1 #include <linux/export.h>
2 #include <linux/bitops.h>
7 #include <linux/sched.h>
8 #include <linux/sched/clock.h>
9 #include <linux/random.h>
10 #include <asm/processor.h>
12 #include <asm/cacheinfo.h>
14 #include <asm/spec-ctrl.h>
16 #include <asm/pci-direct.h>
17 #include <asm/delay.h>
20 # include <asm/mmconfig.h>
21 # include <asm/set_memory.h>
26 static const int amd_erratum_383[];
27 static const int amd_erratum_400[];
28 static const int amd_erratum_1054[];
29 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
32 * nodes_per_socket: Stores the number of nodes per socket.
33 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
34 * Node Identifiers[10:8]
36 static u32 nodes_per_socket = 1;
38 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
43 WARN_ONCE((boot_cpu_data.x86 != 0xf),
44 "%s should only be used on K8!\n", __func__);
49 err = rdmsr_safe_regs(gprs);
51 *p = gprs[0] | ((u64)gprs[2] << 32);
56 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
60 WARN_ONCE((boot_cpu_data.x86 != 0xf),
61 "%s should only be used on K8!\n", __func__);
68 return wrmsr_safe_regs(gprs);
72 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
73 * misexecution of code under Linux. Owners of such processors should
74 * contact AMD for precise details and a CPU swap.
76 * See http://www.multimania.com/poulot/k6bug.html
77 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
78 * (Publication # 21266 Issue Date: August 1998)
80 * The following test is erm.. interesting. AMD neglected to up
81 * the chip setting when fixing the bug but they also tweaked some
82 * performance at the same time..
85 extern __visible void vide(void);
86 __asm__(".globl vide\n"
87 ".type vide, @function\n"
91 static void init_amd_k5(struct cpuinfo_x86 *c)
95 * General Systems BIOSen alias the cpu frequency registers
96 * of the Elan at 0x000df000. Unfortunately, one of the Linux
97 * drivers subsequently pokes it, and changes the CPU speed.
98 * Workaround : Remove the unneeded alias.
100 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
101 #define CBAR_ENB (0x80000000)
102 #define CBAR_KEY (0X000000CB)
103 if (c->x86_model == 9 || c->x86_model == 10) {
104 if (inl(CBAR) & CBAR_ENB)
105 outl(0 | CBAR_KEY, CBAR);
110 static void init_amd_k6(struct cpuinfo_x86 *c)
114 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
116 if (c->x86_model < 6) {
117 /* Based on AMD doc 20734R - June 2000 */
118 if (c->x86_model == 0) {
119 clear_cpu_cap(c, X86_FEATURE_APIC);
120 set_cpu_cap(c, X86_FEATURE_PGE);
125 if (c->x86_model == 6 && c->x86_stepping == 1) {
126 const int K6_BUG_LOOP = 1000000;
128 void (*f_vide)(void);
131 pr_info("AMD K6 stepping B detected - ");
134 * It looks like AMD fixed the 2.6.2 bug and improved indirect
135 * calls at the same time.
140 OPTIMIZER_HIDE_VAR(f_vide);
147 if (d > 20*K6_BUG_LOOP)
148 pr_cont("system stability may be impaired when more than 32 MB are used.\n");
150 pr_cont("probably OK (after B9730xxxx).\n");
153 /* K6 with old style WHCR */
154 if (c->x86_model < 8 ||
155 (c->x86_model == 8 && c->x86_stepping < 8)) {
156 /* We can only write allocate on the low 508Mb */
160 rdmsr(MSR_K6_WHCR, l, h);
161 if ((l&0x0000FFFF) == 0) {
163 l = (1<<0)|((mbytes/4)<<1);
164 local_irq_save(flags);
166 wrmsr(MSR_K6_WHCR, l, h);
167 local_irq_restore(flags);
168 pr_info("Enabling old style K6 write allocation for %d Mb\n",
174 if ((c->x86_model == 8 && c->x86_stepping > 7) ||
175 c->x86_model == 9 || c->x86_model == 13) {
176 /* The more serious chips .. */
181 rdmsr(MSR_K6_WHCR, l, h);
182 if ((l&0xFFFF0000) == 0) {
184 l = ((mbytes>>2)<<22)|(1<<16);
185 local_irq_save(flags);
187 wrmsr(MSR_K6_WHCR, l, h);
188 local_irq_restore(flags);
189 pr_info("Enabling new style K6 write allocation for %d Mb\n",
196 if (c->x86_model == 10) {
197 /* AMD Geode LX is model 10 */
198 /* placeholder for any needed mods */
204 static void init_amd_k7(struct cpuinfo_x86 *c)
210 * Bit 15 of Athlon specific MSR 15, needs to be 0
211 * to enable SSE on Palomino/Morgan/Barton CPU's.
