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.h>
11 #include <linux/init.h>
12 #include <linux/kprobes.h>
13 #include <linux/kgdb.h>
14 #include <linux/smp.h>
16 #include <linux/syscore_ops.h>
18 #include <asm/stackprotector.h>
19 #include <asm/perf_event.h>
20 #include <asm/mmu_context.h>
21 #include <asm/archrandom.h>
22 #include <asm/hypervisor.h>
23 #include <asm/processor.h>
24 #include <asm/tlbflush.h>
25 #include <asm/debugreg.h>
26 #include <asm/sections.h>
27 #include <asm/vsyscall.h>
28 #include <linux/topology.h>
29 #include <linux/cpumask.h>
30 #include <asm/pgtable.h>
31 #include <linux/atomic.h>
32 #include <asm/proto.h>
33 #include <asm/setup.h>
36 #include <asm/fpu/internal.h>
38 #include <linux/numa.h>
45 #include <asm/microcode.h>
46 #include <asm/microcode_intel.h>
48 #ifdef CONFIG_X86_LOCAL_APIC
49 #include <asm/uv/uv.h>
54 /* all of these masks are initialized in setup_cpu_local_masks() */
55 cpumask_var_t cpu_initialized_mask;
56 cpumask_var_t cpu_callout_mask;
57 cpumask_var_t cpu_callin_mask;
59 /* representing cpus for which sibling maps can be computed */
60 cpumask_var_t cpu_sibling_setup_mask;
62 /* correctly size the local cpu masks */
63 void __init setup_cpu_local_masks(void)
65 alloc_bootmem_cpumask_var(&cpu_initialized_mask);
66 alloc_bootmem_cpumask_var(&cpu_callin_mask);
67 alloc_bootmem_cpumask_var(&cpu_callout_mask);
68 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
71 static void default_init(struct cpuinfo_x86 *c)
74 cpu_detect_cache_sizes(c);
76 /* Not much we can do here... */
77 /* Check if at least it has cpuid */
78 if (c->cpuid_level == -1) {
79 /* No cpuid. It must be an ancient CPU */
81 strcpy(c->x86_model_id, "486");
83 strcpy(c->x86_model_id, "386");
88 static const struct cpu_dev default_cpu = {
89 .c_init = default_init,
90 .c_vendor = "Unknown",
91 .c_x86_vendor = X86_VENDOR_UNKNOWN,
94 static const struct cpu_dev *this_cpu = &default_cpu;
96 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
99 * We need valid kernel segments for data and code in long mode too
100 * IRET will check the segment types kkeil 2000/10/28
101 * Also sysret mandates a special GDT layout
103 * TLS descriptors are currently at a different place compared to i386.
104 * Hopefully nobody expects them at a fixed place (Wine?)
106 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
107 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
108 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
109 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
110 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
111 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
113 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
114 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
115 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
116 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
118 * Segments used for calling PnP BIOS have byte granularity.
119 * They code segments and data segments have fixed 64k limits,
120 * the transfer segment sizes are set at run time.
123 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
125 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
127 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(0x0092, 0, 0xffff),
129 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(0x0092, 0, 0),
131 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(0x0092, 0, 0),
133 * The APM segments have byte granularity and their bases
134 * are set at run time. All have 64k limits.
137 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
139 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
141 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(0x4092, 0, 0xffff),
143 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
144 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
145 GDT_STACK_CANARY_INIT
148 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
150 static int __init x86_mpx_setup(char *s)
152 /* require an exact match without trailing characters */
156 /* do not emit a message if the feature is not present */
157 if (!boot_cpu_has(X86_FEATURE_MPX))
160 setup_clear_cpu_cap(X86_FEATURE_MPX);
161 pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
164 __setup("nompx", x86_mpx_setup);
166 static int __init x86_noinvpcid_setup(char *s)
168 /* noinvpcid doesn't accept parameters */
172 /* do not emit a message if the feature is not present */
173 if (!boot_cpu_has(X86_FEATURE_INVPCID))
176 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
177 pr_info("noinvpcid: INVPCID feature disabled\n");
180 early_param("noinvpcid", x86_noinvpcid_setup);
183 static int cachesize_override = -1;
184 static int disable_x86_serial_nr = 1;
186 static int __init cachesize_setup(char *str)
188 get_option(&str, &cachesize_override);
191 __setup("cachesize=", cachesize_setup);
193 static int __init x86_sep_setup(char *s)
195 setup_clear_cpu_cap(X86_FEATURE_SEP);
198 __setup("nosep", x86_sep_setup);
200 /* Standard macro to see if a specific flag is changeable */
201 static inline int flag_is_changeable_p(u32 flag)
206 * Cyrix and IDT cpus allow disabling of CPUID
207 * so the code below may return different results
208 * when it is executed before and after enabling
209 * the CPUID. Add "volatile" to not allow gcc to
210 * optimize the subsequent calls to this function.
