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>
13 #include <asm/spec-ctrl.h>
15 #include <asm/pci-direct.h>
16 #include <asm/delay.h>
19 # include <asm/mmconfig.h>
20 # include <asm/set_memory.h>
25 static const int amd_erratum_383[];
26 static const int amd_erratum_400[];
27 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
30 * nodes_per_socket: Stores the number of nodes per socket.
31 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
32 * Node Identifiers[10:8]
34 static u32 nodes_per_socket = 1;
36 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
41 WARN_ONCE((boot_cpu_data.x86 != 0xf),
42 "%s should only be used on K8!\n", __func__);
47 err = rdmsr_safe_regs(gprs);
49 *p = gprs[0] | ((u64)gprs[2] << 32);
54 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
58 WARN_ONCE((boot_cpu_data.x86 != 0xf),
59 "%s should only be used on K8!\n", __func__);
66 return wrmsr_safe_regs(gprs);
70 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
71 * misexecution of code under Linux. Owners of such processors should
72 * contact AMD for precise details and a CPU swap.
74 * See http://www.multimania.com/poulot/k6bug.html
75 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
76 * (Publication # 21266 Issue Date: August 1998)
78 * The following test is erm.. interesting. AMD neglected to up
79 * the chip setting when fixing the bug but they also tweaked some
80 * performance at the same time..
83 extern __visible void vide(void);
84 __asm__(".globl vide\n"
85 ".type vide, @function\n"
89 static void init_amd_k5(struct cpuinfo_x86 *c)
93 * General Systems BIOSen alias the cpu frequency registers
94 * of the Elan at 0x000df000. Unfortunately, one of the Linux
95 * drivers subsequently pokes it, and changes the CPU speed.
96 * Workaround : Remove the unneeded alias.
98 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
99 #define CBAR_ENB (0x80000000)
100 #define CBAR_KEY (0X000000CB)
101 if (c->x86_model == 9 || c->x86_model == 10) {
102 if (inl(CBAR) & CBAR_ENB)
103 outl(0 | CBAR_KEY, CBAR);
108 static void init_amd_k6(struct cpuinfo_x86 *c)
112 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
114 if (c->x86_model < 6) {
115 /* Based on AMD doc 20734R - June 2000 */
116 if (c->x86_model == 0) {
117 clear_cpu_cap(c, X86_FEATURE_APIC);
118 set_cpu_cap(c, X86_FEATURE_PGE);
123 if (c->x86_model == 6 && c->x86_stepping == 1) {
124 const int K6_BUG_LOOP = 1000000;
126 void (*f_vide)(void);
129 pr_info("AMD K6 stepping B detected - ");
132 * It looks like AMD fixed the 2.6.2 bug and improved indirect
133 * calls at the same time.
138 OPTIMIZER_HIDE_VAR(f_vide);
145 if (d > 20*K6_BUG_LOOP)
146 pr_cont("system stability may be impaired when more than 32 MB are used.\n");
148 pr_cont("probably OK (after B9730xxxx).\n");
151 /* K6 with old style WHCR */
152 if (c->x86_model < 8 ||
153 (c->x86_model == 8 && c->x86_stepping < 8)) {
154 /* We can only write allocate on the low 508Mb */
158 rdmsr(MSR_K6_WHCR, l, h);
159 if ((l&0x0000FFFF) == 0) {
161 l = (1<<0)|((mbytes/4)<<1);
162 local_irq_save(flags);
164 wrmsr(MSR_K6_WHCR, l, h);
165 local_irq_restore(flags);
166 pr_info("Enabling old style K6 write allocation for %d Mb\n",
172 if ((c->x86_model == 8 && c->x86_stepping > 7) ||
173 c->x86_model == 9 || c->x86_model == 13) {
174 /* The more serious chips .. */
179 rdmsr(MSR_K6_WHCR, l, h);
180 if ((l&0xFFFF0000) == 0) {
182 l = ((mbytes>>2)<<22)|(1<<16);
183 local_irq_save(flags);
185 wrmsr(MSR_K6_WHCR, l, h);
186 local_irq_restore(flags);
187 pr_info("Enabling new style K6 write allocation for %d Mb\n",
194 if (c->x86_model == 10) {
195 /* AMD Geode LX is model 10 */
196 /* placeholder for any needed mods */
201 * Work around Erratum 1386. The XSAVES instruction malfunctions in
202 * certain circumstances on Zen1/2 uarch, and not all parts have had
203 * updated microcode at the time of writing (March 2023).