212 * If the BIOS didn't enable it already, enable it here.
214 if (c->x86_model >= 6 && c->x86_model <= 10) {
215 if (!cpu_has(c, X86_FEATURE_XMM)) {
216 pr_info("Enabling disabled K7/SSE Support.\n");
217 msr_clear_bit(MSR_K7_HWCR, 15);
218 set_cpu_cap(c, X86_FEATURE_XMM);
223 * It's been determined by AMD that Athlons since model 8 stepping 1
224 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
225 * As per AMD technical note 27212 0.2
227 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
228 rdmsr(MSR_K7_CLK_CTL, l, h);
229 if ((l & 0xfff00000) != 0x20000000) {
230 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
231 l, ((l & 0x000fffff)|0x20000000));
232 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
236 /* calling is from identify_secondary_cpu() ? */
241 * Certain Athlons might work (for various values of 'work') in SMP
242 * but they are not certified as MP capable.
244 /* Athlon 660/661 is valid. */
245 if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
246 (c->x86_stepping == 1)))
249 /* Duron 670 is valid */
250 if ((c->x86_model == 7) && (c->x86_stepping == 0))
254 * Athlon 662, Duron 671, and Athlon >model 7 have capability
255 * bit. It's worth noting that the A5 stepping (662) of some
256 * Athlon XP's have the MP bit set.
257 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
260 if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
261 ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
263 if (cpu_has(c, X86_FEATURE_MP))
266 /* If we get here, not a certified SMP capable AMD system. */
269 * Don't taint if we are running SMP kernel on a single non-MP
272 WARN_ONCE(1, "WARNING: This combination of AMD"
273 " processors is not suitable for SMP.\n");
274 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
280 * To workaround broken NUMA config. Read the comment in
281 * srat_detect_node().
283 static int nearby_node(int apicid)
287 for (i = apicid - 1; i >= 0; i--) {
288 node = __apicid_to_node[i];
289 if (node != NUMA_NO_NODE && node_online(node))
292 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
293 node = __apicid_to_node[i];
294 if (node != NUMA_NO_NODE && node_online(node))
297 return first_node(node_online_map); /* Shouldn't happen */
302 * Fix up cpu_core_id for pre-F17h systems to be in the
303 * [0 .. cores_per_node - 1] range. Not really needed but
304 * kept so as not to break existing setups.
306 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
313 cus_per_node = c->x86_max_cores / nodes_per_socket;
314 c->cpu_core_id %= cus_per_node;
318 static void amd_get_topology_early(struct cpuinfo_x86 *c)
320 if (cpu_has(c, X86_FEATURE_TOPOEXT))
321 smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
325 * Fixup core topology information for
326 * (1) AMD multi-node processors
327 * Assumption: Number of cores in each internal node is the same.
328 * (2) AMD processors supporting compute units
330 static void amd_get_topology(struct cpuinfo_x86 *c)
333 int cpu = smp_processor_id();
335 /* get information required for multi-node processors */
336 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
338 u32 eax, ebx, ecx, edx;
340 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
342 node_id = ecx & 0xff;
345 c->cu_id = ebx & 0xff;
347 if (c->x86 >= 0x17) {
348 c->cpu_core_id = ebx & 0xff;
350 if (smp_num_siblings > 1)
351 c->x86_max_cores /= smp_num_siblings;
355 * In case leaf B is available, use it to derive
356 * topology information.
358 err = detect_extended_topology(c);
360 c->x86_coreid_bits = get_count_order(c->x86_max_cores);
362 cacheinfo_amd_init_llc_id(c, cpu, node_id);
364 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
367 rdmsrl(MSR_FAM10H_NODE_ID, value);
370 per_cpu(cpu_llc_id, cpu) = node_id;
374 if (nodes_per_socket > 1) {
375 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
376 legacy_fixup_core_id(c);
381 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
382 * Assumes number of cores is a power of two.
