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 /* Additional Intel-defined flags: level 0x00000007 */
671 if (c->cpuid_level >= 0x00000007) {
672 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
674 c->x86_capability[CPUID_7_0_EBX] = ebx;
676 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
677 c->x86_capability[CPUID_7_ECX] = ecx;
680 /* Extended state features: level 0x0000000d */
681 if (c->cpuid_level >= 0x0000000d) {
682 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
684 c->x86_capability[CPUID_D_1_EAX] = eax;
687 /* Additional Intel-defined flags: level 0x0000000F */
688 if (c->cpuid_level >= 0x0000000F) {
690 /* QoS sub-leaf, EAX=0Fh, ECX=0 */
691 cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
692 c->x86_capability[CPUID_F_0_EDX] = edx;
694 if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
695 /* will be overridden if occupancy monitoring exists */
696 c->x86_cache_max_rmid = ebx;
698 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
699 cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
700 c->x86_capability[CPUID_F_1_EDX] = edx;
702 if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
703 ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
704 (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
705 c->x86_cache_max_rmid = ecx;
706 c->x86_cache_occ_scale = ebx;
709 c->x86_cache_max_rmid = -1;
710 c->x86_cache_occ_scale = -1;
714 /* AMD-defined flags: level 0x80000001 */
715 eax = cpuid_eax(0x80000000);
716 c->extended_cpuid_level = eax;
718 if ((eax & 0xffff0000) == 0x80000000) {
719 if (eax >= 0x80000001) {
720 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
722 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
723 c->x86_capability[CPUID_8000_0001_EDX] = edx;
727 if (c->extended_cpuid_level >= 0x80000007) {
728 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
730 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
734 if (c->extended_cpuid_level >= 0x80000008) {
735 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
737 c->x86_virt_bits = (eax >> 8) & 0xff;
738 c->x86_phys_bits = eax & 0xff;
739 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
742 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
743 c->x86_phys_bits = 36;
746 if (c->extended_cpuid_level >= 0x8000000a)
747 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
749 init_scattered_cpuid_features(c);
752 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
758 * First of all, decide if this is a 486 or higher
759 * It's a 486 if we can modify the AC flag
761 if (flag_is_changeable_p(X86_EFLAGS_AC))
766 for (i = 0; i < X86_VENDOR_NUM; i++)
767 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
768 c->x86_vendor_id[0] = 0;
769 cpu_devs[i]->c_identify(c);
770 if (c->x86_vendor_id[0]) {
779 * Do minimum CPU detection early.
780 * Fields really needed: vendor, cpuid_level, family, model, mask,
782 * The others are not touched to avoid unwanted side effects.
784 * WARNING: this function is only called on the BP. Don't add code here
785 * that is supposed to run on all CPUs.
787 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
790 c->x86_clflush_size = 64;
791 c->x86_phys_bits = 36;
792 c->x86_virt_bits = 48;
794 c->x86_clflush_size = 32;
795 c->x86_phys_bits = 32;
796 c->x86_virt_bits = 32;
798 c->x86_cache_alignment = c->x86_clflush_size;
800 memset(&c->x86_capability, 0, sizeof c->x86_capability);
801 c->extended_cpuid_level = 0;
804 identify_cpu_without_cpuid(c);
806 /* cyrix could have cpuid enabled via c_identify()*/
807 if (have_cpuid_p()) {
812 if (this_cpu->c_early_init)
813 this_cpu->c_early_init(c);
816 filter_cpuid_features(c, false);
818 if (this_cpu->c_bsp_init)
819 this_cpu->c_bsp_init(c);
822 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
826 void __init early_cpu_init(void)
828 const struct cpu_dev *const *cdev;
831 #ifdef CONFIG_PROCESSOR_SELECT
832 pr_info("KERNEL supported cpus:\n");
835 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
836 const struct cpu_dev *cpudev = *cdev;
838 if (count >= X86_VENDOR_NUM)
840 cpu_devs[count] = cpudev;
843 #ifdef CONFIG_PROCESSOR_SELECT
847 for (j = 0; j < 2; j++) {
848 if (!cpudev->c_ident[j])
850 pr_info(" %s %s\n", cpudev->c_vendor,
856 early_identify_cpu(&boot_cpu_data);
860 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
861 * unfortunately, that's not true in practice because of early VIA
862 * chips and (more importantly) broken virtualizers that are not easy
863 * to detect. In the latter case it doesn't even *fail* reliably, so
864 * probing for it doesn't even work. Disable it completely on 32-bit
865 * unless we can find a reliable way to detect all the broken cases.