205 * Affected parts all have no supervisor XSAVE states, meaning that
206 * the XSAVEC instruction (which works fine) is equivalent.
208 clear_cpu_cap(c, X86_FEATURE_XSAVES);
211 static void init_amd_k7(struct cpuinfo_x86 *c)
217 * Bit 15 of Athlon specific MSR 15, needs to be 0
218 * to enable SSE on Palomino/Morgan/Barton CPU's.
219 * If the BIOS didn't enable it already, enable it here.
221 if (c->x86_model >= 6 && c->x86_model <= 10) {
222 if (!cpu_has(c, X86_FEATURE_XMM)) {
223 pr_info("Enabling disabled K7/SSE Support.\n");
224 msr_clear_bit(MSR_K7_HWCR, 15);
225 set_cpu_cap(c, X86_FEATURE_XMM);
230 * It's been determined by AMD that Athlons since model 8 stepping 1
231 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
232 * As per AMD technical note 27212 0.2
234 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
235 rdmsr(MSR_K7_CLK_CTL, l, h);
236 if ((l & 0xfff00000) != 0x20000000) {
237 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
238 l, ((l & 0x000fffff)|0x20000000));
239 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
243 set_cpu_cap(c, X86_FEATURE_K7);
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)) {
346 u32 eax, ebx, ecx, edx;
348 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
350 node_id = ecx & 0xff;
353 c->cu_id = ebx & 0xff;
355 if (c->x86 >= 0x17) {
356 c->cpu_core_id = ebx & 0xff;
358 if (smp_num_siblings > 1)
359 c->x86_max_cores /= smp_num_siblings;
363 * We may have multiple LLCs if L3 caches exist, so check if we
364 * have an L3 cache by looking at the L3 cache CPUID leaf.
366 if (cpuid_edx(0x80000006)) {
367 if (c->x86 == 0x17) {
369 * LLC is at the core complex level.
370 * Core complex id is ApicId[3].
372 per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
374 /* LLC is at the node level. */
375 per_cpu(cpu_llc_id, cpu) = node_id;
378 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
381 rdmsrl(MSR_FAM10H_NODE_ID, value);
384 per_cpu(cpu_llc_id, cpu) = node_id;
388 if (nodes_per_socket > 1) {
389 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
390 legacy_fixup_core_id(c);
395 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
396 * Assumes number of cores is a power of two.
398 static void amd_detect_cmp(struct cpuinfo_x86 *c)
401 int cpu = smp_processor_id();
403 bits = c->x86_coreid_bits;
404 /* Low order bits define the core id (index of core in socket) */
405 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
406 /* Convert the initial APIC ID into the socket ID */
407 c->phys_proc_id = c->initial_apicid >> bits;
408 /* use socket ID also for last level cache */
409 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
413 u16 amd_get_nb_id(int cpu)
415 return per_cpu(cpu_llc_id, cpu);
417 EXPORT_SYMBOL_GPL(amd_get_nb_id);
419 u32 amd_get_nodes_per_socket(void)
421 return nodes_per_socket;
423 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
425 static void srat_detect_node(struct cpuinfo_x86 *c)
428 int cpu = smp_processor_id();
430 unsigned apicid = c->apicid;
432 node = numa_cpu_node(cpu);
433 if (node == NUMA_NO_NODE)
434 node = per_cpu(cpu_llc_id, cpu);
437 * On multi-fabric platform (e.g. Numascale NumaChip) a
438 * platform-specific handler needs to be called to fixup some
441 if (x86_cpuinit.fixup_cpu_id)
442 x86_cpuinit.fixup_cpu_id(c, node);
444 if (!node_online(node)) {
446 * Two possibilities here:
448 * - The CPU is missing memory and no node was created. In
449 * that case try picking one from a nearby CPU.