384 static void amd_detect_cmp(struct cpuinfo_x86 *c)
387 int cpu = smp_processor_id();
389 bits = c->x86_coreid_bits;
390 /* Low order bits define the core id (index of core in socket) */
391 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
392 /* Convert the initial APIC ID into the socket ID */
393 c->phys_proc_id = c->initial_apicid >> bits;
394 /* use socket ID also for last level cache */
395 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
398 u16 amd_get_nb_id(int cpu)
400 return per_cpu(cpu_llc_id, cpu);
402 EXPORT_SYMBOL_GPL(amd_get_nb_id);
404 u32 amd_get_nodes_per_socket(void)
406 return nodes_per_socket;
408 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
410 static void srat_detect_node(struct cpuinfo_x86 *c)
413 int cpu = smp_processor_id();
415 unsigned apicid = c->apicid;
417 node = numa_cpu_node(cpu);
418 if (node == NUMA_NO_NODE)
419 node = per_cpu(cpu_llc_id, cpu);
422 * On multi-fabric platform (e.g. Numascale NumaChip) a
423 * platform-specific handler needs to be called to fixup some
426 if (x86_cpuinit.fixup_cpu_id)
427 x86_cpuinit.fixup_cpu_id(c, node);
429 if (!node_online(node)) {
431 * Two possibilities here:
433 * - The CPU is missing memory and no node was created. In
434 * that case try picking one from a nearby CPU.
436 * - The APIC IDs differ from the HyperTransport node IDs
437 * which the K8 northbridge parsing fills in. Assume
438 * they are all increased by a constant offset, but in
439 * the same order as the HT nodeids. If that doesn't
440 * result in a usable node fall back to the path for the
443 * This workaround operates directly on the mapping between
444 * APIC ID and NUMA node, assuming certain relationship
445 * between APIC ID, HT node ID and NUMA topology. As going
446 * through CPU mapping may alter the outcome, directly
447 * access __apicid_to_node[].
449 int ht_nodeid = c->initial_apicid;
451 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
452 node = __apicid_to_node[ht_nodeid];
453 /* Pick a nearby node */
454 if (!node_online(node))
455 node = nearby_node(apicid);
457 numa_set_node(cpu, node);
461 static void early_init_amd_mc(struct cpuinfo_x86 *c)
466 /* Multi core CPU? */
467 if (c->extended_cpuid_level < 0x80000008)
470 ecx = cpuid_ecx(0x80000008);
472 c->x86_max_cores = (ecx & 0xff) + 1;
474 /* CPU telling us the core id bits shift? */
475 bits = (ecx >> 12) & 0xF;
477 /* Otherwise recompute */
479 while ((1 << bits) < c->x86_max_cores)
483 c->x86_coreid_bits = bits;
487 static void bsp_init_amd(struct cpuinfo_x86 *c)
492 unsigned long long tseg;
495 * Split up direct mapping around the TSEG SMM area.
496 * Don't do it for gbpages because there seems very little
497 * benefit in doing so.
499 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
500 unsigned long pfn = tseg >> PAGE_SHIFT;
502 pr_debug("tseg: %010llx\n", tseg);
503 if (pfn_range_is_mapped(pfn, pfn + 1))
504 set_memory_4k((unsigned long)__va(tseg), 1);
509 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
512 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
515 rdmsrl(MSR_K7_HWCR, val);
516 if (!(val & BIT(24)))
517 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
521 if (c->x86 == 0x15) {
522 unsigned long upperbit;
525 cpuid = cpuid_edx(0x80000005);
526 assoc = cpuid >> 16 & 0xff;
527 upperbit = ((cpuid >> 24) << 10) / assoc;
529 va_align.mask = (upperbit - 1) & PAGE_MASK;
530 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
532 /* A random value per boot for bit slice [12:upper_bit) */
533 va_align.bits = get_random_int() & va_align.mask;
536 if (cpu_has(c, X86_FEATURE_MWAITX))
539 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
542 ecx = cpuid_ecx(0x8000001e);
543 nodes_per_socket = ((ecx >> 8) & 7) + 1;
544 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
547 rdmsrl(MSR_FAM10H_NODE_ID, value);
548 nodes_per_socket = ((value >> 3) & 7) + 1;
551 if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
552 !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
553 c->x86 >= 0x15 && c->x86 <= 0x17) {
557 case 0x15: bit = 54; break;
558 case 0x16: bit = 33; break;
559 case 0x17: bit = 10; break;
563 * Try to cache the base value so further operations can
564 * avoid RMW. If that faults, do not enable SSBD.
566 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
567 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
568 setup_force_cpu_cap(X86_FEATURE_SSBD);
569 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
574 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
579 * BIOS support is required for SME and SEV.
580 * For SME: If BIOS has enabled SME then adjust x86_phys_bits by
581 * the SME physical address space reduction value.
