1 #include <linux/bootmem.h>
2 #include <linux/linkage.h>
3 #include <linux/bitops.h>
4 #include <linux/kernel.h>
5 #include <linux/export.h>
6 #include <linux/percpu.h>
7 #include <linux/string.h>
8 #include <linux/ctype.h>
9 #include <linux/delay.h>
10 #include <linux/sched/mm.h>
11 #include <linux/sched/clock.h>
12 #include <linux/sched/task.h>
13 #include <linux/init.h>
14 #include <linux/kprobes.h>
15 #include <linux/kgdb.h>
16 #include <linux/smp.h>
18 #include <linux/syscore_ops.h>
20 #include <asm/stackprotector.h>
21 #include <asm/perf_event.h>
22 #include <asm/mmu_context.h>
23 #include <asm/archrandom.h>
24 #include <asm/hypervisor.h>
25 #include <asm/processor.h>
26 #include <asm/tlbflush.h>
27 #include <asm/debugreg.h>
28 #include <asm/sections.h>
29 #include <asm/vsyscall.h>
30 #include <linux/topology.h>
31 #include <linux/cpumask.h>
32 #include <asm/pgtable.h>
33 #include <linux/atomic.h>
34 #include <asm/proto.h>
35 #include <asm/setup.h>
38 #include <asm/fpu/internal.h>
40 #include <asm/hwcap2.h>
41 #include <linux/numa.h>
48 #include <asm/microcode.h>
49 #include <asm/microcode_intel.h>
50 #include <asm/intel-family.h>
51 #include <asm/cpu_device_id.h>
53 #ifdef CONFIG_X86_LOCAL_APIC
54 #include <asm/uv/uv.h>
59 u32 elf_hwcap2 __read_mostly;
61 /* all of these masks are initialized in setup_cpu_local_masks() */
62 cpumask_var_t cpu_initialized_mask;
63 cpumask_var_t cpu_callout_mask;
64 cpumask_var_t cpu_callin_mask;
66 /* representing cpus for which sibling maps can be computed */
67 cpumask_var_t cpu_sibling_setup_mask;
69 /* Number of siblings per CPU package */
70 int smp_num_siblings = 1;
71 EXPORT_SYMBOL(smp_num_siblings);
73 /* Last level cache ID of each logical CPU */
74 DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
76 /* correctly size the local cpu masks */
77 void __init setup_cpu_local_masks(void)
79 alloc_bootmem_cpumask_var(&cpu_initialized_mask);
80 alloc_bootmem_cpumask_var(&cpu_callin_mask);
81 alloc_bootmem_cpumask_var(&cpu_callout_mask);
82 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
85 static void default_init(struct cpuinfo_x86 *c)
88 cpu_detect_cache_sizes(c);
90 /* Not much we can do here... */
91 /* Check if at least it has cpuid */
92 if (c->cpuid_level == -1) {
93 /* No cpuid. It must be an ancient CPU */
95 strcpy(c->x86_model_id, "486");
97 strcpy(c->x86_model_id, "386");
102 static const struct cpu_dev default_cpu = {
103 .c_init = default_init,
104 .c_vendor = "Unknown",
105 .c_x86_vendor = X86_VENDOR_UNKNOWN,
108 static const struct cpu_dev *this_cpu = &default_cpu;
110 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
113 * We need valid kernel segments for data and code in long mode too
114 * IRET will check the segment types kkeil 2000/10/28
115 * Also sysret mandates a special GDT layout
117 * TLS descriptors are currently at a different place compared to i386.
118 * Hopefully nobody expects them at a fixed place (Wine?)
120 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
121 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
122 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
123 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
124 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
125 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
127 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
128 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
129 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
130 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
132 * Segments used for calling PnP BIOS have byte granularity.
133 * They code segments and data segments have fixed 64k limits,
134 * the transfer segment sizes are set at run time.
137 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
139 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
141 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(0x0092, 0, 0xffff),
143 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(0x0092, 0, 0),
145 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(0x0092, 0, 0),
147 * The APM segments have byte granularity and their bases
148 * are set at run time. All have 64k limits.
151 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
153 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
155 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(0x4092, 0, 0xffff),
157 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
158 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
159 GDT_STACK_CANARY_INIT
162 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
164 static int __init x86_mpx_setup(char *s)
166 /* require an exact match without trailing characters */
170 /* do not emit a message if the feature is not present */
171 if (!boot_cpu_has(X86_FEATURE_MPX))
174 setup_clear_cpu_cap(X86_FEATURE_MPX);
175 pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
178 __setup("nompx", x86_mpx_setup);
181 static int __init x86_nopcid_setup(char *s)
183 /* nopcid doesn't accept parameters */
187 /* do not emit a message if the feature is not present */
188 if (!boot_cpu_has(X86_FEATURE_PCID))
191 setup_clear_cpu_cap(X86_FEATURE_PCID);
192 pr_info("nopcid: PCID feature disabled\n");
195 early_param("nopcid", x86_nopcid_setup);
198 static int __init x86_noinvpcid_setup(char *s)
200 /* noinvpcid doesn't accept parameters */
204 /* do not emit a message if the feature is not present */
205 if (!boot_cpu_has(X86_FEATURE_INVPCID))
208 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
209 pr_info("noinvpcid: INVPCID feature disabled\n");
212 early_param("noinvpcid", x86_noinvpcid_setup);
215 static int cachesize_override = -1;
216 static int disable_x86_serial_nr = 1;
218 static int __init cachesize_setup(char *str)
220 get_option(&str, &cachesize_override);
223 __setup("cachesize=", cachesize_setup);
225 static int __init x86_sep_setup(char *s)
227 setup_clear_cpu_cap(X86_FEATURE_SEP);
230 __setup("nosep", x86_sep_setup);
232 /* Standard macro to see if a specific flag is changeable */
233 static inline int flag_is_changeable_p(u32 flag)
238 * Cyrix and IDT cpus allow disabling of CPUID
239 * so the code below may return different results
240 * when it is executed before and after enabling
241 * the CPUID. Add "volatile" to not allow gcc to
242 * optimize the subsequent calls to this function.
