1 // SPDX-License-Identifier: GPL-2.0-only 2 #include <linux/export.h> 3 #include <linux/bitops.h> 4 #include <linux/elf.h> 5 #include <linux/mm.h> 6 7 #include <linux/io.h> 8 #include <linux/sched.h> 9 #include <linux/sched/clock.h> 10 #include <linux/random.h> 11 #include <asm/processor.h> 12 #include <asm/apic.h> 13 #include <asm/cacheinfo.h> 14 #include <asm/cpu.h> 15 #include <asm/spec-ctrl.h> 16 #include <asm/smp.h> 17 #include <asm/pci-direct.h> 18 #include <asm/delay.h> 19 #include <asm/debugreg.h> 20 21 #ifdef CONFIG_X86_64 22 # include <asm/mmconfig.h> 23 # include <asm/set_memory.h> 24 #endif 25 26 #include "cpu.h" 27 28 static const int amd_erratum_383[]; 29 static const int amd_erratum_400[]; 30 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum); 31 32 /* 33 * nodes_per_socket: Stores the number of nodes per socket. 34 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX 35 * Node Identifiers[10:8] 36 */ 37 static u32 nodes_per_socket = 1; 38 39 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p) 40 { 41 u32 gprs[8] = { 0 }; 42 int err; 43 44 WARN_ONCE((boot_cpu_data.x86 != 0xf), 45 "%s should only be used on K8!\n", __func__); 46 47 gprs[1] = msr; 48 gprs[7] = 0x9c5a203a; 49 50 err = rdmsr_safe_regs(gprs); 51 52 *p = gprs[0] | ((u64)gprs[2] << 32); 53 54 return err; 55 } 56 57 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val) 58 { 59 u32 gprs[8] = { 0 }; 60 61 WARN_ONCE((boot_cpu_data.x86 != 0xf), 62 "%s should only be used on K8!\n", __func__); 63 64 gprs[0] = (u32)val; 65 gprs[1] = msr; 66 gprs[2] = val >> 32; 67 gprs[7] = 0x9c5a203a; 68 69 return wrmsr_safe_regs(gprs); 70 } 71 72 /* 73 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause 74 * misexecution of code under Linux. Owners of such processors should 75 * contact AMD for precise details and a CPU swap. 76 * 77 * See http://www.multimania.com/poulot/k6bug.html 78 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6" 79 * (Publication # 21266 Issue Date: August 1998) 80 * 81 * The following test is erm.. interesting. AMD neglected to up 82 * the chip setting when fixing the bug but they also tweaked some 83 * performance at the same time.. 84 */ 85 86 #ifdef CONFIG_X86_32 87 extern __visible void vide(void); 88 __asm__(".text\n" 89 ".globl vide\n" 90 ".type vide, @function\n" 91 ".align 4\n" 92 "vide: ret\n"); 93 #endif 94 95 static void init_amd_k5(struct cpuinfo_x86 *c) 96 { 97 #ifdef CONFIG_X86_32 98 /* 99 * General Systems BIOSen alias the cpu frequency registers 100 * of the Elan at 0x000df000. Unfortunately, one of the Linux 101 * drivers subsequently pokes it, and changes the CPU speed. 102 * Workaround : Remove the unneeded alias. 103 */ 104 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */ 105 #define CBAR_ENB (0x80000000) 106 #define CBAR_KEY (0X000000CB) 107 if (c->x86_model == 9 || c->x86_model == 10) { 108 if (inl(CBAR) & CBAR_ENB) 109 outl(0 | CBAR_KEY, CBAR); 110 } 111 #endif 112 } 113 114 static void init_amd_k6(struct cpuinfo_x86 *c) 115 { 116 #ifdef CONFIG_X86_32 117 u32 l, h; 118 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT); 119 120 if (c->x86_model < 6) { 121 /* Based on AMD doc 20734R - June 2000 */ 122 if (c->x86_model == 0) { 123 clear_cpu_cap(c, X86_FEATURE_APIC); 124 set_cpu_cap(c, X86_FEATURE_PGE); 125 } 126 return; 127 } 128 129 if (c->x86_model == 6 && c->x86_stepping == 1) { 130 const int K6_BUG_LOOP = 1000000; 131 int n; 132 void (*f_vide)(void); 133 u64 d, d2; 134 135 pr_info("AMD K6 stepping B detected - "); 136 137 /* 138 * It looks like AMD fixed the 2.6.2 bug and improved indirect 139 * calls at the same time. 140 */ 141 142 n = K6_BUG_LOOP; 143 f_vide = vide; 144 OPTIMIZER_HIDE_VAR(f_vide); 145 d = rdtsc(); 146 while (n--) 147 f_vide(); 148 d2 = rdtsc(); 149 d = d2-d; 150 151 if (d > 20*K6_BUG_LOOP) 152 pr_cont("system stability may be impaired when more than 32 MB are used.