582 * If BIOS has not enabled SME then don't advertise the
583 * SME feature (set in scattered.c).
584 * For SEV: If BIOS has not enabled SEV then don't advertise the
585 * SEV feature (set in scattered.c).
587 * In all cases, since support for SME and SEV requires long mode,
588 * don't advertise the feature under CONFIG_X86_32.
590 if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
591 /* Check if memory encryption is enabled */
592 rdmsrl(MSR_K8_SYSCFG, msr);
593 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
597 * Always adjust physical address bits. Even though this
598 * will be a value above 32-bits this is still done for
599 * CONFIG_X86_32 so that accurate values are reported.
601 c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
603 if (IS_ENABLED(CONFIG_X86_32))
606 rdmsrl(MSR_K7_HWCR, msr);
607 if (!(msr & MSR_K7_HWCR_SMMLOCK))
613 setup_clear_cpu_cap(X86_FEATURE_SME);
615 setup_clear_cpu_cap(X86_FEATURE_SEV);
619 static void early_init_amd(struct cpuinfo_x86 *c)
624 early_init_amd_mc(c);
628 set_cpu_cap(c, X86_FEATURE_K7);
632 set_cpu_cap(c, X86_FEATURE_K8);
634 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
637 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
638 * with P/T states and does not stop in deep C-states
640 if (c->x86_power & (1 << 8)) {
641 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
642 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
645 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
646 if (c->x86_power & BIT(12))
647 set_cpu_cap(c, X86_FEATURE_ACC_POWER);
650 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
652 /* Set MTRR capability flag if appropriate */
654 if (c->x86_model == 13 || c->x86_model == 9 ||
655 (c->x86_model == 8 && c->x86_stepping >= 8))
656 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
658 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
660 * ApicID can always be treated as an 8-bit value for AMD APIC versions
661 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
662 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
665 if (boot_cpu_has(X86_FEATURE_APIC)) {
667 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
668 else if (c->x86 >= 0xf) {
669 /* check CPU config space for extended APIC ID */
672 val = read_pci_config(0, 24, 0, 0x68);
673 if ((val >> 17 & 0x3) == 0x3)
674 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
680 * This is only needed to tell the kernel whether to use VMCALL
681 * and VMMCALL. VMMCALL is never executed except under virt, so
682 * we can set it unconditionally.
684 set_cpu_cap(c, X86_FEATURE_VMMCALL);
686 /* F16h erratum 793, CVE-2013-6885 */
687 if (c->x86 == 0x16 && c->x86_model <= 0xf)
688 msr_set_bit(MSR_AMD64_LS_CFG, 15);
691 * Check whether the machine is affected by erratum 400. This is
692 * used to select the proper idle routine and to enable the check
693 * whether the machine is affected in arch_post_acpi_init(), which
694 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
696 if (cpu_has_amd_erratum(c, amd_erratum_400))
697 set_cpu_bug(c, X86_BUG_AMD_E400);
699 early_detect_mem_encrypt(c);
701 /* Re-enable TopologyExtensions if switched off by BIOS */
702 if (c->x86 == 0x15 &&
703 (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
704 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
706 if (msr_set_bit(0xc0011005, 54) > 0) {
707 rdmsrl(0xc0011005, value);
708 if (value & BIT_64(54)) {
709 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
710 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
715 amd_get_topology_early(c);
718 static void init_amd_k8(struct cpuinfo_x86 *c)
723 /* On C+ stepping K8 rep microcode works well for copy/memset */
724 level = cpuid_eax(1);
725 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
726 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
729 * Some BIOSes incorrectly force this feature, but only K8 revision D
730 * (model = 0x14) and later actually support it.
731 * (AMD Erratum #110, docId: 25759).
733 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
734 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
735 if (!rdmsrl_amd_safe(0xc001100d, &value)) {
736 value &= ~BIT_64(32);
737 wrmsrl_amd_safe(0xc001100d, value);
741 if (!c->x86_model_id[0])
742 strcpy(c->x86_model_id, "Hammer");
746 * Disable TLB flush filter by setting HWCR.FFDIS on K8
747 * bit 6 of msr C001_0015
749 * Errata 63 for SH-B3 steppings
750 * Errata 122 for all steppings (F+ have it disabled by default)
752 msr_set_bit(MSR_K7_HWCR, 6);
754 set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
757 static void init_amd_gh(struct cpuinfo_x86 *c)
759 #ifdef CONFIG_MMCONF_FAM10H
760 /* do this for boot cpu */
761 if (c == &boot_cpu_data)
762 check_enable_amd_mmconf_dmi();
764 fam10h_check_enable_mmcfg();
768 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
769 * is always needed when GART is enabled, even in a kernel which has no
770 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
771 * If it doesn't, we do it here as suggested by the BKDG.
