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 <linux/random.h> 9 #include <asm/processor.h> 10 #include <asm/apic.h> 11 #include <asm/cpu.h> 12 #include <asm/smp.h> 13 #include <asm/pci-direct.h> 14 #include <asm/delay.h> 15 16 #ifdef CONFIG_X86_64 17 # include <asm/mmconfig.h> 18 # include <asm/cacheflush.h> 19 #endif 20 21 #include "cpu.h" 22 23 /* 24 * nodes_per_socket: Stores the number of nodes per socket. 25 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX 26 * Node Identifiers[10:8] 27 */ 28 static u32 nodes_per_socket = 1; 29 30 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p) 31 { 32 u32 gprs[8] = { 0 }; 33 int err; 34 35 WARN_ONCE((boot_cpu_data.x86 != 0xf), 36 "%s should only be used on K8!\n", __func__); 37 38 gprs[1] = msr; 39 gprs[7] = 0x9c5a203a; 40 41 err = rdmsr_safe_regs(gprs); 42 43 *p = gprs[0] | ((u64)gprs[2] << 32); 44 45 return err; 46 } 47 48 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val) 49 { 50 u32 gprs[8] = { 0 }; 51 52 WARN_ONCE((boot_cpu_data.x86 != 0xf), 53 "%s should only be used on K8!\n", __func__); 54 55 gprs[0] = (u32)val; 56 gprs[1] = msr; 57 gprs[2] = val >> 32; 58 gprs[7] = 0x9c5a203a; 59 60 return wrmsr_safe_regs(gprs); 61 } 62 63 /* 64 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause 65 * misexecution of code under Linux. Owners of such processors should 66 * contact AMD for precise details and a CPU swap. 67 * 68 * See http://www.multimania.com/poulot/k6bug.html 69 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6" 70 * (Publication # 21266 Issue Date: August 1998) 71 * 72 * The following test is erm.. interesting. AMD neglected to up 73 * the chip setting when fixing the bug but they also tweaked some 74 * performance at the same time.. 75 */ 76 77 extern __visible void vide(void); 78 __asm__(".globl vide\n" 79 ".type vide, @function\n" 80 ".align 4\n" 81 "vide: ret\n"); 82 83 static void init_amd_k5(struct cpuinfo_x86 *c) 84 { 85 #ifdef CONFIG_X86_32 86 /* 87 * General Systems BIOSen alias the cpu frequency registers 88 * of the Elan at 0x000df000. Unfortunately, one of the Linux 89 * drivers subsequently pokes it, and changes the CPU speed. 90 * Workaround : Remove the unneeded alias. 91 */ 92 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */ 93 #define CBAR_ENB (0x80000000) 94 #define CBAR_KEY (0X000000CB) 95 if (c->x86_model == 9 || c->x86_model == 10) { 96 if (inl(CBAR) & CBAR_ENB) 97 outl(0 | CBAR_KEY, CBAR); 98 } 99 #endif 100 } 101 102 static void init_amd_k6(struct cpuinfo_x86 *c) 103 { 104 #ifdef CONFIG_X86_32 105 u32 l, h; 106 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT); 107 108 if (c->x86_model < 6) { 109 /* Based on AMD doc 20734R - June 2000 */ 110 if (c->x86_model == 0) { 111 clear_cpu_cap(c, X86_FEATURE_APIC); 112 set_cpu_cap(c, X86_FEATURE_PGE); 113 } 114 return; 115 } 116 117 if (c->x86_model == 6 && c->x86_mask == 1) { 118 const int K6_BUG_LOOP = 1000000; 119 int n; 120 void (*f_vide)(void); 121 u64 d, d2; 122 123 pr_info("AMD K6 stepping B detected - "); 124 125 /* 126 * It looks like AMD fixed the 2.6.2 bug and improved indirect 127 * calls at the same time. 128 */ 129 130 n = K6_BUG_LOOP; 131 f_vide = vide; 132 d = rdtsc(); 133 while (n--) 134 f_vide(); 135 d2 = rdtsc(); 136 d = d2-d; 137 138 if (d > 20*K6_BUG_LOOP) 139 pr_cont("system stability may be impaired when more than 32 MB are used.