866 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
868 static void detect_nopl(struct cpuinfo_x86 *c)
871 clear_cpu_cap(c, X86_FEATURE_NOPL);
873 set_cpu_cap(c, X86_FEATURE_NOPL);
877 static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
881 * Empirically, writing zero to a segment selector on AMD does
882 * not clear the base, whereas writing zero to a segment
883 * selector on Intel does clear the base. Intel's behavior
884 * allows slightly faster context switches in the common case
885 * where GS is unused by the prev and next threads.
887 * Since neither vendor documents this anywhere that I can see,
888 * detect it directly instead of hardcoding the choice by
891 * I've designated AMD's behavior as the "bug" because it's
892 * counterintuitive and less friendly.
895 unsigned long old_base, tmp;
896 rdmsrl(MSR_FS_BASE, old_base);
897 wrmsrl(MSR_FS_BASE, 1);
899 rdmsrl(MSR_FS_BASE, tmp);
901 set_cpu_bug(c, X86_BUG_NULL_SEG);
902 wrmsrl(MSR_FS_BASE, old_base);
906 static void generic_identify(struct cpuinfo_x86 *c)
908 c->extended_cpuid_level = 0;
911 identify_cpu_without_cpuid(c);
913 /* cyrix could have cpuid enabled via c_identify()*/
923 if (c->cpuid_level >= 0x00000001) {
924 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
927 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
929 c->apicid = c->initial_apicid;
932 c->phys_proc_id = c->initial_apicid;
935 get_model_name(c); /* Default name */
939 detect_null_seg_behavior(c);
942 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
943 * systems that run Linux at CPL > 0 may or may not have the
944 * issue, but, even if they have the issue, there's absolutely
945 * nothing we can do about it because we can't use the real IRET
948 * NB: For the time being, only 32-bit kernels support
949 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
950 * whether to apply espfix using paravirt hooks. If any
951 * non-paravirt system ever shows up that does *not* have the
952 * ESPFIX issue, we can change this.
955 # ifdef CONFIG_PARAVIRT
957 extern void native_iret(void);
958 if (pv_cpu_ops.iret == native_iret)
959 set_cpu_bug(c, X86_BUG_ESPFIX);
962 set_cpu_bug(c, X86_BUG_ESPFIX);
967 static void x86_init_cache_qos(struct cpuinfo_x86 *c)
970 * The heavy lifting of max_rmid and cache_occ_scale are handled
971 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
972 * in case CQM bits really aren't there in this CPU.
974 if (c != &boot_cpu_data) {
975 boot_cpu_data.x86_cache_max_rmid =
976 min(boot_cpu_data.x86_cache_max_rmid,
977 c->x86_cache_max_rmid);
982 * This does the hard work of actually picking apart the CPU stuff...
984 static void identify_cpu(struct cpuinfo_x86 *c)
988 c->loops_per_jiffy = loops_per_jiffy;
989 c->x86_cache_size = -1;
990 c->x86_vendor = X86_VENDOR_UNKNOWN;
991 c->x86_model = c->x86_mask = 0; /* So far unknown... */
992 c->x86_vendor_id[0] = '\0'; /* Unset */
993 c->x86_model_id[0] = '\0'; /* Unset */
994 c->x86_max_cores = 1;
995 c->x86_coreid_bits = 0;
997 c->x86_clflush_size = 64;
998 c->x86_phys_bits = 36;
999 c->x86_virt_bits = 48;
1001 c->cpuid_level = -1; /* CPUID not detected */
1002 c->x86_clflush_size = 32;
1003 c->x86_phys_bits = 32;
1004 c->x86_virt_bits = 32;
1006 c->x86_cache_alignment = c->x86_clflush_size;
1007 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1009 generic_identify(c);
1011 if (this_cpu->c_identify)
1012 this_cpu->c_identify(c);
1014 /* Clear/Set all flags overridden by options, after probe */
1015 for (i = 0; i < NCAPINTS; i++) {
1016 c->x86_capability[i] &= ~cpu_caps_cleared[i];
1017 c->x86_capability[i] |= cpu_caps_set[i];
1020 #ifdef CONFIG_X86_64
1021 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1025 * Vendor-specific initialization. In this section we
1026 * canonicalize the feature flags, meaning if there are
1027 * features a certain CPU supports which CPUID doesn't
1028 * tell us, CPUID claiming incorrect flags, or other bugs,
1029 * we handle them here.