773 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
775 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
778 * On family 10h BIOS may not have properly enabled WC+ support, causing
779 * it to be converted to CD memtype. This may result in performance
780 * degradation for certain nested-paging guests. Prevent this conversion
781 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
783 * NOTE: we want to use the _safe accessors so as not to #GP kvm
784 * guests on older kvm hosts.
786 msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
788 if (cpu_has_amd_erratum(c, amd_erratum_383))
789 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
792 #define MSR_AMD64_DE_CFG 0xC0011029
794 static void init_amd_ln(struct cpuinfo_x86 *c)
797 * Apply erratum 665 fix unconditionally so machines without a BIOS
800 msr_set_bit(MSR_AMD64_DE_CFG, 31);
803 static bool rdrand_force;
805 static int __init rdrand_cmdline(char *str)
810 if (!strcmp(str, "force"))
817 early_param("rdrand", rdrand_cmdline);
819 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
822 * Saving of the MSR used to hide the RDRAND support during
823 * suspend/resume is done by arch/x86/power/cpu.c, which is
824 * dependent on CONFIG_PM_SLEEP.
826 if (!IS_ENABLED(CONFIG_PM_SLEEP))
830 * The nordrand option can clear X86_FEATURE_RDRAND, so check for
831 * RDRAND support using the CPUID function directly.
833 if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
836 msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
839 * Verify that the CPUID change has occurred in case the kernel is
840 * running virtualized and the hypervisor doesn't support the MSR.
842 if (cpuid_ecx(1) & BIT(30)) {
843 pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
847 clear_cpu_cap(c, X86_FEATURE_RDRAND);
848 pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
851 static void init_amd_jg(struct cpuinfo_x86 *c)
854 * Some BIOS implementations do not restore proper RDRAND support
855 * across suspend and resume. Check on whether to hide the RDRAND
856 * instruction support via CPUID.
858 clear_rdrand_cpuid_bit(c);
861 static void init_amd_bd(struct cpuinfo_x86 *c)
866 * The way access filter has a performance penalty on some workloads.
867 * Disable it on the affected CPUs.
869 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
870 if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
872 wrmsrl_safe(MSR_F15H_IC_CFG, value);
877 * Some BIOS implementations do not restore proper RDRAND support
878 * across suspend and resume. Check on whether to hide the RDRAND
879 * instruction support via CPUID.
881 clear_rdrand_cpuid_bit(c);
884 static void init_amd_zn(struct cpuinfo_x86 *c)
886 set_cpu_cap(c, X86_FEATURE_ZEN);
889 * Fix erratum 1076: CPB feature bit not being set in CPUID.
890 * Always set it, except when running under a hypervisor.
892 if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_CPB))
893 set_cpu_cap(c, X86_FEATURE_CPB);
896 static void init_amd(struct cpuinfo_x86 *c)
901 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
902 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
904 clear_cpu_cap(c, 0*32+31);
907 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
909 /* get apicid instead of initial apic id from cpuid */
910 c->apicid = hard_smp_processor_id();
912 /* K6s reports MCEs but don't actually have all the MSRs */
914 clear_cpu_cap(c, X86_FEATURE_MCE);
917 case 4: init_amd_k5(c); break;
918 case 5: init_amd_k6(c); break;
919 case 6: init_amd_k7(c); break;
920 case 0xf: init_amd_k8(c); break;
921 case 0x10: init_amd_gh(c); break;
922 case 0x12: init_amd_ln(c); break;
923 case 0x15: init_amd_bd(c); break;
924 case 0x16: init_amd_jg(c); break;
925 case 0x17: init_amd_zn(c); break;
929 * Enable workaround for FXSAVE leak on CPUs
930 * without a XSaveErPtr feature
932 if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
933 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
935 cpu_detect_cache_sizes(c);
941 init_amd_cacheinfo(c);
943 if (cpu_has(c, X86_FEATURE_XMM2)) {
944 unsigned long long val;
948 * A serializing LFENCE has less overhead than MFENCE, so
949 * use it for execution serialization. On families which
950 * don't have that MSR, LFENCE is already serializing.