\n"); 140 else 141 pr_cont("probably OK (after B9730xxxx).\n"); 142 } 143 144 /* K6 with old style WHCR */ 145 if (c->x86_model < 8 || 146 (c->x86_model == 8 && c->x86_mask < 8)) { 147 /* We can only write allocate on the low 508Mb */ 148 if (mbytes > 508) 149 mbytes = 508; 150 151 rdmsr(MSR_K6_WHCR, l, h); 152 if ((l&0x0000FFFF) == 0) { 153 unsigned long flags; 154 l = (1<<0)|((mbytes/4)<<1); 155 local_irq_save(flags); 156 wbinvd(); 157 wrmsr(MSR_K6_WHCR, l, h); 158 local_irq_restore(flags); 159 pr_info("Enabling old style K6 write allocation for %d Mb\n", 160 mbytes); 161 } 162 return; 163 } 164 165 if ((c->x86_model == 8 && c->x86_mask > 7) || 166 c->x86_model == 9 || c->x86_model == 13) { 167 /* The more serious chips .. */ 168 169 if (mbytes > 4092) 170 mbytes = 4092; 171 172 rdmsr(MSR_K6_WHCR, l, h); 173 if ((l&0xFFFF0000) == 0) { 174 unsigned long flags; 175 l = ((mbytes>>2)<<22)|(1<<16); 176 local_irq_save(flags); 177 wbinvd(); 178 wrmsr(MSR_K6_WHCR, l, h); 179 local_irq_restore(flags); 180 pr_info("Enabling new style K6 write allocation for %d Mb\n", 181 mbytes); 182 } 183 184 return; 185 } 186 187 if (c->x86_model == 10) { 188 /* AMD Geode LX is model 10 */ 189 /* placeholder for any needed mods */ 190 return; 191 } 192 #endif 193 } 194 195 static void init_amd_k7(struct cpuinfo_x86 *c) 196 { 197 #ifdef CONFIG_X86_32 198 u32 l, h; 199 200 /* 201 * Bit 15 of Athlon specific MSR 15, needs to be 0 202 * to enable SSE on Palomino/Morgan/Barton CPU's. 203 * If the BIOS didn't enable it already, enable it here. 204 */ 205 if (c->x86_model >= 6 && c->x86_model <= 10) { 206 if (!cpu_has(c, X86_FEATURE_XMM)) { 207 pr_info("Enabling disabled K7/SSE Support.\n"); 208 msr_clear_bit(MSR_K7_HWCR, 15); 209 set_cpu_cap(c, X86_FEATURE_XMM); 210 } 211 } 212 213 /* 214 * It's been determined by AMD that Athlons since model 8 stepping 1 215 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx 216 * As per AMD technical note 27212 0.2 217 */ 218 if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) { 219 rdmsr(MSR_K7_CLK_CTL, l, h); 220 if ((l & 0xfff00000) != 0x20000000) { 221 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n", 222 l, ((l & 0x000fffff)|0x20000000)); 223 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h); 224 } 225 } 226 227 set_cpu_cap(c, X86_FEATURE_K7); 228 229 /* calling is from identify_secondary_cpu() ? */ 230 if (!c->cpu_index) 231 return; 232 233 /* 234 * Certain Athlons might work (for various values of 'work') in SMP 235 * but they are not certified as MP capable. 236 */ 237 /* Athlon 660/661 is valid. */ 238 if ((c->x86_model == 6) && ((c->x86_mask == 0) || 239 (c->x86_mask == 1))) 240 return; 241 242 /* Duron 670 is valid */ 243 if ((c->x86_model == 7) && (c->x86_mask == 0)) 244 return; 245 246 /* 247 * Athlon 662, Duron 671, and Athlon >model 7 have capability 248 * bit. It's worth noting that the A5 stepping (662) of some 249 * Athlon XP's have the MP bit set. 250 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for 251 * more. 252 */ 253 if (((c->x86_model == 6) && (c->x86_mask >= 2)) || 254 ((c->x86_model == 7) && (c->x86_mask >= 1)) || 255 (c->x86_model > 7)) 256 if (cpu_has(c, X86_FEATURE_MP)) 257 return; 258 259 /* If we get here, not a certified SMP capable AMD system. */ 260 261 /* 262 * Don't taint if we are running SMP kernel on a single non-MP 263 * approved Athlon 264 */ 265 WARN_ONCE(1, "WARNING: This combination of AMD" 266 " processors is not suitable for SMP.