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>
19 #include <asm/pci-direct.h>
20 #include <asm/delay.h>
21 #include <asm/debugreg.h>
22 #include <asm/resctrl.h>
25 # include <asm/mmconfig.h>
31 * nodes_per_socket: Stores the number of nodes per socket.
32 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
33 * Node Identifiers[10:8]
35 static u32 nodes_per_socket = 1;
40 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
41 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
42 * have an OSVW id assigned, which it takes as first argument. Both take a
43 * variable number of family-specific model-stepping ranges created by
48 * const int amd_erratum_319[] =
49 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
50 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
51 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
54 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
55 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
56 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
57 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
58 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
59 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
60 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
62 static const int amd_erratum_400[] =
63 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
64 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
66 static const int amd_erratum_383[] =
67 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
69 /* #1054: Instructions Retired Performance Counter May Be Inaccurate */
70 static const int amd_erratum_1054[] =
71 AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0, 0, 0x2f, 0xf));
73 static const int amd_zenbleed[] =
74 AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x17, 0x30, 0x0, 0x4f, 0xf),
75 AMD_MODEL_RANGE(0x17, 0x60, 0x0, 0x7f, 0xf),
76 AMD_MODEL_RANGE(0x17, 0x90, 0x0, 0x91, 0xf),
77 AMD_MODEL_RANGE(0x17, 0xa0, 0x0, 0xaf, 0xf));
79 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
81 int osvw_id = *erratum++;
85 if (osvw_id >= 0 && osvw_id < 65536 &&
86 cpu_has(cpu, X86_FEATURE_OSVW)) {
89 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
90 if (osvw_id < osvw_len) {
93 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
95 return osvw_bits & (1ULL << (osvw_id & 0x3f));
99 /* OSVW unavailable or ID unknown, match family-model-stepping range */
100 ms = (cpu->x86_model << 4) | cpu->x86_stepping;
101 while ((range = *erratum++))
102 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
103 (ms >= AMD_MODEL_RANGE_START(range)) &&
104 (ms <= AMD_MODEL_RANGE_END(range)))
110 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
115 WARN_ONCE((boot_cpu_data.x86 != 0xf),
116 "%s should only be used on K8!\n", __func__);
119 gprs[7] = 0x9c5a203a;
121 err = rdmsr_safe_regs(gprs);
123 *p = gprs[0] | ((u64)gprs[2] << 32);
128 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
132 WARN_ONCE((boot_cpu_data.x86 != 0xf),
133 "%s should only be used on K8!\n", __func__);
138 gprs[7] = 0x9c5a203a;
140 return wrmsr_safe_regs(gprs);
144 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
145 * misexecution of code under Linux. Owners of such processors should
146 * contact AMD for precise details and a CPU swap.
148 * See http://www.multimania.com/poulot/k6bug.html
149 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
150 * (Publication # 21266 Issue Date: August 1998)
152 * The following test is erm.. interesting. AMD neglected to up
153 * the chip setting when fixing the bug but they also tweaked some
154 * performance at the same time..
158 extern __visible void vide(void);
161 ".type vide, @function\n"
166 static void init_amd_k5(struct cpuinfo_x86 *c)
170 * General Systems BIOSen alias the cpu frequency registers
171 * of the Elan at 0x000df000. Unfortunately, one of the Linux
172 * drivers subsequently pokes it, and changes the CPU speed.
173 * Workaround : Remove the unneeded alias.
175 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
176 #define CBAR_ENB (0x80000000)
177 #define CBAR_KEY (0X000000CB)
178 if (c->x86_model == 9 || c->x86_model == 10) {
179 if (inl(CBAR) & CBAR_ENB)
180 outl(0 | CBAR_KEY, CBAR);
185 static void init_amd_k6(struct cpuinfo_x86 *c)
189 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
191 if (c->x86_model < 6) {
192 /* Based on AMD doc 20734R - June 2000 */
193 if (c->x86_model == 0) {
194 clear_cpu_cap(c, X86_FEATURE_APIC);
195 set_cpu_cap(c, X86_FEATURE_PGE);
200 if (c->x86_model == 6 && c->x86_stepping == 1) {
201 const int K6_BUG_LOOP = 1000000;
203 void (*f_vide)(void);
206 pr_info("AMD K6 stepping B detected - ");
209 * It looks like AMD fixed the 2.6.2 bug and improved indirect
210 * calls at the same time.