244 asm volatile ("pushfl \n\t"
255 : "=&r" (f1), "=&r" (f2)
258 return ((f1^f2) & flag) != 0;
261 /* Probe for the CPUID instruction */
262 int have_cpuid_p(void)
264 return flag_is_changeable_p(X86_EFLAGS_ID);
267 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
269 unsigned long lo, hi;
271 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
274 /* Disable processor serial number: */
276 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
278 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
280 pr_notice("CPU serial number disabled.\n");
281 clear_cpu_cap(c, X86_FEATURE_PN);
283 /* Disabling the serial number may affect the cpuid level */
284 c->cpuid_level = cpuid_eax(0);
287 static int __init x86_serial_nr_setup(char *s)
289 disable_x86_serial_nr = 0;
292 __setup("serialnumber", x86_serial_nr_setup);
294 static inline int flag_is_changeable_p(u32 flag)
298 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
303 static __init int setup_disable_smep(char *arg)
305 setup_clear_cpu_cap(X86_FEATURE_SMEP);
306 /* Check for things that depend on SMEP being enabled: */
307 check_mpx_erratum(&boot_cpu_data);
310 __setup("nosmep", setup_disable_smep);
312 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
314 if (cpu_has(c, X86_FEATURE_SMEP))
315 cr4_set_bits(X86_CR4_SMEP);
318 static __init int setup_disable_smap(char *arg)
320 setup_clear_cpu_cap(X86_FEATURE_SMAP);
323 __setup("nosmap", setup_disable_smap);
325 static __always_inline void setup_smap(struct cpuinfo_x86 *c)
327 unsigned long eflags = native_save_fl();
329 /* This should have been cleared long ago */
330 BUG_ON(eflags & X86_EFLAGS_AC);
332 if (cpu_has(c, X86_FEATURE_SMAP)) {
333 #ifdef CONFIG_X86_SMAP
334 cr4_set_bits(X86_CR4_SMAP);
336 cr4_clear_bits(X86_CR4_SMAP);
342 * Protection Keys are not available in 32-bit mode.
344 static bool pku_disabled;
346 static __always_inline void setup_pku(struct cpuinfo_x86 *c)
348 /* check the boot processor, plus compile options for PKU: */
349 if (!cpu_feature_enabled(X86_FEATURE_PKU))
351 /* checks the actual processor's cpuid bits: */
352 if (!cpu_has(c, X86_FEATURE_PKU))
357 cr4_set_bits(X86_CR4_PKE);
359 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
360 * cpuid bit to be set. We need to ensure that we
361 * update that bit in this CPU's "cpu_info".
363 set_cpu_cap(c, X86_FEATURE_OSPKE);
366 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
367 static __init int setup_disable_pku(char *arg)
370 * Do not clear the X86_FEATURE_PKU bit. All of the
371 * runtime checks are against OSPKE so clearing the
374 * This way, we will see "pku" in cpuinfo, but not
375 * "ospke", which is exactly what we want. It shows
376 * that the CPU has PKU, but the OS has not enabled it.
377 * This happens to be exactly how a system would look
378 * if we disabled the config option.
380 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
384 __setup("nopku", setup_disable_pku);
385 #endif /* CONFIG_X86_64 */
388 * Some CPU features depend on higher CPUID levels, which may not always
389 * be available due to CPUID level capping or broken virtualization
390 * software. Add those features to this table to auto-disable them.
392 struct cpuid_dependent_feature {
397 static const struct cpuid_dependent_feature
398 cpuid_dependent_features[] = {
399 { X86_FEATURE_MWAIT, 0x00000005 },
400 { X86_FEATURE_DCA, 0x00000009 },
401 { X86_FEATURE_XSAVE, 0x0000000d },
405 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
407 const struct cpuid_dependent_feature *df;
409 for (df = cpuid_dependent_features; df->feature; df++) {
411 if (!cpu_has(c, df->feature))
414 * Note: cpuid_level is set to -1 if unavailable, but
415 * extended_extended_level is set to 0 if unavailable
416 * and the legitimate extended levels are all negative
417 * when signed; hence the weird messing around with
420 if (!((s32)df->level < 0 ?
421 (u32)df->level > (u32)c->extended_cpuid_level :
422 (s32)df->level > (s32)c->cpuid_level))
425 clear_cpu_cap(c, df->feature);
429 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
430 x86_cap_flag(df->feature), df->level);
435 * Naming convention should be: <Name> [(<Codename>)]
436 * This table only is used unless init_<vendor>() below doesn't set it;
437 * in particular, if CPUID levels 0x80000002..4 are supported, this
441 /* Look up CPU names by table lookup. */
442 static const char *table_lookup_model(struct cpuinfo_x86 *c)
445 const struct legacy_cpu_model_info *info;
447 if (c->x86_model >= 16)
448 return NULL; /* Range check */
453 info = this_cpu->legacy_models;
455 while (info->family) {
456 if (info->family == c->x86)
457 return info->model_names[c->x86_model];
461 return NULL; /* Not found */
464 __u32 cpu_caps_cleared[NCAPINTS + NBUGINTS];
465 __u32 cpu_caps_set[NCAPINTS + NBUGINTS];
467 void load_percpu_segment(int cpu)
470 loadsegment(fs, __KERNEL_PERCPU);
472 __loadsegment_simple(gs, 0);
473 wrmsrl(MSR_GS_BASE, (unsigned long)per_cpu(irq_stack_union.gs_base, cpu));
475 load_stack_canary_segment();
479 /* The 32-bit entry code needs to find cpu_entry_area. */
480 DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area);
485 * Special IST stacks which the CPU switches to when it calls
486 * an IST-marked descriptor entry. Up to 7 stacks (hardware
487 * limit), all of them are 4K, except the debug stack which
490 static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
491 [0 ... N_EXCEPTION_STACKS - 1] = EXCEPTION_STKSZ,
492 [DEBUG_STACK - 1] = DEBUG_STKSZ
496 /* Load the original GDT from the per-cpu structure */
497 void load_direct_gdt(int cpu)
499 struct desc_ptr gdt_descr;
501 gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
502 gdt_descr.size = GDT_SIZE - 1;
503 load_gdt(&gdt_descr);
505 EXPORT_SYMBOL_GPL(load_direct_gdt);
507 /* Load a fixmap remapping of the per-cpu GDT */
508 void load_fixmap_gdt(int cpu)
510 struct desc_ptr gdt_descr;
512 gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
513 gdt_descr.size = GDT_SIZE - 1;
514 load_gdt(&gdt_descr);
516 EXPORT_SYMBOL_GPL(load_fixmap_gdt);
519 * Current gdt points %fs at the "master" per-cpu area: after this,
520 * it's on the real one.
