1 /* cpu_feature_enabled() cannot be used this early */
2 #define USE_EARLY_PGTABLE_L5
4 #include <linux/memblock.h>
5 #include <linux/linkage.h>
6 #include <linux/bitops.h>
7 #include <linux/kernel.h>
8 #include <linux/export.h>
9 #include <linux/percpu.h>
10 #include <linux/string.h>
11 #include <linux/ctype.h>
12 #include <linux/delay.h>
13 #include <linux/sched/mm.h>
14 #include <linux/sched/clock.h>
15 #include <linux/sched/task.h>
16 #include <linux/init.h>
17 #include <linux/kprobes.h>
18 #include <linux/kgdb.h>
19 #include <linux/smp.h>
21 #include <linux/syscore_ops.h>
23 #include <asm/stackprotector.h>
24 #include <asm/perf_event.h>
25 #include <asm/mmu_context.h>
26 #include <asm/archrandom.h>
27 #include <asm/hypervisor.h>
28 #include <asm/processor.h>
29 #include <asm/tlbflush.h>
30 #include <asm/debugreg.h>
31 #include <asm/sections.h>
32 #include <asm/vsyscall.h>
33 #include <linux/topology.h>
34 #include <linux/cpumask.h>
35 #include <asm/pgtable.h>
36 #include <linux/atomic.h>
37 #include <asm/proto.h>
38 #include <asm/setup.h>
41 #include <asm/fpu/internal.h>
43 #include <asm/hwcap2.h>
44 #include <linux/numa.h>
51 #include <asm/microcode.h>
52 #include <asm/microcode_intel.h>
53 #include <asm/intel-family.h>
54 #include <asm/cpu_device_id.h>
56 #ifdef CONFIG_X86_LOCAL_APIC
57 #include <asm/uv/uv.h>
62 u32 elf_hwcap2 __read_mostly;
64 /* all of these masks are initialized in setup_cpu_local_masks() */
65 cpumask_var_t cpu_initialized_mask;
66 cpumask_var_t cpu_callout_mask;
67 cpumask_var_t cpu_callin_mask;
69 /* representing cpus for which sibling maps can be computed */
70 cpumask_var_t cpu_sibling_setup_mask;
72 /* Number of siblings per CPU package */
73 int smp_num_siblings = 1;
74 EXPORT_SYMBOL(smp_num_siblings);
76 /* Last level cache ID of each logical CPU */
77 DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
79 /* correctly size the local cpu masks */
80 void __init setup_cpu_local_masks(void)
82 alloc_bootmem_cpumask_var(&cpu_initialized_mask);
83 alloc_bootmem_cpumask_var(&cpu_callin_mask);
84 alloc_bootmem_cpumask_var(&cpu_callout_mask);
85 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
88 static void default_init(struct cpuinfo_x86 *c)
91 cpu_detect_cache_sizes(c);
93 /* Not much we can do here... */
94 /* Check if at least it has cpuid */
95 if (c->cpuid_level == -1) {
96 /* No cpuid. It must be an ancient CPU */
98 strcpy(c->x86_model_id, "486");
100 strcpy(c->x86_model_id, "386");
105 static const struct cpu_dev default_cpu = {
106 .c_init = default_init,
107 .c_vendor = "Unknown",
108 .c_x86_vendor = X86_VENDOR_UNKNOWN,
111 static const struct cpu_dev *this_cpu = &default_cpu;
113 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
116 * We need valid kernel segments for data and code in long mode too
117 * IRET will check the segment types kkeil 2000/10/28
118 * Also sysret mandates a special GDT layout
120 * TLS descriptors are currently at a different place compared to i386.
121 * Hopefully nobody expects them at a fixed place (Wine?)
123 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
124 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
125 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
126 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
127 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
128 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
130 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
131 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
132 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
133 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
135 * Segments used for calling PnP BIOS have byte granularity.
136 * They code segments and data segments have fixed 64k limits,
137 * the transfer segment sizes are set at run time.
140 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
142 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
144 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(0x0092, 0, 0xffff),
146 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(0x0092, 0, 0),
148 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(0x0092, 0, 0),
150 * The APM segments have byte granularity and their bases
151 * are set at run time. All have 64k limits.