215 OPTIMIZER_HIDE_VAR(f_vide);
222 if (d > 20*K6_BUG_LOOP)
223 pr_cont("system stability may be impaired when more than 32 MB are used.\n");
225 pr_cont("probably OK (after B9730xxxx).\n");
228 /* K6 with old style WHCR */
229 if (c->x86_model < 8 ||
230 (c->x86_model == 8 && c->x86_stepping < 8)) {
231 /* We can only write allocate on the low 508Mb */
235 rdmsr(MSR_K6_WHCR, l, h);
236 if ((l&0x0000FFFF) == 0) {
238 l = (1<<0)|((mbytes/4)<<1);
239 local_irq_save(flags);
241 wrmsr(MSR_K6_WHCR, l, h);
242 local_irq_restore(flags);
243 pr_info("Enabling old style K6 write allocation for %d Mb\n",
249 if ((c->x86_model == 8 && c->x86_stepping > 7) ||
250 c->x86_model == 9 || c->x86_model == 13) {
251 /* The more serious chips .. */
256 rdmsr(MSR_K6_WHCR, l, h);
257 if ((l&0xFFFF0000) == 0) {
259 l = ((mbytes>>2)<<22)|(1<<16);
260 local_irq_save(flags);
262 wrmsr(MSR_K6_WHCR, l, h);
263 local_irq_restore(flags);
264 pr_info("Enabling new style K6 write allocation for %d Mb\n",
271 if (c->x86_model == 10) {
272 /* AMD Geode LX is model 10 */
273 /* placeholder for any needed mods */
279 static void init_amd_k7(struct cpuinfo_x86 *c)
285 * Bit 15 of Athlon specific MSR 15, needs to be 0
286 * to enable SSE on Palomino/Morgan/Barton CPU's.
287 * If the BIOS didn't enable it already, enable it here.
289 if (c->x86_model >= 6 && c->x86_model <= 10) {
290 if (!cpu_has(c, X86_FEATURE_XMM)) {
291 pr_info("Enabling disabled K7/SSE Support.\n");
292 msr_clear_bit(MSR_K7_HWCR, 15);
293 set_cpu_cap(c, X86_FEATURE_XMM);
298 * It's been determined by AMD that Athlons since model 8 stepping 1
299 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
300 * As per AMD technical note 27212 0.2
302 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
303 rdmsr(MSR_K7_CLK_CTL, l, h);
304 if ((l & 0xfff00000) != 0x20000000) {
305 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
306 l, ((l & 0x000fffff)|0x20000000));
307 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
311 /* calling is from identify_secondary_cpu() ? */
316 * Certain Athlons might work (for various values of 'work') in SMP
317 * but they are not certified as MP capable.
319 /* Athlon 660/661 is valid. */
320 if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
321 (c->x86_stepping == 1)))
324 /* Duron 670 is valid */
325 if ((c->x86_model == 7) && (c->x86_stepping == 0))
329 * Athlon 662, Duron 671, and Athlon >model 7 have capability
330 * bit. It's worth noting that the A5 stepping (662) of some
331 * Athlon XP's have the MP bit set.
332 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
335 if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
336 ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
338 if (cpu_has(c, X86_FEATURE_MP))
341 /* If we get here, not a certified SMP capable AMD system. */
344 * Don't taint if we are running SMP kernel on a single non-MP
347 WARN_ONCE(1, "WARNING: This combination of AMD"
348 " processors is not suitable for SMP.\n");
349 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
355 * To workaround broken NUMA config. Read the comment in
356 * srat_detect_node().