951 * msr_set_bit() uses the safe accessors, too, even if the MSR
954 msr_set_bit(MSR_F10H_DECFG,
955 MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT);
958 * Verify that the MSR write was successful (could be running
959 * under a hypervisor) and only then assume that LFENCE is
962 ret = rdmsrl_safe(MSR_F10H_DECFG, &val);
963 if (!ret && (val & MSR_F10H_DECFG_LFENCE_SERIALIZE)) {
964 /* A serializing LFENCE stops RDTSC speculation */
965 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
967 /* MFENCE stops RDTSC speculation */
968 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
973 * Family 0x12 and above processors have APIC timer
974 * running in deep C states.
977 set_cpu_cap(c, X86_FEATURE_ARAT);
979 /* 3DNow or LM implies PREFETCHW */
980 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
981 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
982 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
984 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
985 if (!cpu_has(c, X86_FEATURE_XENPV))
986 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
989 * Turn on the Instructions Retired free counter on machines not
990 * susceptible to erratum #1054 "Instructions Retired Performance
991 * Counter May Be Inaccurate".
993 if (cpu_has(c, X86_FEATURE_IRPERF) &&
994 !cpu_has_amd_erratum(c, amd_erratum_1054))
995 msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
997 check_null_seg_clears_base(c);
1000 #ifdef CONFIG_X86_32
1001 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1003 /* AMD errata T13 (order #21922) */
1006 if (c->x86_model == 3 && c->x86_stepping == 0)
1008 /* Tbird rev A1/A2 */
1009 if (c->x86_model == 4 &&
1010 (c->x86_stepping == 0 || c->x86_stepping == 1))
1017 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1019 u32 ebx, eax, ecx, edx;
1025 if (c->extended_cpuid_level < 0x80000006)
1028 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1030 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1031 tlb_lli_4k[ENTRIES] = ebx & mask;
1034 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1035 * characteristics from the CPUID function 0x80000005 instead.
1037 if (c->x86 == 0xf) {
1038 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1042 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1043 if (!((eax >> 16) & mask))
1044 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1046 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1048 /* a 4M entry uses two 2M entries */
1049 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1051 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1052 if (!(eax & mask)) {
1054 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1055 tlb_lli_2m[ENTRIES] = 1024;
1057 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1058 tlb_lli_2m[ENTRIES] = eax & 0xff;
1061 tlb_lli_2m[ENTRIES] = eax & mask;
1063 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1066 static const struct cpu_dev amd_cpu_dev = {
1068 .c_ident = { "AuthenticAMD" },
1069 #ifdef CONFIG_X86_32
1071 { .family = 4, .model_names =
1074 [7] = "486 DX/2-WB",
1076 [9] = "486 DX/4-WB",
1082 .legacy_cache_size = amd_size_cache,
1084 .c_early_init = early_init_amd,
1085 .c_detect_tlb = cpu_detect_tlb_amd,
1086 .c_bsp_init = bsp_init_amd,
1088 .c_x86_vendor = X86_VENDOR_AMD,
1091 cpu_dev_register(amd_cpu_dev);
1094 * AMD errata checking
1096 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1097 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1098 * have an OSVW id assigned, which it takes as first argument. Both take a
1099 * variable number of family-specific model-stepping ranges created by
1100 * AMD_MODEL_RANGE().
1104 * const int amd_erratum_319[] =
1105 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1106 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1107 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1110 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
1111 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
1112 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1113 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1114 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
1115 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
1116 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
1118 static const int amd_erratum_400[] =
1119 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1120 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1122 static const int amd_erratum_383[] =
1123 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1125 /* #1054: Instructions Retired Performance Counter May Be Inaccurate */
1126 static const int amd_erratum_1054[] =
1127 AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0, 0, 0x2f, 0xf));
1129 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1131 int osvw_id = *erratum++;
1135 if (osvw_id >= 0 && osvw_id < 65536 &&
1136 cpu_has(cpu, X86_FEATURE_OSVW)) {
1139 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1140 if (osvw_id < osvw_len) {
1143 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1145 return osvw_bits & (1ULL << (osvw_id & 0x3f));
1149 /* OSVW unavailable or ID unknown, match family-model-stepping range */
1150 ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1151 while ((range = *erratum++))
1152 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1153 (ms >= AMD_MODEL_RANGE_START(range)) &&
1154 (ms <= AMD_MODEL_RANGE_END(range)))
1160 void set_dr_addr_mask(unsigned long mask, int dr)
1162 if (!boot_cpu_has(X86_FEATURE_BPEXT))
1167 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1172 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);