1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/export.h>
3 #include <linux/bitops.h>
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <linux/random.h>
11 #include <linux/topology.h>
12 #include <asm/processor.h>
14 #include <asm/cacheinfo.h>
16 #include <asm/spec-ctrl.h>
18 #include <asm/pci-direct.h>
19 #include <asm/delay.h>
20 #include <asm/debugreg.h>
23 # include <asm/mmconfig.h>
24 # include <asm/set_memory.h>
29 static const int amd_erratum_383[];
30 static const int amd_erratum_400[];
31 static const int amd_erratum_1054[];
32 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
35 * nodes_per_socket: Stores the number of nodes per socket.
36 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
37 * Node Identifiers[10:8]
39 static u32 nodes_per_socket = 1;
41 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
46 WARN_ONCE((boot_cpu_data.x86 != 0xf),
47 "%s should only be used on K8!\n", __func__);
52 err = rdmsr_safe_regs(gprs);
54 *p = gprs[0] | ((u64)gprs[2] << 32);
59 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
63 WARN_ONCE((boot_cpu_data.x86 != 0xf),
64 "%s should only be used on K8!\n", __func__);
71 return wrmsr_safe_regs(gprs);
75 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
76 * misexecution of code under Linux. Owners of such processors should
77 * contact AMD for precise details and a CPU swap.
79 * See http://www.multimania.com/poulot/k6bug.html
80 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
81 * (Publication # 21266 Issue Date: August 1998)
83 * The following test is erm.. interesting. AMD neglected to up
84 * the chip setting when fixing the bug but they also tweaked some
85 * performance at the same time..
89 extern __visible void vide(void);
92 ".type vide, @function\n"
97 static void init_amd_k5(struct cpuinfo_x86 *c)
101 * General Systems BIOSen alias the cpu frequency registers
102 * of the Elan at 0x000df000. Unfortunately, one of the Linux
103 * drivers subsequently pokes it, and changes the CPU speed.
104 * Workaround : Remove the unneeded alias.
106 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
107 #define CBAR_ENB (0x80000000)
108 #define CBAR_KEY (0X000000CB)
109 if (c->x86_model == 9 || c->x86_model == 10) {
110 if (inl(CBAR) & CBAR_ENB)
111 outl(0 | CBAR_KEY, CBAR);
116 static void init_amd_k6(struct cpuinfo_x86 *c)
120 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
122 if (c->x86_model < 6) {
123 /* Based on AMD doc 20734R - June 2000 */
124 if (c->x86_model == 0) {
125 clear_cpu_cap(c, X86_FEATURE_APIC);
126 set_cpu_cap(c, X86_FEATURE_PGE);
131 if (c->x86_model == 6 && c->x86_stepping == 1) {
132 const int K6_BUG_LOOP = 1000000;
134 void (*f_vide)(void);
137 pr_info("AMD K6 stepping B detected - ");
140 * It looks like AMD fixed the 2.6.2 bug and improved indirect
141 * calls at the same time.
146 OPTIMIZER_HIDE_VAR(f_vide);
153 if (d > 20*K6_BUG_LOOP)
154 pr_cont("system stability may be impaired when more than 32 MB are used.\n");
156 pr_cont("probably OK (after B9730xxxx).\n");
159 /* K6 with old style WHCR */
160 if (c->x86_model < 8 ||
161 (c->x86_model == 8 && c->x86_stepping < 8)) {
162 /* We can only write allocate on the low 508Mb */
166 rdmsr(MSR_K6_WHCR, l, h);
167 if ((l&0x0000FFFF) == 0) {
169 l = (1<<0)|((mbytes/4)<<1);
170 local_irq_save(flags);
172 wrmsr(MSR_K6_WHCR, l, h);
173 local_irq_restore(flags);
174 pr_info("Enabling old style K6 write allocation for %d Mb\n",
180 if ((c->x86_model == 8 && c->x86_stepping > 7) ||
181 c->x86_model == 9 || c->x86_model == 13) {
182 /* The more serious chips .. */
187 rdmsr(MSR_K6_WHCR, l, h);
188 if ((l&0xFFFF0000) == 0) {
190 l = ((mbytes>>2)<<22)|(1<<16);
191 local_irq_save(flags);
193 wrmsr(MSR_K6_WHCR, l, h);
194 local_irq_restore(flags);
195 pr_info("Enabling new style K6 write allocation for %d Mb\n",
202 if (c->x86_model == 10) {
203 /* AMD Geode LX is model 10 */
204 /* placeholder for any needed mods */
210 static void init_amd_k7(struct cpuinfo_x86 *c)
216 * Bit 15 of Athlon specific MSR 15, needs to be 0
217 * to enable SSE on Palomino/Morgan/Barton CPU's.
