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 */
203 * Work around Erratum 1386. The XSAVES instruction malfunctions in
204 * certain circumstances on Zen1/2 uarch, and not all parts have had
205 * updated microcode at the time of writing (March 2023).
207 * Affected parts all have no supervisor XSAVE states, meaning that
208 * the XSAVEC instruction (which works fine) is equivalent.
210 clear_cpu_cap(c, X86_FEATURE_XSAVES);
213 static void init_amd_k7(struct cpuinfo_x86 *c)
219 * Bit 15 of Athlon specific MSR 15, needs to be 0
220 * to enable SSE on Palomino/Morgan/Barton CPU's.
221 * If the BIOS didn't enable it already, enable it here.
223 if (c->x86_model >= 6 && c->x86_model <= 10) {
224 if (!cpu_has(c, X86_FEATURE_XMM)) {
225 pr_info("Enabling disabled K7/SSE Support.\n");
226 msr_clear_bit(MSR_K7_HWCR, 15);
227 set_cpu_cap(c, X86_FEATURE_XMM);
232 * It's been determined by AMD that Athlons since model 8 stepping 1
233 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
234 * As per AMD technical note 27212 0.2
236 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
237 rdmsr(MSR_K7_CLK_CTL, l, h);
238 if ((l & 0xfff00000) != 0x20000000) {
239 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
240 l, ((l & 0x000fffff)|0x20000000));
241 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
245 /* calling is from identify_secondary_cpu() ? */
250 * Certain Athlons might work (for various values of 'work') in SMP
251 * but they are not certified as MP capable.
253 /* Athlon 660/661 is valid. */
254 if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
255 (c->x86_stepping == 1)))
258 /* Duron 670 is valid */
259 if ((c->x86_model == 7) && (c->x86_stepping == 0))
263 * Athlon 662, Duron 671, and Athlon >model 7 have capability
264 * bit. It's worth noting that the A5 stepping (662) of some
265 * Athlon XP's have the MP bit set.
266 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
269 if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
270 ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
272 if (cpu_has(c, X86_FEATURE_MP))
275 /* If we get here, not a certified SMP capable AMD system. */
278 * Don't taint if we are running SMP kernel on a single non-MP
281 WARN_ONCE(1, "WARNING: This combination of AMD"
282 " processors is not suitable for SMP.\n");
283 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
289 * To workaround broken NUMA config. Read the comment in
290 * srat_detect_node().
292 static int nearby_node(int apicid)
296 for (i = apicid - 1; i >= 0; i--) {
297 node = __apicid_to_node[i];
298 if (node != NUMA_NO_NODE && node_online(node))
301 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
302 node = __apicid_to_node[i];
303 if (node != NUMA_NO_NODE && node_online(node))
306 return first_node(node_online_map); /* Shouldn't happen */
311 * Fix up cpu_core_id for pre-F17h systems to be in the
312 * [0 .. cores_per_node - 1] range. Not really needed but
313 * kept so as not to break existing setups.
315 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
322 cus_per_node = c->x86_max_cores / nodes_per_socket;
323 c->cpu_core_id %= cus_per_node;
327 static void amd_get_topology_early(struct cpuinfo_x86 *c)
329 if (cpu_has(c, X86_FEATURE_TOPOEXT))
330 smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
334 * Fixup core topology information for
335 * (1) AMD multi-node processors
336 * Assumption: Number of cores in each internal node is the same.
337 * (2) AMD processors supporting compute units
339 static void amd_get_topology(struct cpuinfo_x86 *c)
342 int cpu = smp_processor_id();
344 /* get information required for multi-node processors */
345 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
347 u32 eax, ebx, ecx, edx;
349 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
351 node_id = ecx & 0xff;
354 c->cu_id = ebx & 0xff;
356 if (c->x86 >= 0x17) {
357 c->cpu_core_id = ebx & 0xff;
359 if (smp_num_siblings > 1)
360 c->x86_max_cores /= smp_num_siblings;
364 * In case leaf B is available, use it to derive
365 * topology information.
367 err = detect_extended_topology(c);
369 c->x86_coreid_bits = get_count_order(c->x86_max_cores);
371 cacheinfo_amd_init_llc_id(c, cpu, node_id);
373 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
376 rdmsrl(MSR_FAM10H_NODE_ID, value);
379 per_cpu(cpu_llc_id, cpu) = node_id;
383 if (nodes_per_socket > 1) {
384 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
385 legacy_fixup_core_id(c);
390 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
391 * Assumes number of cores is a power of two.