522 void switch_to_new_gdt(int cpu)
524 /* Load the original GDT */
525 load_direct_gdt(cpu);
526 /* Reload the per-cpu base */
527 load_percpu_segment(cpu);
530 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
532 static void get_model_name(struct cpuinfo_x86 *c)
537 if (c->extended_cpuid_level < 0x80000004)
540 v = (unsigned int *)c->x86_model_id;
541 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
542 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
543 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
544 c->x86_model_id[48] = 0;
546 /* Trim whitespace */
547 p = q = s = &c->x86_model_id[0];
553 /* Note the last non-whitespace index */
563 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
565 unsigned int n, dummy, ebx, ecx, edx, l2size;
567 n = c->extended_cpuid_level;
569 if (n >= 0x80000005) {
570 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
571 c->x86_cache_size = (ecx>>24) + (edx>>24);
573 /* On K8 L1 TLB is inclusive, so don't count it */
578 if (n < 0x80000006) /* Some chips just has a large L1. */
581 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
585 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
587 /* do processor-specific cache resizing */
588 if (this_cpu->legacy_cache_size)
589 l2size = this_cpu->legacy_cache_size(c, l2size);
591 /* Allow user to override all this if necessary. */
592 if (cachesize_override != -1)
593 l2size = cachesize_override;
596 return; /* Again, no L2 cache is possible */
599 c->x86_cache_size = l2size;
602 u16 __read_mostly tlb_lli_4k[NR_INFO];
603 u16 __read_mostly tlb_lli_2m[NR_INFO];
604 u16 __read_mostly tlb_lli_4m[NR_INFO];
605 u16 __read_mostly tlb_lld_4k[NR_INFO];
606 u16 __read_mostly tlb_lld_2m[NR_INFO];
607 u16 __read_mostly tlb_lld_4m[NR_INFO];
608 u16 __read_mostly tlb_lld_1g[NR_INFO];
610 static void cpu_detect_tlb(struct cpuinfo_x86 *c)
612 if (this_cpu->c_detect_tlb)
613 this_cpu->c_detect_tlb(c);
615 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
616 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
617 tlb_lli_4m[ENTRIES]);
619 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
620 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
621 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
624 int detect_ht_early(struct cpuinfo_x86 *c)
627 u32 eax, ebx, ecx, edx;
629 if (!cpu_has(c, X86_FEATURE_HT))
632 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
635 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
638 cpuid(1, &eax, &ebx, &ecx, &edx);
640 smp_num_siblings = (ebx & 0xff0000) >> 16;
641 if (smp_num_siblings == 1)
642 pr_info_once("CPU0: Hyper-Threading is disabled\n");
647 void detect_ht(struct cpuinfo_x86 *c)
650 int index_msb, core_bits;
652 if (detect_ht_early(c) < 0)
655 index_msb = get_count_order(smp_num_siblings);
656 c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
658 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
660 index_msb = get_count_order(smp_num_siblings);
662 core_bits = get_count_order(c->x86_max_cores);
664 c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
665 ((1 << core_bits) - 1);
669 static void get_cpu_vendor(struct cpuinfo_x86 *c)
671 char *v = c->x86_vendor_id;
674 for (i = 0; i < X86_VENDOR_NUM; i++) {
678 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
679 (cpu_devs[i]->c_ident[1] &&
680 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
682 this_cpu = cpu_devs[i];
683 c->x86_vendor = this_cpu->c_x86_vendor;
688 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
689 "CPU: Your system may be unstable.\n", v);
691 c->x86_vendor = X86_VENDOR_UNKNOWN;
692 this_cpu = &default_cpu;
695 void cpu_detect(struct cpuinfo_x86 *c)
697 /* Get vendor name */
698 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
699 (unsigned int *)&c->x86_vendor_id[0],
700 (unsigned int *)&c->x86_vendor_id[8],
701 (unsigned int *)&c->x86_vendor_id[4]);
704 /* Intel-defined flags: level 0x00000001 */
705 if (c->cpuid_level >= 0x00000001) {
706 u32 junk, tfms, cap0, misc;
708 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
709 c->x86 = x86_family(tfms);
710 c->x86_model = x86_model(tfms);
711 c->x86_stepping = x86_stepping(tfms);
713 if (cap0 & (1<<19)) {
714 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
715 c->x86_cache_alignment = c->x86_clflush_size;
720 static void apply_forced_caps(struct cpuinfo_x86 *c)
724 for (i = 0; i < NCAPINTS + NBUGINTS; i++) {
725 c->x86_capability[i] &= ~cpu_caps_cleared[i];
726 c->x86_capability[i] |= cpu_caps_set[i];
730 static void init_speculation_control(struct cpuinfo_x86 *c)
733 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support,
734 * and they also have a different bit for STIBP support. Also,
735 * a hypervisor might have set the individual AMD bits even on
736 * Intel CPUs, for finer-grained selection of what's available.
738 if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) {
739 set_cpu_cap(c, X86_FEATURE_IBRS);
740 set_cpu_cap(c, X86_FEATURE_IBPB);
741 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
744 if (cpu_has(c, X86_FEATURE_INTEL_STIBP))
745 set_cpu_cap(c, X86_FEATURE_STIBP);
747 if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) ||
748 cpu_has(c, X86_FEATURE_VIRT_SSBD))
749 set_cpu_cap(c, X86_FEATURE_SSBD);
751 if (cpu_has(c, X86_FEATURE_AMD_IBRS)) {
752 set_cpu_cap(c, X86_FEATURE_IBRS);
753 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
756 if (cpu_has(c, X86_FEATURE_AMD_IBPB))
757 set_cpu_cap(c, X86_FEATURE_IBPB);
759 if (cpu_has(c, X86_FEATURE_AMD_STIBP)) {
760 set_cpu_cap(c, X86_FEATURE_STIBP);
761 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
764 if (cpu_has(c, X86_FEATURE_AMD_SSBD)) {
765 set_cpu_cap(c, X86_FEATURE_SSBD);
766 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
767 clear_cpu_cap(c, X86_FEATURE_VIRT_SSBD);
771 static void init_cqm(struct cpuinfo_x86 *c)
773 if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
774 c->x86_cache_max_rmid = -1;
775 c->x86_cache_occ_scale = -1;
779 /* will be overridden if occupancy monitoring exists */
780 c->x86_cache_max_rmid = cpuid_ebx(0xf);
782 if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
783 cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
784 cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
785 u32 eax, ebx, ecx, edx;
787 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
788 cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
790 c->x86_cache_max_rmid = ecx;
791 c->x86_cache_occ_scale = ebx;
795 void get_cpu_cap(struct cpuinfo_x86 *c)
797 u32 eax, ebx, ecx, edx;
799 /* Intel-defined flags: level 0x00000001 */
800 if (c->cpuid_level >= 0x00000001) {
801 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
803 c->x86_capability[CPUID_1_ECX] = ecx;
804 c->x86_capability[CPUID_1_EDX] = edx;
807 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
808 if (c->cpuid_level >= 0x00000006)
809 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
811 /* Additional Intel-defined flags: level 0x00000007 */
812 if (c->cpuid_level >= 0x00000007) {
813 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
814 c->x86_capability[CPUID_7_0_EBX] = ebx;
815 c->x86_capability[CPUID_7_ECX] = ecx;
816 c->x86_capability[CPUID_7_EDX] = edx;
819 /* Extended state features: level 0x0000000d */
820 if (c->cpuid_level >= 0x0000000d) {
821 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
823 c->x86_capability[CPUID_D_1_EAX] = eax;
826 /* AMD-defined flags: level 0x80000001 */
827 eax = cpuid_eax(0x80000000);
828 c->extended_cpuid_level = eax;
830 if ((eax & 0xffff0000) == 0x80000000) {
831 if (eax >= 0x80000001) {
832 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
834 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
835 c->x86_capability[CPUID_8000_0001_EDX] = edx;
839 if (c->extended_cpuid_level >= 0x80000007) {
840 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
842 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
846 if (c->extended_cpuid_level >= 0x80000008) {
847 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
849 c->x86_virt_bits = (eax >> 8) & 0xff;
850 c->x86_phys_bits = eax & 0xff;
851 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
854 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
855 c->x86_phys_bits = 36;
857 c->x86_cache_bits = c->x86_phys_bits;
859 if (c->extended_cpuid_level >= 0x8000000a)
860 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
862 init_scattered_cpuid_features(c);
863 init_speculation_control(c);
867 * Clear/Set all flags overridden by options, after probe.