212 asm volatile ("pushfl \n\t"
223 : "=&r" (f1), "=&r" (f2)
226 return ((f1^f2) & flag) != 0;
229 /* Probe for the CPUID instruction */
230 int have_cpuid_p(void)
232 return flag_is_changeable_p(X86_EFLAGS_ID);
235 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
237 unsigned long lo, hi;
239 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
242 /* Disable processor serial number: */
244 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
246 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
248 pr_notice("CPU serial number disabled.\n");
249 clear_cpu_cap(c, X86_FEATURE_PN);
251 /* Disabling the serial number may affect the cpuid level */
252 c->cpuid_level = cpuid_eax(0);
255 static int __init x86_serial_nr_setup(char *s)
257 disable_x86_serial_nr = 0;
260 __setup("serialnumber", x86_serial_nr_setup);
262 static inline int flag_is_changeable_p(u32 flag)
266 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
271 static __init int setup_disable_smep(char *arg)
273 setup_clear_cpu_cap(X86_FEATURE_SMEP);
274 /* Check for things that depend on SMEP being enabled: */
275 check_mpx_erratum(&boot_cpu_data);
278 __setup("nosmep", setup_disable_smep);
280 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
282 if (cpu_has(c, X86_FEATURE_SMEP))
283 cr4_set_bits(X86_CR4_SMEP);
286 static __init int setup_disable_smap(char *arg)
288 setup_clear_cpu_cap(X86_FEATURE_SMAP);
291 __setup("nosmap", setup_disable_smap);
293 static __always_inline void setup_smap(struct cpuinfo_x86 *c)
295 unsigned long eflags = native_save_fl();
297 /* This should have been cleared long ago */
298 BUG_ON(eflags & X86_EFLAGS_AC);
300 if (cpu_has(c, X86_FEATURE_SMAP)) {
301 #ifdef CONFIG_X86_SMAP
302 cr4_set_bits(X86_CR4_SMAP);
304 cr4_clear_bits(X86_CR4_SMAP);
310 * Protection Keys are not available in 32-bit mode.
312 static bool pku_disabled;
314 static __always_inline void setup_pku(struct cpuinfo_x86 *c)
316 /* check the boot processor, plus compile options for PKU: */
317 if (!cpu_feature_enabled(X86_FEATURE_PKU))
319 /* checks the actual processor's cpuid bits: */
320 if (!cpu_has(c, X86_FEATURE_PKU))
325 cr4_set_bits(X86_CR4_PKE);
327 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
328 * cpuid bit to be set. We need to ensure that we
329 * update that bit in this CPU's "cpu_info".
334 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
335 static __init int setup_disable_pku(char *arg)
338 * Do not clear the X86_FEATURE_PKU bit. All of the
339 * runtime checks are against OSPKE so clearing the
342 * This way, we will see "pku" in cpuinfo, but not
343 * "ospke", which is exactly what we want. It shows
344 * that the CPU has PKU, but the OS has not enabled it.
345 * This happens to be exactly how a system would look
346 * if we disabled the config option.
348 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
352 __setup("nopku", setup_disable_pku);
353 #endif /* CONFIG_X86_64 */
356 * Some CPU features depend on higher CPUID levels, which may not always
357 * be available due to CPUID level capping or broken virtualization
358 * software. Add those features to this table to auto-disable them.
360 struct cpuid_dependent_feature {
365 static const struct cpuid_dependent_feature
366 cpuid_dependent_features[] = {
367 { X86_FEATURE_MWAIT, 0x00000005 },
368 { X86_FEATURE_DCA, 0x00000009 },
369 { X86_FEATURE_XSAVE, 0x0000000d },
373 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
375 const struct cpuid_dependent_feature *df;
377 for (df = cpuid_dependent_features; df->feature; df++) {
379 if (!cpu_has(c, df->feature))
382 * Note: cpuid_level is set to -1 if unavailable, but
383 * extended_extended_level is set to 0 if unavailable
384 * and the legitimate extended levels are all negative
385 * when signed; hence the weird messing around with
388 if (!((s32)df->level < 0 ?