218 * If the BIOS didn't enable it already, enable it here.
220 if (c->x86_model >= 6 && c->x86_model <= 10) {
221 if (!cpu_has(c, X86_FEATURE_XMM)) {
222 pr_info("Enabling disabled K7/SSE Support.\n");
223 msr_clear_bit(MSR_K7_HWCR, 15);
224 set_cpu_cap(c, X86_FEATURE_XMM);
229 * It's been determined by AMD that Athlons since model 8 stepping 1
230 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
231 * As per AMD technical note 27212 0.2
233 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
234 rdmsr(MSR_K7_CLK_CTL, l, h);
235 if ((l & 0xfff00000) != 0x20000000) {
236 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
237 l, ((l & 0x000fffff)|0x20000000));
238 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
242 /* calling is from identify_secondary_cpu() ? */
247 * Certain Athlons might work (for various values of 'work') in SMP
248 * but they are not certified as MP capable.
250 /* Athlon 660/661 is valid. */
251 if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
252 (c->x86_stepping == 1)))
255 /* Duron 670 is valid */
256 if ((c->x86_model == 7) && (c->x86_stepping == 0))
260 * Athlon 662, Duron 671, and Athlon >model 7 have capability
261 * bit. It's worth noting that the A5 stepping (662) of some
262 * Athlon XP's have the MP bit set.
263 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
266 if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
267 ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
269 if (cpu_has(c, X86_FEATURE_MP))
272 /* If we get here, not a certified SMP capable AMD system. */
275 * Don't taint if we are running SMP kernel on a single non-MP
278 WARN_ONCE(1, "WARNING: This combination of AMD"
279 " processors is not suitable for SMP.\n");
280 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
286 * To workaround broken NUMA config. Read the comment in
287 * srat_detect_node().
289 static int nearby_node(int apicid)
293 for (i = apicid - 1; i >= 0; i--) {
294 node = __apicid_to_node[i];
295 if (node != NUMA_NO_NODE && node_online(node))
298 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
299 node = __apicid_to_node[i];
300 if (node != NUMA_NO_NODE && node_online(node))
303 return first_node(node_online_map); /* Shouldn't happen */
308 * Fix up cpu_core_id for pre-F17h systems to be in the
309 * [0 .. cores_per_node - 1] range. Not really needed but
310 * kept so as not to break existing setups.
312 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
319 cus_per_node = c->x86_max_cores / nodes_per_socket;
320 c->cpu_core_id %= cus_per_node;
324 * Fixup core topology information for
325 * (1) AMD multi-node processors
326 * Assumption: Number of cores in each internal node is the same.
327 * (2) AMD processors supporting compute units
329 static void amd_get_topology(struct cpuinfo_x86 *c)
332 int cpu = smp_processor_id();
334 /* get information required for multi-node processors */
335 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
337 u32 eax, ebx, ecx, edx;
339 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
341 node_id = ecx & 0xff;
344 c->cu_id = ebx & 0xff;
346 if (c->x86 >= 0x17) {
347 c->cpu_core_id = ebx & 0xff;
349 if (smp_num_siblings > 1)
350 c->x86_max_cores /= smp_num_siblings;
354 * In case leaf B is available, use it to derive
355 * topology information.
357 err = detect_extended_topology(c);
359 c->x86_coreid_bits = get_count_order(c->x86_max_cores);
361 cacheinfo_amd_init_llc_id(c, cpu, node_id);
363 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
366 rdmsrl(MSR_FAM10H_NODE_ID, value);
369 per_cpu(cpu_llc_id, cpu) = node_id;
373 if (nodes_per_socket > 1) {
374 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
375 legacy_fixup_core_id(c);
380 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
381 * Assumes number of cores is a power of two.
