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>
51 #ifdef CONFIG_X86_LOCAL_APIC
52 #include <asm/uv/uv.h>
57 u32 elf_hwcap2 __read_mostly;
59 /* all of these masks are initialized in setup_cpu_local_masks() */
60 cpumask_var_t cpu_initialized_mask;
61 cpumask_var_t cpu_callout_mask;
62 cpumask_var_t cpu_callin_mask;
64 /* representing cpus for which sibling maps can be computed */
65 cpumask_var_t cpu_sibling_setup_mask;
67 /* correctly size the local cpu masks */
68 void __init setup_cpu_local_masks(void)
70 alloc_bootmem_cpumask_var(&cpu_initialized_mask);
71 alloc_bootmem_cpumask_var(&cpu_callin_mask);
72 alloc_bootmem_cpumask_var(&cpu_callout_mask);
73 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
76 static void default_init(struct cpuinfo_x86 *c)
79 cpu_detect_cache_sizes(c);
81 /* Not much we can do here... */
82 /* Check if at least it has cpuid */
83 if (c->cpuid_level == -1) {
84 /* No cpuid. It must be an ancient CPU */
86 strcpy(c->x86_model_id, "486");
88 strcpy(c->x86_model_id, "386");
93 static const struct cpu_dev default_cpu = {
94 .c_init = default_init,
95 .c_vendor = "Unknown",
96 .c_x86_vendor = X86_VENDOR_UNKNOWN,
99 static const struct cpu_dev *this_cpu = &default_cpu;
101 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
104 * We need valid kernel segments for data and code in long mode too
105 * IRET will check the segment types kkeil 2000/10/28
106 * Also sysret mandates a special GDT layout
108 * TLS descriptors are currently at a different place compared to i386.
109 * Hopefully nobody expects them at a fixed place (Wine?)
111 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
112 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
113 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
114 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
115 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
116 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
118 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
119 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
120 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
121 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
123 * Segments used for calling PnP BIOS have byte granularity.
124 * They code segments and data segments have fixed 64k limits,
125 * the transfer segment sizes are set at run time.
128 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
130 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
132 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(0x0092, 0, 0xffff),
134 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(0x0092, 0, 0),
136 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(0x0092, 0, 0),
138 * The APM segments have byte granularity and their bases
139 * are set at run time. All have 64k limits.
142 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
144 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
146 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(0x4092, 0, 0xffff),
148 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
149 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
150 GDT_STACK_CANARY_INIT
153 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
155 static int __init x86_mpx_setup(char *s)
157 /* require an exact match without trailing characters */
161 /* do not emit a message if the feature is not present */
162 if (!boot_cpu_has(X86_FEATURE_MPX))
165 setup_clear_cpu_cap(X86_FEATURE_MPX);
166 pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
169 __setup("nompx", x86_mpx_setup);
172 static int __init x86_nopcid_setup(char *s)
174 /* nopcid doesn't accept parameters */
178 /* do not emit a message if the feature is not present */
179 if (!boot_cpu_has(X86_FEATURE_PCID))
182 setup_clear_cpu_cap(X86_FEATURE_PCID);
183 pr_info("nopcid: PCID feature disabled\n");
186 early_param("nopcid", x86_nopcid_setup);
189 static int __init x86_noinvpcid_setup(char *s)
191 /* noinvpcid doesn't accept parameters */
195 /* do not emit a message if the feature is not present */
196 if (!boot_cpu_has(X86_FEATURE_INVPCID))
199 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
200 pr_info("noinvpcid: INVPCID feature disabled\n");
203 early_param("noinvpcid", x86_noinvpcid_setup);
206 static int cachesize_override = -1;
207 static int disable_x86_serial_nr = 1;
209 static int __init cachesize_setup(char *str)
211 get_option(&str, &cachesize_override);
214 __setup("cachesize=", cachesize_setup);
216 static int __init x86_sep_setup(char *s)
218 setup_clear_cpu_cap(X86_FEATURE_SEP);
221 __setup("nosep", x86_sep_setup);
223 /* Standard macro to see if a specific flag is changeable */
224 static inline int flag_is_changeable_p(u32 flag)
229 * Cyrix and IDT cpus allow disabling of CPUID
230 * so the code below may return different results
231 * when it is executed before and after enabling
232 * the CPUID. Add "volatile" to not allow gcc to
233 * optimize the subsequent calls to this function.
