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>
18 #include <asm/debugreg.h>
21 # include <asm/mmconfig.h>
22 # include <asm/set_memory.h>
27 static const int amd_erratum_383[];
28 static const int amd_erratum_400[];
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);
88 ".type vide, @function\n"
92 static void init_amd_k5(struct cpuinfo_x86 *c)
96 * General Systems BIOSen alias the cpu frequency registers
97 * of the Elan at 0x000df000. Unfortunately, one of the Linux
98 * drivers subsequently pokes it, and changes the CPU speed.
99 * Workaround : Remove the unneeded alias.
101 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
102 #define CBAR_ENB (0x80000000)
103 #define CBAR_KEY (0X000000CB)
104 if (c->x86_model == 9 || c->x86_model == 10) {
105 if (inl(CBAR) & CBAR_ENB)
106 outl(0 | CBAR_KEY, CBAR);
111 static void init_amd_k6(struct cpuinfo_x86 *c)
115 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
117 if (c->x86_model < 6) {
118 /* Based on AMD doc 20734R - June 2000 */
119 if (c->x86_model == 0) {
120 clear_cpu_cap(c, X86_FEATURE_APIC);
121 set_cpu_cap(c, X86_FEATURE_PGE);
126 if (c->x86_model == 6 && c->x86_stepping == 1) {
127 const int K6_BUG_LOOP = 1000000;
129 void (*f_vide)(void);
132 pr_info("AMD K6 stepping B detected - ");
135 * It looks like AMD fixed the 2.6.2 bug and improved indirect
136 * calls at the same time.
141 OPTIMIZER_HIDE_VAR(f_vide);
148 if (d > 20*K6_BUG_LOOP)
149 pr_cont("system stability may be impaired when more than 32 MB are used.\n");
151 pr_cont("probably OK (after B9730xxxx).\n");
154 /* K6 with old style WHCR */
155 if (c->x86_model < 8 ||
156 (c->x86_model == 8 && c->x86_stepping < 8)) {
157 /* We can only write allocate on the low 508Mb */
161 rdmsr(MSR_K6_WHCR, l, h);
162 if ((l&0x0000FFFF) == 0) {
164 l = (1<<0)|((mbytes/4)<<1);
165 local_irq_save(flags);
167 wrmsr(MSR_K6_WHCR, l, h);
168 local_irq_restore(flags);
169 pr_info("Enabling old style K6 write allocation for %d Mb\n",
175 if ((c->x86_model == 8 && c->x86_stepping > 7) ||
176 c->x86_model == 9 || c->x86_model == 13) {
177 /* The more serious chips .. */
182 rdmsr(MSR_K6_WHCR, l, h);
183 if ((l&0xFFFF0000) == 0) {
185 l = ((mbytes>>2)<<22)|(1<<16);
186 local_irq_save(flags);
188 wrmsr(MSR_K6_WHCR, l, h);
189 local_irq_restore(flags);
190 pr_info("Enabling new style K6 write allocation for %d Mb\n",
197 if (c->x86_model == 10) {
198 /* AMD Geode LX is model 10 */
199 /* placeholder for any needed mods */
205 static void init_amd_k7(struct cpuinfo_x86 *c)
211 * Bit 15 of Athlon specific MSR 15, needs to be 0
212 * to enable SSE on Palomino/Morgan/Barton CPU's.
213 * If the BIOS didn't enable it already, enable it here.
215 if (c->x86_model >= 6 && c->x86_model <= 10) {
216 if (!cpu_has(c, X86_FEATURE_XMM)) {
217 pr_info("Enabling disabled K7/SSE Support.\n");
218 msr_clear_bit(MSR_K7_HWCR, 15);
219 set_cpu_cap(c, X86_FEATURE_XMM);
224 * It's been determined by AMD that Athlons since model 8 stepping 1
225 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
226 * As per AMD technical note 27212 0.2
228 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
229 rdmsr(MSR_K7_CLK_CTL, l, h);
230 if ((l & 0xfff00000) != 0x20000000) {
231 pr_info("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
232 l, ((l & 0x000fffff)|0x20000000));
233 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
237 /* calling is from identify_secondary_cpu() ? */
242 * Certain Athlons might work (for various values of 'work') in SMP
243 * but they are not certified as MP capable.
