1 #include <linux/export.h>
2 #include <linux/bitops.h>
7 #include <linux/sched.h>
8 #include <linux/random.h>
9 #include <asm/processor.h>
12 #include <asm/spec-ctrl.h>
14 #include <asm/pci-direct.h>
15 #include <asm/delay.h>
18 # include <asm/mmconfig.h>
19 # include <asm/cacheflush.h>
25 * nodes_per_socket: Stores the number of nodes per socket.
26 * Refer to Fam15h Models 00-0fh BKDG - CPUID Fn8000_001E_ECX
27 * Node Identifiers[10:8]
29 static u32 nodes_per_socket = 1;
31 static inline int rdmsrl_amd_safe(unsigned msr, unsigned long long *p)
36 WARN_ONCE((boot_cpu_data.x86 != 0xf),
37 "%s should only be used on K8!\n", __func__);
42 err = rdmsr_safe_regs(gprs);
44 *p = gprs[0] | ((u64)gprs[2] << 32);
49 static inline int wrmsrl_amd_safe(unsigned msr, unsigned long long val)
53 WARN_ONCE((boot_cpu_data.x86 != 0xf),
54 "%s should only be used on K8!\n", __func__);
61 return wrmsr_safe_regs(gprs);
65 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
66 * misexecution of code under Linux. Owners of such processors should
67 * contact AMD for precise details and a CPU swap.
69 * See http://www.multimania.com/poulot/k6bug.html
70 * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
71 * (Publication # 21266 Issue Date: August 1998)
73 * The following test is erm.. interesting. AMD neglected to up
74 * the chip setting when fixing the bug but they also tweaked some
75 * performance at the same time..
78 extern __visible void vide(void);
79 __asm__(".globl vide\n\t.align 4\nvide: ret");
81 static void init_amd_k5(struct cpuinfo_x86 *c)
85 * General Systems BIOSen alias the cpu frequency registers
86 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
87 * drivers subsequently pokes it, and changes the CPU speed.
88 * Workaround : Remove the unneeded alias.
90 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
91 #define CBAR_ENB (0x80000000)
92 #define CBAR_KEY (0X000000CB)
93 if (c->x86_model == 9 || c->x86_model == 10) {
94 if (inl(CBAR) & CBAR_ENB)
95 outl(0 | CBAR_KEY, CBAR);
100 static void init_amd_k6(struct cpuinfo_x86 *c)
104 int mbytes = get_num_physpages() >> (20-PAGE_SHIFT);
106 if (c->x86_model < 6) {
107 /* Based on AMD doc 20734R - June 2000 */
108 if (c->x86_model == 0) {
109 clear_cpu_cap(c, X86_FEATURE_APIC);
110 set_cpu_cap(c, X86_FEATURE_PGE);
115 if (c->x86_model == 6 && c->x86_stepping == 1) {
116 const int K6_BUG_LOOP = 1000000;
118 void (*f_vide)(void);
121 printk(KERN_INFO "AMD K6 stepping B detected - ");
124 * It looks like AMD fixed the 2.6.2 bug and improved indirect
125 * calls at the same time.
136 if (d > 20*K6_BUG_LOOP)
138 "system stability may be impaired when more than 32 MB are used.\n");
140 printk(KERN_CONT "probably OK (after B9730xxxx).\n");
143 /* K6 with old style WHCR */
144 if (c->x86_model < 8 ||
145 (c->x86_model == 8 && c->x86_stepping < 8)) {
146 /* We can only write allocate on the low 508Mb */
150 rdmsr(MSR_K6_WHCR, l, h);
151 if ((l&0x0000FFFF) == 0) {
153 l = (1<<0)|((mbytes/4)<<1);
154 local_irq_save(flags);
156 wrmsr(MSR_K6_WHCR, l, h);
157 local_irq_restore(flags);
158 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
164 if ((c->x86_model == 8 && c->x86_stepping > 7) ||
165 c->x86_model == 9 || c->x86_model == 13) {
166 /* The more serious chips .. */
171 rdmsr(MSR_K6_WHCR, l, h);
172 if ((l&0xFFFF0000) == 0) {
174 l = ((mbytes>>2)<<22)|(1<<16);
175 local_irq_save(flags);
177 wrmsr(MSR_K6_WHCR, l, h);
178 local_irq_restore(flags);
179 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
186 if (c->x86_model == 10) {
187 /* AMD Geode LX is model 10 */
188 /* placeholder for any needed mods */
194 static void init_amd_k7(struct cpuinfo_x86 *c)
200 * Bit 15 of Athlon specific MSR 15, needs to be 0
201 * to enable SSE on Palomino/Morgan/Barton CPU's.