\n"); 153 else 154 pr_cont("probably OK (after B9730xxxx).\n"); 155 } 156 157 /* K6 with old style WHCR */ 158 if (c->x86_model < 8 || 159 (c->x86_model == 8 && c->x86_stepping < 8)) { 160 /* We can only write allocate on the low 508Mb */ 161 if (mbytes > 508) 162 mbytes = 508; 163 164 rdmsr(MSR_K6_WHCR, l, h); 165 if ((l&0x0000FFFF) == 0) { 166 unsigned long flags; 167 l = (1<<0)|((mbytes/4)<<1); 168 local_irq_save(flags); 169 wbinvd(); 170 wrmsr(MSR_K6_WHCR, l, h); 171 local_irq_restore(flags); 172 pr_info("Enabling old style K6 write allocation for %d Mb\n", 173 mbytes); 174 } 175 return; 176 } 177 178 if ((c->x86_model == 8 && c->x86_stepping > 7) || 179 c->x86_model == 9 || c->x86_model == 13) { 180 /* The more serious chips .. */ 181 182 if (mbytes > 4092) 183 mbytes = 4092; 184 185 rdmsr(MSR_K6_WHCR, l, h); 186 if ((l&0xFFFF0000) == 0) { 187 unsigned long flags; 188 l = ((mbytes>>2)<<22)|(1<<16); 189 local_irq_save(flags); 190 wbinvd(); 191 wrmsr(MSR_K6_WHCR, l, h); 192 local_irq_restore(flags); 193 pr_info("Enabling new style K6 write allocation for %d Mb\n", 194 mbytes); 195 } 196 197 return; 198 } 199 200 if (c->x86_model == 10) { 201 /* AMD Geode LX is model 10 */ 202 /* placeholder for any needed mods */ 203 return; 204 } 205 #endif 206 } 207 208 static void init_amd_k7(struct cpuinfo_x86 *c) 209 { 210 #ifdef CONFIG_X86_32 211 u32 l, h; 212 213 /* 214 * Bit 15 of Athlon specific MSR 15, needs to be 0 215 * to enable SSE on Palomino/Morgan/Barton CPU's. 216 * If the BIOS didn't enable it already, enable it here. 217 */ 218 if (c->x86_model >= 6 && c->x86_model <= 10) { 219 if (!cpu_has(c, X86_FEATURE_XMM)) { 220 pr_info("Enabling disabled K7/SSE Support.\n"); 221 msr_clear_bit(MSR_K7_HWCR, 15); 222 set_cpu_cap(c, X86_FEATURE_XMM); 223 } 224 } 225 226 /* 227 * It's been determined by AMD that Athlons since model 8 stepping 1 228 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx 229 * As per AMD technical note 27212 0.2 230 */ 231 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) { 232 rdmsr(MSR_K7_CLK_CTL, l, h); 233 if ((l & 0xfff00000) != 0x20000000) { 234 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n", 235 l, ((l & 0x000fffff)|0x20000000)); 236 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h); 237 } 238 } 239 240 /* calling is from identify_secondary_cpu() ? */ 241 if (!c->cpu_index) 242 return; 243 244 /* 245 * Certain Athlons might work (for various values of 'work') in SMP 246 * but they are not certified as MP capable. 247 */ 248 /* Athlon 660/661 is valid. */ 249 if ((c->x86_model == 6) && ((c->x86_stepping == 0) || 250 (c->x86_stepping == 1))) 251 return; 252 253 /* Duron 670 is valid */ 254 if ((c->x86_model == 7) && (c->x86_stepping == 0)) 255 return; 256 257 /* 258 * Athlon 662, Duron 671, and Athlon >model 7 have capability 259 * bit. It's worth noting that the A5 stepping (662) of some 260 * Athlon XP's have the MP bit set. 261 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for 262 * more. 263 */ 264 if (((c->x86_model == 6) && (c->x86_stepping >= 2)) || 265 ((c->x86_model == 7) && (c->x86_stepping >= 1)) || 266 (c->x86_model > 7)) 267 if (cpu_has(c, X86_FEATURE_MP)) 268 return; 269 270 /* If we get here, not a certified SMP capable AMD system. */ 271 272 /* 273 * Don't taint if we are running SMP kernel on a single non-MP 274 * approved Athlon 275 */ 276 WARN_ONCE(1, "WARNING: This combination of AMD" 277 " processors is not suitable for SMP.\n"); 278 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE); 279 #endif 280 } 281 282 #ifdef CONFIG_NUMA 283 /* 284 * To workaround broken NUMA config. Read the comment in 285 * srat_detect_node(). 