235 asm volatile ("pushfl \n\t"
246 : "=&r" (f1), "=&r" (f2)
249 return ((f1^f2) & flag) != 0;
252 /* Probe for the CPUID instruction */
253 int have_cpuid_p(void)
255 return flag_is_changeable_p(X86_EFLAGS_ID);
258 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
260 unsigned long lo, hi;
262 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
265 /* Disable processor serial number: */
267 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
269 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
271 pr_notice("CPU serial number disabled.\n");
272 clear_cpu_cap(c, X86_FEATURE_PN);
274 /* Disabling the serial number may affect the cpuid level */
275 c->cpuid_level = cpuid_eax(0);
278 static int __init x86_serial_nr_setup(char *s)
280 disable_x86_serial_nr = 0;
283 __setup("serialnumber", x86_serial_nr_setup);
285 static inline int flag_is_changeable_p(u32 flag)
289 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
294 static __init int setup_disable_smep(char *arg)
296 setup_clear_cpu_cap(X86_FEATURE_SMEP);
297 /* Check for things that depend on SMEP being enabled: */
298 check_mpx_erratum(&boot_cpu_data);
301 __setup("nosmep", setup_disable_smep);
303 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
305 if (cpu_has(c, X86_FEATURE_SMEP))
306 cr4_set_bits(X86_CR4_SMEP);
309 static __init int setup_disable_smap(char *arg)
311 setup_clear_cpu_cap(X86_FEATURE_SMAP);
314 __setup("nosmap", setup_disable_smap);
316 static __always_inline void setup_smap(struct cpuinfo_x86 *c)
318 unsigned long eflags = native_save_fl();
320 /* This should have been cleared long ago */
321 BUG_ON(eflags & X86_EFLAGS_AC);
323 if (cpu_has(c, X86_FEATURE_SMAP)) {
324 #ifdef CONFIG_X86_SMAP
325 cr4_set_bits(X86_CR4_SMAP);
327 cr4_clear_bits(X86_CR4_SMAP);
333 * Protection Keys are not available in 32-bit mode.
335 static bool pku_disabled;
337 static __always_inline void setup_pku(struct cpuinfo_x86 *c)
339 /* check the boot processor, plus compile options for PKU: */
340 if (!cpu_feature_enabled(X86_FEATURE_PKU))
342 /* checks the actual processor's cpuid bits: */
343 if (!cpu_has(c, X86_FEATURE_PKU))
348 cr4_set_bits(X86_CR4_PKE);
350 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
351 * cpuid bit to be set. We need to ensure that we
352 * update that bit in this CPU's "cpu_info".
357 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
358 static __init int setup_disable_pku(char *arg)
361 * Do not clear the X86_FEATURE_PKU bit. All of the
362 * runtime checks are against OSPKE so clearing the
365 * This way, we will see "pku" in cpuinfo, but not
366 * "ospke", which is exactly what we want. It shows
367 * that the CPU has PKU, but the OS has not enabled it.
368 * This happens to be exactly how a system would look
369 * if we disabled the config option.
371 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
375 __setup("nopku", setup_disable_pku);
376 #endif /* CONFIG_X86_64 */
379 * Some CPU features depend on higher CPUID levels, which may not always
380 * be available due to CPUID level capping or broken virtualization
381 * software. Add those features to this table to auto-disable them.
383 struct cpuid_dependent_feature {
388 static const struct cpuid_dependent_feature
389 cpuid_dependent_features[] = {
390 { X86_FEATURE_MWAIT, 0x00000005 },
391 { X86_FEATURE_DCA, 0x00000009 },
392 { X86_FEATURE_XSAVE, 0x0000000d },
396 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
398 const struct cpuid_dependent_feature *df;
400 for (df = cpuid_dependent_features; df->feature; df++) {
402 if (!cpu_has(c, df->feature))
405 * Note: cpuid_level is set to -1 if unavailable, but
406 * extended_extended_level is set to 0 if unavailable
407 * and the legitimate extended levels are all negative
408 * when signed; hence the weird messing around with
411 if (!((s32)df->level < 0 ?
412 (u32)df->level > (u32)c->extended_cpuid_level :
413 (s32)df->level > (s32)c->cpuid_level))
416 clear_cpu_cap(c, df->feature);
420 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
421 x86_cap_flag(df->feature), df->level);
426 * Naming convention should be: <Name> [(<Codename>)]
427 * This table only is used unless init_<vendor>() below doesn't set it;
428 * in particular, if CPUID levels 0x80000002..4 are supported, this
432 /* Look up CPU names by table lookup. */
433 static const char *table_lookup_model(struct cpuinfo_x86 *c)
436 const struct legacy_cpu_model_info *info;
438 if (c->x86_model >= 16)
439 return NULL; /* Range check */
444 info = this_cpu->legacy_models;
446 while (info->family) {
447 if (info->family == c->x86)
448 return info->model_names[c->x86_model];
452 return NULL; /* Not found */
455 __u32 cpu_caps_cleared[NCAPINTS];
456 __u32 cpu_caps_set[NCAPINTS];
458 void load_percpu_segment(int cpu)
461 loadsegment(fs, __KERNEL_PERCPU);
463 __loadsegment_simple(gs, 0);
464 wrmsrl(MSR_GS_BASE, (unsigned long)per_cpu(irq_stack_union.gs_base, cpu));
466 load_stack_canary_segment();
469 /* Setup the fixmap mapping only once per-processor */
470 static inline void setup_fixmap_gdt(int cpu)
473 /* On 64-bit systems, we use a read-only fixmap GDT. */
474 pgprot_t prot = PAGE_KERNEL_RO;
477 * On native 32-bit systems, the GDT cannot be read-only because
478 * our double fault handler uses a task gate, and entering through
479 * a task gate needs to change an available TSS to busy. If the GDT
480 * is read-only, that will triple fault.