1031 * At the end of this section, c->x86_capability better
1032 * indicate the features this CPU genuinely supports!
1034 if (this_cpu->c_init)
1035 this_cpu->c_init(c);
1037 /* Disable the PN if appropriate */
1038 squash_the_stupid_serial_number(c);
1040 /* Set up SMEP/SMAP */
1045 * The vendor-specific functions might have changed features.
1046 * Now we do "generic changes."
1049 /* Filter out anything that depends on CPUID levels we don't have */
1050 filter_cpuid_features(c, true);
1052 /* If the model name is still unset, do table lookup. */
1053 if (!c->x86_model_id[0]) {
1055 p = table_lookup_model(c);
1057 strcpy(c->x86_model_id, p);
1059 /* Last resort... */
1060 sprintf(c->x86_model_id, "%02x/%02x",
1061 c->x86, c->x86_model);
1064 #ifdef CONFIG_X86_64
1070 x86_init_cache_qos(c);
1074 * Clear/Set all flags overridden by options, need do it
1075 * before following smp all cpus cap AND.
1077 for (i = 0; i < NCAPINTS; i++) {
1078 c->x86_capability[i] &= ~cpu_caps_cleared[i];
1079 c->x86_capability[i] |= cpu_caps_set[i];
1083 * On SMP, boot_cpu_data holds the common feature set between
1084 * all CPUs; so make sure that we indicate which features are
1085 * common between the CPUs. The first time this routine gets
1086 * executed, c == &boot_cpu_data.
1088 if (c != &boot_cpu_data) {
1089 /* AND the already accumulated flags with these */
1090 for (i = 0; i < NCAPINTS; i++)
1091 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1093 /* OR, i.e. replicate the bug flags */
1094 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1095 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1098 /* Init Machine Check Exception if available. */
1101 select_idle_routine(c);
1104 numa_add_cpu(smp_processor_id());
1106 /* The boot/hotplug time assigment got cleared, restore it */
1107 c->logical_proc_id = topology_phys_to_logical_pkg(c->phys_proc_id);
1111 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1112 * on 32-bit kernels:
1114 #ifdef CONFIG_X86_32
1115 void enable_sep_cpu(void)
1117 struct tss_struct *tss;
1120 if (!boot_cpu_has(X86_FEATURE_SEP))
1124 tss = &per_cpu(cpu_tss, cpu);
1127 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1128 * see the big comment in struct x86_hw_tss's definition.
1131 tss->x86_tss.ss1 = __KERNEL_CS;
1132 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1134 wrmsr(MSR_IA32_SYSENTER_ESP,
1135 (unsigned long)tss + offsetofend(struct tss_struct, SYSENTER_stack),
1138 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1144 void __init identify_boot_cpu(void)
1146 identify_cpu(&boot_cpu_data);
1147 init_amd_e400_c1e_mask();
1148 #ifdef CONFIG_X86_32
1152 cpu_detect_tlb(&boot_cpu_data);
1155 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1157 BUG_ON(c == &boot_cpu_data);
1159 #ifdef CONFIG_X86_32
1170 static const struct msr_range msr_range_array[] = {
1171 { 0x00000000, 0x00000418},
1172 { 0xc0000000, 0xc000040b},
1173 { 0xc0010000, 0xc0010142},
1174 { 0xc0011000, 0xc001103b},
1177 static void __print_cpu_msr(void)
1179 unsigned index_min, index_max;
1184 for (i = 0; i < ARRAY_SIZE(msr_range_array); i++) {
1185 index_min = msr_range_array[i].min;
1186 index_max = msr_range_array[i].max;
1188 for (index = index_min; index < index_max; index++) {
1189 if (rdmsrl_safe(index, &val))
1191 pr_info(" MSR%08x: %016llx\n", index, val);
1196 static int show_msr;
1198 static __init int setup_show_msr(char *arg)
1202 get_option(&arg, &num);
1208 __setup("show_msr=", setup_show_msr);
1210 static __init int setup_noclflush(char *arg)
1212 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1213 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1216 __setup("noclflush", setup_noclflush);
1218 void print_cpu_info(struct cpuinfo_x86 *c)
1220 const char *vendor = NULL;
1222 if (c->x86_vendor < X86_VENDOR_NUM) {
1223 vendor = this_cpu->c_vendor;
1225 if (c->cpuid_level >= 0)
1226 vendor = c->x86_vendor_id;
1229 if (vendor && !strstr(c->x86_model_id, vendor))
1230 pr_cont("%s ", vendor);
1232 if (c->x86_model_id[0])
1233 pr_cont("%s", c->x86_model_id);
1235 pr_cont("%d86", c->x86);
1237 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1239 if (c->x86_mask || c->cpuid_level >= 0)
1240 pr_cont(", stepping: 0x%x)\n", c->x86_mask);
1247 void print_cpu_msr(struct cpuinfo_x86 *c)
1249 if (c->cpu_index < show_msr)
1253 static __init int setup_disablecpuid(char *arg)
1257 if (get_option(&arg, &bit) && bit < NCAPINTS*32)