868 * This needs to happen each time we re-probe, which may happen
869 * several times during CPU initialization.
871 apply_forced_caps(c);
874 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
880 * First of all, decide if this is a 486 or higher
881 * It's a 486 if we can modify the AC flag
883 if (flag_is_changeable_p(X86_EFLAGS_AC))
888 for (i = 0; i < X86_VENDOR_NUM; i++)
889 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
890 c->x86_vendor_id[0] = 0;
891 cpu_devs[i]->c_identify(c);
892 if (c->x86_vendor_id[0]) {
900 #define NO_SPECULATION BIT(0)
901 #define NO_MELTDOWN BIT(1)
902 #define NO_SSB BIT(2)
903 #define NO_L1TF BIT(3)
904 #define NO_MDS BIT(4)
905 #define MSBDS_ONLY BIT(5)
906 #define NO_SWAPGS BIT(6)
907 #define NO_ITLB_MULTIHIT BIT(7)
908 #define NO_MMIO BIT(8)
909 #define NO_EIBRS_PBRSB BIT(9)
911 #define VULNWL(_vendor, _family, _model, _whitelist) \
912 { X86_VENDOR_##_vendor, _family, _model, X86_FEATURE_ANY, _whitelist }
914 #define VULNWL_INTEL(model, whitelist) \
915 VULNWL(INTEL, 6, INTEL_FAM6_##model, whitelist)
917 #define VULNWL_AMD(family, whitelist) \
918 VULNWL(AMD, family, X86_MODEL_ANY, whitelist)
920 static const __initconst struct x86_cpu_id cpu_vuln_whitelist[] = {
921 VULNWL(ANY, 4, X86_MODEL_ANY, NO_SPECULATION),
922 VULNWL(CENTAUR, 5, X86_MODEL_ANY, NO_SPECULATION),
923 VULNWL(INTEL, 5, X86_MODEL_ANY, NO_SPECULATION),
924 VULNWL(NSC, 5, X86_MODEL_ANY, NO_SPECULATION),
927 VULNWL_INTEL(TIGERLAKE, NO_MMIO),
928 VULNWL_INTEL(TIGERLAKE_L, NO_MMIO),
929 VULNWL_INTEL(ALDERLAKE, NO_MMIO),
930 VULNWL_INTEL(ALDERLAKE_L, NO_MMIO),
932 VULNWL_INTEL(ATOM_SALTWELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
933 VULNWL_INTEL(ATOM_SALTWELL_TABLET, NO_SPECULATION | NO_ITLB_MULTIHIT),
934 VULNWL_INTEL(ATOM_SALTWELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
935 VULNWL_INTEL(ATOM_BONNELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
936 VULNWL_INTEL(ATOM_BONNELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
938 VULNWL_INTEL(ATOM_SILVERMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
939 VULNWL_INTEL(ATOM_SILVERMONT_X, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
940 VULNWL_INTEL(ATOM_SILVERMONT_MID, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
941 VULNWL_INTEL(ATOM_AIRMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
942 VULNWL_INTEL(XEON_PHI_KNL, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
943 VULNWL_INTEL(XEON_PHI_KNM, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
945 VULNWL_INTEL(CORE_YONAH, NO_SSB),
947 VULNWL_INTEL(ATOM_AIRMONT_MID, NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
949 VULNWL_INTEL(ATOM_GOLDMONT, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
950 VULNWL_INTEL(ATOM_GOLDMONT_X, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
951 VULNWL_INTEL(ATOM_GOLDMONT_PLUS, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB),
954 * Technically, swapgs isn't serializing on AMD (despite it previously
955 * being documented as such in the APM). But according to AMD, %gs is
956 * updated non-speculatively, and the issuing of %gs-relative memory
957 * operands will be blocked until the %gs update completes, which is
958 * good enough for our purposes.
961 VULNWL_INTEL(ATOM_TREMONT, NO_EIBRS_PBRSB),
962 VULNWL_INTEL(ATOM_TREMONT_L, NO_EIBRS_PBRSB),
963 VULNWL_INTEL(ATOM_TREMONT_X, NO_ITLB_MULTIHIT | NO_EIBRS_PBRSB),
965 /* AMD Family 0xf - 0x12 */
966 VULNWL_AMD(0x0f, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
967 VULNWL_AMD(0x10, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
968 VULNWL_AMD(0x11, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
969 VULNWL_AMD(0x12, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
971 /* FAMILY_ANY must be last, otherwise 0x0f - 0x12 matches won't work */
972 VULNWL_AMD(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
976 #define VULNBL(vendor, family, model, blacklist) \
977 X86_MATCH_VENDOR_FAM_MODEL(vendor, family, model, blacklist)
979 #define VULNBL_INTEL_STEPPINGS(model, steppings, issues) \
980 X86_MATCH_VENDOR_FAM_MODEL_STEPPINGS_FEATURE(INTEL, 6, \
981 INTEL_FAM6_##model, steppings, \
982 X86_FEATURE_ANY, issues)
984 #define VULNBL_AMD(family, blacklist) \
985 VULNBL(AMD, family, X86_MODEL_ANY, blacklist)
988 /* CPU is affected by X86_BUG_MMIO_STALE_DATA */
990 /* CPU is affected by Shared Buffers Data Sampling (SBDS), a variant of X86_BUG_MMIO_STALE_DATA */
991 #define MMIO_SBDS BIT(2)
992 /* CPU is affected by RETbleed, speculating where you would not expect it */
993 #define RETBLEED BIT(3)
995 static const struct x86_cpu_id cpu_vuln_blacklist[] __initconst = {
996 VULNBL_INTEL_STEPPINGS(IVYBRIDGE, X86_STEPPING_ANY, SRBDS),
997 VULNBL_INTEL_STEPPINGS(HASWELL_CORE, X86_STEPPING_ANY, SRBDS),
998 VULNBL_INTEL_STEPPINGS(HASWELL_ULT, X86_STEPPING_ANY, SRBDS),