451 * - The APIC IDs differ from the HyperTransport node IDs
452 * which the K8 northbridge parsing fills in. Assume
453 * they are all increased by a constant offset, but in
454 * the same order as the HT nodeids. If that doesn't
455 * result in a usable node fall back to the path for the
458 * This workaround operates directly on the mapping between
459 * APIC ID and NUMA node, assuming certain relationship
460 * between APIC ID, HT node ID and NUMA topology. As going
461 * through CPU mapping may alter the outcome, directly
462 * access __apicid_to_node[].
464 int ht_nodeid = c->initial_apicid;
466 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
467 node = __apicid_to_node[ht_nodeid];
468 /* Pick a nearby node */
469 if (!node_online(node))
470 node = nearby_node(apicid);
472 numa_set_node(cpu, node);
476 static void early_init_amd_mc(struct cpuinfo_x86 *c)
481 /* Multi core CPU? */
482 if (c->extended_cpuid_level < 0x80000008)
485 ecx = cpuid_ecx(0x80000008);
487 c->x86_max_cores = (ecx & 0xff) + 1;
489 /* CPU telling us the core id bits shift? */
490 bits = (ecx >> 12) & 0xF;
492 /* Otherwise recompute */
494 while ((1 << bits) < c->x86_max_cores)
498 c->x86_coreid_bits = bits;
502 static void bsp_init_amd(struct cpuinfo_x86 *c)
507 unsigned long long tseg;
510 * Split up direct mapping around the TSEG SMM area.
511 * Don't do it for gbpages because there seems very little
512 * benefit in doing so.
514 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
515 unsigned long pfn = tseg >> PAGE_SHIFT;
517 pr_debug("tseg: %010llx\n", tseg);
518 if (pfn_range_is_mapped(pfn, pfn + 1))
519 set_memory_4k((unsigned long)__va(tseg), 1);
524 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
527 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
530 rdmsrl(MSR_K7_HWCR, val);
531 if (!(val & BIT(24)))
532 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
536 if (c->x86 == 0x15) {
537 unsigned long upperbit;
540 cpuid = cpuid_edx(0x80000005);
541 assoc = cpuid >> 16 & 0xff;
542 upperbit = ((cpuid >> 24) << 10) / assoc;
544 va_align.mask = (upperbit - 1) & PAGE_MASK;
545 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
547 /* A random value per boot for bit slice [12:upper_bit) */
548 va_align.bits = get_random_int() & va_align.mask;
551 if (cpu_has(c, X86_FEATURE_MWAITX))
554 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
557 ecx = cpuid_ecx(0x8000001e);
558 nodes_per_socket = ((ecx >> 8) & 7) + 1;
559 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
562 rdmsrl(MSR_FAM10H_NODE_ID, value);
563 nodes_per_socket = ((value >> 3) & 7) + 1;
566 if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
567 !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
568 c->x86 >= 0x15 && c->x86 <= 0x17) {
572 case 0x15: bit = 54; break;
573 case 0x16: bit = 33; break;
574 case 0x17: bit = 10; break;
578 * Try to cache the base value so further operations can
579 * avoid RMW. If that faults, do not enable SSBD.
581 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
582 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
583 setup_force_cpu_cap(X86_FEATURE_SSBD);
584 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
589 static void early_init_amd(struct cpuinfo_x86 *c)
594 early_init_amd_mc(c);
596 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
599 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
600 * with P/T states and does not stop in deep C-states
602 if (c->x86_power & (1 << 8)) {
603 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
604 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
607 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
608 if (c->x86_power & BIT(12))
609 set_cpu_cap(c, X86_FEATURE_ACC_POWER);
612 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
614 /* Set MTRR capability flag if appropriate */
616 if (c->x86_model == 13 || c->x86_model == 9 ||
617 (c->x86_model == 8 && c->x86_stepping >= 8))
618 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
620 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
622 * ApicID can always be treated as an 8-bit value for AMD APIC versions
623 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
624 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
627 if (boot_cpu_has(X86_FEATURE_APIC)) {
629 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
630 else if (c->x86 >= 0xf) {
631 /* check CPU config space for extended APIC ID */
634 val = read_pci_config(0, 24, 0, 0x68);
635 if ((val >> 17 & 0x3) == 0x3)
636 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
642 * This is only needed to tell the kernel whether to use VMCALL
643 * and VMMCALL. VMMCALL is never executed except under virt, so
644 * we can set it unconditionally.