393 static void amd_detect_cmp(struct cpuinfo_x86 *c)
396 int cpu = smp_processor_id();
398 bits = c->x86_coreid_bits;
399 /* Low order bits define the core id (index of core in socket) */
400 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
401 /* Convert the initial APIC ID into the socket ID */
402 c->phys_proc_id = c->initial_apicid >> bits;
403 /* use socket ID also for last level cache */
404 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
407 u16 amd_get_nb_id(int cpu)
409 return per_cpu(cpu_llc_id, cpu);
411 EXPORT_SYMBOL_GPL(amd_get_nb_id);
413 u32 amd_get_nodes_per_socket(void)
415 return nodes_per_socket;
417 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
419 static void srat_detect_node(struct cpuinfo_x86 *c)
422 int cpu = smp_processor_id();
424 unsigned apicid = c->apicid;
426 node = numa_cpu_node(cpu);
427 if (node == NUMA_NO_NODE)
428 node = per_cpu(cpu_llc_id, cpu);
431 * On multi-fabric platform (e.g. Numascale NumaChip) a
432 * platform-specific handler needs to be called to fixup some
435 if (x86_cpuinit.fixup_cpu_id)
436 x86_cpuinit.fixup_cpu_id(c, node);
438 if (!node_online(node)) {
440 * Two possibilities here:
442 * - The CPU is missing memory and no node was created. In
443 * that case try picking one from a nearby CPU.
445 * - The APIC IDs differ from the HyperTransport node IDs
446 * which the K8 northbridge parsing fills in. Assume
447 * they are all increased by a constant offset, but in
448 * the same order as the HT nodeids. If that doesn't
449 * result in a usable node fall back to the path for the
452 * This workaround operates directly on the mapping between
453 * APIC ID and NUMA node, assuming certain relationship
454 * between APIC ID, HT node ID and NUMA topology. As going
455 * through CPU mapping may alter the outcome, directly
456 * access __apicid_to_node[].
458 int ht_nodeid = c->initial_apicid;
460 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
461 node = __apicid_to_node[ht_nodeid];
462 /* Pick a nearby node */
463 if (!node_online(node))
464 node = nearby_node(apicid);
466 numa_set_node(cpu, node);
470 static void early_init_amd_mc(struct cpuinfo_x86 *c)
475 /* Multi core CPU? */
476 if (c->extended_cpuid_level < 0x80000008)
479 ecx = cpuid_ecx(0x80000008);
481 c->x86_max_cores = (ecx & 0xff) + 1;
483 /* CPU telling us the core id bits shift? */
484 bits = (ecx >> 12) & 0xF;
486 /* Otherwise recompute */
488 while ((1 << bits) < c->x86_max_cores)
492 c->x86_coreid_bits = bits;
496 static void bsp_init_amd(struct cpuinfo_x86 *c)
501 unsigned long long tseg;
504 * Split up direct mapping around the TSEG SMM area.
505 * Don't do it for gbpages because there seems very little
506 * benefit in doing so.
508 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
509 unsigned long pfn = tseg >> PAGE_SHIFT;
511 pr_debug("tseg: %010llx\n", tseg);
512 if (pfn_range_is_mapped(pfn, pfn + 1))
513 set_memory_4k((unsigned long)__va(tseg), 1);
518 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
521 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
524 rdmsrl(MSR_K7_HWCR, val);
525 if (!(val & BIT(24)))
526 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
530 if (c->x86 == 0x15) {
531 unsigned long upperbit;
534 cpuid = cpuid_edx(0x80000005);
535 assoc = cpuid >> 16 & 0xff;
536 upperbit = ((cpuid >> 24) << 10) / assoc;
538 va_align.mask = (upperbit - 1) & PAGE_MASK;
539 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
541 /* A random value per boot for bit slice [12:upper_bit) */
542 va_align.bits = get_random_int() & va_align.mask;
545 if (cpu_has(c, X86_FEATURE_MWAITX))
548 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
551 ecx = cpuid_ecx(0x8000001e);
552 nodes_per_socket = ((ecx >> 8) & 7) + 1;
553 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
556 rdmsrl(MSR_FAM10H_NODE_ID, value);
557 nodes_per_socket = ((value >> 3) & 7) + 1;
560 if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
561 !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
562 c->x86 >= 0x15 && c->x86 <= 0x17) {
566 case 0x15: bit = 54; break;
567 case 0x16: bit = 33; break;
568 case 0x17: bit = 10; break;
572 * Try to cache the base value so further operations can
573 * avoid RMW. If that faults, do not enable SSBD.
575 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
576 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
577 setup_force_cpu_cap(X86_FEATURE_SSBD);
578 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
583 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
588 * BIOS support is required for SME and SEV.
