->cpi_maxeax < 1 || cpi->cpi_family < 5) 1716 return ("i486 compatible"); 1717 1718 switch (cpi->cpi_family) { 1719 case 5: 1720 switch (cpi->cpi_model) { 1721 case 0: 1722 case 1: 1723 case 2: 1724 case 3: 1725 case 4: 1726 case 5: 1727 return ("AMD-K5(r)"); 1728 case 6: 1729 case 7: 1730 return ("AMD-K6(r)"); 1731 case 8: 1732 return ("AMD-K6(r)-2"); 1733 case 9: 1734 return ("AMD-K6(r)-III"); 1735 default: 1736 return ("AMD (family 5)"); 1737 } 1738 case 6: 1739 switch (cpi->cpi_model) { 1740 case 1: 1741 return ("AMD-K7(tm)"); 1742 case 0: 1743 case 2: 1744 case 4: 1745 return ("AMD Athlon(tm)"); 1746 case 3: 1747 case 7: 1748 return ("AMD Duron(tm)"); 1749 case 6: 1750 case 8: 1751 case 10: 1752 /* 1753 * Use the L2 cache size to distinguish 1754 */ 1755 return ((cpi->cpi_extd[6].cp_ecx >> 16) >= 256 ? 1756 "AMD Athlon(tm)" : "AMD Duron(tm)"); 1757 default: 1758 return ("AMD (family 6)"); 1759 } 1760 default: 1761 break; 1762 } 1763 1764 if (cpi->cpi_family == 0xf && cpi->cpi_model == 5 && 1765 cpi->cpi_brandid != 0) { 1766 switch (BITX(cpi->cpi_brandid, 7, 5)) { 1767 case 3: 1768 return ("AMD Opteron(tm) UP 1xx"); 1769 case 4: 1770 return ("AMD Opteron(tm) DP 2xx"); 1771 case 5: 1772 return ("AMD Opteron(tm) MP 8xx"); 1773 default: 1774 return ("AMD Opteron(tm)"); 1775 } 1776 } 1777 1778 return (NULL); 1779 } 1780 1781 static const char * 1782 cyrix_cpubrand(struct cpuid_info *cpi, uint_t type) 1783 { 1784 if ((x86_feature & X86_CPUID) == 0 || 1785 cpi->cpi_maxeax < 1 || cpi->cpi_family < 5 || 1786 type == X86_TYPE_CYRIX_486) 1787 return ("i486 compatible"); 1788 1789 switch (type) { 1790 case X86_TYPE_CYRIX_6x86: 1791 return ("Cyrix 6x86"); 1792 case X86_TYPE_CYRIX_6x86L: 1793 return ("Cyrix 6x86L"); 1794 case X86_TYPE_CYRIX_6x86MX: 1795 return ("Cyrix 6x86MX"); 1796 case X86_TYPE_CYRIX_GXm: 1797 return ("Cyrix GXm"); 1798 case X86_TYPE_CYRIX_MediaGX: 1799 return ("Cyrix MediaGX"); 1800 case X86_TYPE_CYRIX_MII: 1801 return ("Cyrix M2"); 1802 case X86_TYPE_VIA_CYRIX_III: 1803 return ("VIA Cyrix M3"); 1804 default: 1805 /* 1806 * Have another wild guess .. 1807 */ 1808 if (cpi->cpi_family == 4 && cpi->cpi_model == 9) 1809 return ("Cyrix 5x86"); 1810 else if (cpi->cpi_family == 5) { 1811 switch (cpi->cpi_model) { 1812 case 2: 1813 return ("Cyrix 6x86"); /* Cyrix M1 */ 1814 case 4: 1815 return ("Cyrix MediaGX"); 1816 default: 1817 break; 1818 } 1819 } else if (cpi->cpi_family == 6) { 1820 switch (cpi->cpi_model) { 1821 case 0: 1822 return ("Cyrix 6x86MX"); /* Cyrix M2? */ 1823 case 5: 1824 case 6: 1825 case 7: 1826 case 8: 1827 case 9: 1828 return ("VIA C3"); 1829 default: 1830 break; 1831 } 1832 } 1833 break; 1834 } 1835 return (NULL); 1836 } 1837 1838 /* 1839 * This only gets called in the case that the CPU extended 1840 * feature brand string (0x80000002, 0x80000003, 0x80000004) 1841 * aren't available, or contain null bytes for some reason. 1842 */ 1843 static void 1844 fabricate_brandstr(struct cpuid_info *cpi) 1845 { 1846 const char *brand = NULL; 1847 1848 switch (cpi->cpi_vendor) { 1849 case X86_VENDOR_Intel: 1850 brand = intel_cpubrand(cpi); 1851 break; 1852 case X86_VENDOR_AMD: 1853 brand = amd_cpubrand(cpi); 1854 break; 1855 case X86_VENDOR_Cyrix: 1856 brand = cyrix_cpubrand(cpi, x86_type); 1857 break; 1858 case X86_VENDOR_NexGen: 1859 if (cpi->cpi_family == 5 && cpi->cpi_model == 0) 1860 brand = "NexGen Nx586"; 1861 break; 1862 case X86_VENDOR_Centaur: 1863 if (cpi->cpi_family == 5) 1864 switch (cpi->cpi_model) { 1865 case 4: 1866 brand = "Centaur C6"; 1867 break; 1868 case 8: 1869 brand = "Centaur C2"; 1870 break; 1871 case 9: 1872 brand = "Centaur C3"; 1873 break; 1874 default: 1875 break; 1876 } 1877 break; 1878 case X86_VENDOR_Rise: 1879 if (cpi->cpi_family == 5 && 1880 (cpi->cpi_model == 0 || cpi->cpi_model == 2)) 1881 brand = "Rise mP6"; 1882 break; 1883 case X86_VENDOR_SiS: 1884 if (cpi->cpi_family == 5 && cpi->cpi_model == 0) 1885 brand = "SiS 55x"; 1886 break; 1887 case X86_VENDOR_TM: 1888 if (cpi->cpi_family == 5 && cpi->cpi_model == 4) 1889 brand = "Transmeta Crusoe TM3x00 or TM5x00"; 1890 break; 1891 case X86_VENDOR_NSC: 1892 case X86_VENDOR_UMC: 1893 default: 1894 break; 1895 } 1896 if (brand) { 1897 (void) strcpy((char *)cpi->cpi_brandstr, brand); 1898 return; 1899 } 1900 1901 /* 1902 * If all else fails ... 1903 */ 1904 (void) snprintf(cpi->cpi_brandstr, sizeof (cpi->cpi_brandstr), 1905 "%s %d.%d.%d", cpi->cpi_vendorstr, cpi->cpi_family, 1906 cpi->cpi_model, cpi->cpi_step); 1907 } 1908 1909 /* 1910 * This routine is called just after kernel memory allocation 1911 * becomes available on cpu0, and as part of mp_startup() on 1912 * the other cpus. 1913 * 1914 * Fixup the brand string, and collect any information from cpuid 1915 * that requires dynamicically allocated storage to represent. 1916 */ 1917 /*ARGSUSED*/ 1918 void 1919 cpuid_pass3(cpu_t *cpu) 1920 { 1921 int i, max, shft, level, size; 1922 struct cpuid_regs regs; 1923 struct cpuid_regs *cp; 1924 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 1925 1926 ASSERT(cpi->cpi_pass == 2); 1927 1928 /* 1929 * Function 4: Deterministic cache parameters 1930 * 1931 * Take this opportunity to detect the number of threads 1932 * sharing the last level cache, and construct a corresponding 1933 * cache id. The respective cpuid_info members are initialized 1934 * to the default case of "no last level cache sharing". 1935 */ 1936 cpi->cpi_ncpu_shr_last_cache = 1; 1937 cpi->cpi_last_lvl_cacheid = cpu->cpu_id; 1938 1939 if (cpi->cpi_maxeax >= 4 && cpi->cpi_vendor == X86_VENDOR_Intel) { 1940 1941 /* 1942 * Find the # of elements (size) returned by fn 4, and along 1943 * the way detect last level cache sharing details. 1944 */ 1945 bzero(®s, sizeof (regs)); 1946 cp = ®s; 1947 for (i = 0, max = 0; i < CPI_FN4_ECX_MAX; i++) { 1948 cp->cp_eax = 4; 1949 cp->cp_ecx = i; 1950 1951 (void) __cpuid_insn(cp); 1952 1953 if (CPI_CACHE_TYPE(cp) == 0) 1954 break; 1955 level = CPI_CACHE_LVL(cp); 1956 if (level > max) { 1957 max = level; 1958 cpi->cpi_ncpu_shr_last_cache = 1959 CPI_NTHR_SHR_CACHE(cp) + 1; 1960 } 1961 } 1962 cpi->cpi_std_4_size = size = i; 1963 1964 /* 1965 * Allocate the cpi_std_4 array. The first element 1966 * references the regs for fn 4, %ecx == 0, which 1967 * cpuid_pass2() stashed in cpi->cpi_std[4]. 1968 */ 1969 if (size > 0) { 1970 cpi->cpi_std_4 = 1971 kmem_alloc(size * sizeof (cp), KM_SLEEP); 1972 cpi->cpi_std_4[0] = &cpi->cpi_std[4]; 1973 1974 /* 1975 * Allocate storage to hold the additional regs 1976 * for function 4, %ecx == 1 .. cpi_std_4_size. 1977 * 1978 * The regs for fn 4, %ecx == 0 has already 1979 * been allocated as indicated above. 1980 */ 1981 for (i = 1; i < size; i++) { 1982 cp = cpi->cpi_std_4[i] = 1983 kmem_zalloc(sizeof (regs), KM_SLEEP); 1984 cp->cp_eax = 4; 1985 cp->cp_ecx = i; 1986 1987 (void) __cpuid_insn(cp); 1988 } 1989 } 1990 /* 1991 * Determine the number of bits needed to represent 1992 * the number of CPUs sharing the last level cache. 