383 static void amd_detect_cmp(struct cpuinfo_x86 *c)
386 int cpu = smp_processor_id();
388 bits = c->x86_coreid_bits;
389 /* Low order bits define the core id (index of core in socket) */
390 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
391 /* Convert the initial APIC ID into the socket ID */
392 c->phys_proc_id = c->initial_apicid >> bits;
393 /* use socket ID also for last level cache */
394 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
397 u16 amd_get_nb_id(int cpu)
399 return per_cpu(cpu_llc_id, cpu);
401 EXPORT_SYMBOL_GPL(amd_get_nb_id);
403 u32 amd_get_nodes_per_socket(void)
405 return nodes_per_socket;
407 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
409 static void srat_detect_node(struct cpuinfo_x86 *c)
412 int cpu = smp_processor_id();
414 unsigned apicid = c->apicid;
416 node = numa_cpu_node(cpu);
417 if (node == NUMA_NO_NODE)
418 node = per_cpu(cpu_llc_id, cpu);
421 * On multi-fabric platform (e.g. Numascale NumaChip) a
422 * platform-specific handler needs to be called to fixup some
425 if (x86_cpuinit.fixup_cpu_id)
426 x86_cpuinit.fixup_cpu_id(c, node);
428 if (!node_online(node)) {
430 * Two possibilities here:
432 * - The CPU is missing memory and no node was created. In
433 * that case try picking one from a nearby CPU.
435 * - The APIC IDs differ from the HyperTransport node IDs
436 * which the K8 northbridge parsing fills in. Assume
437 * they are all increased by a constant offset, but in
438 * the same order as the HT nodeids. If that doesn't
439 * result in a usable node fall back to the path for the
442 * This workaround operates directly on the mapping between
443 * APIC ID and NUMA node, assuming certain relationship
444 * between APIC ID, HT node ID and NUMA topology. As going
445 * through CPU mapping may alter the outcome, directly
446 * access __apicid_to_node[].
448 int ht_nodeid = c->initial_apicid;
450 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
451 node = __apicid_to_node[ht_nodeid];
452 /* Pick a nearby node */
453 if (!node_online(node))
454 node = nearby_node(apicid);
456 numa_set_node(cpu, node);
460 static void early_init_amd_mc(struct cpuinfo_x86 *c)
465 /* Multi core CPU? */
466 if (c->extended_cpuid_level < 0x80000008)
469 ecx = cpuid_ecx(0x80000008);
471 c->x86_max_cores = (ecx & 0xff) + 1;
473 /* CPU telling us the core id bits shift? */
474 bits = (ecx >> 12) & 0xF;
476 /* Otherwise recompute */
478 while ((1 << bits) < c->x86_max_cores)
482 c->x86_coreid_bits = bits;
486 static void bsp_init_amd(struct cpuinfo_x86 *c)
491 unsigned long long tseg;
494 * Split up direct mapping around the TSEG SMM area.
495 * Don't do it for gbpages because there seems very little
496 * benefit in doing so.
498 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
499 unsigned long pfn = tseg >> PAGE_SHIFT;
501 pr_debug("tseg: %010llx\n", tseg);
502 if (pfn_range_is_mapped(pfn, pfn + 1))
503 set_memory_4k((unsigned long)__va(tseg), 1);
508 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
511 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
514 rdmsrl(MSR_K7_HWCR, val);
515 if (!(val & BIT(24)))
516 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
520 if (c->x86 == 0x15) {
521 unsigned long upperbit;
524 cpuid = cpuid_edx(0x80000005);
525 assoc = cpuid >> 16 & 0xff;
526 upperbit = ((cpuid >> 24) << 10) / assoc;
528 va_align.mask = (upperbit - 1) & PAGE_MASK;
529 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
531 /* A random value per boot for bit slice [12:upper_bit) */
532 va_align.bits = get_random_int() & va_align.mask;
535 if (cpu_has(c, X86_FEATURE_MWAITX))
538 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
541 ecx = cpuid_ecx(0x8000001e);
542 nodes_per_socket = ((ecx >> 8) & 7) + 1;
543 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
546 rdmsrl(MSR_FAM10H_NODE_ID, value);
547 nodes_per_socket = ((value >> 3) & 7) + 1;
550 if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
551 !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
552 c->x86 >= 0x15 && c->x86 <= 0x17) {
556 case 0x15: bit = 54; break;
557 case 0x16: bit = 33; break;
558 case 0x17: bit = 10; break;
562 * Try to cache the base value so further operations can
563 * avoid RMW. If that faults, do not enable SSBD.
565 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
566 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
567 setup_force_cpu_cap(X86_FEATURE_SSBD);
568 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
573 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
578 * BIOS support is required for SME and SEV.
579 * For SME: If BIOS has enabled SME then adjust x86_phys_bits by
580 * the SME physical address space reduction value.