389 (u32)df->level > (u32)c->extended_cpuid_level :
390 (s32)df->level > (s32)c->cpuid_level))
393 clear_cpu_cap(c, df->feature);
397 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
398 x86_cap_flag(df->feature), df->level);
403 * Naming convention should be: <Name> [(<Codename>)]
404 * This table only is used unless init_<vendor>() below doesn't set it;
405 * in particular, if CPUID levels 0x80000002..4 are supported, this
409 /* Look up CPU names by table lookup. */
410 static const char *table_lookup_model(struct cpuinfo_x86 *c)
413 const struct legacy_cpu_model_info *info;
415 if (c->x86_model >= 16)
416 return NULL; /* Range check */
421 info = this_cpu->legacy_models;
423 while (info->family) {
424 if (info->family == c->x86)
425 return info->model_names[c->x86_model];
429 return NULL; /* Not found */
432 __u32 cpu_caps_cleared[NCAPINTS];
433 __u32 cpu_caps_set[NCAPINTS];
435 void load_percpu_segment(int cpu)
438 loadsegment(fs, __KERNEL_PERCPU);
440 __loadsegment_simple(gs, 0);
441 wrmsrl(MSR_GS_BASE, (unsigned long)per_cpu(irq_stack_union.gs_base, cpu));
443 load_stack_canary_segment();
447 * Current gdt points %fs at the "master" per-cpu area: after this,
448 * it's on the real one.
450 void switch_to_new_gdt(int cpu)
452 struct desc_ptr gdt_descr;
454 gdt_descr.address = (long)get_cpu_gdt_table(cpu);
455 gdt_descr.size = GDT_SIZE - 1;
456 load_gdt(&gdt_descr);
457 /* Reload the per-cpu base */
459 load_percpu_segment(cpu);
462 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
464 static void get_model_name(struct cpuinfo_x86 *c)
469 if (c->extended_cpuid_level < 0x80000004)
472 v = (unsigned int *)c->x86_model_id;
473 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
474 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
475 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
476 c->x86_model_id[48] = 0;
478 /* Trim whitespace */
479 p = q = s = &c->x86_model_id[0];
485 /* Note the last non-whitespace index */
495 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
497 unsigned int n, dummy, ebx, ecx, edx, l2size;
499 n = c->extended_cpuid_level;
501 if (n >= 0x80000005) {
502 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
503 c->x86_cache_size = (ecx>>24) + (edx>>24);
505 /* On K8 L1 TLB is inclusive, so don't count it */
510 if (n < 0x80000006) /* Some chips just has a large L1. */
513 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
517 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
519 /* do processor-specific cache resizing */
520 if (this_cpu->legacy_cache_size)
521 l2size = this_cpu->legacy_cache_size(c, l2size);
523 /* Allow user to override all this if necessary. */
524 if (cachesize_override != -1)
525 l2size = cachesize_override;
528 return; /* Again, no L2 cache is possible */
531 c->x86_cache_size = l2size;
534 u16 __read_mostly tlb_lli_4k[NR_INFO];
535 u16 __read_mostly tlb_lli_2m[NR_INFO];
536 u16 __read_mostly tlb_lli_4m[NR_INFO];
537 u16 __read_mostly tlb_lld_4k[NR_INFO];
538 u16 __read_mostly tlb_lld_2m[NR_INFO];
539 u16 __read_mostly tlb_lld_4m[NR_INFO];
540 u16 __read_mostly tlb_lld_1g[NR_INFO];
542 static void cpu_detect_tlb(struct cpuinfo_x86 *c)
544 if (this_cpu->c_detect_tlb)
545 this_cpu->c_detect_tlb(c);
547 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
548 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
549 tlb_lli_4m[ENTRIES]);
551 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
552 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
553 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
556 void detect_ht(struct cpuinfo_x86 *c)
559 u32 eax, ebx, ecx, edx;
560 int index_msb, core_bits;
563 if (!cpu_has(c, X86_FEATURE_HT))
566 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
569 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
572 cpuid(1, &eax, &ebx, &ecx, &edx);
574 smp_num_siblings = (ebx & 0xff0000) >> 16;
576 if (smp_num_siblings == 1) {
577 pr_info_once("CPU0: Hyper-Threading is disabled\n");
581 if (smp_num_siblings <= 1)
584 index_msb = get_count_order(smp_num_siblings);
585 c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
587 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
589 index_msb = get_count_order(smp_num_siblings);
591 core_bits = get_count_order(c->x86_max_cores);
593 c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
594 ((1 << core_bits) - 1);
597 if (!printed && (c->x86_max_cores * smp_num_siblings) > 1) {
598 pr_info("CPU: Physical Processor ID: %d\n",
600 pr_info("CPU: Processor Core ID: %d\n",
607 static void get_cpu_vendor(struct cpuinfo_x86 *c)