154 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
156 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
158 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(0x4092, 0, 0xffff),
160 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
161 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
162 GDT_STACK_CANARY_INIT
165 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
167 static int __init x86_mpx_setup(char *s)
169 /* require an exact match without trailing characters */
173 /* do not emit a message if the feature is not present */
174 if (!boot_cpu_has(X86_FEATURE_MPX))
177 setup_clear_cpu_cap(X86_FEATURE_MPX);
178 pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
181 __setup("nompx", x86_mpx_setup);
184 static int __init x86_nopcid_setup(char *s)
186 /* nopcid doesn't accept parameters */
190 /* do not emit a message if the feature is not present */
191 if (!boot_cpu_has(X86_FEATURE_PCID))
194 setup_clear_cpu_cap(X86_FEATURE_PCID);
195 pr_info("nopcid: PCID feature disabled\n");
198 early_param("nopcid", x86_nopcid_setup);
201 static int __init x86_noinvpcid_setup(char *s)
203 /* noinvpcid doesn't accept parameters */
207 /* do not emit a message if the feature is not present */
208 if (!boot_cpu_has(X86_FEATURE_INVPCID))
211 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
212 pr_info("noinvpcid: INVPCID feature disabled\n");
215 early_param("noinvpcid", x86_noinvpcid_setup);
218 static int cachesize_override = -1;
219 static int disable_x86_serial_nr = 1;
221 static int __init cachesize_setup(char *str)
223 get_option(&str, &cachesize_override);
226 __setup("cachesize=", cachesize_setup);
228 static int __init x86_sep_setup(char *s)
230 setup_clear_cpu_cap(X86_FEATURE_SEP);
233 __setup("nosep", x86_sep_setup);
235 /* Standard macro to see if a specific flag is changeable */
236 static inline int flag_is_changeable_p(u32 flag)
241 * Cyrix and IDT cpus allow disabling of CPUID
242 * so the code below may return different results
243 * when it is executed before and after enabling
244 * the CPUID. Add "volatile" to not allow gcc to
245 * optimize the subsequent calls to this function.
247 asm volatile ("pushfl \n\t"
258 : "=&r" (f1), "=&r" (f2)
261 return ((f1^f2) & flag) != 0;
264 /* Probe for the CPUID instruction */
265 int have_cpuid_p(void)
267 return flag_is_changeable_p(X86_EFLAGS_ID);
270 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
272 unsigned long lo, hi;
274 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
277 /* Disable processor serial number: */
279 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
281 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
283 pr_notice("CPU serial number disabled.\n");
284 clear_cpu_cap(c, X86_FEATURE_PN);
286 /* Disabling the serial number may affect the cpuid level */
287 c->cpuid_level = cpuid_eax(0);
290 static int __init x86_serial_nr_setup(char *s)
292 disable_x86_serial_nr = 0;
295 __setup("serialnumber", x86_serial_nr_setup);
297 static inline int flag_is_changeable_p(u32 flag)
301 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
306 static __init int setup_disable_smep(char *arg)
308 setup_clear_cpu_cap(X86_FEATURE_SMEP);
309 /* Check for things that depend on SMEP being enabled: */
310 check_mpx_erratum(&boot_cpu_data);
313 __setup("nosmep", setup_disable_smep);
315 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
317 if (cpu_has(c, X86_FEATURE_SMEP))
318 cr4_set_bits(X86_CR4_SMEP);
321 static __init int setup_disable_smap(char *arg)
323 setup_clear_cpu_cap(X86_FEATURE_SMAP);
326 __setup("nosmap", setup_disable_smap);
328 static __always_inline void setup_smap(struct cpuinfo_x86 *c)
330 unsigned long eflags = native_save_fl();
332 /* This should have been cleared long ago */
333 BUG_ON(eflags & X86_EFLAGS_AC);
335 if (cpu_has(c, X86_FEATURE_SMAP)) {
336 #ifdef CONFIG_X86_SMAP
337 cr4_set_bits(X86_CR4_SMAP);
339 cr4_clear_bits(X86_CR4_SMAP);
344 static __always_inline void setup_umip(struct cpuinfo_x86 *c)
346 /* Check the boot processor, plus build option for UMIP. */
347 if (!cpu_feature_enabled(X86_FEATURE_UMIP))
350 /* Check the current processor's cpuid bits. */
351 if (!cpu_has(c, X86_FEATURE_UMIP))
354 cr4_set_bits(X86_CR4_UMIP);
356 pr_info_once("x86/cpu: User Mode Instruction Prevention (UMIP) activated\n");
362 * Make sure UMIP is disabled in case it was enabled in a
363 * previous boot (e.g., via kexec).
365 cr4_clear_bits(X86_CR4_UMIP);
369 * Protection Keys are not available in 32-bit mode.
371 static bool pku_disabled;
373 static __always_inline void setup_pku(struct cpuinfo_x86 *c)
375 struct pkru_state *pk;
377 /* check the boot processor, plus compile options for PKU: */
378 if (!cpu_feature_enabled(X86_FEATURE_PKU))
380 /* checks the actual processor's cpuid bits: */
381 if (!cpu_has(c, X86_FEATURE_PKU))
386 cr4_set_bits(X86_CR4_PKE);
387 pk = get_xsave_addr(&init_fpstate.xsave, XFEATURE_PKRU);
389 pk->pkru = init_pkru_value;
391 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
392 * cpuid bit to be set. We need to ensure that we
393 * update that bit in this CPU's "cpu_info".
398 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
399 static __init int setup_disable_pku(char *arg)
402 * Do not clear the X86_FEATURE_PKU bit. All of the
403 * runtime checks are against OSPKE so clearing the
406 * This way, we will see "pku" in cpuinfo, but not
407 * "ospke", which is exactly what we want. It shows
408 * that the CPU has PKU, but the OS has not enabled it.