\n"); 267 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE); 268 #endif 269 } 270 271 #ifdef CONFIG_NUMA 272 /* 273 * To workaround broken NUMA config. Read the comment in 274 * srat_detect_node(). 275 */ 276 static int nearby_node(int apicid) 277 { 278 int i, node; 279 280 for (i = apicid - 1; i >= 0; i--) { 281 node = __apicid_to_node[i]; 282 if (node != NUMA_NO_NODE && node_online(node)) 283 return node; 284 } 285 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) { 286 node = __apicid_to_node[i]; 287 if (node != NUMA_NO_NODE && node_online(node)) 288 return node; 289 } 290 return first_node(node_online_map); /* Shouldn't happen */ 291 } 292 #endif 293 294 /* 295 * Fixup core topology information for 296 * (1) AMD multi-node processors 297 * Assumption: Number of cores in each internal node is the same. 298 * (2) AMD processors supporting compute units 299 */ 300 #ifdef CONFIG_SMP 301 static void amd_get_topology(struct cpuinfo_x86 *c) 302 { 303 u8 node_id; 304 int cpu = smp_processor_id(); 305 306 /* get information required for multi-node processors */ 307 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) { 308 u32 eax, ebx, ecx, edx; 309 310 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx); 311 node_id = ecx & 7; 312 313 /* get compute unit information */ 314 smp_num_siblings = ((ebx >> 8) & 3) + 1; 315 c->x86_max_cores /= smp_num_siblings; 316 c->cpu_core_id = ebx & 0xff; 317 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) { 318 u64 value; 319 320 rdmsrl(MSR_FAM10H_NODE_ID, value); 321 node_id = value & 7; 322 } else 323 return; 324 325 /* fixup multi-node processor information */ 326 if (nodes_per_socket > 1) { 327 u32 cus_per_node; 328 329 set_cpu_cap(c, X86_FEATURE_AMD_DCM); 330 cus_per_node = c->x86_max_cores / nodes_per_socket; 331 332 /* store NodeID, use llc_shared_map to store sibling info */ 333 per_cpu(cpu_llc_id, cpu) = node_id; 334 335 /* core id has to be in the [0 .. cores_per_node - 1] range */ 336 c->cpu_core_id %= cus_per_node; 337 } 338 } 339 #endif 340 341 /* 342 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores. 343 * Assumes number of cores is a power of two. 344 */ 345 static void amd_detect_cmp(struct cpuinfo_x86 *c) 346 { 347 #ifdef CONFIG_SMP 348 unsigned bits; 349 int cpu = smp_processor_id(); 350 unsigned int socket_id, core_complex_id; 351 352 bits = c->x86_coreid_bits; 353 /* Low order bits define the core id (index of core in socket) */ 354 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1); 355 /* Convert the initial APIC ID into the socket ID */ 356 c->phys_proc_id = c->initial_apicid >> bits; 357 /* use socket ID also for last level cache */ 358 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id; 359 amd_get_topology(c); 360 361 /* 362 * Fix percpu cpu_llc_id here as LLC topology is different 363 * for Fam17h systems. 