609 char *v = c->x86_vendor_id;
612 for (i = 0; i < X86_VENDOR_NUM; i++) {
616 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
617 (cpu_devs[i]->c_ident[1] &&
618 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
620 this_cpu = cpu_devs[i];
621 c->x86_vendor = this_cpu->c_x86_vendor;
626 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
627 "CPU: Your system may be unstable.\n", v);
629 c->x86_vendor = X86_VENDOR_UNKNOWN;
630 this_cpu = &default_cpu;
633 void cpu_detect(struct cpuinfo_x86 *c)
635 /* Get vendor name */
636 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
637 (unsigned int *)&c->x86_vendor_id[0],
638 (unsigned int *)&c->x86_vendor_id[8],
639 (unsigned int *)&c->x86_vendor_id[4]);
642 /* Intel-defined flags: level 0x00000001 */
643 if (c->cpuid_level >= 0x00000001) {
644 u32 junk, tfms, cap0, misc;
646 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
647 c->x86 = x86_family(tfms);
648 c->x86_model = x86_model(tfms);
649 c->x86_mask = x86_stepping(tfms);
651 if (cap0 & (1<<19)) {
652 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
653 c->x86_cache_alignment = c->x86_clflush_size;
658 void get_cpu_cap(struct cpuinfo_x86 *c)
660 u32 eax, ebx, ecx, edx;
662 /* Intel-defined flags: level 0x00000001 */
663 if (c->cpuid_level >= 0x00000001) {
664 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
666 c->x86_capability[CPUID_1_ECX] = ecx;
667 c->x86_capability[CPUID_1_EDX] = edx;
670 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
671 if (c->cpuid_level >= 0x00000006)
672 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
674 /* Additional Intel-defined flags: level 0x00000007 */
675 if (c->cpuid_level >= 0x00000007) {
676 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
677 c->x86_capability[CPUID_7_0_EBX] = ebx;
678 c->x86_capability[CPUID_7_ECX] = ecx;
681 /* Extended state features: level 0x0000000d */
682 if (c->cpuid_level >= 0x0000000d) {
683 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
685 c->x86_capability[CPUID_D_1_EAX] = eax;
688 /* Additional Intel-defined flags: level 0x0000000F */
689 if (c->cpuid_level >= 0x0000000F) {
691 /* QoS sub-leaf, EAX=0Fh, ECX=0 */
692 cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
693 c->x86_capability[CPUID_F_0_EDX] = edx;
695 if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
696 /* will be overridden if occupancy monitoring exists */
697 c->x86_cache_max_rmid = ebx;
699 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
700 cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
701 c->x86_capability[CPUID_F_1_EDX] = edx;
703 if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
704 ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
705 (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
706 c->x86_cache_max_rmid = ecx;
707 c->x86_cache_occ_scale = ebx;
710 c->x86_cache_max_rmid = -1;
711 c->x86_cache_occ_scale = -1;
715 /* AMD-defined flags: level 0x80000001 */
716 eax = cpuid_eax(0x80000000);
717 c->extended_cpuid_level = eax;
719 if ((eax & 0xffff0000) == 0x80000000) {
720 if (eax >= 0x80000001) {
721 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
723 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
724 c->x86_capability[CPUID_8000_0001_EDX] = edx;
728 if (c->extended_cpuid_level >= 0x80000007) {
729 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
731 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
735 if (c->extended_cpuid_level >= 0x80000008) {
736 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
738 c->x86_virt_bits = (eax >> 8) & 0xff;
739 c->x86_phys_bits = eax & 0xff;
740 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
743 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
744 c->x86_phys_bits = 36;
747 if (c->extended_cpuid_level >= 0x8000000a)
748 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
750 init_scattered_cpuid_features(c);
753 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
759 * First of all, decide if this is a 486 or higher
760 * It's a 486 if we can modify the AC flag
762 if (flag_is_changeable_p(X86_EFLAGS_AC))
767 for (i = 0; i < X86_VENDOR_NUM; i++)
768 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
769 c->x86_vendor_id[0] = 0;
770 cpu_devs[i]->c_identify(c);
771 if (c->x86_vendor_id[0]) {
780 * Do minimum CPU detection early.
781 * Fields really needed: vendor, cpuid_level, family, model, mask,
783 * The others are not touched to avoid unwanted side effects.
785 * WARNING: this function is only called on the BP. Don't add code here
786 * that is supposed to run on all CPUs.