482 * On Xen PV, the GDT must be read-only because the hypervisor requires
485 pgprot_t prot = boot_cpu_has(X86_FEATURE_XENPV) ?
486 PAGE_KERNEL_RO : PAGE_KERNEL;
489 __set_fixmap(get_cpu_gdt_ro_index(cpu), get_cpu_gdt_paddr(cpu), prot);
492 /* Load the original GDT from the per-cpu structure */
493 void load_direct_gdt(int cpu)
495 struct desc_ptr gdt_descr;
497 gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
498 gdt_descr.size = GDT_SIZE - 1;
499 load_gdt(&gdt_descr);
501 EXPORT_SYMBOL_GPL(load_direct_gdt);
503 /* Load a fixmap remapping of the per-cpu GDT */
504 void load_fixmap_gdt(int cpu)
506 struct desc_ptr gdt_descr;
508 gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
509 gdt_descr.size = GDT_SIZE - 1;
510 load_gdt(&gdt_descr);
512 EXPORT_SYMBOL_GPL(load_fixmap_gdt);
515 * Current gdt points %fs at the "master" per-cpu area: after this,
516 * it's on the real one.
518 void switch_to_new_gdt(int cpu)
520 /* Load the original GDT */
521 load_direct_gdt(cpu);
522 /* Reload the per-cpu base */
523 load_percpu_segment(cpu);
526 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
528 static void get_model_name(struct cpuinfo_x86 *c)
533 if (c->extended_cpuid_level < 0x80000004)
536 v = (unsigned int *)c->x86_model_id;
537 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
538 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
539 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
540 c->x86_model_id[48] = 0;
542 /* Trim whitespace */
543 p = q = s = &c->x86_model_id[0];
549 /* Note the last non-whitespace index */
559 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
561 unsigned int n, dummy, ebx, ecx, edx, l2size;
563 n = c->extended_cpuid_level;
565 if (n >= 0x80000005) {
566 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
567 c->x86_cache_size = (ecx>>24) + (edx>>24);
569 /* On K8 L1 TLB is inclusive, so don't count it */
574 if (n < 0x80000006) /* Some chips just has a large L1. */
577 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
581 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
583 /* do processor-specific cache resizing */
584 if (this_cpu->legacy_cache_size)
585 l2size = this_cpu->legacy_cache_size(c, l2size);
587 /* Allow user to override all this if necessary. */
588 if (cachesize_override != -1)
589 l2size = cachesize_override;
592 return; /* Again, no L2 cache is possible */
595 c->x86_cache_size = l2size;
598 u16 __read_mostly tlb_lli_4k[NR_INFO];
599 u16 __read_mostly tlb_lli_2m[NR_INFO];
600 u16 __read_mostly tlb_lli_4m[NR_INFO];
601 u16 __read_mostly tlb_lld_4k[NR_INFO];
602 u16 __read_mostly tlb_lld_2m[NR_INFO];
603 u16 __read_mostly tlb_lld_4m[NR_INFO];
604 u16 __read_mostly tlb_lld_1g[NR_INFO];
606 static void cpu_detect_tlb(struct cpuinfo_x86 *c)
608 if (this_cpu->c_detect_tlb)
609 this_cpu->c_detect_tlb(c);
611 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
612 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
613 tlb_lli_4m[ENTRIES]);
615 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
616 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
617 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
620 void detect_ht(struct cpuinfo_x86 *c)
623 u32 eax, ebx, ecx, edx;
624 int index_msb, core_bits;
627 if (!cpu_has(c, X86_FEATURE_HT))
630 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
633 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
636 cpuid(1, &eax, &ebx, &ecx, &edx);
638 smp_num_siblings = (ebx & 0xff0000) >> 16;
640 if (smp_num_siblings == 1) {
641 pr_info_once("CPU0: Hyper-Threading is disabled\n");
645 if (smp_num_siblings <= 1)
648 index_msb = get_count_order(smp_num_siblings);
649 c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
651 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
653 index_msb = get_count_order(smp_num_siblings);
655 core_bits = get_count_order(c->x86_max_cores);
657 c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
658 ((1 << core_bits) - 1);
661 if (!printed && (c->x86_max_cores * smp_num_siblings) > 1) {
662 pr_info("CPU: Physical Processor ID: %d\n",
664 pr_info("CPU: Processor Core ID: %d\n",
671 static void get_cpu_vendor(struct cpuinfo_x86 *c)
673 char *v = c->x86_vendor_id;
676 for (i = 0; i < X86_VENDOR_NUM; i++) {
680 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
681 (cpu_devs[i]->c_ident[1] &&
682 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
684 this_cpu = cpu_devs[i];