358 static int nearby_node(int apicid)
362 for (i = apicid - 1; i >= 0; i--) {
363 node = __apicid_to_node[i];
364 if (node != NUMA_NO_NODE && node_online(node))
367 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
368 node = __apicid_to_node[i];
369 if (node != NUMA_NO_NODE && node_online(node))
372 return first_node(node_online_map); /* Shouldn't happen */
377 * Fix up cpu_core_id for pre-F17h systems to be in the
378 * [0 .. cores_per_node - 1] range. Not really needed but
379 * kept so as not to break existing setups.
381 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
388 cus_per_node = c->x86_max_cores / nodes_per_socket;
389 c->cpu_core_id %= cus_per_node;
393 * Fixup core topology information for
394 * (1) AMD multi-node processors
395 * Assumption: Number of cores in each internal node is the same.
396 * (2) AMD processors supporting compute units
398 static void amd_get_topology(struct cpuinfo_x86 *c)
400 int cpu = smp_processor_id();
402 /* get information required for multi-node processors */
403 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
405 u32 eax, ebx, ecx, edx;
407 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
409 c->cpu_die_id = ecx & 0xff;
412 c->cu_id = ebx & 0xff;
414 if (c->x86 >= 0x17) {
415 c->cpu_core_id = ebx & 0xff;
417 if (smp_num_siblings > 1)
418 c->x86_max_cores /= smp_num_siblings;
422 * In case leaf B is available, use it to derive
423 * topology information.
425 err = detect_extended_topology(c);
427 c->x86_coreid_bits = get_count_order(c->x86_max_cores);
429 cacheinfo_amd_init_llc_id(c, cpu);
431 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
434 rdmsrl(MSR_FAM10H_NODE_ID, value);
435 c->cpu_die_id = value & 7;
437 per_cpu(cpu_llc_id, cpu) = c->cpu_die_id;
441 if (nodes_per_socket > 1) {
442 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
443 legacy_fixup_core_id(c);
448 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
449 * Assumes number of cores is a power of two.
451 static void amd_detect_cmp(struct cpuinfo_x86 *c)
454 int cpu = smp_processor_id();
456 bits = c->x86_coreid_bits;
457 /* Low order bits define the core id (index of core in socket) */
458 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
459 /* Convert the initial APIC ID into the socket ID */
460 c->phys_proc_id = c->initial_apicid >> bits;
461 /* use socket ID also for last level cache */
462 per_cpu(cpu_llc_id, cpu) = c->cpu_die_id = c->phys_proc_id;
465 u32 amd_get_nodes_per_socket(void)
467 return nodes_per_socket;
469 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
471 static void srat_detect_node(struct cpuinfo_x86 *c)
474 int cpu = smp_processor_id();
476 unsigned apicid = c->apicid;
478 node = numa_cpu_node(cpu);
479 if (node == NUMA_NO_NODE)
480 node = get_llc_id(cpu);
483 * On multi-fabric platform (e.g. Numascale NumaChip) a
484 * platform-specific handler needs to be called to fixup some
487 if (x86_cpuinit.fixup_cpu_id)
488 x86_cpuinit.fixup_cpu_id(c, node);
490 if (!node_online(node)) {
492 * Two possibilities here:
494 * - The CPU is missing memory and no node was created. In
495 * that case try picking one from a nearby CPU.
497 * - The APIC IDs differ from the HyperTransport node IDs
498 * which the K8 northbridge parsing fills in. Assume
499 * they are all increased by a constant offset, but in
500 * the same order as the HT nodeids. If that doesn't
501 * result in a usable node fall back to the path for the
504 * This workaround operates directly on the mapping between
505 * APIC ID and NUMA node, assuming certain relationship
506 * between APIC ID, HT node ID and NUMA topology. As going
507 * through CPU mapping may alter the outcome, directly
508 * access __apicid_to_node[].