202 * If the BIOS didn't enable it already, enable it here.
204 if (c->x86_model >= 6 && c->x86_model <= 10) {
205 if (!cpu_has(c, X86_FEATURE_XMM)) {
206 printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
207 msr_clear_bit(MSR_K7_HWCR, 15);
208 set_cpu_cap(c, X86_FEATURE_XMM);
213 * It's been determined by AMD that Athlons since model 8 stepping 1
214 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
215 * As per AMD technical note 27212 0.2
217 if ((c->x86_model == 8 && c->x86_stepping >= 1) || (c->x86_model > 8)) {
218 rdmsr(MSR_K7_CLK_CTL, l, h);
219 if ((l & 0xfff00000) != 0x20000000) {
221 "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
222 l, ((l & 0x000fffff)|0x20000000));
223 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
227 set_cpu_cap(c, X86_FEATURE_K7);
229 /* calling is from identify_secondary_cpu() ? */
234 * Certain Athlons might work (for various values of 'work') in SMP
235 * but they are not certified as MP capable.
237 /* Athlon 660/661 is valid. */
238 if ((c->x86_model == 6) && ((c->x86_stepping == 0) ||
239 (c->x86_stepping == 1)))
242 /* Duron 670 is valid */
243 if ((c->x86_model == 7) && (c->x86_stepping == 0))
247 * Athlon 662, Duron 671, and Athlon >model 7 have capability
248 * bit. It's worth noting that the A5 stepping (662) of some
249 * Athlon XP's have the MP bit set.
250 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
253 if (((c->x86_model == 6) && (c->x86_stepping >= 2)) ||
254 ((c->x86_model == 7) && (c->x86_stepping >= 1)) ||
256 if (cpu_has(c, X86_FEATURE_MP))
259 /* If we get here, not a certified SMP capable AMD system. */
262 * Don't taint if we are running SMP kernel on a single non-MP
265 WARN_ONCE(1, "WARNING: This combination of AMD"
266 " processors is not suitable for SMP.\n");
267 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_NOW_UNRELIABLE);
273 * To workaround broken NUMA config. Read the comment in
274 * srat_detect_node().
276 static int nearby_node(int apicid)
280 for (i = apicid - 1; i >= 0; i--) {
281 node = __apicid_to_node[i];
282 if (node != NUMA_NO_NODE && node_online(node))
285 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
286 node = __apicid_to_node[i];
287 if (node != NUMA_NO_NODE && node_online(node))
290 return first_node(node_online_map); /* Shouldn't happen */
295 * Fixup core topology information for
296 * (1) AMD multi-node processors
297 * Assumption: Number of cores in each internal node is the same.
298 * (2) AMD processors supporting compute units
301 static void amd_get_topology(struct cpuinfo_x86 *c)
303 u32 cores_per_cu = 1;
305 int cpu = smp_processor_id();
307 /* get information required for multi-node processors */
308 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
309 u32 eax, ebx, ecx, edx;
311 cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
312 nodes_per_socket = ((ecx >> 8) & 7) + 1;
315 /* get compute unit information */
316 smp_num_siblings = ((ebx >> 8) & 3) + 1;
317 c->compute_unit_id = ebx & 0xff;
318 cores_per_cu += ((ebx >> 8) & 3);
319 } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
322 rdmsrl(MSR_FAM10H_NODE_ID, value);
323 nodes_per_socket = ((value >> 3) & 7) + 1;
328 /* fixup multi-node processor information */
329 if (nodes_per_socket > 1) {
333 set_cpu_cap(c, X86_FEATURE_AMD_DCM);
334 cores_per_node = c->x86_max_cores / nodes_per_socket;
335 cus_per_node = cores_per_node / cores_per_cu;
337 /* store NodeID, use llc_shared_map to store sibling info */
338 per_cpu(cpu_llc_id, cpu) = node_id;
340 /* core id has to be in the [0 .. cores_per_node - 1] range */
341 c->cpu_core_id %= cores_per_node;
342 c->compute_unit_id %= cus_per_node;
348 * On a AMD dual core setup the lower bits of the APIC id distinguish the cores.