685 c->x86_vendor = this_cpu->c_x86_vendor;
690 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
691 "CPU: Your system may be unstable.\n", v);
693 c->x86_vendor = X86_VENDOR_UNKNOWN;
694 this_cpu = &default_cpu;
697 void cpu_detect(struct cpuinfo_x86 *c)
699 /* Get vendor name */
700 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
701 (unsigned int *)&c->x86_vendor_id[0],
702 (unsigned int *)&c->x86_vendor_id[8],
703 (unsigned int *)&c->x86_vendor_id[4]);
706 /* Intel-defined flags: level 0x00000001 */
707 if (c->cpuid_level >= 0x00000001) {
708 u32 junk, tfms, cap0, misc;
710 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
711 c->x86 = x86_family(tfms);
712 c->x86_model = x86_model(tfms);
713 c->x86_mask = x86_stepping(tfms);
715 if (cap0 & (1<<19)) {
716 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
717 c->x86_cache_alignment = c->x86_clflush_size;
722 static void apply_forced_caps(struct cpuinfo_x86 *c)
726 for (i = 0; i < NCAPINTS; i++) {
727 c->x86_capability[i] &= ~cpu_caps_cleared[i];
728 c->x86_capability[i] |= cpu_caps_set[i];
732 void get_cpu_cap(struct cpuinfo_x86 *c)
734 u32 eax, ebx, ecx, edx;
736 /* Intel-defined flags: level 0x00000001 */
737 if (c->cpuid_level >= 0x00000001) {
738 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
740 c->x86_capability[CPUID_1_ECX] = ecx;
741 c->x86_capability[CPUID_1_EDX] = edx;
744 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
745 if (c->cpuid_level >= 0x00000006)
746 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
748 /* Additional Intel-defined flags: level 0x00000007 */
749 if (c->cpuid_level >= 0x00000007) {
750 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
751 c->x86_capability[CPUID_7_0_EBX] = ebx;
752 c->x86_capability[CPUID_7_ECX] = ecx;
755 /* Extended state features: level 0x0000000d */
756 if (c->cpuid_level >= 0x0000000d) {
757 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
759 c->x86_capability[CPUID_D_1_EAX] = eax;
762 /* Additional Intel-defined flags: level 0x0000000F */
763 if (c->cpuid_level >= 0x0000000F) {
765 /* QoS sub-leaf, EAX=0Fh, ECX=0 */
766 cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
767 c->x86_capability[CPUID_F_0_EDX] = edx;
769 if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
770 /* will be overridden if occupancy monitoring exists */
771 c->x86_cache_max_rmid = ebx;
773 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
774 cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
775 c->x86_capability[CPUID_F_1_EDX] = edx;
777 if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
778 ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
779 (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
780 c->x86_cache_max_rmid = ecx;
781 c->x86_cache_occ_scale = ebx;
784 c->x86_cache_max_rmid = -1;
785 c->x86_cache_occ_scale = -1;
789 /* AMD-defined flags: level 0x80000001 */
790 eax = cpuid_eax(0x80000000);
791 c->extended_cpuid_level = eax;
793 if ((eax & 0xffff0000) == 0x80000000) {
794 if (eax >= 0x80000001) {
795 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
797 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
798 c->x86_capability[CPUID_8000_0001_EDX] = edx;
802 if (c->extended_cpuid_level >= 0x80000007) {
803 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
805 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
809 if (c->extended_cpuid_level >= 0x80000008) {
810 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
812 c->x86_virt_bits = (eax >> 8) & 0xff;
813 c->x86_phys_bits = eax & 0xff;
814 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
817 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
818 c->x86_phys_bits = 36;
821 if (c->extended_cpuid_level >= 0x8000000a)
822 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
824 init_scattered_cpuid_features(c);
827 * Clear/Set all flags overridden by options, after probe.
828 * This needs to happen each time we re-probe, which may happen
829 * several times during CPU initialization.
831 apply_forced_caps(c);
834 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
840 * First of all, decide if this is a 486 or higher
841 * It's a 486 if we can modify the AC flag
843 if (flag_is_changeable_p(X86_EFLAGS_AC))
848 for (i = 0; i < X86_VENDOR_NUM; i++)
849 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
850 c->x86_vendor_id[0] = 0;
851 cpu_devs[i]->c_identify(c);
852 if (c->x86_vendor_id[0]) {
861 * Do minimum CPU detection early.