510 int ht_nodeid = c->initial_apicid;
512 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
513 node = __apicid_to_node[ht_nodeid];
514 /* Pick a nearby node */
515 if (!node_online(node))
516 node = nearby_node(apicid);
518 numa_set_node(cpu, node);
522 static void early_init_amd_mc(struct cpuinfo_x86 *c)
527 /* Multi core CPU? */
528 if (c->extended_cpuid_level < 0x80000008)
531 ecx = cpuid_ecx(0x80000008);
533 c->x86_max_cores = (ecx & 0xff) + 1;
535 /* CPU telling us the core id bits shift? */
536 bits = (ecx >> 12) & 0xF;
538 /* Otherwise recompute */
540 while ((1 << bits) < c->x86_max_cores)
544 c->x86_coreid_bits = bits;
548 static void bsp_init_amd(struct cpuinfo_x86 *c)
550 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
553 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
556 rdmsrl(MSR_K7_HWCR, val);
557 if (!(val & BIT(24)))
558 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
562 if (c->x86 == 0x15) {
563 unsigned long upperbit;
566 cpuid = cpuid_edx(0x80000005);
567 assoc = cpuid >> 16 & 0xff;
568 upperbit = ((cpuid >> 24) << 10) / assoc;
570 va_align.mask = (upperbit - 1) & PAGE_MASK;
571 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
573 /* A random value per boot for bit slice [12:upper_bit) */
574 va_align.bits = get_random_u32() & va_align.mask;
577 if (cpu_has(c, X86_FEATURE_MWAITX))
580 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
583 ecx = cpuid_ecx(0x8000001e);
584 __max_die_per_package = nodes_per_socket = ((ecx >> 8) & 7) + 1;
585 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
588 rdmsrl(MSR_FAM10H_NODE_ID, value);
589 __max_die_per_package = nodes_per_socket = ((value >> 3) & 7) + 1;
592 if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
593 !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
594 c->x86 >= 0x15 && c->x86 <= 0x17) {
598 case 0x15: bit = 54; break;
599 case 0x16: bit = 33; break;
600 case 0x17: bit = 10; break;
604 * Try to cache the base value so further operations can
605 * avoid RMW. If that faults, do not enable SSBD.
607 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
608 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
609 setup_force_cpu_cap(X86_FEATURE_SSBD);
610 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
614 resctrl_cpu_detect(c);
617 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
622 * BIOS support is required for SME and SEV.
623 * For SME: If BIOS has enabled SME then adjust x86_phys_bits by
624 * the SME physical address space reduction value.
625 * If BIOS has not enabled SME then don't advertise the
626 * SME feature (set in scattered.c).
627 * If the kernel has not enabled SME via any means then
628 * don't advertise the SME feature.
629 * For SEV: If BIOS has not enabled SEV then don't advertise the
630 * SEV and SEV_ES feature (set in scattered.c).
632 * In all cases, since support for SME and SEV requires long mode,
633 * don't advertise the feature under CONFIG_X86_32.
635 if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
636 /* Check if memory encryption is enabled */
637 rdmsrl(MSR_AMD64_SYSCFG, msr);
638 if (!(msr & MSR_AMD64_SYSCFG_MEM_ENCRYPT))
642 * Always adjust physical address bits. Even though this
643 * will be a value above 32-bits this is still done for
644 * CONFIG_X86_32 so that accurate values are reported.
646 c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
648 if (IS_ENABLED(CONFIG_X86_32))
652 setup_clear_cpu_cap(X86_FEATURE_SME);
654 rdmsrl(MSR_K7_HWCR, msr);
655 if (!(msr & MSR_K7_HWCR_SMMLOCK))
661 setup_clear_cpu_cap(X86_FEATURE_SME);
663 setup_clear_cpu_cap(X86_FEATURE_SEV);
664 setup_clear_cpu_cap(X86_FEATURE_SEV_ES);
668 static void early_init_amd(struct cpuinfo_x86 *c)
673 early_init_amd_mc(c);
676 set_cpu_cap(c, X86_FEATURE_K8);
678 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
681 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
682 * with P/T states and does not stop in deep C-states
684 if (c->x86_power & (1 << 8)) {
685 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
686 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
689 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
690 if (c->x86_power & BIT(12))
691 set_cpu_cap(c, X86_FEATURE_ACC_POWER);
693 /* Bit 14 indicates the Runtime Average Power Limit interface. */
694 if (c->x86_power & BIT(14))
695 set_cpu_cap(c, X86_FEATURE_RAPL);
698 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
700 /* Set MTRR capability flag if appropriate */
702 if (c->x86_model == 13 || c->x86_model == 9 ||
703 (c->x86_model == 8 && c->x86_stepping >= 8))
704 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
706 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
708 * ApicID can always be treated as an 8-bit value for AMD APIC versions
709 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
710 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
713 if (boot_cpu_has(X86_FEATURE_APIC)) {
715 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
716 else if (c->x86 >= 0xf) {
717 /* check CPU config space for extended APIC ID */
720 val = read_pci_config(0, 24, 0, 0x68);
721 if ((val >> 17 & 0x3) == 0x3)
722 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
728 * This is only needed to tell the kernel whether to use VMCALL
729 * and VMMCALL. VMMCALL is never executed except under virt, so
730 * we can set it unconditionally.