1993 * 1994 * Shift off that number of bits from the APIC id to 1995 * derive the cache id. 1996 */ 1997 shft = 0; 1998 for (i = 1; i < cpi->cpi_ncpu_shr_last_cache; i <<= 1) 1999 shft++; 2000 cpi->cpi_last_lvl_cacheid = CPI_APIC_ID(cpi) >> shft; 2001 } 2002 2003 /* 2004 * Now fixup the brand string 2005 */ 2006 if ((cpi->cpi_xmaxeax & 0x80000000) == 0) { 2007 fabricate_brandstr(cpi); 2008 } else { 2009 2010 /* 2011 * If we successfully extracted a brand string from the cpuid 2012 * instruction, clean it up by removing leading spaces and 2013 * similar junk. 2014 */ 2015 if (cpi->cpi_brandstr[0]) { 2016 size_t maxlen = sizeof (cpi->cpi_brandstr); 2017 char *src, *dst; 2018 2019 dst = src = (char *)cpi->cpi_brandstr; 2020 src[maxlen - 1] = '\0'; 2021 /* 2022 * strip leading spaces 2023 */ 2024 while (*src == ' ') 2025 src++; 2026 /* 2027 * Remove any 'Genuine' or "Authentic" prefixes 2028 */ 2029 if (strncmp(src, "Genuine ", 8) == 0) 2030 src += 8; 2031 if (strncmp(src, "Authentic ", 10) == 0) 2032 src += 10; 2033 2034 /* 2035 * Now do an in-place copy. 2036 * Map (R) to (r) and (TM) to (tm). 2037 * The era of teletypes is long gone, and there's 2038 * -really- no need to shout. 2039 */ 2040 while (*src != '\0') { 2041 if (src[0] == '(') { 2042 if (strncmp(src + 1, "R)", 2) == 0) { 2043 (void) strncpy(dst, "(r)", 3); 2044 src += 3; 2045 dst += 3; 2046 continue; 2047 } 2048 if (strncmp(src + 1, "TM)", 3) == 0) { 2049 (void) strncpy(dst, "(tm)", 4); 2050 src += 4; 2051 dst += 4; 2052 continue; 2053 } 2054 } 2055 *dst++ = *src++; 2056 } 2057 *dst = '\0'; 2058 2059 /* 2060 * Finally, remove any trailing spaces 2061 */ 2062 while (--dst > cpi->cpi_brandstr) 2063 if (*dst == ' ') 2064 *dst = '\0'; 2065 else 2066 break; 2067 } else 2068 fabricate_brandstr(cpi); 2069 } 2070 cpi->cpi_pass = 3; 2071 } 2072 2073 /* 2074 * This routine is called out of bind_hwcap() much later in the life 2075 * of the kernel (post_startup()). The job of this routine is to resolve 2076 * the hardware feature support and kernel support for those features into 2077 * what we're actually going to tell applications via the aux vector. 2078 */ 2079 uint_t 2080 cpuid_pass4(cpu_t *cpu) 2081 { 2082 struct cpuid_info *cpi; 2083 uint_t hwcap_flags = 0; 2084 2085 if (cpu == NULL) 2086 cpu = CPU; 2087 cpi = cpu->cpu_m.mcpu_cpi; 2088 2089 ASSERT(cpi->cpi_pass == 3); 2090 2091 if (cpi->cpi_maxeax >= 1) { 2092 uint32_t *edx = &cpi->cpi_support[STD_EDX_FEATURES]; 2093 uint32_t *ecx = &cpi->cpi_support[STD_ECX_FEATURES]; 2094 2095 *edx = CPI_FEATURES_EDX(cpi); 2096 *ecx = CPI_FEATURES_ECX(cpi); 2097 2098 /* 2099 * [these require explicit kernel support] 2100 */ 2101 if ((x86_feature & X86_SEP) == 0) 2102 *edx &= ~CPUID_INTC_EDX_SEP; 2103 2104 if ((x86_feature & X86_SSE) == 0) 2105 *edx &= ~(CPUID_INTC_EDX_FXSR|CPUID_INTC_EDX_SSE); 2106 if ((x86_feature & X86_SSE2) == 0) 2107 *edx &= ~CPUID_INTC_EDX_SSE2; 2108 2109 if ((x86_feature & X86_HTT) == 0) 2110 *edx &= ~CPUID_INTC_EDX_HTT; 2111 2112 if ((x86_feature & X86_SSE3) == 0) 2113 *ecx &= ~CPUID_INTC_ECX_SSE3; 2114 2115 if (cpi->cpi_vendor == X86_VENDOR_Intel) { 2116 if ((x86_feature & X86_SSSE3) == 0) 2117 *ecx &= ~CPUID_INTC_ECX_SSSE3; 2118 if ((x86_feature & X86_SSE4_1) == 0) 2119 *ecx &= ~CPUID_INTC_ECX_SSE4_1; 2120 if ((x86_feature & X86_SSE4_2) == 0) 2121 *ecx &= ~CPUID_INTC_ECX_SSE4_2; 2122 } 2123 2124 /* 2125 * [no explicit support required beyond x87 fp context] 2126 */ 2127 if (!fpu_exists) 2128 *edx &= ~(CPUID_INTC_EDX_FPU | CPUID_INTC_EDX_MMX); 2129 2130 /* 2131 * Now map the supported feature vector to things that we 2132 * think userland will care about. 2133 */ 2134 if (*edx & CPUID_INTC_EDX_SEP) 2135 hwcap_flags |= AV_386_SEP; 2136 if (*edx & CPUID_INTC_EDX_SSE) 2137 hwcap_flags |= AV_386_FXSR | AV_386_SSE; 2138 if (*edx & CPUID_INTC_EDX_SSE2) 2139 hwcap_flags |= AV_386_SSE2; 2140 if (*ecx & CPUID_INTC_ECX_SSE3) 2141 hwcap_flags |= AV_386_SSE3; 2142 if (cpi->cpi_vendor == X86_VENDOR_Intel) { 2143 if (*ecx & CPUID_INTC_ECX_SSSE3) 2144 hwcap_flags |= AV_386_SSSE3; 2145 if (*ecx & CPUID_INTC_ECX_SSE4_1) 2146 hwcap_flags |= AV_386_SSE4_1; 2147 if (*ecx & CPUID_INTC_ECX_SSE4_2) 2148 hwcap_flags |= AV_386_SSE4_2; 2149 } 2150 if (*ecx & CPUID_INTC_ECX_POPCNT) 2151 hwcap_flags |= AV_386_POPCNT; 2152 if (*edx & CPUID_INTC_EDX_FPU) 2153 hwcap_flags |= AV_386_FPU; 2154 if (*edx & CPUID_INTC_EDX_MMX) 2155 hwcap_flags |= AV_386_MMX; 2156 2157 if (*edx & CPUID_INTC_EDX_TSC) 2158 hwcap_flags |= AV_386_TSC; 2159 if (*edx & CPUID_INTC_EDX_CX8) 2160 hwcap_flags |= AV_386_CX8; 2161 if (*edx & CPUID_INTC_EDX_CMOV) 2162 hwcap_flags |= AV_386_CMOV; 2163 if (*ecx & CPUID_INTC_ECX_MON) 2164 hwcap_flags |= AV_386_MON; 2165 if (*ecx & CPUID_INTC_ECX_CX16) 2166 hwcap_flags |= AV_386_CX16; 2167 } 2168 2169 if (x86_feature & X86_HTT) 2170 hwcap_flags |= AV_386_PAUSE; 2171 2172 if (cpi->cpi_xmaxeax < 0x80000001) 2173 goto pass4_done; 2174 2175 switch (cpi->cpi_vendor) { 2176 struct cpuid_regs cp; 2177 uint32_t *edx, *ecx; 2178 2179 case X86_VENDOR_Intel: 2180 /* 2181 * Seems like Intel duplicated what we necessary 2182 * here to make the initial crop of 64-bit OS's work. 2183 * Hopefully, those are the only "extended" bits 2184 * they'll add. 2185 */ 2186 /*FALLTHROUGH*/ 2187 2188 case X86_VENDOR_AMD: 2189 edx = &cpi->cpi_support[AMD_EDX_FEATURES]; 2190 ecx = &cpi->cpi_support[AMD_ECX_FEATURES]; 2191 2192 *edx = CPI_FEATURES_XTD_EDX(cpi); 2193 *ecx = CPI_FEATURES_XTD_ECX(cpi); 2194 2195 /* 2196 * [these features require explicit kernel support] 2197 */ 2198 switch (cpi->cpi_vendor) { 2199 case X86_VENDOR_Intel: 2200 break; 2201 2202 case X86_VENDOR_AMD: 2203 if ((x86_feature & X86_TSCP) == 0) 2204 *edx &= ~CPUID_AMD_EDX_TSCP; 2205 if ((x86_feature & X86_SSE4A) == 0) 2206 *ecx &= ~CPUID_AMD_ECX_SSE4A; 2207 break; 2208 2209 default: 2210 break; 2211 } 2212 2213 /* 2214 * [no explicit support required beyond 2215 * x87 fp context and exception handlers] 2216 */ 2217 if (!fpu_exists) 2218 *edx &= ~(CPUID_AMD_EDX_MMXamd | 2219 CPUID_AMD_EDX_3DNow | CPUID_AMD_EDX_3DNowx); 2220 2221 if ((x86_feature & X86_NX) == 0) 2222 *edx &= ~CPUID_AMD_EDX_NX; 2223 #if !defined(__amd64) 2224 *edx &= ~CPUID_AMD_EDX_LM; 2225 #endif 2226 /* 2227 * Now map the supported feature vector to 2228 * things that we think userland will care about. 2229 */ 2230 #if defined(__amd64) 2231 if (*edx & CPUID_AMD_EDX_SYSC) 2232 hwcap_flags |= AV_386_AMD_SYSC; 2233 #endif 2234 if (*edx & CPUID_AMD_EDX_MMXamd) 2235 hwcap_flags |= AV_386_AMD_MMX; 2236 if (*edx & CPUID_AMD_EDX_3DNow) 2237 hwcap_flags |= AV_386_AMD_3DNow; 2238 if (*edx & CPUID_AMD_EDX_3DNowx) 2239 hwcap_flags |= AV_386_AMD_3DNowx; 2240 2241 switch (cpi->cpi_vendor) { 2242 case X86_VENDOR_AMD: 2243 if (*edx & CPUID_AMD_EDX_TSCP) 2244 hwcap_flags |= AV_386_TSCP; 2245 if (*ecx & CPUID_AMD_ECX_AHF64) 2246 hwcap_flags |= AV_386_AHF; 2247 if (*ecx & CPUID_AMD_ECX_SSE4A) 2248 hwcap_flags |= AV_386_AMD_SSE4A; 2249 if (*ecx & CPUID_AMD_ECX_LZCNT) 2250 hwcap_flags |= AV_386_AMD_LZCNT; 2251 break; 2252 2253 case X86_VENDOR_Intel: 2254 /* 2255 * Aarrgh. 2256 * Intel uses a different bit in the same word. 2257 */ 2258 if (*ecx & CPUID_INTC_ECX_AHF64) 2259 hwcap_flags |= AV_386_AHF; 2260 break; 2261 2262 default: 2263 break; 2264 } 2265 break; 2266 2267 case X86_VENDOR_TM: 2268 cp.cp_eax = 0x80860001; 2269 (void) __cpuid_insn(&cp); 2270 cpi->cpi_support[TM_EDX_FEATURES] = cp.cp_edx; 2271 break; 2272 2273 default: 2274 break; 2275 } 2276 2277 pass4_done: 2278 cpi->cpi_pass = 4; 2279 return (hwcap_flags); 2280 } 2281 2282 2283 /* 2284 * Simulate the cpuid instruction using the data we previously 2285 * captured about this CPU. We try our best to return the truth 2286 * about the hardware, independently of kernel support. 