286 */ 287 static int nearby_node(int apicid) 288 { 289 int i, node; 290 291 for (i = apicid - 1; i >= 0; i--) { 292 node = __apicid_to_node[i]; 293 if (node != NUMA_NO_NODE && node_online(node)) 294 return node; 295 } 296 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) { 297 node = __apicid_to_node[i]; 298 if (node != NUMA_NO_NODE && node_online(node)) 299 return node; 300 } 301 return first_node(node_online_map); /* Shouldn't happen */ 302 } 303 #endif 304 305 /* 306 * Fix up cpu_core_id for pre-F17h systems to be in the 307 * [0 .. cores_per_node - 1] range. Not really needed but 308 * kept so as not to break existing setups. 309 */ 310 static void legacy_fixup_core_id(struct cpuinfo_x86 *c) 311 { 312 u32 cus_per_node; 313 314 if (c->x86 >= 0x17) 315 return; 316 317 cus_per_node = c->x86_max_cores / nodes_per_socket; 318 c->cpu_core_id %= cus_per_node; 319 } 320 321 322 static void amd_get_topology_early(struct cpuinfo_x86 *c) 323 { 324 if (cpu_has(c, X86_FEATURE_TOPOEXT)) 325 smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1; 326 } 327 328 /* 329 * Fixup core topology information for 330 * (1) AMD multi-node processors 331 * Assumption: Number of cores in each internal node is the same. 332 * (2) AMD processors supporting compute units 333 */ 334 static void amd_get_topology(struct cpuinfo_x86 *c) 335 { 336 u8 node_id; 337 int cpu = smp_processor_id(); 338 339 /* get information required for multi-node processors */ 340 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) { 341 int err; 342 u32 eax, ebx, ecx, edx; 343 344 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx); 345 346 node_id = ecx & 0xff; 347 348 if (c->x86 == 0x15) 349 c->cu_id = ebx & 0xff; 350 351 if (c->x86 >= 0x17) { 352 c->cpu_core_id = ebx & 0xff; 353 354 if (smp_num_siblings > 1) 355 c->x86_max_cores /= smp_num_siblings; 356 } 357 358 /* 359 * In case leaf B is available, use it to derive 360 * topology information. 361 */ 362 err = detect_extended_topology(c); 363 if (!err) 364 c->x86_coreid_bits = get_count_order(c->x86_max_cores); 365 366 cacheinfo_amd_init_llc_id(c, cpu, node_id); 367 368 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) { 369 u64 value; 370 371 rdmsrl(MSR_FAM10H_NODE_ID, value); 372 node_id = value & 7; 373 374 per_cpu(cpu_llc_id, cpu) = node_id; 375 } else 376 return; 377 378 if (nodes_per_socket > 1) { 379 set_cpu_cap(c, X86_FEATURE_AMD_DCM); 380 legacy_fixup_core_id(c); 381 } 382 } 383 384 /* 385 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores. 386 * Assumes number of cores is a power of two. 387 */ 388 static void amd_detect_cmp(struct cpuinfo_x86 *c) 389 { 390 unsigned bits; 391 int cpu = smp_processor_id(); 392 393 bits = c->x86_coreid_bits; 394 /* Low order bits define the core id (index of core in socket) */ 395 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1); 396 /* Convert the initial APIC ID into the socket ID */ 397 c->phys_proc_id = c->initial_apicid >> bits; 398 /* use socket ID also for last level cache */ 399 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id; 400 } 401 402 u16 amd_get_nb_id(int cpu) 403 { 404 return per_cpu(cpu_llc_id, cpu); 405 } 406 EXPORT_SYMBOL_GPL(amd_get_nb_id); 407 408 u32 amd_get_nodes_per_socket(void) 409 { 410 return nodes_per_socket; 411 } 412 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket); 413 414 static void srat_detect_node(struct cpuinfo_x86 *c) 415 { 416 #ifdef CONFIG_NUMA 417 int cpu = smp_processor_id(); 418 int node; 419 unsigned apicid = c->apicid; 420 421 node = numa_cpu_node(cpu); 422 if (node == NUMA_NO_NODE) 423 node = per_cpu(cpu_llc_id, cpu); 424 425 /* 426 * On multi-fabric platform (e.g. Numascale NumaChip) a 427 * platform-specific handler needs to be called to fixup some 428 * IDs of the CPU. 429 */ 430 if (x86_cpuinit.fixup_cpu_id) 431 x86_cpuinit.fixup_cpu_id(c, node); 432 433 if (!node_online(node)) { 434 /* 435 * Two possibilities here: 436 * 437 * - The CPU is missing memory and no node was created. In 438 * that case try picking one from a