581 * If BIOS has not enabled SME then don't advertise the
582 * SME feature (set in scattered.c).
583 * For SEV: If BIOS has not enabled SEV then don't advertise the
584 * SEV feature (set in scattered.c).
586 * In all cases, since support for SME and SEV requires long mode,
587 * don't advertise the feature under CONFIG_X86_32.
589 if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
590 /* Check if memory encryption is enabled */
591 rdmsrl(MSR_K8_SYSCFG, msr);
592 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
596 * Always adjust physical address bits. Even though this
597 * will be a value above 32-bits this is still done for
598 * CONFIG_X86_32 so that accurate values are reported.
600 c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
602 if (IS_ENABLED(CONFIG_X86_32))
605 rdmsrl(MSR_K7_HWCR, msr);
606 if (!(msr & MSR_K7_HWCR_SMMLOCK))
612 setup_clear_cpu_cap(X86_FEATURE_SME);
614 setup_clear_cpu_cap(X86_FEATURE_SEV);
618 static void early_init_amd(struct cpuinfo_x86 *c)
623 early_init_amd_mc(c);
627 set_cpu_cap(c, X86_FEATURE_K7);
631 set_cpu_cap(c, X86_FEATURE_K8);
633 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
636 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
637 * with P/T states and does not stop in deep C-states
639 if (c->x86_power & (1 << 8)) {
640 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
641 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
644 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
645 if (c->x86_power & BIT(12))
646 set_cpu_cap(c, X86_FEATURE_ACC_POWER);
649 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
651 /* Set MTRR capability flag if appropriate */
653 if (c->x86_model == 13 || c->x86_model == 9 ||
654 (c->x86_model == 8 && c->x86_stepping >= 8))
655 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
657 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
659 * ApicID can always be treated as an 8-bit value for AMD APIC versions
660 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
661 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
664 if (boot_cpu_has(X86_FEATURE_APIC)) {
666 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
667 else if (c->x86 >= 0xf) {
668 /* check CPU config space for extended APIC ID */
671 val = read_pci_config(0, 24, 0, 0x68);
672 if ((val >> 17 & 0x3) == 0x3)
673 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
679 * This is only needed to tell the kernel whether to use VMCALL
680 * and VMMCALL. VMMCALL is never executed except under virt, so
681 * we can set it unconditionally.
683 set_cpu_cap(c, X86_FEATURE_VMMCALL);
685 /* F16h erratum 793, CVE-2013-6885 */
686 if (c->x86 == 0x16 && c->x86_model <= 0xf)
687 msr_set_bit(MSR_AMD64_LS_CFG, 15);
690 * Check whether the machine is affected by erratum 400. This is
691 * used to select the proper idle routine and to enable the check
692 * whether the machine is affected in arch_post_acpi_init(), which
693 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
695 if (cpu_has_amd_erratum(c, amd_erratum_400))
696 set_cpu_bug(c, X86_BUG_AMD_E400);
698 early_detect_mem_encrypt(c);
700 /* Re-enable TopologyExtensions if switched off by BIOS */
701 if (c->x86 == 0x15 &&
702 (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
703 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
705 if (msr_set_bit(0xc0011005, 54) > 0) {
706 rdmsrl(0xc0011005, value);
707 if (value & BIT_64(54)) {
708 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
709 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
714 if (cpu_has(c, X86_FEATURE_TOPOEXT))
715 smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
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 node_reclaim_distance = 32;
893 * Fix erratum 1076: CPB feature bit not being set in CPUID.
894 * Always set it, except when running under a hypervisor.
896 if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_CPB))
897 set_cpu_cap(c, X86_FEATURE_CPB);
900 static void init_amd(struct cpuinfo_x86 *c)
905 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
906 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
908 clear_cpu_cap(c, 0*32+31);
911 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
913 /* get apicid instead of initial apic id from cpuid */
914 c->apicid = hard_smp_processor_id();
916 /* K6s reports MCEs but don't actually have all the MSRs */
918 clear_cpu_cap(c, X86_FEATURE_MCE);
921 case 4: init_amd_k5(c); break;
922 case 5: init_amd_k6(c); break;
923 case 6: init_amd_k7(c); break;
924 case 0xf: init_amd_k8(c); break;
925 case 0x10: init_amd_gh(c); break;
926 case 0x12: init_amd_ln(c); break;
927 case 0x15: init_amd_bd(c); break;
928 case 0x16: init_amd_jg(c); break;
929 case 0x17: init_amd_zn(c); break;
933 * Enable workaround for FXSAVE leak on CPUs
934 * without a XSaveErPtr feature
936 if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
937 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
939 cpu_detect_cache_sizes(c);
945 init_amd_cacheinfo(c);
947 if (cpu_has(c, X86_FEATURE_XMM2)) {
949 * Use LFENCE 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);
957 /* A serializing LFENCE stops RDTSC speculation */
958 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
962 * Family 0x12 and above processors have APIC timer
963 * running in deep C states.
