1 #include <linux/export.h> 2 #include <linux/bitops.h> 3 #include <linux/elf.h> 4 #include <linux/mm.h> 5 6 #include <linux/io.h> 7 #include <linux/sched.h> 8 #include <asm/processor.h> 9 #include <asm/apic.h> 10 #include <asm/cpu.h> 11 #include <asm/smp.h> 12 #include <asm/pci-direct.h> 13 14 #ifdef CONFIG_X86_64 15 # include <asm/mmconfig.h> 16 # include <asm/cacheflush.h> 17 #endif 18 19 #include "cpu.h" 20 21 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p) 22 { 23 u32 gprs[8] = { 0 }; 24 int err; 25 26 WARN_ONCE((boot_cpu_data.x86 != 0xf), 27 "%s should only be used on K8!\n", __func__); 28 29 gprs[1] = msr; 30 gprs[7] = 0x9c5a203a; 31 32 err = rdmsr_safe_regs(gprs); 33 34 *p = gprs[0] | ((u64)gprs[2] << 32); 35 36 return err; 37 } 38 39 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val) 40 { 41 u32 gprs[8] = { 0 }; 42 43 WARN_ONCE((boot_cpu_data.x86 != 0xf), 44 "%s should only be used on K8!\n", __func__); 45 46 gprs[0] = (u32)val; 47 gprs[1] = msr; 48 gprs[2] = val >> 32; 49 gprs[7] = 0x9c5a203a; 50 51 return wrmsr_safe_regs(gprs); 52 } 53 54 /* 55 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause 56 * misexecution of code under Linux. Owners of such processors should 57 * contact AMD for precise details and a CPU swap. 58 * 59 * See http://www.multimania.com/poulot/k6bug.html 60 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6" 61 * (Publication # 21266 Issue Date: August 1998) 62 * 63 * The following test is erm.. interesting. AMD neglected to up 64 * the chip setting when fixing the bug but they also tweaked some 65 * performance at the same time.. 66 */ 67 68 extern __visible void vide(void); 69 __asm__(".globl vide\n\t.align 4\nvide: ret"); 70 71 static void init_amd_k5(struct cpuinfo_x86 *c) 72 { 73 #ifdef CONFIG_X86_32 74 /* 75 * General Systems BIOSen alias the cpu frequency registers 76 * of the Elan at 0x000df000. Unfortuantly, one of the Linux 77 * drivers subsequently pokes it, and changes the CPU speed. 78 * Workaround : Remove the unneeded alias. 79 */ 80 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */ 81 #define CBAR_ENB (0x80000000) 82 #define CBAR_KEY (0X000000CB) 83 if (c->x86_model == 9 || c->x86_model == 10) { 84 if (inl(CBAR) & CBAR_ENB) 85 outl(0 | CBAR_KEY, CBAR); 86 } 87 #endif 88 } 89 90 static void init_amd_k6(struct cpuinfo_x86 *c) 91 { 92 #ifdef CONFIG_X86_32 93 u32 l, h; 94 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT); 95 96 if (c->x86_model < 6) { 97 /* Based on AMD doc 20734R - June 2000 */ 98 if (c->x86_model == 0) { 99 clear_cpu_cap(c, X86_FEATURE_APIC); 100 set_cpu_cap(c, X86_FEATURE_PGE); 101 } 102 return; 103 } 104 105 if (c->x86_model == 6 && c->x86_mask == 1) { 106 const int K6_BUG_LOOP = 1000000; 107 int n; 108 void (*f_vide)(void); 109 unsigned long d, d2; 110 111 printk(KERN_INFO "AMD K6 stepping B detected - "); 112 113 /* 114 * It looks like AMD fixed the 2.6.2 bug and improved indirect 115 * calls at the same time. 116 */ 117 118 n = K6_BUG_LOOP; 119 f_vide = vide; 120 rdtscl(d); 121 while (n--) 122 f_vide(); 123 rdtscl(d2); 124 d = d2-d; 125 126 if (d > 20*K6_BUG_LOOP) 127 printk(KERN_CONT 128 "system stability may be impaired when more than 32 MB are used.\n"); 129 else 130 printk(KERN_CONT "probably OK (after B9730xxxx).