646 set_cpu_cap(c, X86_FEATURE_VMMCALL);
648 /* F16h erratum 793, CVE-2013-6885 */
649 if (c->x86 == 0x16 && c->x86_model <= 0xf)
650 msr_set_bit(MSR_AMD64_LS_CFG, 15);
653 * Check whether the machine is affected by erratum 400. This is
654 * used to select the proper idle routine and to enable the check
655 * whether the machine is affected in arch_post_acpi_init(), which
656 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
658 if (cpu_has_amd_erratum(c, amd_erratum_400))
659 set_cpu_bug(c, X86_BUG_AMD_E400);
662 * BIOS support is required for SME. If BIOS has enabled SME then
663 * adjust x86_phys_bits by the SME physical address space reduction
664 * value. If BIOS has not enabled SME then don't advertise the
665 * feature (set in scattered.c). Also, since the SME support requires
666 * long mode, don't advertise the feature under CONFIG_X86_32.
668 if (cpu_has(c, X86_FEATURE_SME)) {
671 /* Check if SME is enabled */
672 rdmsrl(MSR_K8_SYSCFG, msr);
673 if (msr & MSR_K8_SYSCFG_MEM_ENCRYPT) {
674 c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
675 if (IS_ENABLED(CONFIG_X86_32))
676 clear_cpu_cap(c, X86_FEATURE_SME);
678 clear_cpu_cap(c, X86_FEATURE_SME);
682 /* Re-enable TopologyExtensions if switched off by BIOS */
683 if (c->x86 == 0x15 &&
684 (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
685 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
687 if (msr_set_bit(0xc0011005, 54) > 0) {
688 rdmsrl(0xc0011005, value);
689 if (value & BIT_64(54)) {
690 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
691 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
696 amd_get_topology_early(c);
699 static void init_amd_k8(struct cpuinfo_x86 *c)
704 /* On C+ stepping K8 rep microcode works well for copy/memset */
705 level = cpuid_eax(1);
706 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
707 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
710 * Some BIOSes incorrectly force this feature, but only K8 revision D
711 * (model = 0x14) and later actually support it.
712 * (AMD Erratum #110, docId: 25759).
714 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
715 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
716 if (!rdmsrl_amd_safe(0xc001100d, &value)) {
717 value &= ~BIT_64(32);
718 wrmsrl_amd_safe(0xc001100d, value);
722 if (!c->x86_model_id[0])
723 strcpy(c->x86_model_id, "Hammer");
727 * Disable TLB flush filter by setting HWCR.FFDIS on K8
728 * bit 6 of msr C001_0015
730 * Errata 63 for SH-B3 steppings
731 * Errata 122 for all steppings (F+ have it disabled by default)
733 msr_set_bit(MSR_K7_HWCR, 6);
735 set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
738 static void init_amd_gh(struct cpuinfo_x86 *c)
741 /* do this for boot cpu */
742 if (c == &boot_cpu_data)
743 check_enable_amd_mmconf_dmi();
745 fam10h_check_enable_mmcfg();
749 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
750 * is always needed when GART is enabled, even in a kernel which has no
751 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
752 * If it doesn't, we do it here as suggested by the BKDG.
754 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
756 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
759 * On family 10h BIOS may not have properly enabled WC+ support, causing
760 * it to be converted to CD memtype. This may result in performance
761 * degradation for certain nested-paging guests. Prevent this conversion
762 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
764 * NOTE: we want to use the _safe accessors so as not to #GP kvm
765 * guests on older kvm hosts.