862 * Fields really needed: vendor, cpuid_level, family, model, mask,
864 * The others are not touched to avoid unwanted side effects.
866 * WARNING: this function is only called on the BP. Don't add code here
867 * that is supposed to run on all CPUs.
869 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
872 c->x86_clflush_size = 64;
873 c->x86_phys_bits = 36;
874 c->x86_virt_bits = 48;
876 c->x86_clflush_size = 32;
877 c->x86_phys_bits = 32;
878 c->x86_virt_bits = 32;
880 c->x86_cache_alignment = c->x86_clflush_size;
882 memset(&c->x86_capability, 0, sizeof c->x86_capability);
883 c->extended_cpuid_level = 0;
885 /* cyrix could have cpuid enabled via c_identify()*/
886 if (have_cpuid_p()) {
890 setup_force_cpu_cap(X86_FEATURE_CPUID);
892 if (this_cpu->c_early_init)
893 this_cpu->c_early_init(c);
896 filter_cpuid_features(c, false);
898 if (this_cpu->c_bsp_init)
899 this_cpu->c_bsp_init(c);
901 identify_cpu_without_cpuid(c);
902 setup_clear_cpu_cap(X86_FEATURE_CPUID);
905 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
910 * Regardless of whether PCID is enumerated, the SDM says
911 * that it can't be enabled in 32-bit mode.
913 setup_clear_cpu_cap(X86_FEATURE_PCID);
917 void __init early_cpu_init(void)
919 const struct cpu_dev *const *cdev;
922 #ifdef CONFIG_PROCESSOR_SELECT
923 pr_info("KERNEL supported cpus:\n");
926 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
927 const struct cpu_dev *cpudev = *cdev;
929 if (count >= X86_VENDOR_NUM)
931 cpu_devs[count] = cpudev;
934 #ifdef CONFIG_PROCESSOR_SELECT
938 for (j = 0; j < 2; j++) {
939 if (!cpudev->c_ident[j])
941 pr_info(" %s %s\n", cpudev->c_vendor,
947 early_identify_cpu(&boot_cpu_data);
951 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
952 * unfortunately, that's not true in practice because of early VIA
953 * chips and (more importantly) broken virtualizers that are not easy
954 * to detect. In the latter case it doesn't even *fail* reliably, so
955 * probing for it doesn't even work. Disable it completely on 32-bit
956 * unless we can find a reliable way to detect all the broken cases.
957 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
959 static void detect_nopl(struct cpuinfo_x86 *c)
962 clear_cpu_cap(c, X86_FEATURE_NOPL);
964 set_cpu_cap(c, X86_FEATURE_NOPL);
968 static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
972 * Empirically, writing zero to a segment selector on AMD does
973 * not clear the base, whereas writing zero to a segment
974 * selector on Intel does clear the base. Intel's behavior
975 * allows slightly faster context switches in the common case
976 * where GS is unused by the prev and next threads.
978 * Since neither vendor documents this anywhere that I can see,
979 * detect it directly instead of hardcoding the choice by
982 * I've designated AMD's behavior as the "bug" because it's
983 * counterintuitive and less friendly.
986 unsigned long old_base, tmp;
987 rdmsrl(MSR_FS_BASE, old_base);
988 wrmsrl(MSR_FS_BASE, 1);
990 rdmsrl(MSR_FS_BASE, tmp);
992 set_cpu_bug(c, X86_BUG_NULL_SEG);
993 wrmsrl(MSR_FS_BASE, old_base);
997 static void generic_identify(struct cpuinfo_x86 *c)
999 c->extended_cpuid_level = 0;
1001 if (!have_cpuid_p())
1002 identify_cpu_without_cpuid(c);
1004 /* cyrix could have cpuid enabled via c_identify()*/
1005 if (!have_cpuid_p())
1014 if (c->cpuid_level >= 0x00000001) {
1015 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1016 #ifdef CONFIG_X86_32
1018 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1020 c->apicid = c->initial_apicid;
1023 c->phys_proc_id = c->initial_apicid;
1026 get_model_name(c); /* Default name */
1030 detect_null_seg_behavior(c);
1033 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1034 * systems that run Linux at CPL > 0 may or may not have the
1035 * issue, but, even if they have the issue, there's absolutely
1036 * nothing we can do about it because we can't use the real IRET
1039 * NB: For the time being, only 32-bit kernels support
1040 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1041 * whether to apply espfix using paravirt hooks. If any
1042 * non-paravirt system ever shows up that does *not* have the
1043 * ESPFIX issue, we can change this.