788 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
791 c->x86_clflush_size = 64;
792 c->x86_phys_bits = 36;
793 c->x86_virt_bits = 48;
795 c->x86_clflush_size = 32;
796 c->x86_phys_bits = 32;
797 c->x86_virt_bits = 32;
799 c->x86_cache_alignment = c->x86_clflush_size;
801 memset(&c->x86_capability, 0, sizeof c->x86_capability);
802 c->extended_cpuid_level = 0;
805 identify_cpu_without_cpuid(c);
807 /* cyrix could have cpuid enabled via c_identify()*/
808 if (have_cpuid_p()) {
813 if (this_cpu->c_early_init)
814 this_cpu->c_early_init(c);
817 filter_cpuid_features(c, false);
819 if (this_cpu->c_bsp_init)
820 this_cpu->c_bsp_init(c);
823 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
827 void __init early_cpu_init(void)
829 const struct cpu_dev *const *cdev;
832 #ifdef CONFIG_PROCESSOR_SELECT
833 pr_info("KERNEL supported cpus:\n");
836 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
837 const struct cpu_dev *cpudev = *cdev;
839 if (count >= X86_VENDOR_NUM)
841 cpu_devs[count] = cpudev;
844 #ifdef CONFIG_PROCESSOR_SELECT
848 for (j = 0; j < 2; j++) {
849 if (!cpudev->c_ident[j])
851 pr_info(" %s %s\n", cpudev->c_vendor,
857 early_identify_cpu(&boot_cpu_data);
861 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
862 * unfortunately, that's not true in practice because of early VIA
863 * chips and (more importantly) broken virtualizers that are not easy
864 * to detect. In the latter case it doesn't even *fail* reliably, so
865 * probing for it doesn't even work. Disable it completely on 32-bit
866 * unless we can find a reliable way to detect all the broken cases.
867 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
869 static void detect_nopl(struct cpuinfo_x86 *c)
872 clear_cpu_cap(c, X86_FEATURE_NOPL);
874 set_cpu_cap(c, X86_FEATURE_NOPL);
878 static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
882 * Empirically, writing zero to a segment selector on AMD does
883 * not clear the base, whereas writing zero to a segment
884 * selector on Intel does clear the base. Intel's behavior
885 * allows slightly faster context switches in the common case
886 * where GS is unused by the prev and next threads.
888 * Since neither vendor documents this anywhere that I can see,
889 * detect it directly instead of hardcoding the choice by
892 * I've designated AMD's behavior as the "bug" because it's
893 * counterintuitive and less friendly.
896 unsigned long old_base, tmp;
897 rdmsrl(MSR_FS_BASE, old_base);
898 wrmsrl(MSR_FS_BASE, 1);
900 rdmsrl(MSR_FS_BASE, tmp);
902 set_cpu_bug(c, X86_BUG_NULL_SEG);
903 wrmsrl(MSR_FS_BASE, old_base);
907 static void generic_identify(struct cpuinfo_x86 *c)
909 c->extended_cpuid_level = 0;
912 identify_cpu_without_cpuid(c);
914 /* cyrix could have cpuid enabled via c_identify()*/
924 if (c->cpuid_level >= 0x00000001) {
925 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
928 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
930 c->apicid = c->initial_apicid;
933 c->phys_proc_id = c->initial_apicid;
936 get_model_name(c); /* Default name */
940 detect_null_seg_behavior(c);
943 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
944 * systems that run Linux at CPL > 0 may or may not have the
945 * issue, but, even if they have the issue, there's absolutely
946 * nothing we can do about it because we can't use the real IRET
949 * NB: For the time being, only 32-bit kernels support
950 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
951 * whether to apply espfix using paravirt hooks. If any
952 * non-paravirt system ever shows up that does *not* have the
953 * ESPFIX issue, we can change this.
956 # ifdef CONFIG_PARAVIRT
958 extern void native_iret(void);
959 if (pv_cpu_ops.iret == native_iret)
960 set_cpu_bug(c, X86_BUG_ESPFIX);
963 set_cpu_bug(c, X86_BUG_ESPFIX);
968 static void x86_init_cache_qos(struct cpuinfo_x86 *c)
971 * The heavy lifting of max_rmid and cache_occ_scale are handled
972 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
973 * in case CQM bits really aren't there in this CPU.
975 if (c != &boot_cpu_data) {
976 boot_cpu_data.x86_cache_max_rmid =
977 min(boot_cpu_data.x86_cache_max_rmid,
978 c->x86_cache_max_rmid);
983 * Validate that ACPI/mptables have the same information about the
984 * effective APIC id and update the package map.
986 static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
989 unsigned int apicid, cpu = smp_processor_id();
991 apicid = apic->cpu_present_to_apicid(cpu);
993 if (apicid != c->apicid) {
994 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
995 cpu, apicid, c->initial_apicid);
997 BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
999 c->logical_proc_id = 0;
1004 * This does the hard work of actually picking apart the CPU stuff...