nearby CPU. 439 * 440 * - The APIC IDs differ from the HyperTransport node IDs 441 * which the K8 northbridge parsing fills in. Assume 442 * they are all increased by a constant offset, but in 443 * the same order as the HT nodeids. If that doesn't 444 * result in a usable node fall back to the path for the 445 * previous case. 446 * 447 * This workaround operates directly on the mapping between 448 * APIC ID and NUMA node, assuming certain relationship 449 * between APIC ID, HT node ID and NUMA topology. As going 450 * through CPU mapping may alter the outcome, directly 451 * access __apicid_to_node[]. 452 */ 453 int ht_nodeid = c->initial_apicid; 454 455 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE) 456 node = __apicid_to_node[ht_nodeid]; 457 /* Pick a nearby node */ 458 if (!node_online(node)) 459 node = nearby_node(apicid); 460 } 461 numa_set_node(cpu, node); 462 #endif 463 } 464 465 static void early_init_amd_mc(struct cpuinfo_x86 *c) 466 { 467 #ifdef CONFIG_SMP 468 unsigned bits, ecx; 469 470 /* Multi core CPU? */ 471 if (c->extended_cpuid_level < 0x80000008) 472 return; 473 474 ecx = cpuid_ecx(0x80000008); 475 476 c->x86_max_cores = (ecx & 0xff) + 1; 477 478 /* CPU telling us the core id bits shift? */ 479 bits = (ecx >> 12) & 0xF; 480 481 /* Otherwise recompute */ 482 if (bits == 0) { 483 while ((1 << bits) < c->x86_max_cores) 484 bits++; 485 } 486 487 c->x86_coreid_bits = bits; 488 #endif 489 } 490 491 static void bsp_init_amd(struct cpuinfo_x86 *c) 492 { 493 494 #ifdef CONFIG_X86_64 495 if (c->x86 >= 0xf) { 496 unsigned long long tseg; 497 498 /* 499 * Split up direct mapping around the TSEG SMM area. 500 * Don't do it for gbpages because there seems very little 501 * benefit in doing so. 502 */ 503 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) { 504 unsigned long pfn = tseg >> PAGE_SHIFT; 505 506 pr_debug("tseg: %010llx\n", tseg); 507 if (pfn_range_is_mapped(pfn, pfn + 1)) 508 set_memory_4k((unsigned long)__va(tseg), 1); 509 } 510 } 511 #endif 512 513 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) { 514 515 if (c->x86 > 0x10 || 516 (c->x86 == 0x10 && c->x86_model >= 0x2)) { 517 u64 val; 518 519 rdmsrl(MSR_K7_HWCR, val); 520 if (!(val & BIT(24))) 521 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n"); 522 } 523 } 524 525 if (c->x86 == 0x15) { 526 unsigned long upperbit; 527 u32 cpuid, assoc; 528 529 cpuid = cpuid_edx(0x80000005); 530 assoc = cpuid >> 16 & 0xff; 531 upperbit = ((cpuid >> 24) << 10) / assoc; 532 533 va_align.mask = (upperbit - 1) & PAGE_MASK; 534 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64; 535 536 /* A random value per boot for bit slice [12:upper_bit) */ 537 va_align.bits = get_random_int() & va_align.mask; 538 } 539 540 if (cpu_has(c, X86_FEATURE_MWAITX)) 541 use_mwaitx_delay(); 542 543 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) { 544 u32 ecx; 545 546 ecx = cpuid_ecx(0x8000001e); 547 nodes_per_socket = ((ecx >> 8) & 7) + 1; 548 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) { 549 u64 value; 550 551 rdmsrl(MSR_FAM10H_NODE_ID, value); 552 nodes_per_socket = ((value >> 3) & 7) + 1; 553 } 554 555 if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) && 556 !boot_cpu_has(X86_FEATURE_VIRT_SSBD) && 557 c->x86 >= 0x15 && c->x86 <= 0x17) { 558 unsigned int bit; 559 560 switch (c->x86) { 561 case 0x15: bit = 54; break; 562 case 0x16: bit = 33; break; 563 case 0x17: bit = 10; break; 564 default: return; 565 } 566 /* 567 * Try to cache the base value so further operations can 568 * avoid RMW. If that faults, do not enable SSBD. 569 */ 570 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) { 571 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD); 572 setup_force_cpu_cap(X86_FEATURE_SSBD); 573 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit; 574 } 575 } 576 } 577 578 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c) 579 { 580 u64 msr; 581 582 /* 583 * BIOS support is required for SME and SEV. 584 * For SME: If BIOS has enabled SME then adjust x86_phys_bits by 585 * the SME physical address space reduction value. 