349 * Assumes number of cores is a power of two.
351 static void amd_detect_cmp(struct cpuinfo_x86 *c)
355 int cpu = smp_processor_id();
357 bits = c->x86_coreid_bits;
358 /* Low order bits define the core id (index of core in socket) */
359 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
360 /* Convert the initial APIC ID into the socket ID */
361 c->phys_proc_id = c->initial_apicid >> bits;
362 /* use socket ID also for last level cache */
363 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
367 * Fix percpu cpu_llc_id here as LLC topology is different
368 * for Fam17h systems.
370 if (c->x86 != 0x17 || !cpuid_edx(0x80000006))
373 per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
377 u16 amd_get_nb_id(int cpu)
381 id = per_cpu(cpu_llc_id, cpu);
385 EXPORT_SYMBOL_GPL(amd_get_nb_id);
387 u32 amd_get_nodes_per_socket(void)
389 return nodes_per_socket;
391 EXPORT_SYMBOL_GPL(amd_get_nodes_per_socket);
393 static void srat_detect_node(struct cpuinfo_x86 *c)
396 int cpu = smp_processor_id();
398 unsigned apicid = c->apicid;
400 node = numa_cpu_node(cpu);
401 if (node == NUMA_NO_NODE)
402 node = per_cpu(cpu_llc_id, cpu);
405 * On multi-fabric platform (e.g. Numascale NumaChip) a
406 * platform-specific handler needs to be called to fixup some
409 if (x86_cpuinit.fixup_cpu_id)
410 x86_cpuinit.fixup_cpu_id(c, node);
412 if (!node_online(node)) {
414 * Two possibilities here:
416 * - The CPU is missing memory and no node was created. In
417 * that case try picking one from a nearby CPU.
419 * - The APIC IDs differ from the HyperTransport node IDs
420 * which the K8 northbridge parsing fills in. Assume
421 * they are all increased by a constant offset, but in
422 * the same order as the HT nodeids. If that doesn't
423 * result in a usable node fall back to the path for the
426 * This workaround operates directly on the mapping between
427 * APIC ID and NUMA node, assuming certain relationship
428 * between APIC ID, HT node ID and NUMA topology. As going
429 * through CPU mapping may alter the outcome, directly
430 * access __apicid_to_node[].
432 int ht_nodeid = c->initial_apicid;
434 if (ht_nodeid >= 0 &&
435 __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
436 node = __apicid_to_node[ht_nodeid];
437 /* Pick a nearby node */
438 if (!node_online(node))
439 node = nearby_node(apicid);
441 numa_set_node(cpu, node);
445 static void early_init_amd_mc(struct cpuinfo_x86 *c)
450 /* Multi core CPU? */
451 if (c->extended_cpuid_level < 0x80000008)
454 ecx = cpuid_ecx(0x80000008);
456 c->x86_max_cores = (ecx & 0xff) + 1;
458 /* CPU telling us the core id bits shift? */
459 bits = (ecx >> 12) & 0xF;
461 /* Otherwise recompute */
463 while ((1 << bits) < c->x86_max_cores)
467 c->x86_coreid_bits = bits;
471 static void bsp_init_amd(struct cpuinfo_x86 *c)
476 unsigned long long tseg;
479 * Split up direct mapping around the TSEG SMM area.
480 * Don't do it for gbpages because there seems very little
481 * benefit in doing so.