999 VULNBL_INTEL_STEPPINGS(HASWELL_GT3E, X86_STEPPING_ANY, SRBDS),
1000 VULNBL_INTEL_STEPPINGS(HASWELL_X, X86_STEPPING_ANY, MMIO),
1001 VULNBL_INTEL_STEPPINGS(BROADWELL_XEON_D,X86_STEPPING_ANY, MMIO),
1002 VULNBL_INTEL_STEPPINGS(BROADWELL_GT3E, X86_STEPPING_ANY, SRBDS),
1003 VULNBL_INTEL_STEPPINGS(BROADWELL_X, X86_STEPPING_ANY, MMIO),
1004 VULNBL_INTEL_STEPPINGS(BROADWELL_CORE, X86_STEPPING_ANY, SRBDS),
1005 VULNBL_INTEL_STEPPINGS(SKYLAKE_MOBILE, X86_STEPPING_ANY, SRBDS | MMIO | RETBLEED),
1006 VULNBL_INTEL_STEPPINGS(SKYLAKE_X, X86_STEPPING_ANY, MMIO | RETBLEED),
1007 VULNBL_INTEL_STEPPINGS(SKYLAKE_DESKTOP, X86_STEPPING_ANY, SRBDS | MMIO | RETBLEED),
1008 VULNBL_INTEL_STEPPINGS(KABYLAKE_MOBILE, X86_STEPPING_ANY, SRBDS | MMIO | RETBLEED),
1009 VULNBL_INTEL_STEPPINGS(KABYLAKE_DESKTOP,X86_STEPPING_ANY, SRBDS | MMIO | RETBLEED),
1010 VULNBL_INTEL_STEPPINGS(CANNONLAKE_MOBILE,X86_STEPPING_ANY, RETBLEED),
1011 VULNBL_INTEL_STEPPINGS(ICELAKE_MOBILE, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED),
1012 VULNBL_INTEL_STEPPINGS(ICELAKE_XEON_D, X86_STEPPING_ANY, MMIO),
1013 VULNBL_INTEL_STEPPINGS(ICELAKE_X, X86_STEPPING_ANY, MMIO),
1014 VULNBL_INTEL_STEPPINGS(COMETLAKE, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED),
1015 VULNBL_INTEL_STEPPINGS(COMETLAKE_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED),
1016 VULNBL_INTEL_STEPPINGS(LAKEFIELD, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED),
1017 VULNBL_INTEL_STEPPINGS(ROCKETLAKE, X86_STEPPING_ANY, MMIO | RETBLEED),
1018 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT, X86_STEPPING_ANY, MMIO | MMIO_SBDS),
1019 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT_X, X86_STEPPING_ANY, MMIO),
1020 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS),
1022 VULNBL_AMD(0x15, RETBLEED),
1023 VULNBL_AMD(0x16, RETBLEED),
1024 VULNBL_AMD(0x17, RETBLEED),
1028 static bool __init cpu_matches(const struct x86_cpu_id *table, unsigned long which)
1030 const struct x86_cpu_id *m = x86_match_cpu(table);
1032 return m && !!(m->driver_data & which);
1035 u64 x86_read_arch_cap_msr(void)
1039 if (boot_cpu_has(X86_FEATURE_ARCH_CAPABILITIES))
1040 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
1045 static bool arch_cap_mmio_immune(u64 ia32_cap)
1047 return (ia32_cap & ARCH_CAP_FBSDP_NO &&
1048 ia32_cap & ARCH_CAP_PSDP_NO &&
1049 ia32_cap & ARCH_CAP_SBDR_SSDP_NO);
1052 static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
1054 u64 ia32_cap = x86_read_arch_cap_msr();
1056 /* Set ITLB_MULTIHIT bug if cpu is not in the whitelist and not mitigated */
1057 if (!cpu_matches(cpu_vuln_whitelist, NO_ITLB_MULTIHIT) &&
1058 !(ia32_cap & ARCH_CAP_PSCHANGE_MC_NO))
1059 setup_force_cpu_bug(X86_BUG_ITLB_MULTIHIT);
1061 if (cpu_matches(cpu_vuln_whitelist, NO_SPECULATION))
1064 setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
1065 setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
1067 if (!cpu_matches(cpu_vuln_whitelist, NO_SSB) &&
1068 !(ia32_cap & ARCH_CAP_SSB_NO) &&
1069 !cpu_has(c, X86_FEATURE_AMD_SSB_NO))
1070 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
1072 if (ia32_cap & ARCH_CAP_IBRS_ALL)
1073 setup_force_cpu_cap(X86_FEATURE_IBRS_ENHANCED);
1075 if (!cpu_matches(cpu_vuln_whitelist, NO_MDS) &&
1076 !(ia32_cap & ARCH_CAP_MDS_NO)) {
1077 setup_force_cpu_bug(X86_BUG_MDS);
1078 if (cpu_matches(cpu_vuln_whitelist, MSBDS_ONLY))
1079 setup_force_cpu_bug(X86_BUG_MSBDS_ONLY);
1082 if (!cpu_matches(cpu_vuln_whitelist, NO_SWAPGS))
1083 setup_force_cpu_bug(X86_BUG_SWAPGS);
1086 * When the CPU is not mitigated for TAA (TAA_NO=0) set TAA bug when:
1087 * - TSX is supported or
1088 * - TSX_CTRL is present
1090 * TSX_CTRL check is needed for cases when TSX could be disabled before
1091 * the kernel boot e.g. kexec.
1092 * TSX_CTRL check alone is not sufficient for cases when the microcode
1093 * update is not present or running as guest that don't get TSX_CTRL.
1095 if (!(ia32_cap & ARCH_CAP_TAA_NO) &&
1096 (cpu_has(c, X86_FEATURE_RTM) ||
1097 (ia32_cap & ARCH_CAP_TSX_CTRL_MSR)))
1098 setup_force_cpu_bug(X86_BUG_TAA);
1101 * SRBDS affects CPUs which support RDRAND or RDSEED and are listed
1102 * in the vulnerability blacklist.
1104 * Some of the implications and mitigation of Shared Buffers Data
1105 * Sampling (SBDS) are similar to SRBDS. Give SBDS same treatment as
1108 if ((cpu_has(c, X86_FEATURE_RDRAND) ||
1109 cpu_has(c, X86_FEATURE_RDSEED)) &&
1110 cpu_matches(cpu_vuln_blacklist, SRBDS | MMIO_SBDS))
1111 setup_force_cpu_bug(X86_BUG_SRBDS);
1114 * Processor MMIO Stale Data bug enumeration
1116 * Affected CPU list is generally enough to enumerate the vulnerability,
1117 * but for virtualization case check for ARCH_CAP MSR bits also, VMM may
1118 * not want the guest to enumerate the bug.
1120 * Set X86_BUG_MMIO_UNKNOWN for CPUs that are neither in the blacklist,
1121 * nor in the whitelist and also don't enumerate MSR ARCH_CAP MMIO bits.