1045 #ifdef CONFIG_X86_32
1046 # ifdef CONFIG_PARAVIRT
1048 extern void native_iret(void);
1049 if (pv_cpu_ops.iret == native_iret)
1050 set_cpu_bug(c, X86_BUG_ESPFIX);
1053 set_cpu_bug(c, X86_BUG_ESPFIX);
1058 static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1061 * The heavy lifting of max_rmid and cache_occ_scale are handled
1062 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
1063 * in case CQM bits really aren't there in this CPU.
1065 if (c != &boot_cpu_data) {
1066 boot_cpu_data.x86_cache_max_rmid =
1067 min(boot_cpu_data.x86_cache_max_rmid,
1068 c->x86_cache_max_rmid);
1073 * Validate that ACPI/mptables have the same information about the
1074 * effective APIC id and update the package map.
1076 static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1079 unsigned int apicid, cpu = smp_processor_id();
1081 apicid = apic->cpu_present_to_apicid(cpu);
1083 if (apicid != c->apicid) {
1084 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1085 cpu, apicid, c->initial_apicid);
1087 BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
1089 c->logical_proc_id = 0;
1094 * This does the hard work of actually picking apart the CPU stuff...
1096 static void identify_cpu(struct cpuinfo_x86 *c)
1100 c->loops_per_jiffy = loops_per_jiffy;
1101 c->x86_cache_size = -1;
1102 c->x86_vendor = X86_VENDOR_UNKNOWN;
1103 c->x86_model = c->x86_mask = 0; /* So far unknown... */
1104 c->x86_vendor_id[0] = '\0'; /* Unset */
1105 c->x86_model_id[0] = '\0'; /* Unset */
1106 c->x86_max_cores = 1;
1107 c->x86_coreid_bits = 0;
1109 #ifdef CONFIG_X86_64
1110 c->x86_clflush_size = 64;
1111 c->x86_phys_bits = 36;
1112 c->x86_virt_bits = 48;
1114 c->cpuid_level = -1; /* CPUID not detected */
1115 c->x86_clflush_size = 32;
1116 c->x86_phys_bits = 32;
1117 c->x86_virt_bits = 32;
1119 c->x86_cache_alignment = c->x86_clflush_size;
1120 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1122 generic_identify(c);
1124 if (this_cpu->c_identify)
1125 this_cpu->c_identify(c);
1127 /* Clear/Set all flags overridden by options, after probe */
1128 apply_forced_caps(c);
1130 #ifdef CONFIG_X86_64
1131 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1135 * Vendor-specific initialization. In this section we
1136 * canonicalize the feature flags, meaning if there are
1137 * features a certain CPU supports which CPUID doesn't
1138 * tell us, CPUID claiming incorrect flags, or other bugs,
1139 * we handle them here.
1141 * At the end of this section, c->x86_capability better
1142 * indicate the features this CPU genuinely supports!
1144 if (this_cpu->c_init)
1145 this_cpu->c_init(c);
1147 /* Disable the PN if appropriate */
1148 squash_the_stupid_serial_number(c);
1150 /* Set up SMEP/SMAP */
1155 * The vendor-specific functions might have changed features.
1156 * Now we do "generic changes."
1159 /* Filter out anything that depends on CPUID levels we don't have */
1160 filter_cpuid_features(c, true);
1162 /* If the model name is still unset, do table lookup. */
1163 if (!c->x86_model_id[0]) {
1165 p = table_lookup_model(c);
1167 strcpy(c->x86_model_id, p);
1169 /* Last resort... */
1170 sprintf(c->x86_model_id, "%02x/%02x",
1171 c->x86, c->x86_model);
1174 #ifdef CONFIG_X86_64
1179 x86_init_cache_qos(c);
1183 * Clear/Set all flags overridden by options, need do it
1184 * before following smp all cpus cap AND.
1186 apply_forced_caps(c);
1189 * On SMP, boot_cpu_data holds the common feature set between
1190 * all CPUs; so make sure that we indicate which features are
1191 * common between the CPUs. The first time this routine gets
1192 * executed, c == &boot_cpu_data.
1194 if (c != &boot_cpu_data) {
1195 /* AND the already accumulated flags with these */
1196 for (i = 0; i < NCAPINTS; i++)
1197 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1199 /* OR, i.e. replicate the bug flags */
1200 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1201 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1204 /* Init Machine Check Exception if available. */
1207 select_idle_routine(c);
1210 numa_add_cpu(smp_processor_id());
1215 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1216 * on 32-bit kernels:
1218 #ifdef CONFIG_X86_32
1219 void enable_sep_cpu(void)
1221 struct tss_struct *tss;
1224 if (!boot_cpu_has(X86_FEATURE_SEP))
1228 tss = &per_cpu(cpu_tss, cpu);
1231 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1232 * see the big comment in struct x86_hw_tss's definition.