245 /* Athlon 660/661 is valid. */
246 if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
247 (c->x86_stepping == 1)))
250 /* Duron 670 is valid */
251 if ((c->x86_model == 7) && (c->x86_stepping == 0))
255 * Athlon 662, Duron 671, and Athlon >model 7 have capability
256 * bit. It's worth noting that the A5 stepping (662) of some
257 * Athlon XP's have the MP bit set.
258 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
261 if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
262 ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
264 if (cpu_has(c, X86_FEATURE_MP))
267 /* If we get here, not a certified SMP capable AMD system. */
270 * Don't taint if we are running SMP kernel on a single non-MP
273 WARN_ONCE(1, "WARNING: This combination of AMD"
274 " processors is not suitable for SMP.\n");
275 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
281 * To workaround broken NUMA config. Read the comment in
282 * srat_detect_node().
284 static int nearby_node(int apicid)
288 for (i = apicid - 1; i >= 0; i--) {
289 node = __apicid_to_node[i];
290 if (node != NUMA_NO_NODE && node_online(node))
293 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
294 node = __apicid_to_node[i];
295 if (node != NUMA_NO_NODE && node_online(node))
298 return first_node(node_online_map); /* Shouldn't happen */
303 * Fix up cpu_core_id for pre-F17h systems to be in the
304 * [0 .. cores_per_node - 1] range. Not really needed but
305 * kept so as not to break existing setups.
307 static void legacy_fixup_core_id(struct cpuinfo_x86 *c)
314 cus_per_node = c->x86_max_cores / nodes_per_socket;
315 c->cpu_core_id %= cus_per_node;
319 static void amd_get_topology_early(struct cpuinfo_x86 *c)
321 if (cpu_has(c, X86_FEATURE_TOPOEXT))
322 smp_num_siblings = ((cpuid_ebx(0x8000001e) >> 8) & 0xff) + 1;
326 * Fixup core topology information for
327 * (1) AMD multi-node processors
328 * Assumption: Number of cores in each internal node is the same.
329 * (2) AMD processors supporting compute units
331 static void amd_get_topology(struct cpuinfo_x86 *c)
334 int cpu = smp_processor_id();
336 /* get information required for multi-node processors */
337 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
339 u32 eax, ebx, ecx, edx;
341 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
343 node_id = ecx & 0xff;
346 c->cu_id = ebx & 0xff;
348 if (c->x86 >= 0x17) {
349 c->cpu_core_id = ebx & 0xff;
351 if (smp_num_siblings > 1)
352 c->x86_max_cores /= smp_num_siblings;
356 * In case leaf B is available, use it to derive
357 * topology information.
359 err = detect_extended_topology(c);
361 c->x86_coreid_bits = get_count_order(c->x86_max_cores);
363 cacheinfo_amd_init_llc_id(c, cpu, node_id);
365 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
368 rdmsrl(MSR_FAM10H_NODE_ID, value);
371 per_cpu(cpu_llc_id, cpu) = node_id;
375 if (nodes_per_socket > 1) {
376 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
377 legacy_fixup_core_id(c);
382 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
383 * Assumes number of cores is a power of two.
385 static void amd_detect_cmp(struct cpuinfo_x86 *c)
388 int cpu = smp_processor_id();
390 bits = c->x86_coreid_bits;
391 /* Low order bits define the core id (index of core in socket) */
392 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
393 /* Convert the initial APIC ID into the socket ID */
394 c->phys_proc_id = c->initial_apicid >> bits;
395 /* use socket ID also for last level cache */
396 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
399 u16 amd_get_nb_id(int cpu)
401 return per_cpu(cpu_llc_id, cpu);
403 EXPORT_SYMBOL_GPL(amd_get_nb_id);
405 u32 amd_get_nodes_per_socket(void)
407 return nodes_per_socket;
409 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
411 static void srat_detect_node(struct cpuinfo_x86 *c)
414 int cpu = smp_processor_id();
416 unsigned apicid = c->apicid;
418 node = numa_cpu_node(cpu);
419 if (node == NUMA_NO_NODE)
420 node = per_cpu(cpu_llc_id, cpu);
423 * On multi-fabric platform (e.g. Numascale NumaChip) a
424 * platform-specific handler needs to be called to fixup some
427 if (x86_cpuinit.fixup_cpu_id)
428 x86_cpuinit.fixup_cpu_id(c, node);
430 if (!node_online(node)) {
432 * Two possibilities here:
434 * - The CPU is missing memory and no node was created. In
435 * that case try picking one from a nearby CPU.