589 * For SME: If BIOS has enabled SME then adjust x86_phys_bits by
590 * the SME physical address space reduction value.
591 * If BIOS has not enabled SME then don't advertise the
592 * SME feature (set in scattered.c).
593 * For SEV: If BIOS has not enabled SEV then don't advertise the
594 * SEV feature (set in scattered.c).
596 * In all cases, since support for SME and SEV requires long mode,
597 * don't advertise the feature under CONFIG_X86_32.
599 if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
600 /* Check if memory encryption is enabled */
601 rdmsrl(MSR_K8_SYSCFG, msr);
602 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
606 * Always adjust physical address bits. Even though this
607 * will be a value above 32-bits this is still done for
608 * CONFIG_X86_32 so that accurate values are reported.
610 c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
612 if (IS_ENABLED(CONFIG_X86_32))
615 rdmsrl(MSR_K7_HWCR, msr);
616 if (!(msr & MSR_K7_HWCR_SMMLOCK))
622 setup_clear_cpu_cap(X86_FEATURE_SME);
624 setup_clear_cpu_cap(X86_FEATURE_SEV);
628 static void early_init_amd(struct cpuinfo_x86 *c)
633 early_init_amd_mc(c);
637 set_cpu_cap(c, X86_FEATURE_K7);
641 set_cpu_cap(c, X86_FEATURE_K8);
643 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
646 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
647 * with P/T states and does not stop in deep C-states
649 if (c->x86_power & (1 << 8)) {
650 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
651 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
654 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
655 if (c->x86_power & BIT(12))
656 set_cpu_cap(c, X86_FEATURE_ACC_POWER);
659 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
661 /* Set MTRR capability flag if appropriate */
663 if (c->x86_model == 13 || c->x86_model == 9 ||
664 (c->x86_model == 8 && c->x86_stepping >= 8))
665 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
667 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
669 * ApicID can always be treated as an 8-bit value for AMD APIC versions
670 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
671 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
674 if (boot_cpu_has(X86_FEATURE_APIC)) {
676 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
677 else if (c->x86 >= 0xf) {
678 /* check CPU config space for extended APIC ID */
681 val = read_pci_config(0, 24, 0, 0x68);
682 if ((val >> 17 & 0x3) == 0x3)
683 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
689 * This is only needed to tell the kernel whether to use VMCALL
690 * and VMMCALL. VMMCALL is never executed except under virt, so
691 * we can set it unconditionally.
693 set_cpu_cap(c, X86_FEATURE_VMMCALL);
695 /* F16h erratum 793, CVE-2013-6885 */
696 if (c->x86 == 0x16 && c->x86_model <= 0xf)
697 msr_set_bit(MSR_AMD64_LS_CFG, 15);
700 * Check whether the machine is affected by erratum 400. This is
701 * used to select the proper idle routine and to enable the check
702 * whether the machine is affected in arch_post_acpi_init(), which
703 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
705 if (cpu_has_amd_erratum(c, amd_erratum_400))
706 set_cpu_bug(c, X86_BUG_AMD_E400);
708 early_detect_mem_encrypt(c);
710 /* Re-enable TopologyExtensions if switched off by BIOS */
711 if (c->x86 == 0x15 &&
712 (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
713 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
715 if (msr_set_bit(0xc0011005, 54) > 0) {
716 rdmsrl(0xc0011005, value);
717 if (value & BIT_64(54)) {
718 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
719 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
724 amd_get_topology_early(c);
727 static void init_amd_k8(struct cpuinfo_x86 *c)
732 /* On C+ stepping K8 rep microcode works well for copy/memset */
733 level = cpuid_eax(1);
734 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
735 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
738 * Some BIOSes incorrectly force this feature, but only K8 revision D
739 * (model = 0x14) and later actually support it.
740 * (AMD Erratum #110, docId: 25759).
742 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
743 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
744 if (!rdmsrl_amd_safe(0xc001100d, &value)) {
745 value &= ~BIT_64(32);
746 wrmsrl_amd_safe(0xc001100d, value);
750 if (!c->x86_model_id[0])
751 strcpy(c->x86_model_id, "Hammer");
755 * Disable TLB flush filter by setting HWCR.FFDIS on K8
756 * bit 6 of msr C001_0015
758 * Errata 63 for SH-B3 steppings
759 * Errata 122 for all steppings (F+ have it disabled by default)
761 msr_set_bit(MSR_K7_HWCR, 6);
763 set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
766 static void init_amd_gh(struct cpuinfo_x86 *c)
768 #ifdef CONFIG_MMCONF_FAM10H
769 /* do this for boot cpu */
770 if (c == &boot_cpu_data)
771 check_enable_amd_mmconf_dmi();
773 fam10h_check_enable_mmcfg();
777 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
778 * is always needed when GART is enabled, even in a kernel which has no
779 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
780 * If it doesn't, we do it here as suggested by the BKDG.