364 */ 365 if (c->x86 != 0x17 || !cpuid_edx(0x80000006)) 366 return; 367 368 socket_id = (c->apicid >> bits) - 1; 369 core_complex_id = (c->apicid & ((1 << bits) - 1)) >> 3; 370 371 per_cpu(cpu_llc_id, cpu) = (socket_id << 3) | core_complex_id; 372 #endif 373 } 374 375 u16 amd_get_nb_id(int cpu) 376 { 377 u16 id = 0; 378 #ifdef CONFIG_SMP 379 id = per_cpu(cpu_llc_id, cpu); 380 #endif 381 return id; 382 } 383 EXPORT_SYMBOL_GPL(amd_get_nb_id); 384 385 u32 amd_get_nodes_per_socket(void) 386 { 387 return nodes_per_socket; 388 } 389 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket); 390 391 static void srat_detect_node(struct cpuinfo_x86 *c) 392 { 393 #ifdef CONFIG_NUMA 394 int cpu = smp_processor_id(); 395 int node; 396 unsigned apicid = c->apicid; 397 398 node = numa_cpu_node(cpu); 399 if (node == NUMA_NO_NODE) 400 node = per_cpu(cpu_llc_id, cpu); 401 402 /* 403 * On multi-fabric platform (e.g. Numascale NumaChip) a 404 * platform-specific handler needs to be called to fixup some 405 * IDs of the CPU. 406 */ 407 if (x86_cpuinit.fixup_cpu_id) 408 x86_cpuinit.fixup_cpu_id(c, node); 409 410 if (!node_online(node)) { 411 /* 412 * Two possibilities here: 413 * 414 * - The CPU is missing memory and no node was created. In 415 * that case try picking one from a nearby CPU. 416 * 417 * - The APIC IDs differ from the HyperTransport node IDs 418 * which the K8 northbridge parsing fills in. Assume 419 * they are all increased by a constant offset, but in 420 * the same order as the HT nodeids. If that doesn't 421 * result in a usable node fall back to the path for the 422 * previous case. 423 * 424 * This workaround operates directly on the mapping between 425 * APIC ID and NUMA node, assuming certain relationship 426 * between APIC ID, HT node ID and NUMA topology. As going 427 * through CPU mapping may alter the outcome, directly 428 * access __apicid_to_node[]. 429 */ 430 int ht_nodeid = c->initial_apicid; 431 432 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE) 433 node = __apicid_to_node[ht_nodeid]; 434 /* Pick a nearby node */ 435 if (!node_online(node)) 436 node = nearby_node(apicid); 437 } 438 numa_set_node(cpu, node); 439 #endif 440 } 441 442 static void early_init_amd_mc(struct cpuinfo_x86 *c) 443 { 444 #ifdef CONFIG_SMP 445 unsigned bits, ecx; 446 447 /* Multi core CPU? */ 448 if (c->extended_cpuid_level < 0x80000008) 449 return; 450 451 ecx = cpuid_ecx(0x80000008); 452 453 c->x86_max_cores = (ecx & 0xff) + 1; 454 455 /* CPU telling us the core id bits shift? */ 456 bits = (ecx >> 12) & 0xF; 457 458 /* Otherwise recompute */ 459 if (bits == 0) { 460 while ((1 << bits) < c->x86_max_cores) 461 bits++; 462 } 463 464 c->x86_coreid_bits = bits; 465 #endif 466 } 467 468 static void bsp_init_amd(struct cpuinfo_x86 *c) 469 { 470 471 #ifdef CONFIG_X86_64 472 if (c->x86 >= 0xf) { 473 unsigned long long tseg; 474 475 /* 476 * Split up direct mapping around the TSEG SMM area. 477 * Don't do it for gbpages because there seems very little 478 * benefit in doing so. 479 */ 480 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) { 481 unsigned long pfn = tseg >> PAGE_SHIFT; 482 483 pr_debug("tseg: %010llx\n", tseg); 484 if (pfn_range_is_mapped(pfn, pfn + 1)) 485 set_memory_4k((unsigned long)__va(tseg), 1); 486 } 487 } 488 #endif 489 490 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) { 491 492 if (c->x86 > 0x10 || 493 (c->x86 == 0x10 && c->x86_model >= 0x2)) { 494 u64 val; 495 496 rdmsrl(MSR_K7_HWCR, val); 497 if (!