2287 */ 2288 uint32_t 2289 cpuid_insn(cpu_t *cpu, struct cpuid_regs *cp) 2290 { 2291 struct cpuid_info *cpi; 2292 struct cpuid_regs *xcp; 2293 2294 if (cpu == NULL) 2295 cpu = CPU; 2296 cpi = cpu->cpu_m.mcpu_cpi; 2297 2298 ASSERT(cpuid_checkpass(cpu, 3)); 2299 2300 /* 2301 * CPUID data is cached in two separate places: cpi_std for standard 2302 * CPUID functions, and cpi_extd for extended CPUID functions. 2303 */ 2304 if (cp->cp_eax <= cpi->cpi_maxeax && cp->cp_eax < NMAX_CPI_STD) 2305 xcp = &cpi->cpi_std[cp->cp_eax]; 2306 else if (cp->cp_eax >= 0x80000000 && cp->cp_eax <= cpi->cpi_xmaxeax && 2307 cp->cp_eax < 0x80000000 + NMAX_CPI_EXTD) 2308 xcp = &cpi->cpi_extd[cp->cp_eax - 0x80000000]; 2309 else 2310 /* 2311 * The caller is asking for data from an input parameter which 2312 * the kernel has not cached. In this case we go fetch from 2313 * the hardware and return the data directly to the user. 2314 */ 2315 return (__cpuid_insn(cp)); 2316 2317 cp->cp_eax = xcp->cp_eax; 2318 cp->cp_ebx = xcp->cp_ebx; 2319 cp->cp_ecx = xcp->cp_ecx; 2320 cp->cp_edx = xcp->cp_edx; 2321 return (cp->cp_eax); 2322 } 2323 2324 int 2325 cpuid_checkpass(cpu_t *cpu, int pass) 2326 { 2327 return (cpu != NULL && cpu->cpu_m.mcpu_cpi != NULL && 2328 cpu->cpu_m.mcpu_cpi->cpi_pass >= pass); 2329 } 2330 2331 int 2332 cpuid_getbrandstr(cpu_t *cpu, char *s, size_t n) 2333 { 2334 ASSERT(cpuid_checkpass(cpu, 3)); 2335 2336 return (snprintf(s, n, "%s", cpu->cpu_m.mcpu_cpi->cpi_brandstr)); 2337 } 2338 2339 int 2340 cpuid_is_cmt(cpu_t *cpu) 2341 { 2342 if (cpu == NULL) 2343 cpu = CPU; 2344 2345 ASSERT(cpuid_checkpass(cpu, 1)); 2346 2347 return (cpu->cpu_m.mcpu_cpi->cpi_chipid >= 0); 2348 } 2349 2350 /* 2351 * AMD and Intel both implement the 64-bit variant of the syscall 2352 * instruction (syscallq), so if there's -any- support for syscall, 2353 * cpuid currently says "yes, we support this". 2354 * 2355 * However, Intel decided to -not- implement the 32-bit variant of the 2356 * syscall instruction, so we provide a predicate to allow our caller 2357 * to test that subtlety here. 2358 * 2359 * XXPV Currently, 32-bit syscall instructions don't work via the hypervisor, 2360 * even in the case where the hardware would in fact support it. 2361 */ 2362 /*ARGSUSED*/ 2363 int 2364 cpuid_syscall32_insn(cpu_t *cpu) 2365 { 2366 ASSERT(cpuid_checkpass((cpu == NULL ? CPU : cpu), 1)); 2367 2368 #if !defined(__xpv) 2369 if (cpu == NULL) 2370 cpu = CPU; 2371 2372 /*CSTYLED*/ 2373 { 2374 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 2375 2376 if (cpi->cpi_vendor == X86_VENDOR_AMD && 2377 cpi->cpi_xmaxeax >= 0x80000001 && 2378 (CPI_FEATURES_XTD_EDX(cpi) & CPUID_AMD_EDX_SYSC)) 2379 return (1); 2380 } 2381 #endif 2382 return (0); 2383 } 2384 2385 int 2386 cpuid_getidstr(cpu_t *cpu, char *s, size_t n) 2387 { 2388 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 2389 2390 static const char fmt[] = 2391 "x86 (%s %X family %d model %d step %d clock %d MHz)"; 2392 static const char fmt_ht[] = 2393 "x86 (chipid 0x%x %s %X family %d model %d step %d clock %d MHz)"; 2394 2395 ASSERT(cpuid_checkpass(cpu, 1)); 2396 2397 if (cpuid_is_cmt(cpu)) 2398 return (snprintf(s, n, fmt_ht, cpi->cpi_chipid, 2399 cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax, 2400 cpi->cpi_family, cpi->cpi_model, 2401 cpi->cpi_step, cpu->cpu_type_info.pi_clock)); 2402 return (snprintf(s, n, fmt, 2403 cpi->cpi_vendorstr, cpi->cpi_std[1].cp_eax, 2404 cpi->cpi_family, cpi->cpi_model, 2405 cpi->cpi_step, cpu->cpu_type_info.pi_clock)); 2406 } 2407 2408 const char * 2409 cpuid_getvendorstr(cpu_t *cpu) 2410 { 2411 ASSERT(cpuid_checkpass(cpu, 1)); 2412 return ((const char *)cpu->cpu_m.mcpu_cpi->cpi_vendorstr); 2413 } 2414 2415 uint_t 2416 cpuid_getvendor(cpu_t *cpu) 2417 { 2418 ASSERT(cpuid_checkpass(cpu, 1)); 2419 return (cpu->cpu_m.mcpu_cpi->cpi_vendor); 2420 } 2421 2422 uint_t 2423 cpuid_getfamily(cpu_t *cpu) 2424 { 2425 ASSERT(cpuid_checkpass(cpu, 1)); 2426 return (cpu->cpu_m.mcpu_cpi->cpi_family); 2427 } 2428 2429 uint_t 2430 cpuid_getmodel(cpu_t *cpu) 2431 { 2432 ASSERT(cpuid_checkpass(cpu, 1)); 2433 return (cpu->cpu_m.mcpu_cpi->cpi_model); 2434 } 2435 2436 uint_t 2437 cpuid_get_ncpu_per_chip(cpu_t *cpu) 2438 { 2439 ASSERT(cpuid_checkpass(cpu, 1)); 2440 return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_per_chip); 2441 } 2442 2443 uint_t 2444 cpuid_get_ncore_per_chip(cpu_t *cpu) 2445 { 2446 ASSERT(cpuid_checkpass(cpu, 1)); 2447 return (cpu->cpu_m.mcpu_cpi->cpi_ncore_per_chip); 2448 } 2449 2450 uint_t 2451 cpuid_get_ncpu_sharing_last_cache(cpu_t *cpu) 2452 { 2453 ASSERT(cpuid_checkpass(cpu, 2)); 2454 return (cpu->cpu_m.mcpu_cpi->cpi_ncpu_shr_last_cache); 2455 } 2456 2457 id_t 2458 cpuid_get_last_lvl_cacheid(cpu_t *cpu) 2459 { 2460 ASSERT(cpuid_checkpass(cpu, 2)); 2461 return (cpu->cpu_m.mcpu_cpi->cpi_last_lvl_cacheid); 2462 } 2463 2464 uint_t 2465 cpuid_getstep(cpu_t *cpu) 2466 { 2467 ASSERT(cpuid_checkpass(cpu, 1)); 2468 return (cpu->cpu_m.mcpu_cpi->cpi_step); 2469 } 2470 2471 uint_t 2472 cpuid_getsig(struct cpu *cpu) 2473 { 2474 ASSERT(cpuid_checkpass(cpu, 1)); 2475 return (cpu->cpu_m.mcpu_cpi->cpi_std[1].cp_eax); 2476 } 2477 2478 uint32_t 2479 cpuid_getchiprev(struct cpu *cpu) 2480 { 2481 ASSERT(cpuid_checkpass(cpu, 1)); 2482 return (cpu->cpu_m.mcpu_cpi->cpi_chiprev); 2483 } 2484 2485 const char * 2486 cpuid_getchiprevstr(struct cpu *cpu) 2487 { 2488 ASSERT(cpuid_checkpass(cpu, 1)); 2489 return (cpu->cpu_m.mcpu_cpi->cpi_chiprevstr); 2490 } 2491 2492 uint32_t 2493 cpuid_getsockettype(struct cpu *cpu) 2494 { 2495 ASSERT(cpuid_checkpass(cpu, 1)); 2496 return (cpu->cpu_m.mcpu_cpi->cpi_socket); 2497 } 2498 2499 int 2500 cpuid_get_chipid(cpu_t *cpu) 2501 { 2502 ASSERT(cpuid_checkpass(cpu, 1)); 2503 2504 if (cpuid_is_cmt(cpu)) 2505 return (cpu->cpu_m.mcpu_cpi->cpi_chipid); 2506 return (cpu->cpu_id); 2507 } 2508 2509 id_t 2510 cpuid_get_coreid(cpu_t *cpu) 2511 { 2512 ASSERT(cpuid_checkpass(cpu, 1)); 2513 return (cpu->cpu_m.mcpu_cpi->cpi_coreid); 2514 } 2515 2516 int 2517 cpuid_get_pkgcoreid(cpu_t *cpu) 2518 { 2519 ASSERT(cpuid_checkpass(cpu, 1)); 2520 return (cpu->cpu_m.mcpu_cpi->cpi_pkgcoreid); 2521 } 2522 2523 int 2524 cpuid_get_clogid(cpu_t *cpu) 2525 { 2526 ASSERT(cpuid_checkpass(cpu, 1)); 2527 return (cpu->cpu_m.mcpu_cpi->cpi_clogid); 2528 } 2529 2530 void 2531 cpuid_get_addrsize(cpu_t *cpu, uint_t *pabits, uint_t *vabits) 2532 { 2533 struct cpuid_info *cpi; 2534 2535 if (cpu == NULL) 2536 cpu = CPU; 2537 cpi = cpu->cpu_m.mcpu_cpi; 2538 2539 ASSERT(cpuid_checkpass(cpu, 1)); 2540 2541 if (pabits) 2542 *pabits = cpi->cpi_pabits; 2543 if (vabits) 2544 *vabits = cpi->cpi_vabits; 2545 } 2546 2547 /* 2548 * Returns the number of data TLB entries for a corresponding 2549 * pagesize. If it can't be computed, or isn't known, the 2550 * routine returns zero. If you ask about an architecturally 2551 * impossible pagesize, the routine will panic (so that the 2552 * hat implementor knows that things are inconsistent.) 