\n"); 131 } 132 133 /* K6 with old style WHCR */ 134 if (c->x86_model < 8 || 135 (c->x86_model == 8 && c->x86_mask < 8)) { 136 /* We can only write allocate on the low 508Mb */ 137 if (mbytes > 508) 138 mbytes = 508; 139 140 rdmsr(MSR_K6_WHCR, l, h); 141 if ((l&0x0000FFFF) == 0) { 142 unsigned long flags; 143 l = (1<<0)|((mbytes/4)<<1); 144 local_irq_save(flags); 145 wbinvd(); 146 wrmsr(MSR_K6_WHCR, l, h); 147 local_irq_restore(flags); 148 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n", 149 mbytes); 150 } 151 return; 152 } 153 154 if ((c->x86_model == 8 && c->x86_mask > 7) || 155 c->x86_model == 9 || c->x86_model == 13) { 156 /* The more serious chips .. */ 157 158 if (mbytes > 4092) 159 mbytes = 4092; 160 161 rdmsr(MSR_K6_WHCR, l, h); 162 if ((l&0xFFFF0000) == 0) { 163 unsigned long flags; 164 l = ((mbytes>>2)<<22)|(1<<16); 165 local_irq_save(flags); 166 wbinvd(); 167 wrmsr(MSR_K6_WHCR, l, h); 168 local_irq_restore(flags); 169 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n", 170 mbytes); 171 } 172 173 return; 174 } 175 176 if (c->x86_model == 10) { 177 /* AMD Geode LX is model 10 */ 178 /* placeholder for any needed mods */ 179 return; 180 } 181 #endif 182 } 183 184 static void init_amd_k7(struct cpuinfo_x86 *c) 185 { 186 #ifdef CONFIG_X86_32 187 u32 l, h; 188 189 /* 190 * Bit 15 of Athlon specific MSR 15, needs to be 0 191 * to enable SSE on Palomino/Morgan/Barton CPU's. 192 * If the BIOS didn't enable it already, enable it here. 193 */ 194 if (c->x86_model >= 6 && c->x86_model <= 10) { 195 if (!cpu_has(c, X86_FEATURE_XMM)) { 196 printk(KERN_INFO "Enabling disabled K7/SSE Support.\n"); 197 msr_clear_bit(MSR_K7_HWCR, 15); 198 set_cpu_cap(c, X86_FEATURE_XMM); 199 } 200 } 201 202 /* 203 * It's been determined by AMD that Athlons since model 8 stepping 1 204 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx 205 * As per AMD technical note 27212 0.2 206 */ 207 if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) { 208 rdmsr(MSR_K7_CLK_CTL, l, h); 209 if ((l & 0xfff00000) != 0x20000000) { 210 printk(KERN_INFO 211 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n", 212 l, ((l & 0x000fffff)|0x20000000)); 213 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h); 214 } 215 } 216 217 set_cpu_cap(c, X86_FEATURE_K7); 218 219 /* calling is from identify_secondary_cpu() ? */ 220 if (!c->cpu_index) 221 return; 222 223 /* 224 * Certain Athlons might work (for various values of 'work') in SMP 225 * but they are not certified as MP capable. 226 */ 227 /* Athlon 660/661 is valid. */ 228 if ((c->x86_model == 6) && ((c->x86_mask == 0) || 229 (c->x86_mask == 1))) 230 return; 231 232 /* Duron 670 is valid */ 233 if ((c->x86_model == 7) && (c->x86_mask == 0)) 234 return; 235 236 /* 237 * Athlon 662, Duron 671, and Athlon >model 7 have capability 238 * bit. It's worth noting that the A5 stepping (662) of some 239 * Athlon XP's have the MP bit set. 240 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for 241 * more. 242 */ 243 if (((c->x86_model == 6) && (c->x86_mask >= 2)) || 244 ((c->x86_model == 7) && (c->x86_mask >= 1)) || 245 (c->x86_model > 7)) 246 if (cpu_has(c, X86_FEATURE_MP)) 247 return; 248 249 /* If we get here, not a certified SMP capable AMD system. */ 250 251 /* 252 * Don't taint if we are running SMP kernel on a single non-MP 253 * approved Athlon 254 */ 255 WARN_ONCE(1, "WARNING: This combination of AMD" 256 " processors is not suitable for SMP.\n"); 257 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE); 258 #endif 259 } 260 261 #ifdef CONFIG_NUMA 262 /* 263 * To workaround broken NUMA config. Read the comment in 264 * srat_detect_node(). 