586 * If BIOS has not enabled SME then don't advertise the 587 * SME feature (set in scattered.c). 588 * For SEV: If BIOS has not enabled SEV then don't advertise the 589 * SEV feature (set in scattered.c). 590 * 591 * In all cases, since support for SME and SEV requires long mode, 592 * don't advertise the feature under CONFIG_X86_32. 593 */ 594 if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) { 595 /* Check if memory encryption is enabled */ 596 rdmsrl(MSR_K8_SYSCFG, msr); 597 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT)) 598 goto clear_all; 599 600 /* 601 * Always adjust physical address bits. Even though this 602 * will be a value above 32-bits this is still done for 603 * CONFIG_X86_32 so that accurate values are reported. 604 */ 605 c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f; 606 607 if (IS_ENABLED(CONFIG_X86_32)) 608 goto clear_all; 609 610 rdmsrl(MSR_K7_HWCR, msr); 611 if (!(msr & MSR_K7_HWCR_SMMLOCK)) 612 goto clear_sev; 613 614 return; 615 616 clear_all: 617 clear_cpu_cap(c, X86_FEATURE_SME); 618 clear_sev: 619 clear_cpu_cap(c, X86_FEATURE_SEV); 620 } 621 } 622 623 static void early_init_amd(struct cpuinfo_x86 *c) 624 { 625 u64 value; 626 u32 dummy; 627 628 early_init_amd_mc(c); 629 630 #ifdef CONFIG_X86_32 631 if (c->x86 == 6) 632 set_cpu_cap(c, X86_FEATURE_K7); 633 #endif 634 635 if (c->x86 >= 0xf) 636 set_cpu_cap(c, X86_FEATURE_K8); 637 638 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy); 639 640 /* 641 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate 642 * with P/T states and does not stop in deep C-states 643 */ 644 if (c->x86_power & (1 << 8)) { 645 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); 646 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); 647 } 648 649 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */ 650 if (c->x86_power & BIT(12)) 651 set_cpu_cap(c, X86_FEATURE_ACC_POWER); 652 653 #ifdef CONFIG_X86_64 654 set_cpu_cap(c, X86_FEATURE_SYSCALL32); 655 #else 656 /* Set MTRR capability flag if appropriate */ 657 if (c->x86 == 5) 658 if (c->x86_model == 13 || c->x86_model == 9 || 659 (c->x86_model == 8 && c->x86_stepping >= 8)) 660 set_cpu_cap(c, X86_FEATURE_K6_MTRR); 661 #endif 662 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI) 663 /* 664 * ApicID can always be treated as an 8-bit value for AMD APIC versions 665 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we 666 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families 667 * after 16h. 668 */ 669 if (boot_cpu_has(X86_FEATURE_APIC)) { 670 if (c->x86 > 0x16) 671 set_cpu_cap(c, X86_FEATURE_EXTD_APICID); 672 else if (c->x86 >= 0xf) { 673 /* check CPU config space for extended APIC ID */ 674 unsigned int val; 675 676 val = read_pci_config(0, 24, 0, 0x68); 677 if ((val >> 17 & 0x3) == 0x3) 678 set_cpu_cap(c, X86_FEATURE_EXTD_APICID); 679 } 680 } 681 #endif 682 683 /* 684 * This is only needed to tell the kernel whether to use VMCALL 685 * and VMMCALL. VMMCALL is never executed except under virt, so 686 * we can set it unconditionally. 687 */ 688 set_cpu_cap(c, X86_FEATURE_VMMCALL); 689 690 /* F16h erratum 793, CVE-2013-6885 */ 691 if (c->x86 == 0x16 && c->x86_model <= 0xf) 692 msr_set_bit(MSR_AMD64_LS_CFG, 15); 693 694 /* 695 * Check whether the machine is affected by erratum 400. This is 696 * used to select the proper idle routine and to enable the check 697 * whether the machine is affected in arch_post_acpi_init(), which 698 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check. 