2553 */ 2554 uint_t 2555 cpuid_get_dtlb_nent(cpu_t *cpu, size_t pagesize) 2556 { 2557 struct cpuid_info *cpi; 2558 uint_t dtlb_nent = 0; 2559 2560 if (cpu == NULL) 2561 cpu = CPU; 2562 cpi = cpu->cpu_m.mcpu_cpi; 2563 2564 ASSERT(cpuid_checkpass(cpu, 1)); 2565 2566 /* 2567 * Check the L2 TLB info 2568 */ 2569 if (cpi->cpi_xmaxeax >= 0x80000006) { 2570 struct cpuid_regs *cp = &cpi->cpi_extd[6]; 2571 2572 switch (pagesize) { 2573 2574 case 4 * 1024: 2575 /* 2576 * All zero in the top 16 bits of the register 2577 * indicates a unified TLB. Size is in low 16 bits. 2578 */ 2579 if ((cp->cp_ebx & 0xffff0000) == 0) 2580 dtlb_nent = cp->cp_ebx & 0x0000ffff; 2581 else 2582 dtlb_nent = BITX(cp->cp_ebx, 27, 16); 2583 break; 2584 2585 case 2 * 1024 * 1024: 2586 if ((cp->cp_eax & 0xffff0000) == 0) 2587 dtlb_nent = cp->cp_eax & 0x0000ffff; 2588 else 2589 dtlb_nent = BITX(cp->cp_eax, 27, 16); 2590 break; 2591 2592 default: 2593 panic("unknown L2 pagesize"); 2594 /*NOTREACHED*/ 2595 } 2596 } 2597 2598 if (dtlb_nent != 0) 2599 return (dtlb_nent); 2600 2601 /* 2602 * No L2 TLB support for this size, try L1. 2603 */ 2604 if (cpi->cpi_xmaxeax >= 0x80000005) { 2605 struct cpuid_regs *cp = &cpi->cpi_extd[5]; 2606 2607 switch (pagesize) { 2608 case 4 * 1024: 2609 dtlb_nent = BITX(cp->cp_ebx, 23, 16); 2610 break; 2611 case 2 * 1024 * 1024: 2612 dtlb_nent = BITX(cp->cp_eax, 23, 16); 2613 break; 2614 default: 2615 panic("unknown L1 d-TLB pagesize"); 2616 /*NOTREACHED*/ 2617 } 2618 } 2619 2620 return (dtlb_nent); 2621 } 2622 2623 uint_t 2624 cpuid_get_apicid(cpu_t *cpu) 2625 { 2626 struct cpuid_info *cpi; 2627 2628 if (cpu == NULL) 2629 cpu = CPU; 2630 cpi = cpu->cpu_m.mcpu_cpi; 2631 2632 ASSERT(cpuid_checkpass(cpu, 1)); 2633 return (CPI_APIC_ID(cpi)); 2634 } 2635 2636 /* 2637 * Return 0 if the erratum is not present or not applicable, positive 2638 * if it is, and negative if the status of the erratum is unknown. 2639 * 2640 * See "Revision Guide for AMD Athlon(tm) 64 and AMD Opteron(tm) 2641 * Processors" #25759, Rev 3.57, August 2005 2642 */ 2643 int 2644 cpuid_opteron_erratum(cpu_t *cpu, uint_t erratum) 2645 { 2646 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 2647 uint_t eax; 2648 2649 /* 2650 * Bail out if this CPU isn't an AMD CPU, or if it's 2651 * a legacy (32-bit) AMD CPU. 2652 */ 2653 if (cpi->cpi_vendor != X86_VENDOR_AMD || 2654 cpi->cpi_family == 4 || cpi->cpi_family == 5 || 2655 cpi->cpi_family == 6) 2656 2657 return (0); 2658 2659 eax = cpi->cpi_std[1].cp_eax; 2660 2661 #define SH_B0(eax) (eax == 0xf40 || eax == 0xf50) 2662 #define SH_B3(eax) (eax == 0xf51) 2663 #define B(eax) (SH_B0(eax) || SH_B3(eax)) 2664 2665 #define SH_C0(eax) (eax == 0xf48 || eax == 0xf58) 2666 2667 #define SH_CG(eax) (eax == 0xf4a || eax == 0xf5a || eax == 0xf7a) 2668 #define DH_CG(eax) (eax == 0xfc0 || eax == 0xfe0 || eax == 0xff0) 2669 #define CH_CG(eax) (eax == 0xf82 || eax == 0xfb2) 2670 #define CG(eax) (SH_CG(eax) || DH_CG(eax) || CH_CG(eax)) 2671 2672 #define SH_D0(eax) (eax == 0x10f40 || eax == 0x10f50 || eax == 0x10f70) 2673 #define DH_D0(eax) (eax == 0x10fc0 || eax == 0x10ff0) 2674 #define CH_D0(eax) (eax == 0x10f80 || eax == 0x10fb0) 2675 #define D0(eax) (SH_D0(eax) || DH_D0(eax) || CH_D0(eax)) 2676 2677 #define SH_E0(eax) (eax == 0x20f50 || eax == 0x20f40 || eax == 0x20f70) 2678 #define JH_E1(eax) (eax == 0x20f10) /* JH8_E0 had 0x20f30 */ 2679 #define DH_E3(eax) (eax == 0x20fc0 || eax == 0x20ff0) 2680 #define SH_E4(eax) (eax == 0x20f51 || eax == 0x20f71) 2681 #define BH_E4(eax) (eax == 0x20fb1) 2682 #define SH_E5(eax) (eax == 0x20f42) 2683 #define DH_E6(eax) (eax == 0x20ff2 || eax == 0x20fc2) 2684 #define JH_E6(eax) (eax == 0x20f12 || eax == 0x20f32) 2685 #define EX(eax) (SH_E0(eax) || JH_E1(eax) || DH_E3(eax) || \ 2686 SH_E4(eax) || BH_E4(eax) || SH_E5(eax) || \ 2687 DH_E6(eax) || JH_E6(eax)) 2688 2689 switch (erratum) { 2690 case 1: 2691 return (cpi->cpi_family < 0x10); 2692 case 51: /* what does the asterisk mean? */ 2693 return (B(eax) || SH_C0(eax) || CG(eax)); 2694 case 52: 2695 return (B(eax)); 2696 case 57: 2697 return (cpi->cpi_family <= 0x10); 2698 case 58: 2699 return (B(eax)); 2700 case 60: 2701 return (cpi->cpi_family <= 0x10); 2702 case 61: 2703 case 62: 2704 case 63: 2705 case 64: 2706 case 65: 2707 case 66: 2708 case 68: 2709 case 69: 2710 case 70: 2711 case 71: 2712 return (B(eax)); 2713 case 72: 2714 return (SH_B0(eax)); 2715 case 74: 2716 return (B(eax)); 2717 case 75: 2718 return (cpi->cpi_family < 0x10); 2719 case 76: 2720 return (B(eax)); 2721 case 77: 2722 return (cpi->cpi_family <= 0x10); 2723 case 78: 2724 return (B(eax) || SH_C0(eax)); 2725 case 79: 2726 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax)); 2727 case 80: 2728 case 81: 2729 case 82: 2730 return (B(eax)); 2731 case 83: 2732 return (B(eax) || SH_C0(eax) || CG(eax)); 2733 case 85: 2734 return (cpi->cpi_family < 0x10); 2735 case 86: 2736 return (SH_C0(eax) || CG(eax)); 2737 case 88: 2738 #if !defined(__amd64) 2739 return (0); 2740 #else 2741 return (B(eax) || SH_C0(eax)); 2742 #endif 2743 case 89: 2744 return (cpi->cpi_family < 0x10); 2745 case 90: 2746 return (B(eax) || SH_C0(eax) || CG(eax)); 2747 case 91: 2748 case 92: 2749 return (B(eax) || SH_C0(eax)); 2750 case 93: 2751 return (SH_C0(eax)); 2752 case 94: 2753 return (B(eax) || SH_C0(eax) || CG(eax)); 2754 case 95: 2755 #if !defined(__amd64) 2756 return (0); 2757 #else 2758 return (B(eax) || SH_C0(eax)); 2759 #endif 2760 case 96: 2761 return (B(eax) || SH_C0(eax) || CG(eax)); 2762 case 97: 2763 case 98: 2764 return (SH_C0(eax) || CG(eax)); 2765 case 99: 2766 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax)); 2767 case 100: 2768 return (B(eax) || SH_C0(eax)); 2769 case 101: 2770 case 103: 2771 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax)); 2772 case 104: 2773 return (SH_C0(eax) || CG(eax) || D0(eax)); 2774 case 105: 2775 case 106: 2776 case 107: 2777 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax)); 2778 case 108: 2779 return (DH_CG(eax)); 2780 case 109: 2781 return (SH_C0(eax) || CG(eax) || D0(eax)); 2782 case 110: 2783 return (D0(eax) || EX(eax)); 2784 case 111: 2785 return (CG(eax)); 2786 case 112: 2787 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax)); 2788 case 113: 2789 return (eax == 0x20fc0); 2790 case 114: 2791 return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax)); 2792 case 115: 2793 return (SH_E0(eax) || JH_E1(eax)); 2794 case 116: 2795 return (SH_E0(eax) || JH_E1(eax) || DH_E3(eax)); 2796 case 117: 2797 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax)); 2798 case 118: 2799 return (SH_E0(eax) || JH_E1(eax) || SH_E4(eax) || BH_E4(eax) || 2800 JH_E6(eax)); 2801 case 121: 2802 return (B(eax) || SH_C0(eax) || CG(eax) || D0(eax) || EX(eax)); 2803 case 122: 2804 return (cpi->cpi_family < 0x10); 2805 case 123: 2806 return (JH_E1(eax) || BH_E4(eax) || JH_E6(eax)); 2807 case 131: 2808 return (cpi->cpi_family < 0x10); 2809 case 6336786: 2810 /* 2811 * Test for AdvPowerMgmtInfo.TscPStateInvariant 2812 * if this is a K8 family or newer processor 2813 */ 2814 if (CPI_FAMILY(cpi) == 0xf) { 2815 struct cpuid_regs regs; 2816 regs.cp_eax = 0x80000007; 2817 (void) __cpuid_insn(®s); 2818 return (!