437 * - The APIC IDs differ from the HyperTransport node IDs
438 * which the K8 northbridge parsing fills in. Assume
439 * they are all increased by a constant offset, but in
440 * the same order as the HT nodeids. If that doesn't
441 * result in a usable node fall back to the path for the
444 * This workaround operates directly on the mapping between
445 * APIC ID and NUMA node, assuming certain relationship
446 * between APIC ID, HT node ID and NUMA topology. As going
447 * through CPU mapping may alter the outcome, directly
448 * access __apicid_to_node[].
450 int ht_nodeid = c->initial_apicid;
452 if (__apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
453 node = __apicid_to_node[ht_nodeid];
454 /* Pick a nearby node */
455 if (!node_online(node))
456 node = nearby_node(apicid);
458 numa_set_node(cpu, node);
462 static void early_init_amd_mc(struct cpuinfo_x86 *c)
467 /* Multi core CPU? */
468 if (c->extended_cpuid_level < 0x80000008)
471 ecx = cpuid_ecx(0x80000008);
473 c->x86_max_cores = (ecx & 0xff) + 1;
475 /* CPU telling us the core id bits shift? */
476 bits = (ecx >> 12) & 0xF;
478 /* Otherwise recompute */
480 while ((1 << bits) < c->x86_max_cores)
484 c->x86_coreid_bits = bits;
488 static void bsp_init_amd(struct cpuinfo_x86 *c)
493 unsigned long long tseg;
496 * Split up direct mapping around the TSEG SMM area.
497 * Don't do it for gbpages because there seems very little
498 * benefit in doing so.
500 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
501 unsigned long pfn = tseg >> PAGE_SHIFT;
503 pr_debug("tseg: %010llx\n", tseg);
504 if (pfn_range_is_mapped(pfn, pfn + 1))
505 set_memory_4k((unsigned long)__va(tseg), 1);
510 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
513 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
516 rdmsrl(MSR_K7_HWCR, val);
517 if (!(val & BIT(24)))
518 pr_warn(FW_BUG "TSC doesn't count with P0 frequency!\n");
522 if (c->x86 == 0x15) {
523 unsigned long upperbit;
526 cpuid = cpuid_edx(0x80000005);
527 assoc = cpuid >> 16 & 0xff;
528 upperbit = ((cpuid >> 24) << 10) / assoc;
530 va_align.mask = (upperbit - 1) & PAGE_MASK;
531 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
533 /* A random value per boot for bit slice [12:upper_bit) */
534 va_align.bits = get_random_int() & va_align.mask;
537 if (cpu_has(c, X86_FEATURE_MWAITX))
540 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
543 ecx = cpuid_ecx(0x8000001e);
544 nodes_per_socket = ((ecx >> 8) & 7) + 1;
545 } else if (boot_cpu_has(X86_FEATURE_NODEID_MSR)) {
548 rdmsrl(MSR_FAM10H_NODE_ID, value);
549 nodes_per_socket = ((value >> 3) & 7) + 1;
552 if (!boot_cpu_has(X86_FEATURE_AMD_SSBD) &&
553 !boot_cpu_has(X86_FEATURE_VIRT_SSBD) &&
554 c->x86 >= 0x15 && c->x86 <= 0x17) {
558 case 0x15: bit = 54; break;
559 case 0x16: bit = 33; break;
560 case 0x17: bit = 10; break;
564 * Try to cache the base value so further operations can
565 * avoid RMW. If that faults, do not enable SSBD.