699 */ 700 if (cpu_has_amd_erratum(c, amd_erratum_400)) 701 set_cpu_bug(c, X86_BUG_AMD_E400); 702 703 early_detect_mem_encrypt(c); 704 705 /* Re-enable TopologyExtensions if switched off by BIOS */ 706 if (c->x86 == 0x15 && 707 (c->x86_model >= 0x10 && c->x86_model <= 0x6f) && 708 !cpu_has(c, X86_FEATURE_TOPOEXT)) { 709 710 if (msr_set_bit(0xc0011005, 54) > 0) { 711 rdmsrl(0xc0011005, value); 712 if (value & BIT_64(54)) { 713 set_cpu_cap(c, X86_FEATURE_TOPOEXT); 714 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n"); 715 } 716 } 717 } 718 719 amd_get_topology_early(c); 720 } 721 722 static void init_amd_k8(struct cpuinfo_x86 *c) 723 { 724 u32 level; 725 u64 value; 726 727 /* On C+ stepping K8 rep microcode works well for copy/memset */ 728 level = cpuid_eax(1); 729 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58) 730 set_cpu_cap(c, X86_FEATURE_REP_GOOD); 731 732 /* 733 * Some BIOSes incorrectly force this feature, but only K8 revision D 734 * (model = 0x14) and later actually support it. 735 * (AMD Erratum #110, docId: 25759). 736 */ 737 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) { 738 clear_cpu_cap(c, X86_FEATURE_LAHF_LM); 739 if (!rdmsrl_amd_safe(0xc001100d, &value)) { 740 value &= ~BIT_64(32); 741 wrmsrl_amd_safe(0xc001100d, value); 742 } 743 } 744 745 if (!c->x86_model_id[0]) 746 strcpy(c->x86_model_id, "Hammer"); 747 748 #ifdef CONFIG_SMP 749 /* 750 * Disable TLB flush filter by setting HWCR.FFDIS on K8 751 * bit 6 of msr C001_0015 752 * 753 * Errata 63 for SH-B3 steppings 754 * Errata 122 for all steppings (F+ have it disabled by default) 755 */ 756 msr_set_bit(MSR_K7_HWCR, 6); 757 #endif 758 set_cpu_bug(c, X86_BUG_SWAPGS_FENCE); 759 } 760 761 static void init_amd_gh(struct cpuinfo_x86 *c) 762 { 763 #ifdef CONFIG_MMCONF_FAM10H 764 /* do this for boot cpu */ 765 if (c == &boot_cpu_data) 766 check_enable_amd_mmconf_dmi(); 767 768 fam10h_check_enable_mmcfg(); 769 #endif 770 771 /* 772 * Disable GART TLB Walk Errors on Fam10h. We do this here because this 773 * is always needed when GART is enabled, even in a kernel which has no 774 * MCE support built in. BIOS should disable GartTlbWlk Errors already. 775 * If it doesn't, we do it here as suggested by the BKDG. 776 * 777 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012 778 */ 779 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10); 780 781 /* 782 * On family 10h BIOS may not have properly enabled WC+ support, causing 783 * it to be converted to CD memtype. This may result in performance 784 * degradation for certain nested-paging guests. Prevent this conversion 785 * by clearing bit 24 in MSR_AMD64_BU_CFG2. 786 * 787 * NOTE: we want to use the _safe accessors so as not to #GP kvm 788 * guests on older kvm hosts. 789 */ 790 msr_clear_bit(MSR_AMD64_BU_CFG2, 24); 791 792 if (cpu_has_amd_erratum(c, amd_erratum_383)) 793 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH); 794 } 795 796 #define MSR_AMD64_DE_CFG 0xC0011029 797 798 static void init_amd_ln(struct cpuinfo_x86 *c) 799 { 800 /* 801 * Apply erratum 665 fix unconditionally so machines without a BIOS 802 * fix work. 803 */ 804 msr_set_bit(MSR_AMD64_DE_CFG, 31); 805 } 806 807 static void init_amd_bd(struct cpuinfo_x86 *c) 808 { 809 u64 value; 810 811 /* 812 * The way access filter has a performance penalty on some workloads. 813 * Disable it on the affected CPUs. 814 */ 815 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) { 816 if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) { 817 value |= 0x1E; 818 wrmsrl_safe(MSR_F15H_IC_CFG, value); 819 } 820 } 821 } 822 823 static void init_amd_zn(struct cpuinfo_x86 *c) 824 { 825 set_cpu_cap(c, X86_FEATURE_ZEN); 826 827 /* 828 * Fix erratum 1076: CPB feature bit not being set in CPUID. 829 * Always set it, except when running under a hypervisor. 