(regs.cp_edx & 0x100)); 2819 } 2820 return (0); 2821 case 6323525: 2822 return (((((eax >> 12) & 0xff00) + (eax & 0xf00)) | 2823 (((eax >> 4) & 0xf) | ((eax >> 12) & 0xf0))) < 0xf40); 2824 2825 default: 2826 return (-1); 2827 } 2828 } 2829 2830 static const char assoc_str[] = "associativity"; 2831 static const char line_str[] = "line-size"; 2832 static const char size_str[] = "size"; 2833 2834 static void 2835 add_cache_prop(dev_info_t *devi, const char *label, const char *type, 2836 uint32_t val) 2837 { 2838 char buf[128]; 2839 2840 /* 2841 * ndi_prop_update_int() is used because it is desirable for 2842 * DDI_PROP_HW_DEF and DDI_PROP_DONTSLEEP to be set. 2843 */ 2844 if (snprintf(buf, sizeof (buf), "%s-%s", label, type) < sizeof (buf)) 2845 (void) ndi_prop_update_int(DDI_DEV_T_NONE, devi, buf, val); 2846 } 2847 2848 /* 2849 * Intel-style cache/tlb description 2850 * 2851 * Standard cpuid level 2 gives a randomly ordered 2852 * selection of tags that index into a table that describes 2853 * cache and tlb properties. 2854 */ 2855 2856 static const char l1_icache_str[] = "l1-icache"; 2857 static const char l1_dcache_str[] = "l1-dcache"; 2858 static const char l2_cache_str[] = "l2-cache"; 2859 static const char l3_cache_str[] = "l3-cache"; 2860 static const char itlb4k_str[] = "itlb-4K"; 2861 static const char dtlb4k_str[] = "dtlb-4K"; 2862 static const char itlb4M_str[] = "itlb-4M"; 2863 static const char dtlb4M_str[] = "dtlb-4M"; 2864 static const char dtlb24_str[] = "dtlb0-2M-4M"; 2865 static const char itlb424_str[] = "itlb-4K-2M-4M"; 2866 static const char itlb24_str[] = "itlb-2M-4M"; 2867 static const char dtlb44_str[] = "dtlb-4K-4M"; 2868 static const char sl1_dcache_str[] = "sectored-l1-dcache"; 2869 static const char sl2_cache_str[] = "sectored-l2-cache"; 2870 static const char itrace_str[] = "itrace-cache"; 2871 static const char sl3_cache_str[] = "sectored-l3-cache"; 2872 static const char sh_l2_tlb4k_str[] = "shared-l2-tlb-4k"; 2873 2874 static const struct cachetab { 2875 uint8_t ct_code; 2876 uint8_t ct_assoc; 2877 uint16_t ct_line_size; 2878 size_t ct_size; 2879 const char *ct_label; 2880 } intel_ctab[] = { 2881 /* maintain descending order! */ 2882 { 0xe4, 16, 64, 8*1024*1024, l3_cache_str}, 2883 { 0xe3, 16, 64, 4*1024*1024, l3_cache_str}, 2884 { 0xe2, 16, 64, 2*1024*1024, l3_cache_str}, 2885 { 0xde, 12, 64, 6*1024*1024, l3_cache_str}, 2886 { 0xdd, 12, 64, 3*1024*1024, l3_cache_str}, 2887 { 0xdc, 12, 64, ((1*1024*1024)+(512*1024)), l3_cache_str}, 2888 { 0xd8, 8, 64, 4*1024*1024, l3_cache_str}, 2889 { 0xd7, 8, 64, 2*1024*1024, l3_cache_str}, 2890 { 0xd6, 8, 64, 1*1024*1024, l3_cache_str}, 2891 { 0xd2, 4, 64, 2*1024*1024, l3_cache_str}, 2892 { 0xd1, 4, 64, 1*1024*1024, l3_cache_str}, 2893 { 0xd0, 4, 64, 512*1024, l3_cache_str}, 2894 { 0xca, 4, 0, 512, sh_l2_tlb4k_str}, 2895 { 0xb4, 4, 0, 256, dtlb4k_str }, 2896 { 0xb3, 4, 0, 128, dtlb4k_str }, 2897 { 0xb2, 4, 0, 64, itlb4k_str }, 2898 { 0xb0, 4, 0, 128, itlb4k_str }, 2899 { 0x87, 8, 64, 1024*1024, l2_cache_str}, 2900 { 0x86, 4, 64, 512*1024, l2_cache_str}, 2901 { 0x85, 8, 32, 2*1024*1024, l2_cache_str}, 2902 { 0x84, 8, 32, 1024*1024, l2_cache_str}, 2903 { 0x83, 8, 32, 512*1024, l2_cache_str}, 2904 { 0x82, 8, 32, 256*1024, l2_cache_str}, 2905 { 0x7f, 2, 64, 512*1024, l2_cache_str}, 2906 { 0x7d, 8, 64, 2*1024*1024, sl2_cache_str}, 2907 { 0x7c, 8, 64, 1024*1024, sl2_cache_str}, 2908 { 0x7b, 8, 64, 512*1024, sl2_cache_str}, 2909 { 0x7a, 8, 64, 256*1024, sl2_cache_str}, 2910 { 0x79, 8, 64, 128*1024, sl2_cache_str}, 2911 { 0x78, 8, 64, 1024*1024, l2_cache_str}, 2912 { 0x73, 8, 0, 64*1024, itrace_str}, 2913 { 0x72, 8, 0, 32*1024, itrace_str}, 2914 { 0x71, 8, 0, 16*1024, itrace_str}, 2915 { 0x70, 8, 0, 12*1024, itrace_str}, 2916 { 0x68, 4, 64, 32*1024, sl1_dcache_str}, 2917 { 0x67, 4, 64, 16*1024, sl1_dcache_str}, 2918 { 0x66, 4, 64, 8*1024, sl1_dcache_str}, 2919 { 0x60, 8, 64, 16*1024, sl1_dcache_str}, 2920 { 0x5d, 0, 0, 256, dtlb44_str}, 2921 { 0x5c, 0, 0, 128, dtlb44_str}, 2922 { 0x5b, 0, 0, 64, dtlb44_str}, 2923 { 0x5a, 4, 0, 32, dtlb24_str}, 2924 { 0x55, 0, 0, 7, itlb24_str}, 2925 { 0x52, 0, 0, 256, itlb424_str}, 2926 { 0x51, 0, 0, 128, itlb424_str}, 2927 { 0x50, 0, 0, 64, itlb424_str}, 2928 { 0x4d, 16, 64, 16*1024*1024, l3_cache_str}, 2929 { 0x4c, 12, 64, 12*1024*1024, l3_cache_str}, 2930 { 0x4b, 16, 64, 8*1024*1024, l3_cache_str}, 2931 { 0x4a, 12, 64, 6*1024*1024, l3_cache_str}, 2932 { 0x49, 16, 64, 4*1024*1024, l3_cache_str}, 2933 { 0x47, 8, 64, 8*1024*1024, l3_cache_str}, 2934 { 0x46, 4, 64, 4*1024*1024, l3_cache_str}, 2935 { 0x45, 4, 32, 2*1024*1024, l2_cache_str}, 2936 { 0x44, 4, 32, 1024*1024, l2_cache_str}, 2937 { 0x43, 4, 32, 512*1024, l2_cache_str}, 2938 { 0x42, 4, 32, 256*1024, l2_cache_str}, 2939 { 0x41, 4, 32, 128*1024, l2_cache_str}, 2940 { 0x3e, 4, 64, 512*1024, sl2_cache_str}, 2941 { 0x3d, 6, 64, 384*1024, sl2_cache_str}, 2942 { 0x3c, 4, 64, 256*1024, sl2_cache_str}, 2943 { 0x3b, 2, 64, 128*1024, sl2_cache_str}, 2944 { 0x3a, 6, 64, 192*1024, sl2_cache_str}, 2945 { 0x39, 4, 64, 128*1024, sl2_cache_str}, 2946 { 0x30, 8, 64, 32*1024, l1_icache_str}, 2947 { 0x2c, 8, 64, 32*1024, l1_dcache_str}, 2948 { 0x29, 8, 64, 4096*1024, sl3_cache_str}, 2949 { 0x25, 8, 64, 2048*1024, sl3_cache_str}, 2950 { 0x23, 8, 64, 1024*1024, sl3_cache_str}, 2951 { 0x22, 4, 64, 512*1024, sl3_cache_str}, 2952 { 0x0d, 4, 32, 16*1024, l1_dcache_str}, 2953 { 0x0c, 4, 32, 16*1024, l1_dcache_str}, 2954 { 0x0b, 4, 0, 4, itlb4M_str}, 2955 { 0x0a, 2, 32, 8*1024, l1_dcache_str}, 2956 { 0x08, 4, 32, 16*1024, l1_icache_str}, 2957 { 0x06, 4, 32, 8*1024, l1_icache_str}, 2958 { 0x04, 4, 0, 8, dtlb4M_str}, 2959 { 0x03, 4, 0, 64, dtlb4k_str}, 2960 { 0x02, 4, 0, 2, itlb4M_str}, 2961 { 0x01, 4, 0, 32, itlb4k_str}, 2962 { 0 } 2963 }; 2964 2965 static const struct cachetab cyrix_ctab[] = { 2966 { 0x70, 4, 0, 32, "tlb-4K" }, 2967 { 0x80, 4, 16, 16*1024, "l1-cache" }, 2968 { 0 } 2969 }; 2970 2971 /* 2972 * Search a cache table for a matching entry 2973 */ 2974 static const struct cachetab * 2975 find_cacheent(const struct cachetab *ct, uint_t code) 2976 { 2977 if (code != 0) { 2978 for (; ct->ct_code != 0; ct++) 2979 if (ct->ct_code <= code) 2980 break; 2981 if (ct->ct_code == code) 2982 return (ct); 2983 } 2984 return (NULL); 2985 } 2986 2987 /* 2988 * Populate cachetab entry with L2 or L3 cache-information using 2989 * cpuid function 4. This function is called from intel_walk_cacheinfo() 2990 * when descriptor 0x49 is encountered. It returns 0 if no such cache 2991 * information is found. 2992 */ 2993 static int 2994 intel_cpuid_4_cache_info(struct cachetab *ct, struct cpuid_info *cpi) 2995 { 2996 uint32_t level, i; 2997 int ret = 0; 2998 2999 for (i = 0; i < cpi->cpi_std_4_size; i++) { 3000 level = CPI_CACHE_LVL(cpi->cpi_std_4[i]); 3001 3002 if (level == 2 || level == 3) { 3003 ct->ct_assoc = CPI_CACHE_WAYS(cpi->cpi_std_4[i]) + 1; 3004 ct->ct_line_size = 3005 CPI_CACHE_COH_LN_SZ(cpi->cpi_std_4[i]) + 1; 3006 ct->ct_size = ct->ct_assoc * 3007 (CPI_CACHE_PARTS(cpi->cpi_std_4[i]) + 1) * 3008 ct->ct_line_size * 3009 (cpi->cpi_std_4[i]->cp_ecx + 1); 3010 3011 if (level == 2) { 3012 ct->ct_label = l2_cache_str; 3013 } else if (level == 3) { 3014 ct->ct_label = l3_cache_str; 3015 } 3016 ret = 1; 3017 } 3018 } 3019 3020 return (ret); 3021 } 3022 3023 /* 3024 * Walk the cacheinfo descriptor, applying 'func' to every valid element 3025 * The walk is terminated if the walker returns non-zero. 3026 */ 3027 static void 3028 intel_walk_cacheinfo(struct cpuid_info *cpi, 3029 void *arg, int (*func)(void *, const struct cachetab *)) 3030 { 3031 const struct cachetab *ct; 3032 struct cachetab des_49_ct; 3033 uint8_t *dp; 3034 int i; 3035 3036 if ((dp = cpi->cpi_cacheinfo) == NULL) 3037 return; 3038 for (i = 0; i < cpi->cpi_ncache; i++, dp++) { 3039 /* 3040 * For overloaded descriptor 0x49 we use cpuid function 4 3041 * if supported by the current processor, to create 3042 * cache information. 