1006 static void identify_cpu(struct cpuinfo_x86 *c)
1010 c->loops_per_jiffy = loops_per_jiffy;
1011 c->x86_cache_size = -1;
1012 c->x86_vendor = X86_VENDOR_UNKNOWN;
1013 c->x86_model = c->x86_mask = 0; /* So far unknown... */
1014 c->x86_vendor_id[0] = '\0'; /* Unset */
1015 c->x86_model_id[0] = '\0'; /* Unset */
1016 c->x86_max_cores = 1;
1017 c->x86_coreid_bits = 0;
1018 #ifdef CONFIG_X86_64
1019 c->x86_clflush_size = 64;
1020 c->x86_phys_bits = 36;
1021 c->x86_virt_bits = 48;
1023 c->cpuid_level = -1; /* CPUID not detected */
1024 c->x86_clflush_size = 32;
1025 c->x86_phys_bits = 32;
1026 c->x86_virt_bits = 32;
1028 c->x86_cache_alignment = c->x86_clflush_size;
1029 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1031 generic_identify(c);
1033 if (this_cpu->c_identify)
1034 this_cpu->c_identify(c);
1036 /* Clear/Set all flags overridden by options, after probe */
1037 for (i = 0; i < NCAPINTS; i++) {
1038 c->x86_capability[i] &= ~cpu_caps_cleared[i];
1039 c->x86_capability[i] |= cpu_caps_set[i];
1042 #ifdef CONFIG_X86_64
1043 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1047 * Vendor-specific initialization. In this section we
1048 * canonicalize the feature flags, meaning if there are
1049 * features a certain CPU supports which CPUID doesn't
1050 * tell us, CPUID claiming incorrect flags, or other bugs,
1051 * we handle them here.
1053 * At the end of this section, c->x86_capability better
1054 * indicate the features this CPU genuinely supports!
1056 if (this_cpu->c_init)
1057 this_cpu->c_init(c);
1059 /* Disable the PN if appropriate */
1060 squash_the_stupid_serial_number(c);
1062 /* Set up SMEP/SMAP */
1067 * The vendor-specific functions might have changed features.
1068 * Now we do "generic changes."
1071 /* Filter out anything that depends on CPUID levels we don't have */
1072 filter_cpuid_features(c, true);
1074 /* If the model name is still unset, do table lookup. */
1075 if (!c->x86_model_id[0]) {
1077 p = table_lookup_model(c);
1079 strcpy(c->x86_model_id, p);
1081 /* Last resort... */
1082 sprintf(c->x86_model_id, "%02x/%02x",
1083 c->x86, c->x86_model);
1086 #ifdef CONFIG_X86_64
1092 x86_init_cache_qos(c);
1096 * Clear/Set all flags overridden by options, need do it
1097 * before following smp all cpus cap AND.
1099 for (i = 0; i < NCAPINTS; i++) {
1100 c->x86_capability[i] &= ~cpu_caps_cleared[i];
1101 c->x86_capability[i] |= cpu_caps_set[i];
1105 * On SMP, boot_cpu_data holds the common feature set between
1106 * all CPUs; so make sure that we indicate which features are
1107 * common between the CPUs. The first time this routine gets
1108 * executed, c == &boot_cpu_data.
1110 if (c != &boot_cpu_data) {
1111 /* AND the already accumulated flags with these */
1112 for (i = 0; i < NCAPINTS; i++)
1113 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1115 /* OR, i.e. replicate the bug flags */
1116 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1117 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1120 /* Init Machine Check Exception if available. */
1123 select_idle_routine(c);
1126 numa_add_cpu(smp_processor_id());
1131 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1132 * on 32-bit kernels:
1134 #ifdef CONFIG_X86_32
1135 void enable_sep_cpu(void)
1137 struct tss_struct *tss;
1140 if (!boot_cpu_has(X86_FEATURE_SEP))
1144 tss = &per_cpu(cpu_tss, cpu);
1147 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1148 * see the big comment in struct x86_hw_tss's definition.