483 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
484 unsigned long pfn = tseg >> PAGE_SHIFT;
486 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
487 if (pfn_range_is_mapped(pfn, pfn + 1))
488 set_memory_4k((unsigned long)__va(tseg), 1);
493 if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
496 (c->x86 == 0x10 && c->x86_model >= 0x2)) {
499 rdmsrl(MSR_K7_HWCR, val);
500 if (!(val & BIT(24)))
501 printk(KERN_WARNING FW_BUG "TSC doesn't count "
502 "with P0 frequency!\n");
506 if (c->x86 == 0x15) {
507 unsigned long upperbit;
510 cpuid = cpuid_edx(0x80000005);
511 assoc = cpuid >> 16 & 0xff;
512 upperbit = ((cpuid >> 24) << 10) / assoc;
514 va_align.mask = (upperbit - 1) & PAGE_MASK;
515 va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
517 /* A random value per boot for bit slice [12:upper_bit) */
518 va_align.bits = get_random_int() & va_align.mask;
521 if (cpu_has(c, X86_FEATURE_MWAITX))
524 if (c->x86 >= 0x15 && c->x86 <= 0x17) {
528 case 0x15: bit = 54; break;
529 case 0x16: bit = 33; break;
530 case 0x17: bit = 10; break;
534 * Try to cache the base value so further operations can
535 * avoid RMW. If that faults, do not enable SSBD.
537 if (!rdmsrl_safe(MSR_AMD64_LS_CFG, &x86_amd_ls_cfg_base)) {
538 setup_force_cpu_cap(X86_FEATURE_LS_CFG_SSBD);
539 setup_force_cpu_cap(X86_FEATURE_SSBD);
540 x86_amd_ls_cfg_ssbd_mask = 1ULL << bit;
545 static void early_init_amd(struct cpuinfo_x86 *c)
547 early_init_amd_mc(c);
550 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
551 * with P/T states and does not stop in deep C-states
553 if (c->x86_power & (1 << 8)) {
554 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
555 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
556 if (!check_tsc_unstable())
557 set_sched_clock_stable();
561 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
563 /* Set MTRR capability flag if appropriate */
565 if (c->x86_model == 13 || c->x86_model == 9 ||
566 (c->x86_model == 8 && c->x86_stepping >= 8))
567 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
569 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
571 * ApicID can always be treated as an 8-bit value for AMD APIC versions
572 * >= 0x10, but even old K8s came out of reset with version 0x10. So, we
573 * can safely set X86_FEATURE_EXTD_APICID unconditionally for families
576 if (cpu_has_apic && c->x86 > 0x16) {
577 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
578 } else if (cpu_has_apic && c->x86 >= 0xf) {
579 /* check CPU config space for extended APIC ID */
581 val = read_pci_config(0, 24, 0, 0x68);
582 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
583 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
588 * This is only needed to tell the kernel whether to use VMCALL
589 * and VMMCALL. VMMCALL is never executed except under virt, so
590 * we can set it unconditionally.
592 set_cpu_cap(c, X86_FEATURE_VMMCALL);
594 /* F16h erratum 793, CVE-2013-6885 */
595 if (c->x86 == 0x16 && c->x86_model <= 0xf)
596 msr_set_bit(MSR_AMD64_LS_CFG, 15);
599 static const int amd_erratum_383[];
600 static const int amd_erratum_400[];
601 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum);
603 static void init_amd_k8(struct cpuinfo_x86 *c)
608 /* On C+ stepping K8 rep microcode works well for copy/memset */
609 level = cpuid_eax(1);
610 if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
611 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
614 * Some BIOSes incorrectly force this feature, but only K8 revision D
615 * (model = 0x14) and later actually support it.
616 * (AMD Erratum #110, docId: 25759).
618 if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
619 clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
620 if (!rdmsrl_amd_safe(0xc001100d, &value)) {
621 value &= ~BIT_64(32);
622 wrmsrl_amd_safe(0xc001100d, value);
626 if (!c->x86_model_id[0])
627 strcpy(c->x86_model_id, "Hammer");
631 * Disable TLB flush filter by setting HWCR.FFDIS on K8
632 * bit 6 of msr C001_0015
634 * Errata 63 for SH-B3 steppings
635 * Errata 122 for all steppings (F+ have it disabled by default)
637 msr_set_bit(MSR_K7_HWCR, 6);
641 static void init_amd_gh(struct cpuinfo_x86 *c)
644 /* do this for boot cpu */
645 if (c == &boot_cpu_data)
646 check_enable_amd_mmconf_dmi();
648 fam10h_check_enable_mmcfg();
652 * Disable GART TLB Walk Errors on Fam10h. We do this here because this
653 * is always needed when GART is enabled, even in a kernel which has no
654 * MCE support built in. BIOS should disable GartTlbWlk Errors already.