3043 */ 3044 if (*dp == 0x49 && cpi->cpi_maxeax >= 0x4 && 3045 intel_cpuid_4_cache_info(&des_49_ct, cpi) == 1) { 3046 ct = &des_49_ct; 3047 } else { 3048 if ((ct = find_cacheent(intel_ctab, *dp)) == NULL) { 3049 continue; 3050 } 3051 } 3052 3053 if (func(arg, ct) != 0) { 3054 break; 3055 } 3056 } 3057 } 3058 3059 /* 3060 * (Like the Intel one, except for Cyrix CPUs) 3061 */ 3062 static void 3063 cyrix_walk_cacheinfo(struct cpuid_info *cpi, 3064 void *arg, int (*func)(void *, const struct cachetab *)) 3065 { 3066 const struct cachetab *ct; 3067 uint8_t *dp; 3068 int i; 3069 3070 if ((dp = cpi->cpi_cacheinfo) == NULL) 3071 return; 3072 for (i = 0; i < cpi->cpi_ncache; i++, dp++) { 3073 /* 3074 * Search Cyrix-specific descriptor table first .. 3075 */ 3076 if ((ct = find_cacheent(cyrix_ctab, *dp)) != NULL) { 3077 if (func(arg, ct) != 0) 3078 break; 3079 continue; 3080 } 3081 /* 3082 * .. else fall back to the Intel one 3083 */ 3084 if ((ct = find_cacheent(intel_ctab, *dp)) != NULL) { 3085 if (func(arg, ct) != 0) 3086 break; 3087 continue; 3088 } 3089 } 3090 } 3091 3092 /* 3093 * A cacheinfo walker that adds associativity, line-size, and size properties 3094 * to the devinfo node it is passed as an argument. 3095 */ 3096 static int 3097 add_cacheent_props(void *arg, const struct cachetab *ct) 3098 { 3099 dev_info_t *devi = arg; 3100 3101 add_cache_prop(devi, ct->ct_label, assoc_str, ct->ct_assoc); 3102 if (ct->ct_line_size != 0) 3103 add_cache_prop(devi, ct->ct_label, line_str, 3104 ct->ct_line_size); 3105 add_cache_prop(devi, ct->ct_label, size_str, ct->ct_size); 3106 return (0); 3107 } 3108 3109 3110 static const char fully_assoc[] = "fully-associative?"; 3111 3112 /* 3113 * AMD style cache/tlb description 3114 * 3115 * Extended functions 5 and 6 directly describe properties of 3116 * tlbs and various cache levels. 3117 */ 3118 static void 3119 add_amd_assoc(dev_info_t *devi, const char *label, uint_t assoc) 3120 { 3121 switch (assoc) { 3122 case 0: /* reserved; ignore */ 3123 break; 3124 default: 3125 add_cache_prop(devi, label, assoc_str, assoc); 3126 break; 3127 case 0xff: 3128 add_cache_prop(devi, label, fully_assoc, 1); 3129 break; 3130 } 3131 } 3132 3133 static void 3134 add_amd_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size) 3135 { 3136 if (size == 0) 3137 return; 3138 add_cache_prop(devi, label, size_str, size); 3139 add_amd_assoc(devi, label, assoc); 3140 } 3141 3142 static void 3143 add_amd_cache(dev_info_t *devi, const char *label, 3144 uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size) 3145 { 3146 if (size == 0 || line_size == 0) 3147 return; 3148 add_amd_assoc(devi, label, assoc); 3149 /* 3150 * Most AMD parts have a sectored cache. Multiple cache lines are 3151 * associated with each tag. A sector consists of all cache lines 3152 * associated with a tag. For example, the AMD K6-III has a sector 3153 * size of 2 cache lines per tag. 3154 */ 3155 if (lines_per_tag != 0) 3156 add_cache_prop(devi, label, "lines-per-tag", lines_per_tag); 3157 add_cache_prop(devi, label, line_str, line_size); 3158 add_cache_prop(devi, label, size_str, size * 1024); 3159 } 3160 3161 static void 3162 add_amd_l2_assoc(dev_info_t *devi, const char *label, uint_t assoc) 3163 { 3164 switch (assoc) { 3165 case 0: /* off */ 3166 break; 3167 case 1: 3168 case 2: 3169 case 4: 3170 add_cache_prop(devi, label, assoc_str, assoc); 3171 break; 3172 case 6: 3173 add_cache_prop(devi, label, assoc_str, 8); 3174 break; 3175 case 8: 3176 add_cache_prop(devi, label, assoc_str, 16); 3177 break; 3178 case 0xf: 3179 add_cache_prop(devi, label, fully_assoc, 1); 3180 break; 3181 default: /* reserved; ignore */ 3182 break; 3183 } 3184 } 3185 3186 static void 3187 add_amd_l2_tlb(dev_info_t *devi, const char *label, uint_t assoc, uint_t size) 3188 { 3189 if (size == 0 || assoc == 0) 3190 return; 3191 add_amd_l2_assoc(devi, label, assoc); 3192 add_cache_prop(devi, label, size_str, size); 3193 } 3194 3195 static void 3196 add_amd_l2_cache(dev_info_t *devi, const char *label, 3197 uint_t size, uint_t assoc, uint_t lines_per_tag, uint_t line_size) 3198 { 3199 if (size == 0 || assoc == 0 || line_size == 0) 3200 return; 3201 add_amd_l2_assoc(devi, label, assoc); 3202 if (lines_per_tag != 0) 3203 add_cache_prop(devi, label, "lines-per-tag", lines_per_tag); 3204 add_cache_prop(devi, label, line_str, line_size); 3205 add_cache_prop(devi, label, size_str, size * 1024); 3206 } 3207 3208 static void 3209 amd_cache_info(struct cpuid_info *cpi, dev_info_t *devi) 3210 { 3211 struct cpuid_regs *cp; 3212 3213 if (cpi->cpi_xmaxeax < 0x80000005) 3214 return; 3215 cp = &cpi->cpi_extd[5]; 3216 3217 /* 3218 * 4M/2M L1 TLB configuration 3219 * 3220 * We report the size for 2M pages because AMD uses two 3221 * TLB entries for one 4M page. 3222 */ 3223 add_amd_tlb(devi, "dtlb-2M", 3224 BITX(cp->cp_eax, 31, 24), BITX(cp->cp_eax, 23, 16)); 3225 add_amd_tlb(devi, "itlb-2M", 3226 BITX(cp->cp_eax, 15, 8), BITX(cp->cp_eax, 7, 0)); 3227 3228 /* 3229 * 4K L1 TLB configuration 3230 */ 3231 3232 switch (cpi->cpi_vendor) { 3233 uint_t nentries; 3234 case X86_VENDOR_TM: 3235 if (cpi->cpi_family >= 5) { 3236 /* 3237 * Crusoe processors have 256 TLB entries, but 3238 * cpuid data format constrains them to only 3239 * reporting 255 of them. 3240 */ 3241 if ((nentries = BITX(cp->cp_ebx, 23, 16)) == 255) 3242 nentries = 256; 3243 /* 3244 * Crusoe processors also have a unified TLB 3245 */ 3246 add_amd_tlb(devi, "tlb-4K", BITX(cp->cp_ebx, 31, 24), 3247 nentries); 3248 break; 3249 } 3250 /*FALLTHROUGH*/ 3251 default: 3252 add_amd_tlb(devi, itlb4k_str, 3253 BITX(cp->cp_ebx, 31, 24), BITX(cp->cp_ebx, 23, 16)); 3254 add_amd_tlb(devi, dtlb4k_str, 3255 BITX(cp->cp_ebx, 15, 8), BITX(cp->cp_ebx, 7, 0)); 3256 break; 3257 } 3258 3259 /* 3260 * data L1 cache configuration 3261 */ 3262 3263 add_amd_cache(devi, l1_dcache_str, 3264 BITX(cp->cp_ecx, 31, 24), BITX(cp->cp_ecx, 23, 16), 3265 BITX(cp->cp_ecx, 15, 8), BITX(cp->cp_ecx, 7, 0)); 3266 3267 /* 3268 * code L1 cache configuration 3269 */ 3270 3271 add_amd_cache(devi, l1_icache_str, 3272 BITX(cp->cp_edx, 31, 24), BITX(cp->cp_edx, 23, 16), 3273 BITX(cp->cp_edx, 15, 8), BITX(cp->cp_edx, 7, 0)); 3274 3275 if (cpi->cpi_xmaxeax < 0x80000006) 3276 return; 3277 cp = &cpi->cpi_extd[6]; 3278 3279 /* Check for a unified L2 TLB for large pages */ 3280 3281 if (BITX(cp->cp_eax, 31, 16) == 0) 3282 add_amd_l2_tlb(devi, "l2-tlb-2M", 3283 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0)); 3284 else { 3285 add_amd_l2_tlb(devi, "l2-dtlb-2M", 3286 BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16)); 