567 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
568 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
569 setup_force_cpu_cap(X86_FEATURE_SSBD);
570 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
575 static void early_detect_mem_encrypt(struct cpuinfo_x86 *c)
580 * BIOS support is required for SME and SEV.
581 * For SME: If BIOS has enabled SME then adjust x86_phys_bits by
582 * the SME physical address space reduction value.
583 * If BIOS has not enabled SME then don't advertise the
584 * SME feature (set in scattered.c).
585 * For SEV: If BIOS has not enabled SEV then don't advertise the
586 * SEV feature (set in scattered.c).
588 * In all cases, since support for SME and SEV requires long mode,
589 * don't advertise the feature under CONFIG_X86_32.
591 if (cpu_has(c, X86_FEATURE_SME) || cpu_has(c, X86_FEATURE_SEV)) {
592 /* Check if memory encryption is enabled */
593 rdmsrl(MSR_K8_SYSCFG, msr);
594 if (!(msr & MSR_K8_SYSCFG_MEM_ENCRYPT))
598 * Always adjust physical address bits. Even though this
599 * will be a value above 32-bits this is still done for
600 * CONFIG_X86_32 so that accurate values are reported.
602 c->x86_phys_bits -= (cpuid_ebx(0x8000001f) >> 6) & 0x3f;
604 if (IS_ENABLED(CONFIG_X86_32))
607 rdmsrl(MSR_K7_HWCR, msr);
608 if (!(msr & MSR_K7_HWCR_SMMLOCK))
614 clear_cpu_cap(c, X86_FEATURE_SME);
616 clear_cpu_cap(c, X86_FEATURE_SEV);
620 static void early_init_amd(struct cpuinfo_x86 *c)
625 early_init_amd_mc(c);
629 set_cpu_cap(c, X86_FEATURE_K7);
633 set_cpu_cap(c, X86_FEATURE_K8);
635 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
638 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
639 * with P/T states and does not stop in deep C-states
641 if (c->x86_power & (1 << 8)) {
642 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
643 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
646 /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
647 if (c->x86_power & BIT(12))
648 set_cpu_cap(c, X86_FEATURE_ACC_POWER);
651 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
653 /* Set MTRR capability flag if appropriate */
655 if (c->x86_model == 13 || c->x86_model == 9 ||
656 (c->x86_model == 8 && c->x86_stepping >= 8))
657 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
659 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
661 * ApicID can always be treated as an 8-bit value for AMD APIC versions
662 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
663 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
666 if (boot_cpu_has(X86_FEATURE_APIC)) {
668 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
669 else if (c->x86 >= 0xf) {
670 /* check CPU config space for extended APIC ID */
673 val = read_pci_config(0, 24, 0, 0x68);
674 if ((val >> 17 & 0x3) == 0x3)
675 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
681 * This is only needed to tell the kernel whether to use VMCALL
682 * and VMMCALL. VMMCALL is never executed except under virt, so
683 * we can set it unconditionally.
685 set_cpu_cap(c, X86_FEATURE_VMMCALL);
687 /* F16h erratum 793, CVE-2013-6885 */
688 if (c->x86 == 0x16 && c->x86_model <= 0xf)
689 msr_set_bit(MSR_AMD64_LS_CFG, 15);
692 * Check whether the machine is affected by erratum 400. This is
693 * used to select the proper idle routine and to enable the check
694 * whether the machine is affected in arch_post_acpi_init(), which
695 * sets the X86_BUG_AMD_APIC_C1E bug depending on the MSR check.
697 if (cpu_has_amd_erratum(c, amd_erratum_400))
698 set_cpu_bug(c, X86_BUG_AMD_E400);
700 early_detect_mem_encrypt(c);
702 /* Re-enable TopologyExtensions if switched off by BIOS */
703 if (c->x86 == 0x15 &&
704 (c->x86_model >= 0x10 && c->x86_model <= 0x6f) &&
705 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
707 if (msr_set_bit(0xc0011005, 54) > 0) {
708 rdmsrl(0xc0011005, value);
709 if (value & BIT_64(54)) {
710 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
711 pr_info_once(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
716 amd_get_topology_early(c);
719 static void init_amd_k8(struct cpuinfo_x86 *c)
724 /* On C+ stepping K8 rep microcode works well for copy/memset */
725 level = cpuid_eax(1);
726 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
727 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
730 * Some BIOSes incorrectly force this feature, but only K8 revision D
731 * (model = 0x14) and later actually support it.