782 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
784 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
787 * On family 10h BIOS may not have properly enabled WC+ support, causing
788 * it to be converted to CD memtype. This may result in performance
789 * degradation for certain nested-paging guests. Prevent this conversion
790 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
792 * NOTE: we want to use the _safe accessors so as not to #GP kvm
793 * guests on older kvm hosts.
795 msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
797 if (cpu_has_amd_erratum(c, amd_erratum_383))
798 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
801 static void init_amd_ln(struct cpuinfo_x86 *c)
804 * Apply erratum 665 fix unconditionally so machines without a BIOS
807 msr_set_bit(MSR_AMD64_DE_CFG, 31);
810 static bool rdrand_force;
812 static int __init rdrand_cmdline(char *str)
817 if (!strcmp(str, "force"))
824 early_param("rdrand", rdrand_cmdline);
826 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
829 * Saving of the MSR used to hide the RDRAND support during
830 * suspend/resume is done by arch/x86/power/cpu.c, which is
831 * dependent on CONFIG_PM_SLEEP.
833 if (!IS_ENABLED(CONFIG_PM_SLEEP))
837 * The nordrand option can clear X86_FEATURE_RDRAND, so check for
838 * RDRAND support using the CPUID function directly.
840 if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
843 msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
846 * Verify that the CPUID change has occurred in case the kernel is
847 * running virtualized and the hypervisor doesn't support the MSR.
849 if (cpuid_ecx(1) & BIT(30)) {
850 pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
854 clear_cpu_cap(c, X86_FEATURE_RDRAND);
855 pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
858 static void init_amd_jg(struct cpuinfo_x86 *c)
861 * Some BIOS implementations do not restore proper RDRAND support
862 * across suspend and resume. Check on whether to hide the RDRAND
863 * instruction support via CPUID.
865 clear_rdrand_cpuid_bit(c);
868 static void init_amd_bd(struct cpuinfo_x86 *c)
873 * The way access filter has a performance penalty on some workloads.
874 * Disable it on the affected CPUs.
876 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
877 if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
879 wrmsrl_safe(MSR_F15H_IC_CFG, value);
884 * Some BIOS implementations do not restore proper RDRAND support
885 * across suspend and resume. Check on whether to hide the RDRAND
886 * instruction support via CPUID.
888 clear_rdrand_cpuid_bit(c);
891 static void init_amd_zn(struct cpuinfo_x86 *c)
893 set_cpu_cap(c, X86_FEATURE_ZEN);
895 /* Fix up CPUID bits, but only if not virtualised. */
896 if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
898 /* Erratum 1076: CPB feature bit not being set in CPUID. */
899 if (!cpu_has(c, X86_FEATURE_CPB))
900 set_cpu_cap(c, X86_FEATURE_CPB);
903 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
904 * Branch Type Confusion, but predate the allocation of the
907 if (c->x86 == 0x19 && !cpu_has(c, X86_FEATURE_BTC_NO))
908 set_cpu_cap(c, X86_FEATURE_BTC_NO);
912 static void init_amd(struct cpuinfo_x86 *c)
917 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
918 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
920 clear_cpu_cap(c, 0*32+31);
923 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
925 /* get apicid instead of initial apic id from cpuid */
926 c->apicid = hard_smp_processor_id();
928 /* K6s reports MCEs but don't actually have all the MSRs */
930 clear_cpu_cap(c, X86_FEATURE_MCE);
933 case 4: init_amd_k5(c); break;
934 case 5: init_amd_k6(c); break;
935 case 6: init_amd_k7(c); break;
936 case 0xf: init_amd_k8(c); break;
937 case 0x10: init_amd_gh(c); break;
938 case 0x12: init_amd_ln(c); break;
939 case 0x15: init_amd_bd(c); break;
940 case 0x16: init_amd_jg(c); break;
941 case 0x17: init_amd_zn(c); break;
945 * Enable workaround for FXSAVE leak on CPUs
946 * without a XSaveErPtr feature
948 if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
949 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
951 cpu_detect_cache_sizes(c);
957 init_amd_cacheinfo(c);
959 if (cpu_has(c, X86_FEATURE_XMM2)) {
960 unsigned long long val;
964 * A serializing LFENCE has less overhead than MFENCE, so
965 * use it for execution serialization. On families which
966 * don't have that MSR, LFENCE is already serializing.