3287 add_amd_l2_tlb(devi, "l2-itlb-2M", 3288 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0)); 3289 } 3290 3291 /* Check for a unified L2 TLB for 4K pages */ 3292 3293 if (BITX(cp->cp_ebx, 31, 16) == 0) { 3294 add_amd_l2_tlb(devi, "l2-tlb-4K", 3295 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0)); 3296 } else { 3297 add_amd_l2_tlb(devi, "l2-dtlb-4K", 3298 BITX(cp->cp_eax, 31, 28), BITX(cp->cp_eax, 27, 16)); 3299 add_amd_l2_tlb(devi, "l2-itlb-4K", 3300 BITX(cp->cp_eax, 15, 12), BITX(cp->cp_eax, 11, 0)); 3301 } 3302 3303 add_amd_l2_cache(devi, l2_cache_str, 3304 BITX(cp->cp_ecx, 31, 16), BITX(cp->cp_ecx, 15, 12), 3305 BITX(cp->cp_ecx, 11, 8), BITX(cp->cp_ecx, 7, 0)); 3306 } 3307 3308 /* 3309 * There are two basic ways that the x86 world describes it cache 3310 * and tlb architecture - Intel's way and AMD's way. 3311 * 3312 * Return which flavor of cache architecture we should use 3313 */ 3314 static int 3315 x86_which_cacheinfo(struct cpuid_info *cpi) 3316 { 3317 switch (cpi->cpi_vendor) { 3318 case X86_VENDOR_Intel: 3319 if (cpi->cpi_maxeax >= 2) 3320 return (X86_VENDOR_Intel); 3321 break; 3322 case X86_VENDOR_AMD: 3323 /* 3324 * The K5 model 1 was the first part from AMD that reported 3325 * cache sizes via extended cpuid functions. 3326 */ 3327 if (cpi->cpi_family > 5 || 3328 (cpi->cpi_family == 5 && cpi->cpi_model >= 1)) 3329 return (X86_VENDOR_AMD); 3330 break; 3331 case X86_VENDOR_TM: 3332 if (cpi->cpi_family >= 5) 3333 return (X86_VENDOR_AMD); 3334 /*FALLTHROUGH*/ 3335 default: 3336 /* 3337 * If they have extended CPU data for 0x80000005 3338 * then we assume they have AMD-format cache 3339 * information. 3340 * 3341 * If not, and the vendor happens to be Cyrix, 3342 * then try our-Cyrix specific handler. 3343 * 3344 * If we're not Cyrix, then assume we're using Intel's 3345 * table-driven format instead. 3346 */ 3347 if (cpi->cpi_xmaxeax >= 0x80000005) 3348 return (X86_VENDOR_AMD); 3349 else if (cpi->cpi_vendor == X86_VENDOR_Cyrix) 3350 return (X86_VENDOR_Cyrix); 3351 else if (cpi->cpi_maxeax >= 2) 3352 return (X86_VENDOR_Intel); 3353 break; 3354 } 3355 return (-1); 3356 } 3357 3358 /* 3359 * create a node for the given cpu under the prom root node. 3360 * Also, create a cpu node in the device tree. 3361 */ 3362 static dev_info_t *cpu_nex_devi = NULL; 3363 static kmutex_t cpu_node_lock; 3364 3365 /* 3366 * Called from post_startup() and mp_startup() 3367 */ 3368 void 3369 add_cpunode2devtree(processorid_t cpu_id, struct cpuid_info *cpi) 3370 { 3371 dev_info_t *cpu_devi; 3372 int create; 3373 3374 mutex_enter(&cpu_node_lock); 3375 3376 /* 3377 * create a nexus node for all cpus identified as 'cpu_id' under 3378 * the root node. 3379 */ 3380 if (cpu_nex_devi == NULL) { 3381 if (ndi_devi_alloc(ddi_root_node(), "cpus", 3382 (pnode_t)DEVI_SID_NODEID, &cpu_nex_devi) != NDI_SUCCESS) { 3383 mutex_exit(&cpu_node_lock); 3384 return; 3385 } 3386 (void) ndi_devi_online(cpu_nex_devi, 0); 3387 } 3388 3389 /* 3390 * create a child node for cpu identified as 'cpu_id' 3391 */ 3392 cpu_devi = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID, 3393 cpu_id); 3394 if (cpu_devi == NULL) { 3395 mutex_exit(&cpu_node_lock); 3396 return; 3397 } 3398 3399 /* device_type */ 3400 3401 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi, 3402 "device_type", "cpu"); 3403 3404 /* reg */ 3405 3406 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3407 "reg", cpu_id); 3408 3409 /* cpu-mhz, and clock-frequency */ 3410 3411 if (cpu_freq > 0) { 3412 long long mul; 3413 3414 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3415 "cpu-mhz", cpu_freq); 3416 3417 if ((mul = cpu_freq * 1000000LL) <= INT_MAX) 3418 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3419 "clock-frequency", (int)mul); 3420 } 3421 3422 (void) ndi_devi_online(cpu_devi, 0); 3423 3424 if ((x86_feature & X86_CPUID) == 0) { 3425 mutex_exit(&cpu_node_lock); 3426 return; 3427 } 3428 3429 /* vendor-id */ 3430 3431 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi, 3432 "vendor-id", cpi->cpi_vendorstr); 3433 3434 if (cpi->cpi_maxeax == 0) { 3435 mutex_exit(&cpu_node_lock); 3436 return; 3437 } 3438 3439 /* 3440 * family, model, and step 3441 */ 3442 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3443 "family", CPI_FAMILY(cpi)); 3444 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3445 "cpu-model", CPI_MODEL(cpi)); 3446 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3447 "stepping-id", CPI_STEP(cpi)); 3448 3449 /* type */ 3450 3451 switch (cpi->cpi_vendor) { 3452 case X86_VENDOR_Intel: 3453 create = 1; 3454 break; 3455 default: 3456 create = 0; 3457 break; 3458 } 3459 if (create) 3460 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3461 "type", CPI_TYPE(cpi)); 3462 3463 /* ext-family */ 3464 3465 switch (cpi->cpi_vendor) { 3466 case X86_VENDOR_Intel: 3467 case X86_VENDOR_AMD: 3468 create = cpi->cpi_family >= 0xf; 3469 break; 3470 default: 3471 create = 0; 3472 break; 3473 } 3474 if (create) 3475 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3476 "ext-family", CPI_FAMILY_XTD(cpi)); 3477 3478 /* ext-model */ 3479 3480 switch (cpi->cpi_vendor) { 3481 case X86_VENDOR_Intel: 3482 create = IS_EXTENDED_MODEL_INTEL(cpi); 3483 break; 3484 case X86_VENDOR_AMD: 3485 create = CPI_FAMILY(cpi) == 0xf; 3486 break; 3487 default: 3488 create = 0; 3489 break; 3490 } 3491 if (create) 3492 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3493 "ext-model", CPI_MODEL_XTD(cpi)); 3494 3495 /* generation */ 3496 3497 switch (cpi->cpi_vendor) { 3498 case X86_VENDOR_AMD: 3499 /* 3500 * AMD K5 model 1 was the first part to support this 3501 */ 3502 create = cpi->cpi_xmaxeax >= 0x80000001; 3503 break; 3504 default: 3505 create = 0; 3506 break; 3507 } 3508 if (create) 3509 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3510 "generation", BITX((cpi)->cpi_extd[1].cp_eax, 11, 8)); 3511 3512 /* brand-id */ 3513 3514 switch (cpi->cpi_vendor) { 3515 case X86_VENDOR_Intel: 3516 /* 3517 * brand id first appeared on Pentium III Xeon model 8, 3518 * and Celeron model 8 processors and Opteron 3519 */ 3520 create = cpi->cpi_family > 6 || 3521 (cpi->cpi_family == 6 && cpi->cpi_model >= 8); 3522 break; 3523 case X86_VENDOR_AMD: 3524 create = cpi->cpi_family >= 0xf; 3525 break; 3526 default: 3527 create = 0; 3528 break; 3529 } 3530 if (create && cpi->cpi_brandid != 0) { 3531 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3532 "brand-id", cpi->cpi_brandid); 3533 } 3534 3535 /* chunks, and apic-id */ 3536 3537 switch (cpi->cpi_vendor) { 3538 /* 3539 * first available on Pentium IV and Opteron (K8) 3540 */ 3541 case X86_VENDOR_Intel: 3542 create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf; 3543 break; 3544 case X86_VENDOR_AMD: 3545 create = cpi->cpi_family >= 0xf; 3546 break; 3547 default: 3548 create = 0; 3549 break; 3550 } 3551 if (create) { 3552 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3553 "chunks", CPI_CHUNKS(cpi)); 3554 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3555 "apic-id", CPI_APIC_ID(cpi)); 3556 if (cpi->cpi_chipid >= 0) { 3557 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3558 "chip#", cpi->cpi_chipid); 3559 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3560 "clog#", cpi->cpi_clogid); 3561 } 3562 } 3563 3564 /* cpuid-features */ 3565 3566 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3567 "cpuid-features", CPI_FEATURES_EDX(cpi)); 3568 3569 3570 /* cpuid-features-ecx */ 3571 3572 switch (cpi->cpi_vendor) { 3573 case X86_VENDOR_Intel: 3574 create = IS_NEW_F6(cpi) || cpi->cpi_family >= 0xf; 3575 break; 3576 default: 3577 create = 0; 3578 break; 3579 } 3580 if (create) 3581 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3582 "cpuid-features-ecx", CPI_FEATURES_ECX(cpi)); 3583 3584 /* ext-cpuid-features */ 3585 3586 switch (cpi->cpi_vendor) { 3587 case X86_VENDOR_Intel: 3588 case X86_VENDOR_AMD: 3589 case X86_VENDOR_Cyrix: 3590 case X86_VENDOR_TM: 3591 case X86_VENDOR_Centaur: 3592 create = cpi->cpi_xmaxeax >= 0x80000001; 3593 break; 3594 default: 3595 create = 0; 3596 break; 3597 } 3598 if (create) { 3599 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3600 "ext-cpuid-features", CPI_FEATURES_XTD_EDX(cpi)); 3601 (void) ndi_prop_update_int(DDI_DEV_T_NONE, cpu_devi, 3602 "ext-cpuid-features-ecx", CPI_FEATURES_XTD_ECX(cpi)); 3603 } 3604 3605 /* 3606 * Brand String first appeared in Intel Pentium IV, AMD K5 3607 * model 1, and Cyrix GXm. On earlier models we try and 3608 * simulate something similar .. so this string should always 3609 * same -something- about the processor, however lame. 3610 */ 3611 (void) ndi_prop_update_string(DDI_DEV_T_NONE, cpu_devi, 3612 "brand-string", cpi->cpi_brandstr); 3613 3614 /* 3615 * Finally, cache and tlb information 3616 */ 3617 switch (x86_which_cacheinfo(cpi)) { 3618 case X86_VENDOR_Intel: 3619 intel_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props); 3620 break; 3621 case X86_VENDOR_Cyrix: 3622 cyrix_walk_cacheinfo(cpi, cpu_devi, add_cacheent_props); 3623 break; 3624 case X86_VENDOR_AMD: 3625 amd_cache_info(cpi, cpu_devi); 3626 break; 3627 default: 3628 break; 3629 } 3630 3631 mutex_exit(&cpu_node_lock); 3632 } 3633 3634 struct l2info { 3635 int *l2i_csz; 3636 int *l2i_lsz; 3637 int *l2i_assoc; 3638 int l2i_ret; 3639 }; 3640 3641 /* 3642 * A cacheinfo walker that fetches the size, line-size and associativity 3643 * of the L2 cache 3644 */ 3645 static int 3646 intel_l2cinfo(void *arg, const struct cachetab *ct) 3647 { 3648 struct l2info *l2i = arg; 3649 int *ip; 3650 3651 if (ct->ct_label != l2_cache_str && 3652 ct->ct_label != sl2_cache_str) 3653 return (0); /* not an L2 -- keep walking */ 3654 3655 if ((ip = l2i->l2i_csz) != NULL) 3656 *ip = ct->ct_size; 3657 if ((ip = l2i->l2i_lsz) != NULL) 3658 *ip = ct->ct_line_size; 3659 if ((ip = l2i->l2i_assoc) != NULL) 3660 *ip = ct->ct_assoc; 3661 l2i->l2i_ret = ct->ct_size; 3662 return (1); /* was an L2 -- terminate walk */ 3663 } 3664 3665 /* 3666 * AMD L2/L3 Cache and TLB Associativity Field Definition: 3667 * 3668 * Unlike the associativity for the L1 cache and tlb where the 8 bit 3669 * value is the associativity, the associativity for the L2 cache and 3670 * tlb is encoded in the following table. The 4 bit L2 value serves as 3671 * an index into the amd_afd[] array to determine the associativity. 3672 * -1 is undefined. 0 is fully associative. 3673 */ 3674 3675 static int amd_afd[] = 3676 {-1, 1, 2, -1, 4, -1, 8, -1, 16, -1, 32, 48, 64, 96, 128, 0}; 3677 3678 static void 3679 amd_l2cacheinfo(struct cpuid_info *cpi, struct l2info *l2i) 3680 { 3681 struct cpuid_regs *cp; 3682 uint_t size, assoc; 3683 int i; 3684 int *ip; 3685 3686 if (cpi->cpi_xmaxeax < 0x80000006) 3687 return; 3688 cp = &cpi->cpi_extd[6]; 3689 3690 if ((i = BITX(cp->cp_ecx, 15, 12)) != 0 && 3691 (size = BITX(cp->cp_ecx, 31, 16)) != 0) { 3692 uint_t cachesz = size * 1024; 3693 assoc = amd_afd[i]; 3694 3695 ASSERT(assoc != -1); 3696 3697 if ((ip = l2i->l2i_csz) != NULL) 3698 *ip = cachesz; 3699 if ((ip = l2i->l2i_lsz) != NULL) 3700 *ip = BITX(cp->cp_ecx, 7, 0); 3701 if ((ip = l2i->l2i_assoc) != NULL) 3702 *ip = assoc; 3703 l2i->l2i_ret = cachesz; 3704 } 3705 } 3706 3707 int 3708 getl2cacheinfo(cpu_t *cpu, int *csz, int *lsz, int *assoc) 3709 { 3710 struct cpuid_info *cpi = cpu->cpu_m.mcpu_cpi; 3711 struct l2info __l2info, *l2i = &__l2info; 3712 3713 l2i->l2i_csz = csz; 3714 l2i->l2i_lsz = lsz; 3715 l2i->l2i_assoc = assoc; 3716 l2i->l2i_ret = -1; 3717 3718 switch (x86_which_cacheinfo(cpi)) { 3719 case X86_VENDOR_Intel: 3720 intel_walk_cacheinfo(cpi, l2i, intel_l2cinfo); 3721 break; 3722 case X86_VENDOR_Cyrix: 3723 cyrix_walk_cacheinfo(cpi, l2i, intel_l2cinfo); 3724 break; 3725 case X86_VENDOR_AMD: 3726 amd_l2cacheinfo(cpi, l2i); 3727 break; 3728 default: 3729 break; 3730 } 3731 return (l2i->l2i_ret); 3732 } 3733 3734 #if !defined(__xpv) 3735 3736 uint32_t * 3737 cpuid_mwait_alloc(cpu_t *cpu) 3738 { 3739 uint32_t *ret; 3740 size_t mwait_size; 3741 3742 ASSERT(cpuid_checkpass(cpu, 2)); 3743 3744 mwait_size = cpu->cpu_m.mcpu_cpi->cpi_mwait.mon_max; 3745 if (mwait_size == 0) 3746 return (NULL); 3747 3748 /* 3749 * kmem_alloc() returns cache line size aligned data for mwait_size 3750 * allocations. mwait_size is currently cache line sized. Neither 3751 * of these implementation details are guarantied to be true in the 3752 * future. 3753 * 3754 * First try allocating mwait_size as kmem_alloc() currently returns 3755 * correctly aligned memory. If kmem_alloc() does not return 3756 * mwait_size aligned memory, then use mwait_size ROUNDUP. 3757 * 3758 * Set cpi_mwait.buf_actual and cpi_mwait.size_actual in case we 3759 * decide to free this memory. 3760 */ 3761 ret = kmem_zalloc(mwait_size, KM_SLEEP); 3762 if (ret == (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size)) { 3763 cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret; 3764 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size; 3765 *ret = MWAIT_RUNNING; 3766 return (ret); 3767 } else { 3768 kmem_free(ret, mwait_size); 3769 ret = kmem_zalloc(mwait_size * 2, KM_SLEEP); 3770 cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = ret; 3771 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = mwait_size * 2; 3772 ret = (uint32_t *)P2ROUNDUP((uintptr_t)ret, mwait_size); 3773 *ret = MWAIT_RUNNING; 3774 return (ret); 3775 } 3776 } 3777 3778 void 3779 cpuid_mwait_free(cpu_t *cpu) 3780 { 3781 ASSERT(cpuid_checkpass(cpu, 2)); 3782 3783 if (cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual != NULL && 3784 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual > 0) { 3785 kmem_free(cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual, 3786 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual); 3787 } 3788 3789 cpu->cpu_m.mcpu_cpi->cpi_mwait.buf_actual = NULL; 3790 cpu->cpu_m.mcpu_cpi->cpi_mwait.size_actual = 0; 3791 } 3792 3793 void 3794 patch_tsc_read(int flag) 3795 { 3796 size_t cnt; 3797 switch (flag) { 3798 case X86_NO_TSC: 3799 cnt = &_no_rdtsc_end - &_no_rdtsc_start; 3800 (void) memcpy((void *)tsc_read, (void *)&_no_rdtsc_start, cnt); 3801 break; 3802 case X86_HAVE_TSCP: 3803 cnt = &_tscp_end - &_tscp_start; 3804 (void) memcpy((void *)tsc_read, (void *)&_tscp_start, cnt); 3805 break; 3806 case X86_TSC_MFENCE: 3807 cnt = &_tsc_mfence_end - &_tsc_mfence_start; 3808 (void) memcpy((void *)tsc_read, 3809 (void *)&_tsc_mfence_start, cnt); 3810 break; 3811 default: 3812 break; 3813 } 3814 } 3815 3816 #endif /* !__xpv */ 3817