732 * (AMD Erratum #110, docId: 25759).
734 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
735 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
736 if (!rdmsrl_amd_safe(0xc001100d, &value)) {
737 value &= ~BIT_64(32);
738 wrmsrl_amd_safe(0xc001100d, value);
742 if (!c->x86_model_id[0])
743 strcpy(c->x86_model_id, "Hammer");
747 * Disable TLB flush filter by setting HWCR.FFDIS on K8
748 * bit 6 of msr C001_0015
750 * Errata 63 for SH-B3 steppings
751 * Errata 122 for all steppings (F+ have it disabled by default)
753 msr_set_bit(MSR_K7_HWCR, 6);
755 set_cpu_bug(c, X86_BUG_SWAPGS_FENCE);
758 static void init_amd_gh(struct cpuinfo_x86 *c)
760 #ifdef CONFIG_MMCONF_FAM10H
761 /* do this for boot cpu */
762 if (c == &boot_cpu_data)
763 check_enable_amd_mmconf_dmi();
765 fam10h_check_enable_mmcfg();
769 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
770 * is always needed when GART is enabled, even in a kernel which has no
771 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
772 * If it doesn't, we do it here as suggested by the BKDG.
774 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
776 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
779 * On family 10h BIOS may not have properly enabled WC+ support, causing
780 * it to be converted to CD memtype. This may result in performance
781 * degradation for certain nested-paging guests. Prevent this conversion
782 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
784 * NOTE: we want to use the _safe accessors so as not to #GP kvm
785 * guests on older kvm hosts.
787 msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
789 if (cpu_has_amd_erratum(c, amd_erratum_383))
790 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
793 #define MSR_AMD64_DE_CFG 0xC0011029
795 static void init_amd_ln(struct cpuinfo_x86 *c)
798 * Apply erratum 665 fix unconditionally so machines without a BIOS
801 msr_set_bit(MSR_AMD64_DE_CFG, 31);
804 static void init_amd_bd(struct cpuinfo_x86 *c)
809 * The way access filter has a performance penalty on some workloads.
810 * Disable it on the affected CPUs.
812 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
813 if (!rdmsrl_safe(MSR_F15H_IC_CFG, &value) && !(value & 0x1E)) {
815 wrmsrl_safe(MSR_F15H_IC_CFG, value);
820 static void init_amd_zn(struct cpuinfo_x86 *c)
822 set_cpu_cap(c, X86_FEATURE_ZEN);
824 /* Fix erratum 1076: CPB feature bit not being set in CPUID. */
825 if (!cpu_has(c, X86_FEATURE_CPB))
826 set_cpu_cap(c, X86_FEATURE_CPB);
829 static void init_amd(struct cpuinfo_x86 *c)
834 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
835 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
837 clear_cpu_cap(c, 0*32+31);
840 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
842 /* get apicid instead of initial apic id from cpuid */
843 c->apicid = hard_smp_processor_id();
845 /* K6s reports MCEs but don't actually have all the MSRs */
847 clear_cpu_cap(c, X86_FEATURE_MCE);
850 case 4: init_amd_k5(c); break;
851 case 5: init_amd_k6(c); break;
852 case 6: init_amd_k7(c); break;
853 case 0xf: init_amd_k8(c); break;
854 case 0x10: init_amd_gh(c); break;
855 case 0x12: init_amd_ln(c); break;
856 case 0x15: init_amd_bd(c); break;
857 case 0x17: init_amd_zn(c); break;
861 * Enable workaround for FXSAVE leak on CPUs
862 * without a XSaveErPtr feature
864 if ((c->x86 >= 6) && (!cpu_has(c, X86_FEATURE_XSAVEERPTR)))
865 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
867 cpu_detect_cache_sizes(c);
873 init_amd_cacheinfo(c);
875 if (cpu_has(c, X86_FEATURE_XMM2)) {
876 unsigned long long val;
880 * A serializing LFENCE has less overhead than MFENCE, so
881 * use it for execution serialization. On families which
882 * don't have that MSR, LFENCE is already serializing.