967 * msr_set_bit() uses the safe accessors, too, even if the MSR
970 msr_set_bit(MSR_AMD64_DE_CFG,
971 MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
974 * Verify that the MSR write was successful (could be running
975 * under a hypervisor) and only then assume that LFENCE is
978 ret = rdmsrl_safe(MSR_AMD64_DE_CFG, &val);
979 if (!ret && (val & MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT)) {
980 /* A serializing LFENCE stops RDTSC speculation */
981 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
983 /* MFENCE stops RDTSC speculation */
984 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
989 * Family 0x12 and above processors have APIC timer
990 * running in deep C states.
993 set_cpu_cap(c, X86_FEATURE_ARAT);
995 /* 3DNow or LM implies PREFETCHW */
996 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
997 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
998 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1000 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1001 if (!cpu_has(c, X86_FEATURE_XENPV))
1002 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1005 * Turn on the Instructions Retired free counter on machines not
1006 * susceptible to erratum #1054 "Instructions Retired Performance
1007 * Counter May Be Inaccurate".
1009 if (cpu_has(c, X86_FEATURE_IRPERF) &&
1010 !cpu_has_amd_erratum(c, amd_erratum_1054))
1011 msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1013 check_null_seg_clears_base(c);
1016 #ifdef CONFIG_X86_32
1017 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1019 /* AMD errata T13 (order #21922) */
1022 if (c->x86_model == 3 && c->x86_stepping == 0)
1024 /* Tbird rev A1/A2 */
1025 if (c->x86_model == 4 &&
1026 (c->x86_stepping == 0 || c->x86_stepping == 1))
1033 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1035 u32 ebx, eax, ecx, edx;
1041 if (c->extended_cpuid_level < 0x80000006)
1044 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1046 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1047 tlb_lli_4k[ENTRIES] = ebx & mask;
1050 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1051 * characteristics from the CPUID function 0x80000005 instead.
1053 if (c->x86 == 0xf) {
1054 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1058 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1059 if (!((eax >> 16) & mask))
1060 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1062 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1064 /* a 4M entry uses two 2M entries */
1065 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1067 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1068 if (!(eax & mask)) {
1070 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1071 tlb_lli_2m[ENTRIES] = 1024;
1073 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1074 tlb_lli_2m[ENTRIES] = eax & 0xff;
1077 tlb_lli_2m[ENTRIES] = eax & mask;
1079 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1082 static const struct cpu_dev amd_cpu_dev = {
1084 .c_ident = { "AuthenticAMD" },
1085 #ifdef CONFIG_X86_32
1087 { .family = 4, .model_names =
1090 [7] = "486 DX/2-WB",
1092 [9] = "486 DX/4-WB",
1098 .legacy_cache_size = amd_size_cache,
1100 .c_early_init = early_init_amd,
1101 .c_detect_tlb = cpu_detect_tlb_amd,
1102 .c_bsp_init = bsp_init_amd,
1104 .c_x86_vendor = X86_VENDOR_AMD,
1107 cpu_dev_register(amd_cpu_dev);
1110 * AMD errata checking
1112 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1113 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1114 * have an OSVW id assigned, which it takes as first argument. Both take a
1115 * variable number of family-specific model-stepping ranges created by
1116 * AMD_MODEL_RANGE().
1120 * const int amd_erratum_319[] =
1121 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1122 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1123 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1126 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
1127 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
1128 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1129 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1130 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
1131 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
1132 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
1134 static const int amd_erratum_400[] =
1135 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1136 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1138 static const int amd_erratum_383[] =
1139 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1141 /* #1054: Instructions Retired Performance Counter May Be Inaccurate */
1142 static const int amd_erratum_1054[] =
1143 AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0, 0, 0x2f, 0xf));
1145 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1147 int osvw_id = *erratum++;
1151 if (osvw_id >= 0 && osvw_id < 65536 &&
1152 cpu_has(cpu, X86_FEATURE_OSVW)) {
1155 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1156 if (osvw_id < osvw_len) {
1159 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1161 return osvw_bits & (1ULL << (osvw_id & 0x3f));
1165 /* OSVW unavailable or ID unknown, match family-model-stepping range */
1166 ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1167 while ((range = *erratum++))
1168 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1169 (ms >= AMD_MODEL_RANGE_START(range)) &&
1170 (ms <= AMD_MODEL_RANGE_END(range)))
1176 void set_dr_addr_mask(unsigned long mask, int dr)
1178 if (!boot_cpu_has(X86_FEATURE_BPEXT))
1183 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1188 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);