265 */ 266 static int nearby_node(int apicid) 267 { 268 int i, node; 269 270 for (i = apicid - 1; i >= 0; i--) { 271 node = __apicid_to_node[i]; 272 if (node != NUMA_NO_NODE && node_online(node)) 273 return node; 274 } 275 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) { 276 node = __apicid_to_node[i]; 277 if (node != NUMA_NO_NODE && node_online(node)) 278 return node; 279 } 280 return first_node(node_online_map); /* Shouldn't happen */ 281 } 282 #endif 283 284 /* 285 * Fixup core topology information for 286 * (1) AMD multi-node processors 287 * Assumption: Number of cores in each internal node is the same. 288 * (2) AMD processors supporting compute units 289 */ 290 #ifdef CONFIG_X86_HT 291 static void amd_get_topology(struct cpuinfo_x86 *c) 292 { 293 u32 nodes, cores_per_cu = 1; 294 u8 node_id; 295 int cpu = smp_processor_id(); 296 297 /* get information required for multi-node processors */ 298 if (cpu_has_topoext) { 299 u32 eax, ebx, ecx, edx; 300 301 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx); 302 nodes = ((ecx >> 8) & 7) + 1; 303 node_id = ecx & 7; 304 305 /* get compute unit information */ 306 smp_num_siblings = ((ebx >> 8) & 3) + 1; 307 c->compute_unit_id = ebx & 0xff; 308 cores_per_cu += ((ebx >> 8) & 3); 309 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) { 310 u64 value; 311 312 rdmsrl(MSR_FAM10H_NODE_ID, value); 313 nodes = ((value >> 3) & 7) + 1; 314 node_id = value & 7; 315 } else 316 return; 317 318 /* fixup multi-node processor information */ 319 if (nodes > 1) { 320 u32 cores_per_node; 321 u32 cus_per_node; 322 323 set_cpu_cap(c, X86_FEATURE_AMD_DCM); 324 cores_per_node = c->x86_max_cores / nodes; 325 cus_per_node = cores_per_node / cores_per_cu; 326 327 /* store NodeID, use llc_shared_map to store sibling info */ 328 per_cpu(cpu_llc_id, cpu) = node_id; 329 330 /* core id has to be in the [0 .. cores_per_node - 1] range */ 331 c->cpu_core_id %= cores_per_node; 332 c->compute_unit_id %= cus_per_node; 333 } 334 } 335 #endif 336 337 /* 338 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores. 339 * Assumes number of cores is a power of two. 340 */ 341 static void amd_detect_cmp(struct cpuinfo_x86 *c) 342 { 343 #ifdef CONFIG_X86_HT 344 unsigned bits; 345 int cpu = smp_processor_id(); 346 347 bits = c->x86_coreid_bits; 348 /* Low order bits define the core id (index of core in socket) */ 349 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1); 350 /* Convert the initial APIC ID into the socket ID */ 351 c->phys_proc_id = c->initial_apicid >> bits; 352 /* use socket ID also for last level cache */ 353 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id; 354 amd_get_topology(c); 355 #endif 356 } 357 358 u16 amd_get_nb_id(int cpu) 359 { 360 u16 id = 0; 361 #ifdef CONFIG_SMP 362 id = per_cpu(cpu_llc_id, cpu); 363 #endif 364 return id; 365 } 366 EXPORT_SYMBOL_GPL(amd_get_nb_id); 367 368 static void srat_detect_node(struct cpuinfo_x86 *c) 369 { 370 #ifdef CONFIG_NUMA 371 int cpu = smp_processor_id(); 372 int node; 373 unsigned apicid = c->apicid; 374 375 node = numa_cpu_node(cpu); 376 if (node == NUMA_NO_NODE) 377 node = per_cpu(cpu_llc_id, cpu); 378 379 /* 380 * On multi-fabric platform (e.g. Numascale NumaChip) a 381 * platform-specific handler needs to be called to fixup some 382 * IDs of the CPU. 383 */ 384 if (x86_cpuinit.fixup_cpu_id) 385 x86_cpuinit.fixup_cpu_id(c, node); 386 387 if (!node_online(node)) { 388 /* 389 * Two possibilities here: 390 * 391 * - The CPU is missing memory and no node was created. In 392 * that case try picking one from a nearby