1151 tss->x86_tss.ss1 = __KERNEL_CS;
1152 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1154 wrmsr(MSR_IA32_SYSENTER_ESP,
1155 (unsigned long)tss + offsetofend(struct tss_struct, SYSENTER_stack),
1158 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1164 void __init identify_boot_cpu(void)
1166 identify_cpu(&boot_cpu_data);
1167 #ifdef CONFIG_X86_32
1171 cpu_detect_tlb(&boot_cpu_data);
1174 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1176 BUG_ON(c == &boot_cpu_data);
1178 #ifdef CONFIG_X86_32
1182 validate_apic_and_package_id(c);
1185 static __init int setup_noclflush(char *arg)
1187 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1188 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1191 __setup("noclflush", setup_noclflush);
1193 void print_cpu_info(struct cpuinfo_x86 *c)
1195 const char *vendor = NULL;
1197 if (c->x86_vendor < X86_VENDOR_NUM) {
1198 vendor = this_cpu->c_vendor;
1200 if (c->cpuid_level >= 0)
1201 vendor = c->x86_vendor_id;
1204 if (vendor && !strstr(c->x86_model_id, vendor))
1205 pr_cont("%s ", vendor);
1207 if (c->x86_model_id[0])
1208 pr_cont("%s", c->x86_model_id);
1210 pr_cont("%d86", c->x86);
1212 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1214 if (c->x86_mask || c->cpuid_level >= 0)
1215 pr_cont(", stepping: 0x%x)\n", c->x86_mask);
1220 static __init int setup_disablecpuid(char *arg)
1224 if (get_option(&arg, &bit) && bit < NCAPINTS*32)
1225 setup_clear_cpu_cap(bit);
1231 __setup("clearcpuid=", setup_disablecpuid);
1233 #ifdef CONFIG_X86_64
1234 struct desc_ptr idt_descr __ro_after_init = {
1235 .size = NR_VECTORS * 16 - 1,
1236 .address = (unsigned long) idt_table,
1238 const struct desc_ptr debug_idt_descr = {
1239 .size = NR_VECTORS * 16 - 1,
1240 .address = (unsigned long) debug_idt_table,
1243 DEFINE_PER_CPU_FIRST(union irq_stack_union,
1244 irq_stack_union) __aligned(PAGE_SIZE) __visible;
1247 * The following percpu variables are hot. Align current_task to
1248 * cacheline size such that they fall in the same cacheline.
1250 DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1252 EXPORT_PER_CPU_SYMBOL(current_task);
1254 DEFINE_PER_CPU(char *, irq_stack_ptr) =
1255 init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE;
1257 DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1259 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1260 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1263 * Special IST stacks which the CPU switches to when it calls
1264 * an IST-marked descriptor entry. Up to 7 stacks (hardware
1265 * limit), all of them are 4K, except the debug stack which
1268 static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
1269 [0 ... N_EXCEPTION_STACKS - 1] = EXCEPTION_STKSZ,
1270 [DEBUG_STACK - 1] = DEBUG_STKSZ
1273 static DEFINE_PER_CPU_PAGE_ALIGNED(char, exception_stacks
1274 [(N_EXCEPTION_STACKS - 1) * EXCEPTION_STKSZ + DEBUG_STKSZ]);
1276 /* May not be marked __init: used by software suspend */
1277 void syscall_init(void)
1279 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1280 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1282 #ifdef CONFIG_IA32_EMULATION
1283 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1285 * This only works on Intel CPUs.
1286 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1287 * This does not cause SYSENTER to jump to the wrong location, because
1288 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1290 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1291 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1292 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1294 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1295 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1296 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1297 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1300 /* Flags to clear on syscall */
1301 wrmsrl(MSR_SYSCALL_MASK,
1302 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1303 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1307 * Copies of the original ist values from the tss are only accessed during
1308 * debugging, no special alignment required.
1310 DEFINE_PER_CPU(struct orig_ist, orig_ist);
1312 static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1313 DEFINE_PER_CPU(int, debug_stack_usage);
1315 int is_debug_stack(unsigned long addr)
1317 return __this_cpu_read(debug_stack_usage) ||
1318 (addr <= __this_cpu_read(debug_stack_addr) &&
1319 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1321 NOKPROBE_SYMBOL(is_debug_stack);
1323 DEFINE_PER_CPU(u32, debug_idt_ctr);
1325 void debug_stack_set_zero(void)
1327 this_cpu_inc(debug_idt_ctr);
1330 NOKPROBE_SYMBOL(debug_stack_set_zero);
1332 void debug_stack_reset(void)
1334 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1336 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1339 NOKPROBE_SYMBOL(debug_stack_reset);
1341 #else /* CONFIG_X86_64 */
1343 DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1344 EXPORT_PER_CPU_SYMBOL(current_task);
1345 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1346 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1349 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1350 * the top of the kernel stack. Use an extra percpu variable to track the
1351 * top of the kernel stack directly.