767 msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
769 if (cpu_has_amd_erratum(c, amd_erratum_383))
770 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
773 static void init_amd_ln(struct cpuinfo_x86 *c)
776 * Apply erratum 665 fix unconditionally so machines without a BIOS
779 msr_set_bit(MSR_AMD64_DE_CFG, 31);
782 static bool rdrand_force;
784 static int __init rdrand_cmdline(char *str)
789 if (!strcmp(str, "force"))
796 early_param("rdrand", rdrand_cmdline);
798 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
801 * Saving of the MSR used to hide the RDRAND support during
802 * suspend/resume is done by arch/x86/power/cpu.c, which is
803 * dependent on CONFIG_PM_SLEEP.
805 if (!IS_ENABLED(CONFIG_PM_SLEEP))
809 * The nordrand option can clear X86_FEATURE_RDRAND, so check for
810 * RDRAND support using the CPUID function directly.
812 if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
815 msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
818 * Verify that the CPUID change has occurred in case the kernel is
819 * running virtualized and the hypervisor doesn't support the MSR.
821 if (cpuid_ecx(1) & BIT(30)) {
822 pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
826 clear_cpu_cap(c, X86_FEATURE_RDRAND);
827 pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
830 static void init_amd_jg(struct cpuinfo_x86 *c)
833 * Some BIOS implementations do not restore proper RDRAND support
834 * across suspend and resume. Check on whether to hide the RDRAND
835 * instruction support via CPUID.
837 clear_rdrand_cpuid_bit(c);
840 static void init_amd_bd(struct cpuinfo_x86 *c)
845 * The way access filter has a performance penalty on some workloads.
846 * Disable it on the affected CPUs.
848 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
849 if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
851 wrmsrl_safe(MSR_F15H_IC_CFG, value);
856 * Some BIOS implementations do not restore proper RDRAND support
857 * across suspend and resume. Check on whether to hide the RDRAND
858 * instruction support via CPUID.
860 clear_rdrand_cpuid_bit(c);
863 static void init_amd_zn(struct cpuinfo_x86 *c)
865 set_cpu_cap(c, X86_FEATURE_ZEN);
867 /* Fix up CPUID bits, but only if not virtualised. */
868 if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
870 /* Erratum 1076: CPB feature bit not being set in CPUID. */
871 if (!cpu_has(c, X86_FEATURE_CPB))
872 set_cpu_cap(c, X86_FEATURE_CPB);
875 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
876 * Branch Type Confusion, but predate the allocation of the
879 if (c->x86 == 0x19 && !cpu_has(c, X86_FEATURE_BTC_NO))
880 set_cpu_cap(c, X86_FEATURE_BTC_NO);
884 static void init_amd(struct cpuinfo_x86 *c)
889 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
890 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
892 clear_cpu_cap(c, 0*32+31);
895 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
897 /* get apicid instead of initial apic id from cpuid */
898 c->apicid = hard_smp_processor_id();
900 /* K6s reports MCEs but don't actually have all the MSRs */
902 clear_cpu_cap(c, X86_FEATURE_MCE);
905 case 4: init_amd_k5(c); break;
906 case 5: init_amd_k6(c); break;
907 case 6: init_amd_k7(c); break;
908 case 0xf: init_amd_k8(c); break;
909 case 0x10: init_amd_gh(c); break;
910 case 0x12: init_amd_ln(c); break;
911 case 0x15: init_amd_bd(c); break;
912 case 0x16: init_amd_jg(c); break;
913 case 0x17: init_amd_zn(c); break;
917 * Enable workaround for FXSAVE leak on CPUs
918 * without a XSaveErPtr feature
920 if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
921 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
923 cpu_detect_cache_sizes(c);
928 init_amd_cacheinfo(c);
931 set_cpu_cap(c, X86_FEATURE_K8);
933 if (cpu_has(c, X86_FEATURE_XMM2)) {
934 unsigned long long val;
938 * A serializing LFENCE has less overhead than MFENCE, so
939 * use it for execution serialization. On families which
940 * don't have that MSR, LFENCE is already serializing.