CPU. 393 * 394 * - The APIC IDs differ from the HyperTransport node IDs 395 * which the K8 northbridge parsing fills in. Assume 396 * they are all increased by a constant offset, but in 397 * the same order as the HT nodeids. If that doesn't 398 * result in a usable node fall back to the path for the 399 * previous case. 400 * 401 * This workaround operates directly on the mapping between 402 * APIC ID and NUMA node, assuming certain relationship 403 * between APIC ID, HT node ID and NUMA topology. As going 404 * through CPU mapping may alter the outcome, directly 405 * access __apicid_to_node[]. 406 */ 407 int ht_nodeid = c->initial_apicid; 408 409 if (ht_nodeid >= 0 && 410 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE) 411 node = __apicid_to_node[ht_nodeid]; 412 /* Pick a nearby node */ 413 if (!node_online(node)) 414 node = nearby_node(apicid); 415 } 416 numa_set_node(cpu, node); 417 #endif 418 } 419 420 static void early_init_amd_mc(struct cpuinfo_x86 *c) 421 { 422 #ifdef CONFIG_X86_HT 423 unsigned bits, ecx; 424 425 /* Multi core CPU? */ 426 if (c->extended_cpuid_level < 0x80000008) 427 return; 428 429 ecx = cpuid_ecx(0x80000008); 430 431 c->x86_max_cores = (ecx & 0xff) + 1; 432 433 /* CPU telling us the core id bits shift? */ 434 bits = (ecx >> 12) & 0xF; 435 436 /* Otherwise recompute */ 437 if (bits == 0) { 438 while ((1 << bits) < c->x86_max_cores) 439 bits++; 440 } 441 442 c->x86_coreid_bits = bits; 443 #endif 444 } 445 446 static void bsp_init_amd(struct cpuinfo_x86 *c) 447 { 448 449 #ifdef CONFIG_X86_64 450 if (c->x86 >= 0xf) { 451 unsigned long long tseg; 452 453 /* 454 * Split up direct mapping around the TSEG SMM area. 455 * Don't do it for gbpages because there seems very little 456 * benefit in doing so. 457 */ 458 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) { 459 unsigned long pfn = tseg >> PAGE_SHIFT; 460 461 printk(KERN_DEBUG "tseg: %010llx\n", tseg); 462 if (pfn_range_is_mapped(pfn, pfn + 1)) 463 set_memory_4k((unsigned long)__va(tseg), 1); 464 } 465 } 466 #endif 467 468 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) { 469 470 if (c->x86 > 0x10 || 471 (c->x86 == 0x10 && c->x86_model >= 0x2)) { 472 u64 val; 473 474 rdmsrl(MSR_K7_HWCR, val); 475 if (!(val & BIT(24))) 476 printk(KERN_WARNING FW_BUG "TSC doesn't count " 477 "with P0 frequency!\n"); 478 } 479 } 480 481 if (c->x86 == 0x15) { 482 unsigned long upperbit; 483 u32 cpuid, assoc; 484 485 cpuid = cpuid_edx(0x80000005); 486 assoc = cpuid >> 16 & 0xff; 487 upperbit = ((cpuid >> 24) << 10) / assoc; 488 489 va_align.mask = (upperbit - 1) & PAGE_MASK; 490 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64; 491 } 492 } 493 494 static void early_init_amd(struct cpuinfo_x86 *c) 495 { 496 early_init_amd_mc(c); 497 498 /* 499 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate 500 * with P/T states and does not stop in deep C-states 501 */ 502 if (c->x86_power & (1 << 8)) { 503 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); 504 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); 505 if (!check_tsc_unstable()) 506 set_sched_clock_stable(); 507 } 508 509 #ifdef CONFIG_X86_64 510 set_cpu_cap(c, X86_FEATURE_SYSCALL32); 511 #else 512 /* Set MTRR capability flag if appropriate */ 513 if (c->x86 == 5) 514 if (c->x86_model == 13 || c->x86_model == 9 || 515 (c->x86_model == 8 && c->x86_mask >= 8)) 516 set_cpu_cap(c, X86_FEATURE_K6_MTRR); 517 #endif 518 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI) 519 /* check CPU config space for extended APIC ID */ 520 if (cpu_has_apic && c->x86 >= 0xf) { 521 unsigned int val; 522 val = read_pci_config(0, 24, 0, 0x68); 523 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18))) 524 set_cpu_cap(c, X86_FEATURE_EXTD_APICID); 525 } 526 #endif 527 528 /* 529 * This is only needed to tell the kernel whether to use VMCALL 530 * and VMMCALL. VMMCALL is never executed except under virt, so 531 * we can set it unconditionally. 