1123 if (!arch_cap_mmio_immune(ia32_cap)) {
1124 if (cpu_matches(cpu_vuln_blacklist, MMIO))
1125 setup_force_cpu_bug(X86_BUG_MMIO_STALE_DATA);
1126 else if (!cpu_matches(cpu_vuln_whitelist, NO_MMIO))
1127 setup_force_cpu_bug(X86_BUG_MMIO_UNKNOWN);
1130 if (!cpu_has(c, X86_FEATURE_BTC_NO)) {
1131 if (cpu_matches(cpu_vuln_blacklist, RETBLEED) || (ia32_cap & ARCH_CAP_RSBA))
1132 setup_force_cpu_bug(X86_BUG_RETBLEED);
1135 if (cpu_has(c, X86_FEATURE_IBRS_ENHANCED) &&
1136 !cpu_matches(cpu_vuln_whitelist, NO_EIBRS_PBRSB) &&
1137 !(ia32_cap & ARCH_CAP_PBRSB_NO))
1138 setup_force_cpu_bug(X86_BUG_EIBRS_PBRSB);
1140 if (cpu_matches(cpu_vuln_whitelist, NO_MELTDOWN))
1143 /* Rogue Data Cache Load? No! */
1144 if (ia32_cap & ARCH_CAP_RDCL_NO)
1147 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
1149 if (cpu_matches(cpu_vuln_whitelist, NO_L1TF))
1152 setup_force_cpu_bug(X86_BUG_L1TF);
1156 * Do minimum CPU detection early.
1157 * Fields really needed: vendor, cpuid_level, family, model, mask,
1159 * The others are not touched to avoid unwanted side effects.
1161 * WARNING: this function is only called on the BP. Don't add code here
1162 * that is supposed to run on all CPUs.
1164 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
1166 #ifdef CONFIG_X86_64
1167 c->x86_clflush_size = 64;
1168 c->x86_phys_bits = 36;
1169 c->x86_virt_bits = 48;
1171 c->x86_clflush_size = 32;
1172 c->x86_phys_bits = 32;
1173 c->x86_virt_bits = 32;
1175 c->x86_cache_alignment = c->x86_clflush_size;
1177 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1178 c->extended_cpuid_level = 0;
1180 if (!have_cpuid_p())
1181 identify_cpu_without_cpuid(c);
1183 /* cyrix could have cpuid enabled via c_identify()*/
1184 if (have_cpuid_p()) {
1188 setup_force_cpu_cap(X86_FEATURE_CPUID);
1190 if (this_cpu->c_early_init)
1191 this_cpu->c_early_init(c);
1194 filter_cpuid_features(c, false);
1196 if (this_cpu->c_bsp_init)
1197 this_cpu->c_bsp_init(c);
1199 setup_clear_cpu_cap(X86_FEATURE_CPUID);
1202 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1204 cpu_set_bug_bits(c);
1206 fpu__init_system(c);
1208 #ifdef CONFIG_X86_32
1210 * Regardless of whether PCID is enumerated, the SDM says
1211 * that it can't be enabled in 32-bit mode.
1213 setup_clear_cpu_cap(X86_FEATURE_PCID);
1217 void __init early_cpu_init(void)
1219 const struct cpu_dev *const *cdev;
1222 #ifdef CONFIG_PROCESSOR_SELECT
1223 pr_info("KERNEL supported cpus:\n");
1226 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1227 const struct cpu_dev *cpudev = *cdev;
1229 if (count >= X86_VENDOR_NUM)
1231 cpu_devs[count] = cpudev;
1234 #ifdef CONFIG_PROCESSOR_SELECT
1238 for (j = 0; j < 2; j++) {
1239 if (!cpudev->c_ident[j])
1241 pr_info(" %s %s\n", cpudev->c_vendor,
1242 cpudev->c_ident[j]);
1247 early_identify_cpu(&boot_cpu_data);
1251 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1252 * unfortunately, that's not true in practice because of early VIA
1253 * chips and (more importantly) broken virtualizers that are not easy
1254 * to detect. In the latter case it doesn't even *fail* reliably, so
1255 * probing for it doesn't even work. Disable it completely on 32-bit
1256 * unless we can find a reliable way to detect all the broken cases.
1257 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1259 static void detect_nopl(struct cpuinfo_x86 *c)
1261 #ifdef CONFIG_X86_32
1262 clear_cpu_cap(c, X86_FEATURE_NOPL);
1264 set_cpu_cap(c, X86_FEATURE_NOPL);
1268 static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
1270 #ifdef CONFIG_X86_64
1272 * Empirically, writing zero to a segment selector on AMD does
1273 * not clear the base, whereas writing zero to a segment
1274 * selector on Intel does clear the base. Intel's behavior
1275 * allows slightly faster context switches in the common case
1276 * where GS is unused by the prev and next threads.
1278 * Since neither vendor documents this anywhere that I can see,
1279 * detect it directly instead of hardcoding the choice by
1282 * I've designated AMD's behavior as the "bug" because it's
1283 * counterintuitive and less friendly.
1286 unsigned long old_base, tmp;
1287 rdmsrl(MSR_FS_BASE, old_base);
1288 wrmsrl(MSR_FS_BASE, 1);
1290 rdmsrl(MSR_FS_BASE, tmp);
1292 set_cpu_bug(c, X86_BUG_NULL_SEG);
1293 wrmsrl(MSR_FS_BASE, old_base);
1297 static void generic_identify(struct cpuinfo_x86 *c)
1299 c->extended_cpuid_level = 0;
1301 if (!have_cpuid_p())
1302 identify_cpu_without_cpuid(c);
1304 /* cyrix could have cpuid enabled via c_identify()*/
1305 if (!have_cpuid_p())
1314 if (c->cpuid_level >= 0x00000001) {
1315 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1316 #ifdef CONFIG_X86_32
1318 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1320 c->apicid = c->initial_apicid;
1323 c->phys_proc_id = c->initial_apicid;
1326 get_model_name(c); /* Default name */
1330 detect_null_seg_behavior(c);
1333 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1334 * systems that run Linux at CPL > 0 may or may not have the
1335 * issue, but, even if they have the issue, there's absolutely
1336 * nothing we can do about it because we can't use the real IRET
1339 * NB: For the time being, only 32-bit kernels support
1340 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1341 * whether to apply espfix using paravirt hooks. If any
1342 * non-paravirt system ever shows up that does *not* have the
1343 * ESPFIX issue, we can change this.
1345 #ifdef CONFIG_X86_32
1346 # ifdef CONFIG_PARAVIRT
1348 extern void native_iret(void);
1349 if (pv_cpu_ops.iret == native_iret)
1350 set_cpu_bug(c, X86_BUG_ESPFIX);
1353 set_cpu_bug(c, X86_BUG_ESPFIX);
1358 static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1361 * The heavy lifting of max_rmid and cache_occ_scale are handled
1362 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
1363 * in case CQM bits really aren't there in this CPU.
1365 if (c != &boot_cpu_data) {
1366 boot_cpu_data.x86_cache_max_rmid =
1367 min(boot_cpu_data.x86_cache_max_rmid,
1368 c->x86_cache_max_rmid);
1373 * Validate that ACPI/mptables have the same information about the
1374 * effective APIC id and update the package map.