941 * msr_set_bit() uses the safe accessors, too, even if the MSR
944 msr_set_bit(MSR_AMD64_DE_CFG,
945 MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
948 * Verify that the MSR write was successful (could be running
949 * under a hypervisor) and only then assume that LFENCE is
952 ret = rdmsrl_safe(MSR_AMD64_DE_CFG, &val);
953 if (!ret && (val & MSR_AMD64_DE_CFG_LFENCE_SERIALIZE)) {
954 /* A serializing LFENCE stops RDTSC speculation */
955 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
957 /* MFENCE stops RDTSC speculation */
958 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
963 * Family 0x12 and above processors have APIC timer
964 * running in deep C states.
967 set_cpu_cap(c, X86_FEATURE_ARAT);
969 /* 3DNow or LM implies PREFETCHW */
970 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
971 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
972 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
974 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
975 if (!cpu_has(c, X86_FEATURE_XENPV))
976 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
980 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
982 /* AMD errata T13 (order #21922) */
985 if (c->x86_model == 3 && c->x86_stepping == 0)
987 /* Tbird rev A1/A2 */
988 if (c->x86_model == 4 &&
989 (c->x86_stepping == 0 || c->x86_stepping == 1))
996 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
998 u32 ebx, eax, ecx, edx;
1004 if (c->extended_cpuid_level < 0x80000006)
1007 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1009 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1010 tlb_lli_4k[ENTRIES] = ebx & mask;
1013 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1014 * characteristics from the CPUID function 0x80000005 instead.
1016 if (c->x86 == 0xf) {
1017 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1021 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1022 if (!((eax >> 16) & mask))
1023 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1025 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1027 /* a 4M entry uses two 2M entries */
1028 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1030 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1031 if (!(eax & mask)) {
1033 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1034 tlb_lli_2m[ENTRIES] = 1024;
1036 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1037 tlb_lli_2m[ENTRIES] = eax & 0xff;
1040 tlb_lli_2m[ENTRIES] = eax & mask;
1042 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1045 static const struct cpu_dev amd_cpu_dev = {
1047 .c_ident = { "AuthenticAMD" },
1048 #ifdef CONFIG_X86_32
1050 { .family = 4, .model_names =
1053 [7] = "486 DX/2-WB",
1055 [9] = "486 DX/4-WB",
1061 .legacy_cache_size = amd_size_cache,
1063 .c_early_init = early_init_amd,
1064 .c_detect_tlb = cpu_detect_tlb_amd,
1065 .c_bsp_init = bsp_init_amd,
1067 .c_x86_vendor = X86_VENDOR_AMD,
1070 cpu_dev_register(amd_cpu_dev);
1073 * AMD errata checking
1075 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1076 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1077 * have an OSVW id assigned, which it takes as first argument. Both take a
1078 * variable number of family-specific model-stepping ranges created by
1079 * AMD_MODEL_RANGE().
1083 * const int amd_erratum_319[] =
1084 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1085 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1086 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1089 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
1090 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
1091 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1092 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1093 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
1094 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
1095 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
1097 static const int amd_erratum_400[] =
1098 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1099 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1101 static const int amd_erratum_383[] =
1102 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1105 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1107 int osvw_id = *erratum++;
1111 if (osvw_id >= 0 && osvw_id < 65536 &&
1112 cpu_has(cpu, X86_FEATURE_OSVW)) {
1115 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1116 if (osvw_id < osvw_len) {
1119 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1121 return osvw_bits & (1ULL << (osvw_id & 0x3f));
1125 /* OSVW unavailable or ID unknown, match family-model-stepping range */
1126 ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1127 while ((range = *erratum++))
1128 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1129 (ms >= AMD_MODEL_RANGE_START(range)) &&
1130 (ms <= AMD_MODEL_RANGE_END(range)))
1136 void set_dr_addr_mask(unsigned long mask, int dr)
1138 if (!boot_cpu_has(X86_FEATURE_BPEXT))
1143 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1148 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);