830 */ 831 if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_CPB)) 832 set_cpu_cap(c, X86_FEATURE_CPB); 833 } 834 835 static void init_amd(struct cpuinfo_x86 *c) 836 { 837 early_init_amd(c); 838 839 /* 840 * Bit 31 in normal CPUID used for nonstandard 3DNow ID; 841 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway 842 */ 843 clear_cpu_cap(c, 0*32+31); 844 845 if (c->x86 >= 0x10) 846 set_cpu_cap(c, X86_FEATURE_REP_GOOD); 847 848 /* get apicid instead of initial apic id from cpuid */ 849 c->apicid = hard_smp_processor_id(); 850 851 /* K6s reports MCEs but don't actually have all the MSRs */ 852 if (c->x86 < 6) 853 clear_cpu_cap(c, X86_FEATURE_MCE); 854 855 switch (c->x86) { 856 case 4: init_amd_k5(c); break; 857 case 5: init_amd_k6(c); break; 858 case 6: init_amd_k7(c); break; 859 case 0xf: init_amd_k8(c); break; 860 case 0x10: init_amd_gh(c); break; 861 case 0x12: init_amd_ln(c); break; 862 case 0x15: init_amd_bd(c); break; 863 case 0x17: init_amd_zn(c); break; 864 } 865 866 /* 867 * Enable workaround for FXSAVE leak on CPUs 868 * without a XSaveErPtr feature 869 */ 870 if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR))) 871 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK); 872 873 cpu_detect_cache_sizes(c); 874 875 amd_detect_cmp(c); 876 amd_get_topology(c); 877 srat_detect_node(c); 878 879 init_amd_cacheinfo(c); 880 881 if (cpu_has(c, X86_FEATURE_XMM2)) { 882 unsigned long long val; 883 int ret; 884 885 /* 886 * A serializing LFENCE has less overhead than MFENCE, so 887 * use it for execution serialization. On families which 888 * don't have that MSR, LFENCE is already serializing. 889 * msr_set_bit() uses the safe accessors, too, even if the MSR 890 * is not present. 891 */ 892 msr_set_bit(MSR_F10H_DECFG, 893 MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT); 894 895 /* 896 * Verify that the MSR write was successful (could be running 897 * under a hypervisor) and only then assume that LFENCE is 898 * serializing. 899 */ 900 ret = rdmsrl_safe(MSR_F10H_DECFG, &val); 901 if (!ret && (val & MSR_F10H_DECFG_LFENCE_SERIALIZE)) { 902 /* A serializing LFENCE stops RDTSC speculation */ 903 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC); 904 } else { 905 /* MFENCE stops RDTSC speculation */ 906 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC); 907 } 908 } 909 910 /* 911 * Family 0x12 and above processors have APIC timer 912 * running in deep C states. 913 */ 914 if (c->x86 > 0x11) 915 set_cpu_cap(c, X86_FEATURE_ARAT); 916 917 /* 3DNow or LM implies PREFETCHW */ 918 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH)) 919 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM)) 920 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH); 921 922 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */ 923 if (!cpu_has(c, X86_FEATURE_XENPV)) 924 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS); 925 } 926 927 #ifdef CONFIG_X86_32 928 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size) 929 { 930 /* AMD errata T13 (order #21922) */ 931 if (c->x86 == 6) { 932 /* Duron Rev A0 */ 933 if (c->x86_model == 3 && c->x86_stepping == 0) 934 size = 64; 935 /* Tbird rev A1/A2 */ 936 if (c->x86_model == 4 && 937 (c->x86_stepping == 0 || c->x86_stepping == 1)) 938 size = 256; 939 } 940 return size; 941 } 942 #endif 943 944 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c) 945 { 946 u32 ebx, eax, ecx, edx; 947 u16 mask = 0xfff; 948 949 if (c->x86 < 0xf) 950 return; 951 952 if (c->extended_cpuid_level < 0x80000006) 953 return; 954 955 cpuid(0x80000006, &eax, &ebx, &ecx, &edx); 956 957 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask; 958 tlb_lli_4k[ENTRIES] = ebx & mask; 959 960 /* 961 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB 962 * characteristics from the CPUID function 0x80000005 instead. 963 */ 964 if (c->x86 == 0xf) { 965 cpuid(0x80000005, &eax, &ebx, &ecx, &edx); 966 mask = 0xff; 967 } 968 969 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ 970 if (!