732 set_cpu_cap(c, X86_FEATURE_VMMCALL);
734 /* F16h erratum 793, CVE-2013-6885 */
735 if (c->x86 == 0x16 && c->x86_model <= 0xf)
736 msr_set_bit(MSR_AMD64_LS_CFG, 15);
739 * Check whether the machine is affected by erratum 400. This is
740 * used to select the proper idle routine and to enable the check
741 * whether the machine is affected in arch_post_acpi_init(), which
742 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
744 if (cpu_has_amd_erratum(c, amd_erratum_400))
745 set_cpu_bug(c, X86_BUG_AMD_E400);
747 early_detect_mem_encrypt(c);
749 /* Re-enable TopologyExtensions if switched off by BIOS */
750 if (c->x86 == 0x15 &&
751 (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
752 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
754 if (msr_set_bit(0xc0011005, 54) > 0) {
755 rdmsrl(0xc0011005, value);
756 if (value & BIT_64(54)) {
757 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
758 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
763 if (cpu_has(c, X86_FEATURE_TOPOEXT))
764 smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
767 static void init_amd_k8(struct cpuinfo_x86 *c)
772 /* On C+ stepping K8 rep microcode works well for copy/memset */
773 level = cpuid_eax(1);
774 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
775 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
778 * Some BIOSes incorrectly force this feature, but only K8 revision D
779 * (model = 0x14) and later actually support it.
780 * (AMD Erratum #110, docId: 25759).
782 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
783 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
784 if (!rdmsrl_amd_safe(0xc001100d, &value)) {
785 value &= ~BIT_64(32);
786 wrmsrl_amd_safe(0xc001100d, value);
790 if (!c->x86_model_id[0])
791 strcpy(c->x86_model_id, "Hammer");
795 * Disable TLB flush filter by setting HWCR.FFDIS on K8
796 * bit 6 of msr C001_0015
798 * Errata 63 for SH-B3 steppings
799 * Errata 122 for all steppings (F+ have it disabled by default)
801 msr_set_bit(MSR_K7_HWCR, 6);
803 set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
806 static void init_amd_gh(struct cpuinfo_x86 *c)
808 #ifdef CONFIG_MMCONF_FAM10H
809 /* do this for boot cpu */
810 if (c == &boot_cpu_data)
811 check_enable_amd_mmconf_dmi();
813 fam10h_check_enable_mmcfg();
817 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
818 * is always needed when GART is enabled, even in a kernel which has no
819 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
820 * If it doesn't, we do it here as suggested by the BKDG.
822 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
824 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
827 * On family 10h BIOS may not have properly enabled WC+ support, causing
828 * it to be converted to CD memtype. This may result in performance
829 * degradation for certain nested-paging guests. Prevent this conversion
830 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
832 * NOTE: we want to use the _safe accessors so as not to #GP kvm
833 * guests on older kvm hosts.
835 msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
837 if (cpu_has_amd_erratum(c, amd_erratum_383))
838 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
841 static void init_amd_ln(struct cpuinfo_x86 *c)
844 * Apply erratum 665 fix unconditionally so machines without a BIOS
847 msr_set_bit(MSR_AMD64_DE_CFG, 31);
850 static bool rdrand_force;
852 static int __init rdrand_cmdline(char *str)
857 if (!strcmp(str, "force"))
864 early_param("rdrand", rdrand_cmdline);
866 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
869 * Saving of the MSR used to hide the RDRAND support during
870 * suspend/resume is done by arch/x86/power/cpu.c, which is
871 * dependent on CONFIG_PM_SLEEP.