655 * If it doesn't, we do it here as suggested by the BKDG.
657 * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
659 msr_set_bit(MSR_AMD64_MCx_MASK(4), 10);
662 * On family 10h BIOS may not have properly enabled WC+ support, causing
663 * it to be converted to CD memtype. This may result in performance
664 * degradation for certain nested-paging guests. Prevent this conversion
665 * by clearing bit 24 in MSR_AMD64_BU_CFG2.
667 * NOTE: we want to use the _safe accessors so as not to #GP kvm
668 * guests on older kvm hosts.
670 msr_clear_bit(MSR_AMD64_BU_CFG2, 24);
672 if (cpu_has_amd_erratum(c, amd_erratum_383))
673 set_cpu_bug(c, X86_BUG_AMD_TLB_MMATCH);
676 #define MSR_AMD64_DE_CFG 0xC0011029
678 static void init_amd_ln(struct cpuinfo_x86 *c)
681 * Apply erratum 665 fix unconditionally so machines without a BIOS
684 msr_set_bit(MSR_AMD64_DE_CFG, 31);
687 static bool rdrand_force;
689 static int __init rdrand_cmdline(char *str)
694 if (!strcmp(str, "force"))
701 early_param("rdrand", rdrand_cmdline);
703 static void clear_rdrand_cpuid_bit(struct cpuinfo_x86 *c)
706 * Saving of the MSR used to hide the RDRAND support during
707 * suspend/resume is done by arch/x86/power/cpu.c, which is
708 * dependent on CONFIG_PM_SLEEP.
710 if (!IS_ENABLED(CONFIG_PM_SLEEP))
714 * The nordrand option can clear X86_FEATURE_RDRAND, so check for
715 * RDRAND support using the CPUID function directly.
717 if (!(cpuid_ecx(1) & BIT(30)) || rdrand_force)
720 msr_clear_bit(MSR_AMD64_CPUID_FN_1, 62);
723 * Verify that the CPUID change has occurred in case the kernel is
724 * running virtualized and the hypervisor doesn't support the MSR.
726 if (cpuid_ecx(1) & BIT(30)) {
727 pr_info_once("BIOS may not properly restore RDRAND after suspend, but hypervisor does not support hiding RDRAND via CPUID.\n");
731 clear_cpu_cap(c, X86_FEATURE_RDRAND);
732 pr_info_once("BIOS may not properly restore RDRAND after suspend, hiding RDRAND via CPUID. Use rdrand=force to reenable.\n");
735 static void init_amd_jg(struct cpuinfo_x86 *c)
738 * Some BIOS implementations do not restore proper RDRAND support
739 * across suspend and resume. Check on whether to hide the RDRAND
740 * instruction support via CPUID.
742 clear_rdrand_cpuid_bit(c);
745 static void init_amd_bd(struct cpuinfo_x86 *c)
749 /* re-enable TopologyExtensions if switched off by BIOS */
750 if ((c->x86_model >= 0x10) && (c->x86_model <= 0x1f) &&
751 !cpu_has(c, X86_FEATURE_TOPOEXT)) {
753 if (msr_set_bit(0xc0011005, 54) > 0) {
754 rdmsrl(0xc0011005, value);
755 if (value & BIT_64(54)) {
756 set_cpu_cap(c, X86_FEATURE_TOPOEXT);
757 pr_info(FW_INFO "CPU: Re-enabling disabled Topology Extensions Support.\n");
763 * The way access filter has a performance penalty on some workloads.
764 * Disable it on the affected CPUs.
766 if ((c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
767 if (!rdmsrl_safe(0xc0011021, &value) && !(value & 0x1E)) {
769 wrmsrl_safe(0xc0011021, value);
774 * Some BIOS implementations do not restore proper RDRAND support
775 * across suspend and resume. Check on whether to hide the RDRAND
776 * instruction support via CPUID.