966 set_cpu_cap(c, X86_FEATURE_ARAT);
968 /* 3DNow or LM implies PREFETCHW */
969 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
970 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
971 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
973 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
974 if (!cpu_has(c, X86_FEATURE_XENPV))
975 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
978 * Turn on the Instructions Retired free counter on machines not
979 * susceptible to erratum #1054 "Instructions Retired Performance
980 * Counter May Be Inaccurate".
982 if (cpu_has(c, X86_FEATURE_IRPERF) &&
983 !cpu_has_amd_erratum(c, amd_erratum_1054))
984 msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
988 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
990 /* AMD errata T13 (order #21922) */
993 if (c->x86_model == 3 && c->x86_stepping == 0)
995 /* Tbird rev A1/A2 */
996 if (c->x86_model == 4 &&
997 (c->x86_stepping == 0 || c->x86_stepping == 1))
1004 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1006 u32 ebx, eax, ecx, edx;
1012 if (c->extended_cpuid_level < 0x80000006)
1015 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1017 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1018 tlb_lli_4k[ENTRIES] = ebx & mask;
1021 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1022 * characteristics from the CPUID function 0x80000005 instead.
1024 if (c->x86 == 0xf) {
1025 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1029 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1030 if (!((eax >> 16) & mask))
1031 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1033 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1035 /* a 4M entry uses two 2M entries */
1036 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1038 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1039 if (!(eax & mask)) {
1041 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1042 tlb_lli_2m[ENTRIES] = 1024;
1044 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1045 tlb_lli_2m[ENTRIES] = eax & 0xff;
1048 tlb_lli_2m[ENTRIES] = eax & mask;
1050 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1053 static const struct cpu_dev amd_cpu_dev = {
1055 .c_ident = { "AuthenticAMD" },
1056 #ifdef CONFIG_X86_32
1058 { .family = 4, .model_names =
1061 [7] = "486 DX/2-WB",
1063 [9] = "486 DX/4-WB",
1069 .legacy_cache_size = amd_size_cache,
1071 .c_early_init = early_init_amd,
1072 .c_detect_tlb = cpu_detect_tlb_amd,
1073 .c_bsp_init = bsp_init_amd,
1075 .c_x86_vendor = X86_VENDOR_AMD,
1078 cpu_dev_register(amd_cpu_dev);
1081 * AMD errata checking
1083 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1084 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1085 * have an OSVW id assigned, which it takes as first argument. Both take a
1086 * variable number of family-specific model-stepping ranges created by
1087 * AMD_MODEL_RANGE().
1091 * const int amd_erratum_319[] =
1092 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1093 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1094 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1097 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
1098 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
1099 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1100 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1101 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
1102 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
1103 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
1105 static const int amd_erratum_400[] =
1106 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1107 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1109 static const int amd_erratum_383[] =
1110 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1112 /* #1054: Instructions Retired Performance Counter May Be Inaccurate */
1113 static const int amd_erratum_1054[] =
1114 AMD_OSVW_ERRATUM(0, AMD_MODEL_RANGE(0x17, 0, 0, 0x2f, 0xf));
1117 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1119 int osvw_id = *erratum++;
1123 if (osvw_id >= 0 && osvw_id < 65536 &&
1124 cpu_has(cpu, X86_FEATURE_OSVW)) {
1127 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1128 if (osvw_id < osvw_len) {
1131 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1133 return osvw_bits & (1ULL << (osvw_id & 0x3f));
1137 /* OSVW unavailable or ID unknown, match family-model-stepping range */
1138 ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1139 while ((range = *erratum++))
1140 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1141 (ms >= AMD_MODEL_RANGE_START(range)) &&
1142 (ms <= AMD_MODEL_RANGE_END(range)))
1148 void set_dr_addr_mask(unsigned long mask, int dr)
1150 if (!boot_cpu_has(X86_FEATURE_BPEXT))
1155 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1160 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);