532 */ 533 set_cpu_cap(c, X86_FEATURE_VMMCALL); 534 535 /* F16h erratum 793, CVE-2013-6885 */ 536 if (c->x86 == 0x16 && c->x86_model <= 0xf) 537 msr_set_bit(MSR_AMD64_LS_CFG, 15); 538 } 539 540 static const int amd_erratum_383[]; 541 static const int amd_erratum_400[]; 542 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum); 543 544 static void init_amd_k8(struct cpuinfo_x86 *c) 545 { 546 u32 level; 547 u64 value; 548 549 /* On C+ stepping K8 rep microcode works well for copy/memset */ 550 level = cpuid_eax(1); 551 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58) 552 set_cpu_cap(c, X86_FEATURE_REP_GOOD); 553 554 /* 555 * Some BIOSes incorrectly force this feature, but only K8 revision D 556 * (model = 0x14) and later actually support it. 557 * (AMD Erratum #110, docId: 25759). 558 */ 559 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) { 560 clear_cpu_cap(c, X86_FEATURE_LAHF_LM); 561 if (!rdmsrl_amd_safe(0xc001100d, &value)) { 562 value &= ~BIT_64(32); 563 wrmsrl_amd_safe(0xc001100d, value); 564 } 565 } 566 567 if (!c->x86_model_id[0]) 568 strcpy(c->x86_model_id, "Hammer"); 569 570 #ifdef CONFIG_SMP 571 /* 572 * Disable TLB flush filter by setting HWCR.FFDIS on K8 573 * bit 6 of msr C001_0015 574 * 575 * Errata 63 for SH-B3 steppings 576 * Errata 122 for all steppings (F+ have it disabled by default) 577 */ 578 msr_set_bit(MSR_K7_HWCR, 6); 579 #endif 580 } 581 582 static void init_amd_gh(struct cpuinfo_x86 *c) 583 { 584 #ifdef CONFIG_X86_64 585 /* do this for boot cpu */ 586 if (c == &boot_cpu_data) 587 check_enable_amd_mmconf_dmi(); 588 589 fam10h_check_enable_mmcfg(); 590 #endif 591 592 /* 593 * Disable GART TLB Walk Errors on Fam10h. We do this here because this 594 * is always needed when GART is enabled, even in a kernel which has no 595 * MCE support built in. BIOS should disable GartTlbWlk Errors already. 596 * If it doesn't, we do it here as suggested by the BKDG. 597 * 598 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012 599 */ 600 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10); 601 602 /* 603 * On family 10h BIOS may not have properly enabled WC+ support, causing 604 * it to be converted to CD memtype. This may result in performance 605 * degradation for certain nested-paging guests. Prevent this conversion 606 * by clearing bit 24 in MSR_AMD64_BU_CFG2. 607 * 608 * NOTE: we want to use the _safe accessors so as not to #GP kvm 609 * guests on older kvm hosts. 610 */ 611 msr_clear_bit(MSR_AMD64_BU_CFG2, 24); 612 613 if (cpu_has_amd_erratum(c, amd_erratum_383)) 614 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH); 615 } 616 617 static void init_amd_bd(struct cpuinfo_x86 *c) 618 { 619 u64 value; 620 621 /* re-enable TopologyExtensions if switched off by BIOS */ 622 if ((c->x86_model >= 0x10) && (c->x86_model <= 0x1f) && 623 !cpu_has(c, X86_FEATURE_TOPOEXT)) { 624 625 if (msr_set_bit(0xc0011005, 54) > 0) { 626 rdmsrl(0xc0011005, value); 627 if (value & BIT_64(54)) { 628 set_cpu_cap(c, X86_FEATURE_TOPOEXT); 629 pr_info(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n"); 630 } 631 } 632 } 633 634 /* 635 * The way access filter has a performance penalty on some workloads. 