1376 static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1379 unsigned int apicid, cpu = smp_processor_id();
1381 apicid = apic->cpu_present_to_apicid(cpu);
1383 if (apicid != c->apicid) {
1384 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1385 cpu, apicid, c->initial_apicid);
1387 BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
1389 c->logical_proc_id = 0;
1394 * This does the hard work of actually picking apart the CPU stuff...
1396 static void identify_cpu(struct cpuinfo_x86 *c)
1400 c->loops_per_jiffy = loops_per_jiffy;
1401 c->x86_cache_size = 0;
1402 c->x86_vendor = X86_VENDOR_UNKNOWN;
1403 c->x86_model = c->x86_stepping = 0; /* So far unknown... */
1404 c->x86_vendor_id[0] = '\0'; /* Unset */
1405 c->x86_model_id[0] = '\0'; /* Unset */
1406 c->x86_max_cores = 1;
1407 c->x86_coreid_bits = 0;
1409 #ifdef CONFIG_X86_64
1410 c->x86_clflush_size = 64;
1411 c->x86_phys_bits = 36;
1412 c->x86_virt_bits = 48;
1414 c->cpuid_level = -1; /* CPUID not detected */
1415 c->x86_clflush_size = 32;
1416 c->x86_phys_bits = 32;
1417 c->x86_virt_bits = 32;
1419 c->x86_cache_alignment = c->x86_clflush_size;
1420 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1422 generic_identify(c);
1424 if (this_cpu->c_identify)
1425 this_cpu->c_identify(c);
1427 /* Clear/Set all flags overridden by options, after probe */
1428 apply_forced_caps(c);
1430 #ifdef CONFIG_X86_64
1431 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1435 * Vendor-specific initialization. In this section we
1436 * canonicalize the feature flags, meaning if there are
1437 * features a certain CPU supports which CPUID doesn't
1438 * tell us, CPUID claiming incorrect flags, or other bugs,
1439 * we handle them here.
1441 * At the end of this section, c->x86_capability better
1442 * indicate the features this CPU genuinely supports!
1444 if (this_cpu->c_init)
1445 this_cpu->c_init(c);
1447 /* Disable the PN if appropriate */
1448 squash_the_stupid_serial_number(c);
1450 /* Set up SMEP/SMAP */
1455 * The vendor-specific functions might have changed features.
1456 * Now we do "generic changes."
1459 /* Filter out anything that depends on CPUID levels we don't have */
1460 filter_cpuid_features(c, true);
1462 /* If the model name is still unset, do table lookup. */
1463 if (!c->x86_model_id[0]) {
1465 p = table_lookup_model(c);
1467 strcpy(c->x86_model_id, p);
1469 /* Last resort... */
1470 sprintf(c->x86_model_id, "%02x/%02x",
1471 c->x86, c->x86_model);
1474 #ifdef CONFIG_X86_64
1479 x86_init_cache_qos(c);
1483 * Clear/Set all flags overridden by options, need do it
1484 * before following smp all cpus cap AND.
1486 apply_forced_caps(c);
1489 * On SMP, boot_cpu_data holds the common feature set between
1490 * all CPUs; so make sure that we indicate which features are
1491 * common between the CPUs. The first time this routine gets
1492 * executed, c == &boot_cpu_data.
1494 if (c != &boot_cpu_data) {
1495 /* AND the already accumulated flags with these */
1496 for (i = 0; i < NCAPINTS; i++)
1497 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1499 /* OR, i.e. replicate the bug flags */
1500 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1501 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1504 /* Init Machine Check Exception if available. */
1507 select_idle_routine(c);
1510 numa_add_cpu(smp_processor_id());
1515 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1516 * on 32-bit kernels:
1518 #ifdef CONFIG_X86_32
1519 void enable_sep_cpu(void)
1521 struct tss_struct *tss;
1524 if (!boot_cpu_has(X86_FEATURE_SEP))
1528 tss = &per_cpu(cpu_tss_rw, cpu);
1531 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1532 * see the big comment in struct x86_hw_tss's definition.
1535 tss->x86_tss.ss1 = __KERNEL_CS;
1536 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1537 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
1538 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1544 void __init identify_boot_cpu(void)
1546 identify_cpu(&boot_cpu_data);
1547 #ifdef CONFIG_X86_32
1551 cpu_detect_tlb(&boot_cpu_data);
1555 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1557 BUG_ON(c == &boot_cpu_data);
1559 #ifdef CONFIG_X86_32
1563 validate_apic_and_package_id(c);
1564 x86_spec_ctrl_setup_ap();
1568 static __init int setup_noclflush(char *arg)
1570 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1571 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1574 __setup("noclflush", setup_noclflush);
1576 void print_cpu_info(struct cpuinfo_x86 *c)
1578 const char *vendor = NULL;
1580 if (c->x86_vendor < X86_VENDOR_NUM) {
1581 vendor = this_cpu->c_vendor;
1583 if (c->cpuid_level >= 0)
1584 vendor = c->x86_vendor_id;
1587 if (vendor && !strstr(c->x86_model_id, vendor))
1588 pr_cont("%s ", vendor);
1590 if (c->x86_model_id[0])
1591 pr_cont("%s", c->x86_model_id);
1593 pr_cont("%d86", c->x86);
1595 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1597 if (c->x86_stepping || c->cpuid_level >= 0)
1598 pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
1604 * clearcpuid= was already parsed in fpu__init_parse_early_param.
1605 * But we need to keep a dummy __setup around otherwise it would
1606 * show up as an environment variable for init.
1608 static __init int setup_clearcpuid(char *arg)
1612 __setup("clearcpuid=", setup_clearcpuid);
1614 #ifdef CONFIG_X86_64
1615 DEFINE_PER_CPU_FIRST(union irq_stack_union,
1616 irq_stack_union) __aligned(PAGE_SIZE) __visible;
1619 * The following percpu variables are hot. Align current_task to
1620 * cacheline size such that they fall in the same cacheline.
1622 DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1624 EXPORT_PER_CPU_SYMBOL(current_task);
1626 DEFINE_PER_CPU(char *, irq_stack_ptr) =
1627 init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE;
1629 DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1631 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1632 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1634 /* May not be marked __init: used by software suspend */
1635 void syscall_init(void)
1637 extern char _entry_trampoline[];
1638 extern char entry_SYSCALL_64_trampoline[];
1640 int cpu = smp_processor_id();
1641 unsigned long SYSCALL64_entry_trampoline =
1642 (unsigned long)get_cpu_entry_area(cpu)->entry_trampoline +
1643 (entry_SYSCALL_64_trampoline - _entry_trampoline);
1645 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1646 if (static_cpu_has(X86_FEATURE_PTI))
1647 wrmsrl(MSR_LSTAR, SYSCALL64_entry_trampoline);
1649 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1651 #ifdef CONFIG_IA32_EMULATION
1652 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1654 * This only works on Intel CPUs.