873 if (!IS_ENABLED(CONFIG_PM_SLEEP))
877 * The self-test can clear X86_FEATURE_RDRAND, so check for
878 * RDRAND support using the CPUID function directly.
880 if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
883 msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
886 * Verify that the CPUID change has occurred in case the kernel is
887 * running virtualized and the hypervisor doesn't support the MSR.
889 if (cpuid_ecx(1) & BIT(30)) {
890 pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
894 clear_cpu_cap(c, X86_FEATURE_RDRAND);
895 pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
898 static void init_amd_jg(struct cpuinfo_x86 *c)
901 * Some BIOS implementations do not restore proper RDRAND support
902 * across suspend and resume. Check on whether to hide the RDRAND
903 * instruction support via CPUID.
905 clear_rdrand_cpuid_bit(c);
908 static void init_amd_bd(struct cpuinfo_x86 *c)
913 * The way access filter has a performance penalty on some workloads.
914 * Disable it on the affected CPUs.
916 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
917 if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
919 wrmsrl_safe(MSR_F15H_IC_CFG, value);
924 * Some BIOS implementations do not restore proper RDRAND support
925 * across suspend and resume. Check on whether to hide the RDRAND
926 * instruction support via CPUID.
928 clear_rdrand_cpuid_bit(c);
931 void init_spectral_chicken(struct cpuinfo_x86 *c)
933 #ifdef CONFIG_CPU_UNRET_ENTRY
937 * On Zen2 we offer this chicken (bit) on the altar of Speculation.
939 * This suppresses speculation from the middle of a basic block, i.e. it
940 * suppresses non-branch predictions.
942 * We use STIBP as a heuristic to filter out Zen2 from the rest of F17H
944 if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && cpu_has(c, X86_FEATURE_AMD_STIBP)) {
945 if (!rdmsrl_safe(MSR_ZEN2_SPECTRAL_CHICKEN, &value)) {
946 value |= MSR_ZEN2_SPECTRAL_CHICKEN_BIT;
947 wrmsrl_safe(MSR_ZEN2_SPECTRAL_CHICKEN, value);
952 * Work around Erratum 1386. The XSAVES instruction malfunctions in
953 * certain circumstances on Zen1/2 uarch, and not all parts have had
954 * updated microcode at the time of writing (March 2023).
956 * Affected parts all have no supervisor XSAVE states, meaning that
957 * the XSAVEC instruction (which works fine) is equivalent.
959 clear_cpu_cap(c, X86_FEATURE_XSAVES);
962 static void init_amd_zn(struct cpuinfo_x86 *c)
964 set_cpu_cap(c, X86_FEATURE_ZEN);
967 node_reclaim_distance = 32;
970 /* Fix up CPUID bits, but only if not virtualised. */
971 if (!cpu_has(c, X86_FEATURE_HYPERVISOR)) {
973 /* Erratum 1076: CPB feature bit not being set in CPUID. */
974 if (!cpu_has(c, X86_FEATURE_CPB))
975 set_cpu_cap(c, X86_FEATURE_CPB);
978 * Zen3 (Fam19 model < 0x10) parts are not susceptible to
979 * Branch Type Confusion, but predate the allocation of the
982 if (c->x86 == 0x19 && !cpu_has(c, X86_FEATURE_BTC_NO))
983 set_cpu_cap(c, X86_FEATURE_BTC_NO);
987 static bool cpu_has_zenbleed_microcode(void)
991 switch (boot_cpu_data.x86_model) {
992 case 0x30 ... 0x3f: good_rev = 0x0830107a; break;
993 case 0x60 ... 0x67: good_rev = 0x0860010b; break;
994 case 0x68 ... 0x6f: good_rev = 0x08608105; break;
995 case 0x70 ... 0x7f: good_rev = 0x08701032; break;
996 case 0xa0 ... 0xaf: good_rev = 0x08a00008; break;
1003 if (boot_cpu_data.microcode < good_rev)
1009 static void zenbleed_check(struct cpuinfo_x86 *c)