(val & BIT(24))) 498 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n"); 499 } 500 } 501 502 if (c->x86 == 0x15) { 503 unsigned long upperbit; 504 u32 cpuid, assoc; 505 506 cpuid = cpuid_edx(0x80000005); 507 assoc = cpuid >> 16 & 0xff; 508 upperbit = ((cpuid >> 24) << 10) / assoc; 509 510 va_align.mask = (upperbit - 1) & PAGE_MASK; 511 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64; 512 513 /* A random value per boot for bit slice [12:upper_bit) */ 514 va_align.bits = get_random_int() & va_align.mask; 515 } 516 517 if (cpu_has(c, X86_FEATURE_MWAITX)) 518 use_mwaitx_delay(); 519 520 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) { 521 u32 ecx; 522 523 ecx = cpuid_ecx(0x8000001e); 524 nodes_per_socket = ((ecx >> 8) & 7) + 1; 525 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) { 526 u64 value; 527 528 rdmsrl(MSR_FAM10H_NODE_ID, value); 529 nodes_per_socket = ((value >> 3) & 7) + 1; 530 } 531 } 532 533 static void early_init_amd(struct cpuinfo_x86 *c) 534 { 535 early_init_amd_mc(c); 536 537 /* 538 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate 539 * with P/T states and does not stop in deep C-states 540 */ 541 if (c->x86_power & (1 << 8)) { 542 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC); 543 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC); 544 if (!check_tsc_unstable()) 545 set_sched_clock_stable(); 546 } 547 548 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */ 549 if (c->x86_power & BIT(12)) 550 set_cpu_cap(c, X86_FEATURE_ACC_POWER); 551 552 #ifdef CONFIG_X86_64 553 set_cpu_cap(c, X86_FEATURE_SYSCALL32); 554 #else 555 /* Set MTRR capability flag if appropriate */ 556 if (c->x86 == 5) 557 if (c->x86_model == 13 || c->x86_model == 9 || 558 (c->x86_model == 8 && c->x86_mask >= 8)) 559 set_cpu_cap(c, X86_FEATURE_K6_MTRR); 560 #endif 561 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI) 562 /* 563 * ApicID can always be treated as an 8-bit value for AMD APIC versions 564 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we 565 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families 566 * after 16h. 567 */ 568 if (boot_cpu_has(X86_FEATURE_APIC)) { 569 if (c->x86 > 0x16) 570 set_cpu_cap(c, X86_FEATURE_EXTD_APICID); 571 else if (c->x86 >= 0xf) { 572 /* check CPU config space for extended APIC ID */ 573 unsigned int val; 574 575 val = read_pci_config(0, 24, 0, 0x68); 576 if ((val >> 17 & 0x3) == 0x3) 577 set_cpu_cap(c, X86_FEATURE_EXTD_APICID); 578 } 579 } 580 #endif 581 582 /* 583 * This is only needed to tell the kernel whether to use VMCALL 584 * and VMMCALL. VMMCALL is never executed except under virt, so 585 * we can set it unconditionally. 586 */ 587 set_cpu_cap(c, X86_FEATURE_VMMCALL); 588 589 /* F16h erratum 793, CVE-2013-6885 */ 590 if (c->x86 == 0x16 && c->x86_model <= 0xf) 591 msr_set_bit(MSR_AMD64_LS_CFG, 15); 592 } 593 594 static const int amd_erratum_383[]; 595 static const int amd_erratum_400[]; 596 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum); 597 598 static void init_amd_k8(struct cpuinfo_x86 *c) 599 { 600 u32 level; 601 u64 value; 602 603 /* On C+ stepping K8 rep microcode works well for copy/memset */ 604 level = cpuid_eax(1); 605 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58) 606 set_cpu_cap(c, X86_FEATURE_REP_GOOD); 607 608 /* 609 * Some BIOSes incorrectly force this feature, but only K8 revision D 610 * (model = 0x14) and later actually support it. 611 * (AMD Erratum #110, docId: 25759). 