1258 setup_clear_cpu_cap(bit);
1264 __setup("clearcpuid=", setup_disablecpuid);
1266 #ifdef CONFIG_X86_64
1267 struct desc_ptr idt_descr __ro_after_init = {
1268 .size = NR_VECTORS * 16 - 1,
1269 .address = (unsigned long) idt_table,
1271 const struct desc_ptr debug_idt_descr = {
1272 .size = NR_VECTORS * 16 - 1,
1273 .address = (unsigned long) debug_idt_table,
1276 DEFINE_PER_CPU_FIRST(union irq_stack_union,
1277 irq_stack_union) __aligned(PAGE_SIZE) __visible;
1280 * The following percpu variables are hot. Align current_task to
1281 * cacheline size such that they fall in the same cacheline.
1283 DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1285 EXPORT_PER_CPU_SYMBOL(current_task);
1287 DEFINE_PER_CPU(char *, irq_stack_ptr) =
1288 init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE;
1290 DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1292 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1293 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1296 * Special IST stacks which the CPU switches to when it calls
1297 * an IST-marked descriptor entry. Up to 7 stacks (hardware
1298 * limit), all of them are 4K, except the debug stack which
1301 static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
1302 [0 ... N_EXCEPTION_STACKS - 1] = EXCEPTION_STKSZ,
1303 [DEBUG_STACK - 1] = DEBUG_STKSZ
1306 static DEFINE_PER_CPU_PAGE_ALIGNED(char, exception_stacks
1307 [(N_EXCEPTION_STACKS - 1) * EXCEPTION_STKSZ + DEBUG_STKSZ]);
1309 /* May not be marked __init: used by software suspend */
1310 void syscall_init(void)
1312 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1313 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1315 #ifdef CONFIG_IA32_EMULATION
1316 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1318 * This only works on Intel CPUs.
1319 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1320 * This does not cause SYSENTER to jump to the wrong location, because
1321 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1323 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1324 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1325 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1327 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1328 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1329 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1330 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1333 /* Flags to clear on syscall */
1334 wrmsrl(MSR_SYSCALL_MASK,
1335 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1336 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1340 * Copies of the original ist values from the tss are only accessed during
1341 * debugging, no special alignment required.
1343 DEFINE_PER_CPU(struct orig_ist, orig_ist);
1345 static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1346 DEFINE_PER_CPU(int, debug_stack_usage);
1348 int is_debug_stack(unsigned long addr)
1350 return __this_cpu_read(debug_stack_usage) ||
1351 (addr <= __this_cpu_read(debug_stack_addr) &&
1352 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1354 NOKPROBE_SYMBOL(is_debug_stack);
1356 DEFINE_PER_CPU(u32, debug_idt_ctr);
1358 void debug_stack_set_zero(void)
1360 this_cpu_inc(debug_idt_ctr);
1363 NOKPROBE_SYMBOL(debug_stack_set_zero);
1365 void debug_stack_reset(void)
1367 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1369 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1372 NOKPROBE_SYMBOL(debug_stack_reset);
1374 #else /* CONFIG_X86_64 */
1376 DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1377 EXPORT_PER_CPU_SYMBOL(current_task);
1378 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1379 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1382 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1383 * the top of the kernel stack. Use an extra percpu variable to track the
1384 * top of the kernel stack directly.
1386 DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1387 (unsigned long)&init_thread_union + THREAD_SIZE;
1388 EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1390 #ifdef CONFIG_CC_STACKPROTECTOR
1391 DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1394 #endif /* CONFIG_X86_64 */
1397 * Clear all 6 debug registers:
1399 static void clear_all_debug_regs(void)
1403 for (i = 0; i < 8; i++) {
1404 /* Ignore db4, db5 */
1405 if ((i == 4) || (i == 5))
1414 * Restore debug regs if using kgdbwait and you have a kernel debugger
1415 * connection established.