778 clear_rdrand_cpuid_bit(c);
781 static void init_amd_zn(struct cpuinfo_x86 *c)
783 set_cpu_cap(c, X86_FEATURE_ZEN);
786 * Fix erratum 1076: CPB feature bit not being set in CPUID.
787 * Always set it, except when running under a hypervisor.
789 if (!cpu_has(c, X86_FEATURE_HYPERVISOR) && !cpu_has(c, X86_FEATURE_CPB))
790 set_cpu_cap(c, X86_FEATURE_CPB);
793 static void init_amd(struct cpuinfo_x86 *c)
800 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
801 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
803 clear_cpu_cap(c, 0*32+31);
806 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
808 /* get apicid instead of initial apic id from cpuid */
809 c->apicid = hard_smp_processor_id();
811 /* K6s reports MCEs but don't actually have all the MSRs */
813 clear_cpu_cap(c, X86_FEATURE_MCE);
816 case 4: init_amd_k5(c); break;
817 case 5: init_amd_k6(c); break;
818 case 6: init_amd_k7(c); break;
819 case 0xf: init_amd_k8(c); break;
820 case 0x10: init_amd_gh(c); break;
821 case 0x12: init_amd_ln(c); break;
822 case 0x15: init_amd_bd(c); break;
823 case 0x16: init_amd_jg(c); break;
824 case 0x17: init_amd_zn(c); break;
827 /* Enable workaround for FXSAVE leak */
829 set_cpu_bug(c, X86_BUG_FXSAVE_LEAK);
831 cpu_detect_cache_sizes(c);
833 /* Multi core CPU? */
834 if (c->extended_cpuid_level >= 0x80000008) {
843 init_amd_cacheinfo(c);
846 set_cpu_cap(c, X86_FEATURE_K8);
849 unsigned long long val;
853 * A serializing LFENCE has less overhead than MFENCE, so
854 * use it for execution serialization. On families which
855 * don't have that MSR, LFENCE is already serializing.
856 * msr_set_bit() uses the safe accessors, too, even if the MSR
859 msr_set_bit(MSR_F10H_DECFG,
860 MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT);
863 * Verify that the MSR write was successful (could be running
864 * under a hypervisor) and only then assume that LFENCE is
867 ret = rdmsrl_safe(MSR_F10H_DECFG, &val);
868 if (!ret && (val & MSR_F10H_DECFG_LFENCE_SERIALIZE)) {
869 /* A serializing LFENCE stops RDTSC speculation */
870 set_cpu_cap(c, X86_FEATURE_LFENCE_RDTSC);
872 /* MFENCE stops RDTSC speculation */
873 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
878 * Family 0x12 and above processors have APIC timer
879 * running in deep C states.
882 set_cpu_cap(c, X86_FEATURE_ARAT);
884 if (cpu_has_amd_erratum(c, amd_erratum_400))
885 set_cpu_bug(c, X86_BUG_AMD_APIC_C1E);
887 rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
889 /* 3DNow or LM implies PREFETCHW */
890 if (!cpu_has(c, X86_FEATURE_3DNOWPREFETCH))
891 if (cpu_has(c, X86_FEATURE_3DNOW) || cpu_has(c, X86_FEATURE_LM))
892 set_cpu_cap(c, X86_FEATURE_3DNOWPREFETCH);
894 /* AMD CPUs don't reset SS attributes on SYSRET, Xen does. */
895 if (!cpu_has(c, X86_FEATURE_XENPV))
896 set_cpu_bug(c, X86_BUG_SYSRET_SS_ATTRS);
900 static unsigned int amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
902 /* AMD errata T13 (order #21922) */
905 if (c->x86_model == 3 && c->x86_stepping == 0)
907 /* Tbird rev A1/A2 */
908 if (c->x86_model == 4 &&
909 (c->x86_stepping == 0 || c->x86_stepping == 1))
916 static void cpu_detect_tlb_amd(struct cpuinfo_x86 *c)
918 u32 ebx, eax, ecx, edx;
924 if (c->extended_cpuid_level < 0x80000006)
927 cpuid(0x80000006, &eax, &ebx, &ecx, &edx);
929 tlb_lld_4k[ENTRIES] = (ebx >> 16) & mask;
930 tlb_lli_4k[ENTRIES] = ebx & mask;
933 * K8 doesn't have 2M/4M entries in the L2 TLB so read out the L1 TLB
934 * characteristics from the CPUID function 0x80000005 instead.