612 */ 613 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) { 614 clear_cpu_cap(c, X86_FEATURE_LAHF_LM); 615 if (!rdmsrl_amd_safe(0xc001100d, &value)) { 616 value &= ~BIT_64(32); 617 wrmsrl_amd_safe(0xc001100d, value); 618 } 619 } 620 621 if (!c->x86_model_id[0]) 622 strcpy(c->x86_model_id, "Hammer"); 623 624 #ifdef CONFIG_SMP 625 /* 626 * Disable TLB flush filter by setting HWCR.FFDIS on K8 627 * bit 6 of msr C001_0015 628 * 629 * Errata 63 for SH-B3 steppings 630 * Errata 122 for all steppings (F+ have it disabled by default) 631 */ 632 msr_set_bit(MSR_K7_HWCR, 6); 633 #endif 634 set_cpu_bug(c, X86_BUG_SWAPGS_FENCE); 635 } 636 637 static void init_amd_gh(struct cpuinfo_x86 *c) 638 { 639 #ifdef CONFIG_X86_64 640 /* do this for boot cpu */ 641 if (c == &boot_cpu_data) 642 check_enable_amd_mmconf_dmi(); 643 644 fam10h_check_enable_mmcfg(); 645 #endif 646 647 /* 648 * Disable GART TLB Walk Errors on Fam10h. We do this here because this 649 * is always needed when GART is enabled, even in a kernel which has no 650 * MCE support built in. BIOS should disable GartTlbWlk Errors already. 651 * If it doesn't, we do it here as suggested by the BKDG. 652 * 653 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012 654 */ 655 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10); 656 657 /* 658 * On family 10h BIOS may not have properly enabled WC+ support, causing 659 * it to be converted to CD memtype. This may result in performance 660 * degradation for certain nested-paging guests. Prevent this conversion 661 * by clearing bit 24 in MSR_AMD64_BU_CFG2. 662 * 663 * NOTE: we want to use the _safe accessors so as not to #GP kvm 664 * guests on older kvm hosts. 665 */ 666 msr_clear_bit(MSR_AMD64_BU_CFG2, 24); 667 668 if (cpu_has_amd_erratum(c, amd_erratum_383)) 669 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH); 670 } 671 672 #define MSR_AMD64_DE_CFG 0xC0011029 673 674 static void init_amd_ln(struct cpuinfo_x86 *c) 675 { 676 /* 677 * Apply erratum 665 fix unconditionally so machines without a BIOS 678 * fix work. 679 */ 680 msr_set_bit(MSR_AMD64_DE_CFG, 31); 681 } 682 683 static void init_amd_bd(struct cpuinfo_x86 *c) 684 { 685 u64 value; 686 687 /* re-enable TopologyExtensions if switched off by BIOS */ 688 if ((c->x86_model >= 0x10) && (c->x86_model <= 0x6f) && 689 !cpu_has(c, X86_FEATURE_TOPOEXT)) { 690 691 if (msr_set_bit(0xc0011005, 54) > 0) { 692 rdmsrl(0xc0011005, value); 693 if (value & BIT_64(54)) { 694 set_cpu_cap(c, X86_FEATURE_TOPOEXT); 695 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n"); 696 } 697 } 698 } 699 700 /* 701 * The way access filter has a performance penalty on some workloads. 702 * Disable it on the affected CPUs. 703 */ 704 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) { 705 if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) { 706 value |= 0x1E; 707 wrmsrl_safe(MSR_F15H_IC_CFG, value); 708 } 709 } 710 } 711 712 static void init_amd(struct cpuinfo_x86 *c) 713 { 714 u32 dummy; 715 716 early_init_amd(c); 717 718 /* 719 * Bit 31 in normal CPUID used for nonstandard 3DNow ID; 720 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway 721 */ 722 clear_cpu_cap(c, 0*32+31); 723 724 if (c->x86 >= 0x10) 725 set_cpu_cap(c, X86_FEATURE_REP_GOOD); 726 727 /* get apicid instead of initial apic id