636 * Disable it on the affected CPUs. 637 */ 638 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) { 639 if (!rdmsrl_safe(0xc0011021, &value) && !(value & 0x1E)) { 640 value |= 0x1E; 641 wrmsrl_safe(0xc0011021, value); 642 } 643 } 644 } 645 646 static void init_amd(struct cpuinfo_x86 *c) 647 { 648 u32 dummy; 649 650 early_init_amd(c); 651 652 /* 653 * Bit 31 in normal CPUID used for nonstandard 3DNow ID; 654 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway 655 */ 656 clear_cpu_cap(c, 0*32+31); 657 658 if (c->x86 >= 0x10) 659 set_cpu_cap(c, X86_FEATURE_REP_GOOD); 660 661 /* get apicid instead of initial apic id from cpuid */ 662 c->apicid = hard_smp_processor_id(); 663 664 /* K6s reports MCEs but don't actually have all the MSRs */ 665 if (c->x86 < 6) 666 clear_cpu_cap(c, X86_FEATURE_MCE); 667 668 switch (c->x86) { 669 case 4: init_amd_k5(c); break; 670 case 5: init_amd_k6(c); break; 671 case 6: init_amd_k7(c); break; 672 case 0xf: init_amd_k8(c); break; 673 case 0x10: init_amd_gh(c); break; 674 case 0x15: init_amd_bd(c); break; 675 } 676 677 /* Enable workaround for FXSAVE leak */ 678 if (c->x86 >= 6) 679 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK); 680 681 cpu_detect_cache_sizes(c); 682 683 /* Multi core CPU? */ 684 if (c->extended_cpuid_level >= 0x80000008) { 685 amd_detect_cmp(c); 686 srat_detect_node(c); 687 } 688 689 #ifdef CONFIG_X86_32 690 detect_ht(c); 691 #endif 692 693 init_amd_cacheinfo(c); 694 695 if (c->x86 >= 0xf) 696 set_cpu_cap(c, X86_FEATURE_K8); 697 698 if (cpu_has_xmm2) { 699 /* MFENCE stops RDTSC speculation */ 700 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC); 701 } 702 703 /* 704 * Family 0x12 and above processors have APIC timer 705 * running in deep C states. 706 */ 707 if (c->x86 > 0x11) 708 set_cpu_cap(c, X86_FEATURE_ARAT); 709 710 if (cpu_has_amd_erratum(c, amd_erratum_400)) 711 set_cpu_bug(c, X86_BUG_AMD_APIC_C1E); 712 713 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy); 714 } 715 716 #ifdef CONFIG_X86_32 717 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size) 718 { 719 /* AMD errata T13 (order #21922) */ 720 if ((c->x86 == 6)) { 721 /* Duron Rev A0 */ 722 if (c->x86_model == 3 && c->x86_mask == 0) 723 size = 64; 724 /* Tbird rev A1/A2 */ 725 if (c->x86_model == 4 && 726 (c->x86_mask == 0 || c->x86_mask == 1)) 727 size = 256; 728 } 729 return size; 730 } 731 #endif 732 733 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c) 734 { 735 u32 ebx, eax, ecx, edx; 736 u16 mask = 0xfff; 737 738 if (c->x86 < 0xf) 739 return; 740 741 if (c->extended_cpuid_level < 0x80000006) 742 return; 743 744 cpuid(0x80000006, &eax, &ebx, &ecx, &edx); 745 746 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask; 747 tlb_lli_4k[ENTRIES] = ebx & mask; 748 749 /* 750 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB 751 * characteristics from the CPUID function 0x80000005 instead. 752 */ 753 if (c->x86 == 0xf) { 754 cpuid(0x80000005, &eax, &ebx, &ecx, &edx); 755 mask = 0xff; 756 } 757 758 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ 759 if (!((eax >> 16) & mask)) 760 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff; 761 else 762 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask; 763 764 /* a 4M entry uses two 2M entries */ 765 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1; 766 767 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ 768 if (!