937 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
941 /* Handle DTLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
942 if (!((eax >> 16) & mask))
943 tlb_lld_2m[ENTRIES] = (cpuid_eax(0x80000005) >> 16) & 0xff;
945 tlb_lld_2m[ENTRIES] = (eax >> 16) & mask;
947 /* a 4M entry uses two 2M entries */
948 tlb_lld_4m[ENTRIES] = tlb_lld_2m[ENTRIES] >> 1;
950 /* Handle ITLB 2M and 4M sizes, fall back to L1 if L2 is disabled */
953 if (c->x86 == 0x15 && c->x86_model <= 0x1f) {
954 tlb_lli_2m[ENTRIES] = 1024;
956 cpuid(0x80000005, &eax, &ebx, &ecx, &edx);
957 tlb_lli_2m[ENTRIES] = eax & 0xff;
960 tlb_lli_2m[ENTRIES] = eax & mask;
962 tlb_lli_4m[ENTRIES] = tlb_lli_2m[ENTRIES] >> 1;
965 static const struct cpu_dev amd_cpu_dev = {
967 .c_ident = { "AuthenticAMD" },
970 { .family = 4, .model_names =
981 .legacy_cache_size = amd_size_cache,
983 .c_early_init = early_init_amd,
984 .c_detect_tlb = cpu_detect_tlb_amd,
985 .c_bsp_init = bsp_init_amd,
987 .c_x86_vendor = X86_VENDOR_AMD,
990 cpu_dev_register(amd_cpu_dev);
993 * AMD errata checking
995 * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
996 * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
997 * have an OSVW id assigned, which it takes as first argument. Both take a
998 * variable number of family-specific model-stepping ranges created by
1003 * const int amd_erratum_319[] =
1004 * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
1005 * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
1006 * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
1009 #define AMD_LEGACY_ERRATUM(...) { -1, __VA_ARGS__, 0 }
1010 #define AMD_OSVW_ERRATUM(osvw_id, ...) { osvw_id, __VA_ARGS__, 0 }
1011 #define AMD_MODEL_RANGE(f, m_start, s_start, m_end, s_end) \
1012 ((f << 24) | (m_start << 16) | (s_start << 12) | (m_end << 4) | (s_end))
1013 #define AMD_MODEL_RANGE_FAMILY(range) (((range) >> 24) & 0xff)
1014 #define AMD_MODEL_RANGE_START(range) (((range) >> 12) & 0xfff)
1015 #define AMD_MODEL_RANGE_END(range) ((range) & 0xfff)
1017 static const int amd_erratum_400[] =
1018 AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
1019 AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
1021 static const int amd_erratum_383[] =
1022 AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
1025 static bool cpu_has_amd_erratum(struct cpuinfo_x86 *cpu, const int *erratum)
1027 int osvw_id = *erratum++;
1031 if (osvw_id >= 0 && osvw_id < 65536 &&
1032 cpu_has(cpu, X86_FEATURE_OSVW)) {
1035 rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
1036 if (osvw_id < osvw_len) {
1039 rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
1041 return osvw_bits & (1ULL << (osvw_id & 0x3f));
1045 /* OSVW unavailable or ID unknown, match family-model-stepping range */
1046 ms = (cpu->x86_model << 4) | cpu->x86_stepping;
1047 while ((range = *erratum++))
1048 if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
1049 (ms >= AMD_MODEL_RANGE_START(range)) &&
1050 (ms <= AMD_MODEL_RANGE_END(range)))
1056 void set_dr_addr_mask(unsigned long mask, int dr)
1058 if (!boot_cpu_has(X86_FEATURE_BPEXT))
1063 wrmsr(MSR_F16H_DR0_ADDR_MASK, mask, 0);
1068 wrmsr(MSR_F16H_DR1_ADDR_MASK - 1 + dr, mask, 0);