((eax >> 16) & mask)) 971 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff; 972 else 973 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask; 974 975 /* a 4M entry uses two 2M entries */ 976 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1; 977 978 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ 979 if (!(eax & mask)) { 980 /* Erratum 658 */ 981 if (c->x86 == 0x15 && c->x86_model <= 0x1f) { 982 tlb_lli_2m[ENTRIES] = 1024; 983 } else { 984 cpuid(0x80000005, &eax, &ebx, &ecx, &edx); 985 tlb_lli_2m[ENTRIES] = eax & 0xff; 986 } 987 } else 988 tlb_lli_2m[ENTRIES] = eax & mask; 989 990 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1; 991 } 992 993 static const struct cpu_dev amd_cpu_dev = { 994 .c_vendor = "AMD", 995 .c_ident = { "AuthenticAMD" }, 996 #ifdef CONFIG_X86_32 997 .legacy_models = { 998 { .family = 4, .model_names = 999 { 1000 [3] = "486 DX/2", 1001 [7] = "486 DX/2-WB", 1002 [8] = "486 DX/4", 1003 [9] = "486 DX/4-WB", 1004 [14] = "Am5x86-WT", 1005 [15] = "Am5x86-WB" 1006 } 1007 }, 1008 }, 1009 .legacy_cache_size = amd_size_cache, 1010 #endif 1011 .c_early_init = early_init_amd, 1012 .c_detect_tlb = cpu_detect_tlb_amd, 1013 .c_bsp_init = bsp_init_amd, 1014 .c_init = init_amd, 1015 .c_x86_vendor = X86_VENDOR_AMD, 1016 }; 1017 1018 cpu_dev_register(amd_cpu_dev); 1019 1020 /* 1021 * AMD errata checking 1022 * 1023 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or 1024 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that 1025 * have an OSVW id assigned, which it takes as first argument. Both take a 1026 * variable number of family-specific model-stepping ranges created by 1027 * AMD_MODEL_RANGE(). 1028 * 1029 * Example: 1030 * 1031 * const int amd_erratum_319[] = 1032 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2), 1033 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0), 1034 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0)); 1035 */ 1036 1037 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 } 1038 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 } 1039 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \ 1040 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end)) 1041 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff) 1042 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff) 1043 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff) 1044 1045 static const int amd_erratum_400[] = 1046 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf), 1047 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf)); 1048 1049 static const int amd_erratum_383[] = 1050 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf)); 1051 1052 1053 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum) 1054 { 1055 int osvw_id = *erratum++; 1056 u32 range; 1057 u32 ms; 1058 1059 if (osvw_id >= 0 && osvw_id < 65536 && 1060 cpu_has(cpu, X86_FEATURE_OSVW)) { 1061 u64 osvw_len; 1062 1063 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len); 1064 if (osvw_id < osvw_len) { 1065 u64 osvw_bits; 1066 1067 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6), 1068 osvw_bits); 1069 return osvw_bits & (1ULL << (osvw_id & 0x3f)); 1070 } 1071 } 1072 1073 /* OSVW unavailable or ID unknown, match family-model-stepping range */ 1074 ms = (cpu->x86_model << 4) | cpu->x86_stepping; 1075 while ((range = *erratum++)) 1076 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) && 1077 (ms >= AMD_MODEL_RANGE_START(range)) && 1078 (ms <= AMD_MODEL_RANGE_END(range))) 1079 return true; 1080 1081 return false; 1082 } 1083 1084 void set_dr_addr_mask(unsigned long mask, int dr) 1085 { 1086 if (!boot_cpu_has(X86_FEATURE_BPEXT)) 1087 return; 1088 1089 switch (dr) { 1090 case 0: 1091 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0); 1092 break; 1093 case 1: 1094 case 2: 1095 case 3: 1096 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0); 1097 break; 1098 default: 1099 break; 1100 } 1101 } 1102