1235 tss->x86_tss.ss1 = __KERNEL_CS;
1236 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1238 wrmsr(MSR_IA32_SYSENTER_ESP,
1239 (unsigned long)tss + offsetofend(struct tss_struct, SYSENTER_stack),
1242 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1248 void __init identify_boot_cpu(void)
1250 identify_cpu(&boot_cpu_data);
1251 #ifdef CONFIG_X86_32
1255 cpu_detect_tlb(&boot_cpu_data);
1258 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1260 BUG_ON(c == &boot_cpu_data);
1262 #ifdef CONFIG_X86_32
1266 validate_apic_and_package_id(c);
1269 static __init int setup_noclflush(char *arg)
1271 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1272 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1275 __setup("noclflush", setup_noclflush);
1277 void print_cpu_info(struct cpuinfo_x86 *c)
1279 const char *vendor = NULL;
1281 if (c->x86_vendor < X86_VENDOR_NUM) {
1282 vendor = this_cpu->c_vendor;
1284 if (c->cpuid_level >= 0)
1285 vendor = c->x86_vendor_id;
1288 if (vendor && !strstr(c->x86_model_id, vendor))
1289 pr_cont("%s ", vendor);
1291 if (c->x86_model_id[0])
1292 pr_cont("%s", c->x86_model_id);
1294 pr_cont("%d86", c->x86);
1296 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1298 if (c->x86_mask || c->cpuid_level >= 0)
1299 pr_cont(", stepping: 0x%x)\n", c->x86_mask);
1304 static __init int setup_disablecpuid(char *arg)
1308 if (get_option(&arg, &bit) && bit >= 0 && bit < NCAPINTS * 32)
1309 setup_clear_cpu_cap(bit);
1315 __setup("clearcpuid=", setup_disablecpuid);
1317 #ifdef CONFIG_X86_64
1318 DEFINE_PER_CPU_FIRST(union irq_stack_union,
1319 irq_stack_union) __aligned(PAGE_SIZE) __visible;
1322 * The following percpu variables are hot. Align current_task to
1323 * cacheline size such that they fall in the same cacheline.
1325 DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1327 EXPORT_PER_CPU_SYMBOL(current_task);
1329 DEFINE_PER_CPU(char *, irq_stack_ptr) =
1330 init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE;
1332 DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1334 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1335 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1338 * Special IST stacks which the CPU switches to when it calls
1339 * an IST-marked descriptor entry. Up to 7 stacks (hardware
1340 * limit), all of them are 4K, except the debug stack which
1343 static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
1344 [0 ... N_EXCEPTION_STACKS - 1] = EXCEPTION_STKSZ,
1345 [DEBUG_STACK - 1] = DEBUG_STKSZ
1348 static DEFINE_PER_CPU_PAGE_ALIGNED(char, exception_stacks
1349 [(N_EXCEPTION_STACKS - 1) * EXCEPTION_STKSZ + DEBUG_STKSZ]);
1351 /* May not be marked __init: used by software suspend */
1352 void syscall_init(void)
1354 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1355 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1357 #ifdef CONFIG_IA32_EMULATION
1358 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1360 * This only works on Intel CPUs.
1361 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1362 * This does not cause SYSENTER to jump to the wrong location, because
1363 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1365 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1366 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1367 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1369 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1370 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1371 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1372 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1375 /* Flags to clear on syscall */
1376 wrmsrl(MSR_SYSCALL_MASK,
1377 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1378 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1382 * Copies of the original ist values from the tss are only accessed during
1383 * debugging, no special alignment required.
1385 DEFINE_PER_CPU(struct orig_ist, orig_ist);
1387 static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1388 DEFINE_PER_CPU(int, debug_stack_usage);
1390 int is_debug_stack(unsigned long addr)
1392 return __this_cpu_read(debug_stack_usage) ||
1393 (addr <= __this_cpu_read(debug_stack_addr) &&
1394 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1396 NOKPROBE_SYMBOL(is_debug_stack);
1398 DEFINE_PER_CPU(u32, debug_idt_ctr);
1400 void debug_stack_set_zero(void)
1402 this_cpu_inc(debug_idt_ctr);
1405 NOKPROBE_SYMBOL(debug_stack_set_zero);
1407 void debug_stack_reset(void)
1409 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1411 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1414 NOKPROBE_SYMBOL(debug_stack_reset);
1416 #else /* CONFIG_X86_64 */
1418 DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1419 EXPORT_PER_CPU_SYMBOL(current_task);
1420 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1421 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1424 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1425 * the top of the kernel stack. Use an extra percpu variable to track the
1426 * top of the kernel stack directly.