1011 if (!cpu_has_amd_erratum(c, amd_zenbleed))
1014 if (cpu_has(c, X86_FEATURE_HYPERVISOR))
1017 if (!cpu_has(c, X86_FEATURE_AVX))
1020 if (!cpu_has_zenbleed_microcode()) {
1021 pr_notice_once("Zenbleed: please update your microcode for the most optimal fix\n");
1022 msr_set_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
1024 msr_clear_bit(MSR_AMD64_DE_CFG, MSR_AMD64_DE_CFG_ZEN2_FP_BACKUP_FIX_BIT);
1028 static void init_amd(struct cpuinfo_x86 *c)
1033 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
1034 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
1036 clear_cpu_cap(c, 0*32+31);
1039 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
1041 /* AMD FSRM also implies FSRS */
1042 if (cpu_has(c, X86_FEATURE_FSRM))
1043 set_cpu_cap(c, X86_FEATURE_FSRS);
1045 /* get apicid instead of initial apic id from cpuid */
1046 c->apicid = hard_smp_processor_id();
1048 /* K6s reports MCEs but don't actually have all the MSRs */
1050 clear_cpu_cap(c, X86_FEATURE_MCE);
1053 case 4: init_amd_k5(c); break;
1054 case 5: init_amd_k6(c); break;
1055 case 6: init_amd_k7(c); break;
1056 case 0xf: init_amd_k8(c); break;
1057 case 0x10: init_amd_gh(c); break;
1058 case 0x12: init_amd_ln(c); break;
1059 case 0x15: init_amd_bd(c); break;
1060 case 0x16: init_amd_jg(c); break;
1061 case 0x17: init_spectral_chicken(c);
1063 case 0x19: init_amd_zn(c); break;
1067 * Enable workaround for FXSAVE leak on CPUs
1068 * without a XSaveErPtr feature
1070 if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
1071 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
1073 cpu_detect_cache_sizes(c);
1076 amd_get_topology(c);
1077 srat_detect_node(c);
1079 init_amd_cacheinfo(c);
1081 if (!cpu_has(c, X86_FEATURE_LFENCE_RDTSC) && cpu_has(c, X86_FEATURE_XMM2)) {
1083 * Use LFENCE for execution serialization. On families which
1084 * don't have that MSR, LFENCE is already serializing.
1085 * msr_set_bit() uses the safe accessors, too, even if the MSR
1088 msr_set_bit(MSR_AMD64_DE_CFG,
1089 MSR_AMD64_DE_CFG_LFENCE_SERIALIZE_BIT);
1091 /* A serializing LFENCE stops RDTSC speculation */
1092 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
1096 * Family 0x12 and above processors have APIC timer
1097 * running in deep C states.
1100 set_cpu_cap(c, X86_FEATURE_ARAT);
1102 /* 3DNow or LM implies PREFETCHW */
1103 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
1104 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
1105 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
1107 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
1108 if (!cpu_feature_enabled(X86_FEATURE_XENPV))
1109 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
1112 * Turn on the Instructions Retired free counter on machines not
1113 * susceptible to erratum #1054 "Instructions Retired Performance
1114 * Counter May Be Inaccurate".
1116 if (cpu_has(c, X86_FEATURE_IRPERF) &&
1117 !cpu_has_amd_erratum(c, amd_erratum_1054))
1118 msr_set_bit(MSR_K7_HWCR, MSR_K7_HWCR_IRPERF_EN_BIT);
1120 check_null_seg_clears_base(c);
1123 * Make sure EFER[AIBRSE - Automatic IBRS Enable] is set. The APs are brought up
1124 * using the trampoline code and as part of it, MSR_EFER gets prepared there in
1125 * order to be replicated onto them. Regardless, set it here again, if not set,
1126 * to protect against any future refactoring/code reorganization which might
1127 * miss setting this important bit.