from cpuid */ 728 c->apicid = hard_smp_processor_id(); 729 730 /* K6s reports MCEs but don't actually have all the MSRs */ 731 if (c->x86 < 6) 732 clear_cpu_cap(c, X86_FEATURE_MCE); 733 734 switch (c->x86) { 735 case 4: init_amd_k5(c); break; 736 case 5: init_amd_k6(c); break; 737 case 6: init_amd_k7(c); break; 738 case 0xf: init_amd_k8(c); break; 739 case 0x10: init_amd_gh(c); break; 740 case 0x12: init_amd_ln(c); break; 741 case 0x15: init_amd_bd(c); break; 742 } 743 744 /* Enable workaround for FXSAVE leak */ 745 if (c->x86 >= 6) 746 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK); 747 748 cpu_detect_cache_sizes(c); 749 750 /* Multi core CPU? */ 751 if (c->extended_cpuid_level >= 0x80000008) { 752 amd_detect_cmp(c); 753 srat_detect_node(c); 754 } 755 756 #ifdef CONFIG_X86_32 757 detect_ht(c); 758 #endif 759 760 init_amd_cacheinfo(c); 761 762 if (c->x86 >= 0xf) 763 set_cpu_cap(c, X86_FEATURE_K8); 764 765 if (cpu_has(c, X86_FEATURE_XMM2)) { 766 /* MFENCE stops RDTSC speculation */ 767 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC); 768 } 769 770 /* 771 * Family 0x12 and above processors have APIC timer 772 * running in deep C states. 773 */ 774 if (c->x86 > 0x11) 775 set_cpu_cap(c, X86_FEATURE_ARAT); 776 777 if (cpu_has_amd_erratum(c, amd_erratum_400)) 778 set_cpu_bug(c, X86_BUG_AMD_APIC_C1E); 779 780 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy); 781 782 /* 3DNow or LM implies PREFETCHW */ 783 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH)) 784 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM)) 785 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH); 786 787 /* AMD CPUs don't reset SS attributes on SYSRET */ 788 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS); 789 } 790 791 #ifdef CONFIG_X86_32 792 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size) 793 { 794 /* AMD errata T13 (order #21922) */ 795 if ((c->x86 == 6)) { 796 /* Duron Rev A0 */ 797 if (c->x86_model == 3 && c->x86_mask == 0) 798 size = 64; 799 /* Tbird rev A1/A2 */ 800 if (c->x86_model == 4 && 801 (c->x86_mask == 0 || c->x86_mask == 1)) 802 size = 256; 803 } 804 return size; 805 } 806 #endif 807 808 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c) 809 { 810 u32 ebx, eax, ecx, edx; 811 u16 mask = 0xfff; 812 813 if (c->x86 < 0xf) 814 return; 815 816 if (c->extended_cpuid_level < 0x80000006) 817 return; 818 819 cpuid(0x80000006, &eax, &ebx, &ecx, &edx); 820 821 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask; 822 tlb_lli_4k[ENTRIES] = ebx & mask; 823 824 /* 825 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB 826 * characteristics from the CPUID function 0x80000005 instead. 827 */ 828 if (c->x86 == 0xf) { 829 cpuid(0x80000005, &eax, &ebx, &ecx, &edx); 830 mask = 0xff; 831 } 832 833 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ 834 if (!((eax >> 16) & mask)) 835 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff; 836 else 837 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask; 838 839 /* a 4M entry uses two 2M entries */ 840 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1; 841 842 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */ 843 if (!