883 * msr_set_bit() uses the safe accessors, too, even if the MSR
886 msr_set_bit(MSR_F10H_DECFG,
887 MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT);
890 * Verify that the MSR write was successful (could be running
891 * under a hypervisor) and only then assume that LFENCE is
894 ret = rdmsrl_safe(MSR_F10H_DECFG, &val);
895 if (!ret && (val & MSR_F10H_DECFG_LFENCE_SERIALIZE)) {
896 /* A serializing LFENCE stops RDTSC speculation */
897 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
899 /* MFENCE stops RDTSC speculation */
900 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
905 * Family 0x12 and above processors have APIC timer
906 * running in deep C states.
909 set_cpu_cap(c, X86_FEATURE_ARAT);
911 /* 3DNow or LM implies PREFETCHW */
912 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
913 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
914 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
916 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
917 if (!cpu_has(c, X86_FEATURE_XENPV))
918 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
922 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
924 /* AMD errata T13 (order #21922) */
927 if (c->x86_model == 3 && c->x86_stepping == 0)
929 /* Tbird rev A1/A2 */
930 if (c->x86_model == 4 &&
931 (c->x86_stepping == 0 || c->x86_stepping == 1))
938 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
940 u32 ebx, eax, ecx, edx;
946 if (c->extended_cpuid_level < 0x80000006)
949 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
951 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
952 tlb_lli_4k[ENTRIES] = ebx & mask;
955 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
956 * characteristics from the CPUID function 0x80000005 instead.
959 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
963 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
964 if (!((eax >> 16) & mask))
965 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
967 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
969 /* a 4M entry uses two 2M entries */
970 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
972 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
975 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
976 tlb_lli_2m[ENTRIES] = 1024;
978 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
979 tlb_lli_2m[ENTRIES] = eax & 0xff;
982 tlb_lli_2m[ENTRIES] = eax & mask;
984 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
987 static const struct cpu_dev amd_cpu_dev = {
989 .c_ident = { "AuthenticAMD" },
992 { .family = 4, .model_names =
1003 .legacy_cache_size = amd_size_cache,
1005 .c_early_init = early_init_amd,
1006 .c_detect_tlb = cpu_detect_tlb_amd,
1007 .c_bsp_init = bsp_init_amd,
1009 .c_x86_vendor = X86_VENDOR_AMD,
1012 cpu_dev_register(amd_cpu_dev);
1015 * AMD errata checking
1017 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
1018 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
1019 * have an OSVW id assigned, which it takes as first argument. Both take a
1020 * variable number of family-specific model-stepping ranges created by
1021 * AMD_MODEL_RANGE().
1025 * const int amd_erratum_319[] =
1026 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1027 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1028 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1031 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
1032 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
1033 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1034 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1035 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
1036 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
1037 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
1039 static const int amd_erratum_400[] =
1040 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1041 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1043 static const int amd_erratum_383[] =
1044 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1047 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1049 int osvw_id = *erratum++;
1053 if (osvw_id >= 0 && osvw_id < 65536 &&
1054 cpu_has(cpu, X86_FEATURE_OSVW)) {
1057 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1058 if (osvw_id < osvw_len) {
1061 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1063 return osvw_bits & (1ULL << (osvw_id & 0x3f));
1067 /* OSVW unavailable or ID unknown, match family-model-stepping range */
1068 ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1069 while ((range = *erratum++))
1070 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1071 (ms >= AMD_MODEL_RANGE_START(range)) &&
1072 (ms <= AMD_MODEL_RANGE_END(range)))
1078 void set_dr_addr_mask(unsigned long mask, int dr)
1080 if (!boot_cpu_has(X86_FEATURE_BPEXT))
1085 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1090 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);