1129 if (spectre_v2_in_eibrs_mode(spectre_v2_enabled) &&
1130 cpu_has(c, X86_FEATURE_AUTOIBRS))
1131 WARN_ON_ONCE(msr_set_bit(MSR_EFER, _EFER_AUTOIBRS));
1136 #ifdef CONFIG_X86_32
1137 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
1139 /* AMD errata T13 (order #21922) */
1142 if (c->x86_model == 3 && c->x86_stepping == 0)
1144 /* Tbird rev A1/A2 */
1145 if (c->x86_model == 4 &&
1146 (c->x86_stepping == 0 || c->x86_stepping == 1))
1153 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
1155 u32 ebx, eax, ecx, edx;
1161 if (c->extended_cpuid_level < 0x80000006)
1164 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
1166 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
1167 tlb_lli_4k[ENTRIES] = ebx & mask;
1170 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
1171 * characteristics from the CPUID function 0x80000005 instead.
1173 if (c->x86 == 0xf) {
1174 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1178 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1179 if (!((eax >> 16) & mask))
1180 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
1182 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
1184 /* a 4M entry uses two 2M entries */
1185 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
1187 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
1188 if (!(eax & mask)) {
1190 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
1191 tlb_lli_2m[ENTRIES] = 1024;
1193 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
1194 tlb_lli_2m[ENTRIES] = eax & 0xff;
1197 tlb_lli_2m[ENTRIES] = eax & mask;
1199 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
1202 static const struct cpu_dev amd_cpu_dev = {
1204 .c_ident = { "AuthenticAMD" },
1205 #ifdef CONFIG_X86_32
1207 { .family = 4, .model_names =
1210 [7] = "486 DX/2-WB",
1212 [9] = "486 DX/4-WB",
1218 .legacy_cache_size = amd_size_cache,
1220 .c_early_init = early_init_amd,
1221 .c_detect_tlb = cpu_detect_tlb_amd,
1222 .c_bsp_init = bsp_init_amd,
1224 .c_x86_vendor = X86_VENDOR_AMD,
1227 cpu_dev_register(amd_cpu_dev);
1229 static DEFINE_PER_CPU_READ_MOSTLY(unsigned long[4], amd_dr_addr_mask);
1231 static unsigned int amd_msr_dr_addr_masks[] = {
1232 MSR_F16H_DR0_ADDR_MASK,
1233 MSR_F16H_DR1_ADDR_MASK,
1234 MSR_F16H_DR1_ADDR_MASK + 1,
1235 MSR_F16H_DR1_ADDR_MASK + 2
1238 void amd_set_dr_addr_mask(unsigned long mask, unsigned int dr)
1240 int cpu = smp_processor_id();
1242 if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1245 if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1248 if (per_cpu(amd_dr_addr_mask, cpu)[dr] == mask)
1251 wrmsr(amd_msr_dr_addr_masks[dr], mask, 0);
1252 per_cpu(amd_dr_addr_mask, cpu)[dr] = mask;
1255 unsigned long amd_get_dr_addr_mask(unsigned int dr)
1257 if (!cpu_feature_enabled(X86_FEATURE_BPEXT))
1260 if (WARN_ON_ONCE(dr >= ARRAY_SIZE(amd_msr_dr_addr_masks)))
1263 return per_cpu(amd_dr_addr_mask[dr], smp_processor_id());
1265 EXPORT_SYMBOL_GPL(amd_get_dr_addr_mask);
1267 u32 amd_get_highest_perf(void)
1269 struct cpuinfo_x86 *c = &boot_cpu_data;
1271 if (c->x86 == 0x17 && ((c->x86_model >= 0x30 && c->x86_model < 0x40) ||
1272 (c->x86_model >= 0x70 && c->x86_model < 0x80)))
1275 if (c->x86 == 0x19 && ((c->x86_model >= 0x20 && c->x86_model < 0x30) ||
1276 (c->x86_model >= 0x40 && c->x86_model < 0x70)))
1281 EXPORT_SYMBOL_GPL(amd_get_highest_perf);
1283 static void zenbleed_check_cpu(void *unused)
1285 struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
1290 void amd_check_microcode(void)
1292 on_each_cpu(zenbleed_check_cpu, NULL, 1);