1417 static void dbg_restore_debug_regs(void)
1419 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1420 arch_kgdb_ops.correct_hw_break();
1422 #else /* ! CONFIG_KGDB */
1423 #define dbg_restore_debug_regs()
1424 #endif /* ! CONFIG_KGDB */
1426 static void wait_for_master_cpu(int cpu)
1430 * wait for ACK from master CPU before continuing
1431 * with AP initialization
1433 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1434 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1440 * cpu_init() initializes state that is per-CPU. Some data is already
1441 * initialized (naturally) in the bootstrap process, such as the GDT
1442 * and IDT. We reload them nevertheless, this function acts as a
1443 * 'CPU state barrier', nothing should get across.
1444 * A lot of state is already set up in PDA init for 64 bit
1446 #ifdef CONFIG_X86_64
1450 struct orig_ist *oist;
1451 struct task_struct *me;
1452 struct tss_struct *t;
1454 int cpu = raw_smp_processor_id();
1457 wait_for_master_cpu(cpu);
1460 * Initialize the CR4 shadow before doing anything that could
1466 * Load microcode on this cpu if a valid microcode is available.
1467 * This is early microcode loading procedure.
1471 t = &per_cpu(cpu_tss, cpu);
1472 oist = &per_cpu(orig_ist, cpu);
1475 if (this_cpu_read(numa_node) == 0 &&
1476 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1477 set_numa_node(early_cpu_to_node(cpu));
1482 pr_debug("Initializing CPU#%d\n", cpu);
1484 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1487 * Initialize the per-CPU GDT with the boot GDT,
1488 * and set up the GDT descriptor:
1491 switch_to_new_gdt(cpu);
1496 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1499 wrmsrl(MSR_FS_BASE, 0);
1500 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1507 * set up and load the per-CPU TSS
1509 if (!oist->ist[0]) {
1510 char *estacks = per_cpu(exception_stacks, cpu);
1512 for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1513 estacks += exception_stack_sizes[v];
1514 oist->ist[v] = t->x86_tss.ist[v] =
1515 (unsigned long)estacks;
1516 if (v == DEBUG_STACK-1)
1517 per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1521 t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1524 * <= is required because the CPU will access up to
1525 * 8 bits beyond the end of the IO permission bitmap.
1527 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1528 t->io_bitmap[i] = ~0UL;
1530 atomic_inc(&init_mm.mm_count);
1531 me->active_mm = &init_mm;
1533 enter_lazy_tlb(&init_mm, me);
1535 load_sp0(t, ¤t->thread);
1536 set_tss_desc(cpu, t);
1538 load_mm_ldt(&init_mm);
1540 clear_all_debug_regs();
1541 dbg_restore_debug_regs();
1553 int cpu = smp_processor_id();
1554 struct task_struct *curr = current;
1555 struct tss_struct *t = &per_cpu(cpu_tss, cpu);
1556 struct thread_struct *thread = &curr->thread;
1558 wait_for_master_cpu(cpu);
1561 * Initialize the CR4 shadow before doing anything that could
1566 show_ucode_info_early();
1568 pr_info("Initializing CPU#%d\n", cpu);
1570 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1571 boot_cpu_has(X86_FEATURE_TSC) ||
1572 boot_cpu_has(X86_FEATURE_DE))
1573 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1576 switch_to_new_gdt(cpu);
1579 * Set up and load the per-CPU TSS and LDT
1581 atomic_inc(&init_mm.mm_count);
1582 curr->active_mm = &init_mm;
1584 enter_lazy_tlb(&init_mm, curr);
1586 load_sp0(t, thread);
1587 set_tss_desc(cpu, t);
1589 load_mm_ldt(&init_mm);
1591 t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1593 #ifdef CONFIG_DOUBLEFAULT
1594 /* Set up doublefault TSS pointer in the GDT */
1595 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1598 clear_all_debug_regs();
1599 dbg_restore_debug_regs();
1605 static void bsp_resume(void)
1607 if (this_cpu->c_bsp_resume)
1608 this_cpu->c_bsp_resume(&boot_cpu_data);
1611 static struct syscore_ops cpu_syscore_ops = {
1612 .resume = bsp_resume,
1615 static int __init init_cpu_syscore(void)
1617 register_syscore_ops(&cpu_syscore_ops);
1620 core_initcall(init_cpu_syscore);