1353 DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1354 (unsigned long)&init_thread_union + THREAD_SIZE;
1355 EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1357 #ifdef CONFIG_CC_STACKPROTECTOR
1358 DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1361 #endif /* CONFIG_X86_64 */
1364 * Clear all 6 debug registers:
1366 static void clear_all_debug_regs(void)
1370 for (i = 0; i < 8; i++) {
1371 /* Ignore db4, db5 */
1372 if ((i == 4) || (i == 5))
1381 * Restore debug regs if using kgdbwait and you have a kernel debugger
1382 * connection established.
1384 static void dbg_restore_debug_regs(void)
1386 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1387 arch_kgdb_ops.correct_hw_break();
1389 #else /* ! CONFIG_KGDB */
1390 #define dbg_restore_debug_regs()
1391 #endif /* ! CONFIG_KGDB */
1393 static void wait_for_master_cpu(int cpu)
1397 * wait for ACK from master CPU before continuing
1398 * with AP initialization
1400 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1401 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1407 * cpu_init() initializes state that is per-CPU. Some data is already
1408 * initialized (naturally) in the bootstrap process, such as the GDT
1409 * and IDT. We reload them nevertheless, this function acts as a
1410 * 'CPU state barrier', nothing should get across.
1411 * A lot of state is already set up in PDA init for 64 bit
1413 #ifdef CONFIG_X86_64
1417 struct orig_ist *oist;
1418 struct task_struct *me;
1419 struct tss_struct *t;
1421 int cpu = raw_smp_processor_id();
1424 wait_for_master_cpu(cpu);
1427 * Initialize the CR4 shadow before doing anything that could
1435 t = &per_cpu(cpu_tss, cpu);
1436 oist = &per_cpu(orig_ist, cpu);
1439 if (this_cpu_read(numa_node) == 0 &&
1440 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1441 set_numa_node(early_cpu_to_node(cpu));
1446 pr_debug("Initializing CPU#%d\n", cpu);
1448 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1451 * Initialize the per-CPU GDT with the boot GDT,
1452 * and set up the GDT descriptor:
1455 switch_to_new_gdt(cpu);
1460 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1463 wrmsrl(MSR_FS_BASE, 0);
1464 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1471 * set up and load the per-CPU TSS
1473 if (!oist->ist[0]) {
1474 char *estacks = per_cpu(exception_stacks, cpu);
1476 for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1477 estacks += exception_stack_sizes[v];
1478 oist->ist[v] = t->x86_tss.ist[v] =
1479 (unsigned long)estacks;
1480 if (v == DEBUG_STACK-1)
1481 per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1485 t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1488 * <= is required because the CPU will access up to
1489 * 8 bits beyond the end of the IO permission bitmap.
1491 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1492 t->io_bitmap[i] = ~0UL;
1494 atomic_inc(&init_mm.mm_count);
1495 me->active_mm = &init_mm;
1497 enter_lazy_tlb(&init_mm, me);
1499 load_sp0(t, ¤t->thread);
1500 set_tss_desc(cpu, t);
1502 load_mm_ldt(&init_mm);
1504 clear_all_debug_regs();
1505 dbg_restore_debug_regs();
1517 int cpu = smp_processor_id();
1518 struct task_struct *curr = current;
1519 struct tss_struct *t = &per_cpu(cpu_tss, cpu);
1520 struct thread_struct *thread = &curr->thread;
1522 wait_for_master_cpu(cpu);
1525 * Initialize the CR4 shadow before doing anything that could
1530 show_ucode_info_early();
1532 pr_info("Initializing CPU#%d\n", cpu);
1534 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1535 boot_cpu_has(X86_FEATURE_TSC) ||
1536 boot_cpu_has(X86_FEATURE_DE))
1537 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1540 switch_to_new_gdt(cpu);
1543 * Set up and load the per-CPU TSS and LDT
1545 atomic_inc(&init_mm.mm_count);
1546 curr->active_mm = &init_mm;
1548 enter_lazy_tlb(&init_mm, curr);
1550 load_sp0(t, thread);
1551 set_tss_desc(cpu, t);
1553 load_mm_ldt(&init_mm);
1555 t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1557 #ifdef CONFIG_DOUBLEFAULT
1558 /* Set up doublefault TSS pointer in the GDT */
1559 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1562 clear_all_debug_regs();
1563 dbg_restore_debug_regs();
1569 static void bsp_resume(void)
1571 if (this_cpu->c_bsp_resume)
1572 this_cpu->c_bsp_resume(&boot_cpu_data);
1575 static struct syscore_ops cpu_syscore_ops = {
1576 .resume = bsp_resume,
1579 static int __init init_cpu_syscore(void)
1581 register_syscore_ops(&cpu_syscore_ops);
1584 core_initcall(init_cpu_syscore);