1655 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1656 * This does not cause SYSENTER to jump to the wrong location, because
1657 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1659 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1660 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1));
1661 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1663 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1664 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1665 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1666 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1669 /* Flags to clear on syscall */
1670 wrmsrl(MSR_SYSCALL_MASK,
1671 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1672 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1676 * Copies of the original ist values from the tss are only accessed during
1677 * debugging, no special alignment required.
1679 DEFINE_PER_CPU(struct orig_ist, orig_ist);
1681 static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1682 DEFINE_PER_CPU(int, debug_stack_usage);
1684 int is_debug_stack(unsigned long addr)
1686 return __this_cpu_read(debug_stack_usage) ||
1687 (addr <= __this_cpu_read(debug_stack_addr) &&
1688 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1690 NOKPROBE_SYMBOL(is_debug_stack);
1692 DEFINE_PER_CPU(u32, debug_idt_ctr);
1694 void debug_stack_set_zero(void)
1696 this_cpu_inc(debug_idt_ctr);
1699 NOKPROBE_SYMBOL(debug_stack_set_zero);
1701 void debug_stack_reset(void)
1703 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1705 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1708 NOKPROBE_SYMBOL(debug_stack_reset);
1710 #else /* CONFIG_X86_64 */
1712 DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1713 EXPORT_PER_CPU_SYMBOL(current_task);
1714 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1715 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1718 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1719 * the top of the kernel stack. Use an extra percpu variable to track the
1720 * top of the kernel stack directly.
1722 DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1723 (unsigned long)&init_thread_union + THREAD_SIZE;
1724 EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1726 #ifdef CONFIG_CC_STACKPROTECTOR
1727 DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1730 #endif /* CONFIG_X86_64 */
1733 * Clear all 6 debug registers:
1735 static void clear_all_debug_regs(void)
1739 for (i = 0; i < 8; i++) {
1740 /* Ignore db4, db5 */
1741 if ((i == 4) || (i == 5))
1750 * Restore debug regs if using kgdbwait and you have a kernel debugger
1751 * connection established.
1753 static void dbg_restore_debug_regs(void)
1755 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1756 arch_kgdb_ops.correct_hw_break();
1758 #else /* ! CONFIG_KGDB */
1759 #define dbg_restore_debug_regs()
1760 #endif /* ! CONFIG_KGDB */
1762 static void wait_for_master_cpu(int cpu)
1766 * wait for ACK from master CPU before continuing
1767 * with AP initialization
1769 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1770 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1776 * cpu_init() initializes state that is per-CPU. Some data is already
1777 * initialized (naturally) in the bootstrap process, such as the GDT
1778 * and IDT. We reload them nevertheless, this function acts as a
1779 * 'CPU state barrier', nothing should get across.
1780 * A lot of state is already set up in PDA init for 64 bit
1782 #ifdef CONFIG_X86_64
1786 struct orig_ist *oist;
1787 struct task_struct *me;
1788 struct tss_struct *t;
1790 int cpu = raw_smp_processor_id();
1793 wait_for_master_cpu(cpu);
1796 * Initialize the CR4 shadow before doing anything that could
1804 t = &per_cpu(cpu_tss_rw, cpu);
1805 oist = &per_cpu(orig_ist, cpu);
1808 if (this_cpu_read(numa_node) == 0 &&
1809 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1810 set_numa_node(early_cpu_to_node(cpu));
1815 pr_debug("Initializing CPU#%d\n", cpu);
1817 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1820 * Initialize the per-CPU GDT with the boot GDT,
1821 * and set up the GDT descriptor:
1824 switch_to_new_gdt(cpu);
1829 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1832 wrmsrl(MSR_FS_BASE, 0);
1833 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1840 * set up and load the per-CPU TSS
1842 if (!oist->ist[0]) {
1843 char *estacks = get_cpu_entry_area(cpu)->exception_stacks;
1845 for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1846 estacks += exception_stack_sizes[v];
1847 oist->ist[v] = t->x86_tss.ist[v] =
1848 (unsigned long)estacks;
1849 if (v == DEBUG_STACK-1)
1850 per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1854 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1857 * <= is required because the CPU will access up to
1858 * 8 bits beyond the end of the IO permission bitmap.
1860 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1861 t->io_bitmap[i] = ~0UL;
1864 me->active_mm = &init_mm;
1866 initialize_tlbstate_and_flush();
1867 enter_lazy_tlb(&init_mm, me);
1870 * Initialize the TSS. sp0 points to the entry trampoline stack
1871 * regardless of what task is running.
1873 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1875 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1877 load_mm_ldt(&init_mm);
1879 clear_all_debug_regs();
1880 dbg_restore_debug_regs();
1887 load_fixmap_gdt(cpu);
1894 int cpu = smp_processor_id();
1895 struct task_struct *curr = current;
1896 struct tss_struct *t = &per_cpu(cpu_tss_rw, cpu);
1898 wait_for_master_cpu(cpu);
1901 * Initialize the CR4 shadow before doing anything that could
1906 show_ucode_info_early();
1908 pr_info("Initializing CPU#%d\n", cpu);
1910 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1911 boot_cpu_has(X86_FEATURE_TSC) ||
1912 boot_cpu_has(X86_FEATURE_DE))
1913 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1916 switch_to_new_gdt(cpu);
1919 * Set up and load the per-CPU TSS and LDT
1922 curr->active_mm = &init_mm;
1924 initialize_tlbstate_and_flush();
1925 enter_lazy_tlb(&init_mm, curr);
1928 * Initialize the TSS. Don't bother initializing sp0, as the initial
1929 * task never enters user mode.
1931 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1934 load_mm_ldt(&init_mm);
1936 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1938 #ifdef CONFIG_DOUBLEFAULT
1939 /* Set up doublefault TSS pointer in the GDT */
1940 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1943 clear_all_debug_regs();
1944 dbg_restore_debug_regs();
1948 load_fixmap_gdt(cpu);
1952 static void bsp_resume(void)
1954 if (this_cpu->c_bsp_resume)
1955 this_cpu->c_bsp_resume(&boot_cpu_data);
1958 static struct syscore_ops cpu_syscore_ops = {
1959 .resume = bsp_resume,
1962 static int __init init_cpu_syscore(void)
1964 register_syscore_ops(&cpu_syscore_ops);
1967 core_initcall(init_cpu_syscore);
1970 * The microcode loader calls this upon late microcode load to recheck features,
1971 * only when microcode has been updated. Caller holds microcode_mutex and CPU
1974 void microcode_check(void)
1976 struct cpuinfo_x86 info;
1978 perf_check_microcode();
1980 /* Reload CPUID max function as it might've changed. */
1981 info.cpuid_level = cpuid_eax(0);
1984 * Copy all capability leafs to pick up the synthetic ones so that
1985 * memcmp() below doesn't fail on that. The ones coming from CPUID will
1986 * get overwritten in get_cpu_cap().
1988 memcpy(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability));
1992 if (!memcmp(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability)))
1995 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
1996 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");