(eax & mask)) { 769 /* Erratum 658 */ 770 if (c->x86 == 0x15 && c->x86_model <= 0x1f) { 771 tlb_lli_2m[ENTRIES] = 1024; 772 } else { 773 cpuid(0x80000005, &eax, &ebx, &ecx, &edx); 774 tlb_lli_2m[ENTRIES] = eax & 0xff; 775 } 776 } else 777 tlb_lli_2m[ENTRIES] = eax & mask; 778 779 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1; 780 } 781 782 static const struct cpu_dev amd_cpu_dev = { 783 .c_vendor = "AMD", 784 .c_ident = { "AuthenticAMD" }, 785 #ifdef CONFIG_X86_32 786 .legacy_models = { 787 { .family = 4, .model_names = 788 { 789 [3] = "486 DX/2", 790 [7] = "486 DX/2-WB", 791 [8] = "486 DX/4", 792 [9] = "486 DX/4-WB", 793 [14] = "Am5x86-WT", 794 [15] = "Am5x86-WB" 795 } 796 }, 797 }, 798 .legacy_cache_size = amd_size_cache, 799 #endif 800 .c_early_init = early_init_amd, 801 .c_detect_tlb = cpu_detect_tlb_amd, 802 .c_bsp_init = bsp_init_amd, 803 .c_init = init_amd, 804 .c_x86_vendor = X86_VENDOR_AMD, 805 }; 806 807 cpu_dev_register(amd_cpu_dev); 808 809 /* 810 * AMD errata checking 811 * 812 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or 813 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that 814 * have an OSVW id assigned, which it takes as first argument. Both take a 815 * variable number of family-specific model-stepping ranges created by 816 * AMD_MODEL_RANGE(). 817 * 818 * Example: 819 * 820 * const int amd_erratum_319[] = 821 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2), 822 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0), 823 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0)); 824 */ 825 826 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 } 827 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 } 828 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \ 829 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end)) 830 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff) 831 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff) 832 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff) 833 834 static const int amd_erratum_400[] = 835 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf), 836 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf)); 837 838 static const int amd_erratum_383[] = 839 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf)); 840 841 842 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum) 843 { 844 int osvw_id = *erratum++; 845 u32 range; 846 u32 ms; 847 848 if (osvw_id >= 0 && osvw_id < 65536 && 849 cpu_has(cpu, X86_FEATURE_OSVW)) { 850 u64 osvw_len; 851 852 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len); 853 if (osvw_id < osvw_len) { 854 u64 osvw_bits; 855 856 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6), 857 osvw_bits); 858 return osvw_bits & (1ULL << (osvw_id & 0x3f)); 859 } 860 } 861 862 /* OSVW unavailable or ID unknown, match family-model-stepping range */ 863 ms = (cpu->x86_model << 4) | cpu->x86_mask; 864 while ((range = *erratum++)) 865 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) && 866 (ms >= AMD_MODEL_RANGE_START(range)) && 867 (ms <= AMD_MODEL_RANGE_END(range))) 868 return true; 869 870 return false; 871 } 872 873 void set_dr_addr_mask(unsigned long mask, int dr) 874 { 875 if (!cpu_has_bpext) 876 return; 877 878 switch (dr) { 879 case 0: 880 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0); 881 break; 882 case 1: 883 case 2: 884 case 3: 885 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0); 886 break; 887 default: 888 break; 889 } 890 } 891