(eax & mask)) { 844 /* Erratum 658 */ 845 if (c->x86 == 0x15 && c->x86_model <= 0x1f) { 846 tlb_lli_2m[ENTRIES] = 1024; 847 } else { 848 cpuid(0x80000005, &eax, &ebx, &ecx, &edx); 849 tlb_lli_2m[ENTRIES] = eax & 0xff; 850 } 851 } else 852 tlb_lli_2m[ENTRIES] = eax & mask; 853 854 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1; 855 } 856 857 static const struct cpu_dev amd_cpu_dev = { 858 .c_vendor = "AMD", 859 .c_ident = { "AuthenticAMD" }, 860 #ifdef CONFIG_X86_32 861 .legacy_models = { 862 { .family = 4, .model_names = 863 { 864 [3] = "486 DX/2", 865 [7] = "486 DX/2-WB", 866 [8] = "486 DX/4", 867 [9] = "486 DX/4-WB", 868 [14] = "Am5x86-WT", 869 [15] = "Am5x86-WB" 870 } 871 }, 872 }, 873 .legacy_cache_size = amd_size_cache, 874 #endif 875 .c_early_init = early_init_amd, 876 .c_detect_tlb = cpu_detect_tlb_amd, 877 .c_bsp_init = bsp_init_amd, 878 .c_init = init_amd, 879 .c_x86_vendor = X86_VENDOR_AMD, 880 }; 881 882 cpu_dev_register(amd_cpu_dev); 883 884 /* 885 * AMD errata checking 886 * 887 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or 888 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that 889 * have an OSVW id assigned, which it takes as first argument. Both take a 890 * variable number of family-specific model-stepping ranges created by 891 * AMD_MODEL_RANGE(). 892 * 893 * Example: 894 * 895 * const int amd_erratum_319[] = 896 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2), 897 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0), 898 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0)); 899 */ 900 901 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 } 902 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 } 903 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \ 904 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end)) 905 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff) 906 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff) 907 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff) 908 909 static const int amd_erratum_400[] = 910 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf), 911 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf)); 912 913 static const int amd_erratum_383[] = 914 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf)); 915 916 917 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum) 918 { 919 int osvw_id = *erratum++; 920 u32 range; 921 u32 ms; 922 923 if (osvw_id >= 0 && osvw_id < 65536 && 924 cpu_has(cpu, X86_FEATURE_OSVW)) { 925 u64 osvw_len; 926 927 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len); 928 if (osvw_id < osvw_len) { 929 u64 osvw_bits; 930 931 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6), 932 osvw_bits); 933 return osvw_bits & (1ULL << (osvw_id & 0x3f)); 934 } 935 } 936 937 /* OSVW unavailable or ID unknown, match family-model-stepping range */ 938 ms = (cpu->x86_model << 4) | cpu->x86_mask; 939 while ((range = *erratum++)) 940 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) && 941 (ms >= AMD_MODEL_RANGE_START(range)) && 942 (ms <= AMD_MODEL_RANGE_END(range))) 943 return true; 944 945 return false; 946 } 947 948 void set_dr_addr_mask(unsigned long mask, int dr) 949 { 950 if (!boot_cpu_has(X86_FEATURE_BPEXT)) 951 return; 952 953 switch (dr) { 954 case 0: 955 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0); 956 break; 957 case 1: 958 case 2: 959 case 3: 960 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0); 961 break; 962 default: 963 break; 964 } 965 } 966