1428 DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1429 (unsigned long)&init_thread_union + THREAD_SIZE;
1430 EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1432 #ifdef CONFIG_CC_STACKPROTECTOR
1433 DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1436 #endif /* CONFIG_X86_64 */
1439 * Clear all 6 debug registers:
1441 static void clear_all_debug_regs(void)
1445 for (i = 0; i < 8; i++) {
1446 /* Ignore db4, db5 */
1447 if ((i == 4) || (i == 5))
1456 * Restore debug regs if using kgdbwait and you have a kernel debugger
1457 * connection established.
1459 static void dbg_restore_debug_regs(void)
1461 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1462 arch_kgdb_ops.correct_hw_break();
1464 #else /* ! CONFIG_KGDB */
1465 #define dbg_restore_debug_regs()
1466 #endif /* ! CONFIG_KGDB */
1468 static void wait_for_master_cpu(int cpu)
1472 * wait for ACK from master CPU before continuing
1473 * with AP initialization
1475 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1476 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1482 * cpu_init() initializes state that is per-CPU. Some data is already
1483 * initialized (naturally) in the bootstrap process, such as the GDT
1484 * and IDT. We reload them nevertheless, this function acts as a
1485 * 'CPU state barrier', nothing should get across.
1486 * A lot of state is already set up in PDA init for 64 bit
1488 #ifdef CONFIG_X86_64
1492 struct orig_ist *oist;
1493 struct task_struct *me;
1494 struct tss_struct *t;
1496 int cpu = raw_smp_processor_id();
1499 wait_for_master_cpu(cpu);
1502 * Initialize the CR4 shadow before doing anything that could
1510 t = &per_cpu(cpu_tss, cpu);
1511 oist = &per_cpu(orig_ist, cpu);
1514 if (this_cpu_read(numa_node) == 0 &&
1515 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1516 set_numa_node(early_cpu_to_node(cpu));
1521 pr_debug("Initializing CPU#%d\n", cpu);
1523 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1526 * Initialize the per-CPU GDT with the boot GDT,
1527 * and set up the GDT descriptor:
1530 switch_to_new_gdt(cpu);
1535 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1538 wrmsrl(MSR_FS_BASE, 0);
1539 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1546 * set up and load the per-CPU TSS
1548 if (!oist->ist[0]) {
1549 char *estacks = per_cpu(exception_stacks, cpu);
1551 for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1552 estacks += exception_stack_sizes[v];
1553 oist->ist[v] = t->x86_tss.ist[v] =
1554 (unsigned long)estacks;
1555 if (v == DEBUG_STACK-1)
1556 per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1560 t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1563 * <= is required because the CPU will access up to
1564 * 8 bits beyond the end of the IO permission bitmap.
1566 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1567 t->io_bitmap[i] = ~0UL;
1570 me->active_mm = &init_mm;
1572 initialize_tlbstate_and_flush();
1573 enter_lazy_tlb(&init_mm, me);
1575 load_sp0(t, ¤t->thread);
1576 set_tss_desc(cpu, t);
1578 load_mm_ldt(&init_mm);
1580 clear_all_debug_regs();
1581 dbg_restore_debug_regs();
1588 setup_fixmap_gdt(cpu);
1589 load_fixmap_gdt(cpu);
1596 int cpu = smp_processor_id();
1597 struct task_struct *curr = current;
1598 struct tss_struct *t = &per_cpu(cpu_tss, cpu);
1599 struct thread_struct *thread = &curr->thread;
1601 wait_for_master_cpu(cpu);
1604 * Initialize the CR4 shadow before doing anything that could
1609 show_ucode_info_early();
1611 pr_info("Initializing CPU#%d\n", cpu);
1613 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1614 boot_cpu_has(X86_FEATURE_TSC) ||
1615 boot_cpu_has(X86_FEATURE_DE))
1616 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1619 switch_to_new_gdt(cpu);
1622 * Set up and load the per-CPU TSS and LDT
1625 curr->active_mm = &init_mm;
1627 initialize_tlbstate_and_flush();
1628 enter_lazy_tlb(&init_mm, curr);
1630 load_sp0(t, thread);
1631 set_tss_desc(cpu, t);
1633 load_mm_ldt(&init_mm);
1635 t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1637 #ifdef CONFIG_DOUBLEFAULT
1638 /* Set up doublefault TSS pointer in the GDT */
1639 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1642 clear_all_debug_regs();
1643 dbg_restore_debug_regs();
1647 setup_fixmap_gdt(cpu);
1648 load_fixmap_gdt(cpu);
1652 static void bsp_resume(void)
1654 if (this_cpu->c_bsp_resume)
1655 this_cpu->c_bsp_resume(&boot_cpu_data);
1658 static struct syscore_ops cpu_syscore_ops = {
1659 .resume = bsp_resume,
1662 static int __init init_cpu_syscore(void)
1664 register_syscore_ops(&cpu_syscore_ops);
1667 core_initcall(init_cpu_syscore);