409 * This happens to be exactly how a system would look
410 * if we disabled the config option.
412 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
416 __setup("nopku", setup_disable_pku);
417 #endif /* CONFIG_X86_64 */
420 * Some CPU features depend on higher CPUID levels, which may not always
421 * be available due to CPUID level capping or broken virtualization
422 * software. Add those features to this table to auto-disable them.
424 struct cpuid_dependent_feature {
429 static const struct cpuid_dependent_feature
430 cpuid_dependent_features[] = {
431 { X86_FEATURE_MWAIT, 0x00000005 },
432 { X86_FEATURE_DCA, 0x00000009 },
433 { X86_FEATURE_XSAVE, 0x0000000d },
437 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
439 const struct cpuid_dependent_feature *df;
441 for (df = cpuid_dependent_features; df->feature; df++) {
443 if (!cpu_has(c, df->feature))
446 * Note: cpuid_level is set to -1 if unavailable, but
447 * extended_extended_level is set to 0 if unavailable
448 * and the legitimate extended levels are all negative
449 * when signed; hence the weird messing around with
452 if (!((s32)df->level < 0 ?
453 (u32)df->level > (u32)c->extended_cpuid_level :
454 (s32)df->level > (s32)c->cpuid_level))
457 clear_cpu_cap(c, df->feature);
461 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
462 x86_cap_flag(df->feature), df->level);
467 * Naming convention should be: <Name> [(<Codename>)]
468 * This table only is used unless init_<vendor>() below doesn't set it;
469 * in particular, if CPUID levels 0x80000002..4 are supported, this
473 /* Look up CPU names by table lookup. */
474 static const char *table_lookup_model(struct cpuinfo_x86 *c)
477 const struct legacy_cpu_model_info *info;
479 if (c->x86_model >= 16)
480 return NULL; /* Range check */
485 info = this_cpu->legacy_models;
487 while (info->family) {
488 if (info->family == c->x86)
489 return info->model_names[c->x86_model];
493 return NULL; /* Not found */
496 __u32 cpu_caps_cleared[NCAPINTS + NBUGINTS];
497 __u32 cpu_caps_set[NCAPINTS + NBUGINTS];
499 void load_percpu_segment(int cpu)
502 loadsegment(fs, __KERNEL_PERCPU);
504 __loadsegment_simple(gs, 0);
505 wrmsrl(MSR_GS_BASE, cpu_kernelmode_gs_base(cpu));
507 load_stack_canary_segment();
511 /* The 32-bit entry code needs to find cpu_entry_area. */
512 DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area);
515 /* Load the original GDT from the per-cpu structure */
516 void load_direct_gdt(int cpu)
518 struct desc_ptr gdt_descr;
520 gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
521 gdt_descr.size = GDT_SIZE - 1;
522 load_gdt(&gdt_descr);
524 EXPORT_SYMBOL_GPL(load_direct_gdt);
526 /* Load a fixmap remapping of the per-cpu GDT */
527 void load_fixmap_gdt(int cpu)
529 struct desc_ptr gdt_descr;
531 gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
532 gdt_descr.size = GDT_SIZE - 1;
533 load_gdt(&gdt_descr);
535 EXPORT_SYMBOL_GPL(load_fixmap_gdt);
538 * Current gdt points %fs at the "master" per-cpu area: after this,
539 * it's on the real one.
541 void switch_to_new_gdt(int cpu)
543 /* Load the original GDT */
544 load_direct_gdt(cpu);
545 /* Reload the per-cpu base */
546 load_percpu_segment(cpu);
549 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
551 static void get_model_name(struct cpuinfo_x86 *c)
556 if (c->extended_cpuid_level < 0x80000004)
559 v = (unsigned int *)c->x86_model_id;
560 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
561 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
562 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
563 c->x86_model_id[48] = 0;
565 /* Trim whitespace */
566 p = q = s = &c->x86_model_id[0];
572 /* Note the last non-whitespace index */
582 void detect_num_cpu_cores(struct cpuinfo_x86 *c)
584 unsigned int eax, ebx, ecx, edx;
586 c->x86_max_cores = 1;
587 if (!IS_ENABLED(CONFIG_SMP) || c->cpuid_level < 4)
590 cpuid_count(4, 0, &eax, &ebx, &ecx, &edx);
592 c->x86_max_cores = (eax >> 26) + 1;
595 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
597 unsigned int n, dummy, ebx, ecx, edx, l2size;
599 n = c->extended_cpuid_level;
601 if (n >= 0x80000005) {
602 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
603 c->x86_cache_size = (ecx>>24) + (edx>>24);
605 /* On K8 L1 TLB is inclusive, so don't count it */
610 if (n < 0x80000006) /* Some chips just has a large L1. */
613 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
617 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
619 /* do processor-specific cache resizing */
620 if (this_cpu->legacy_cache_size)
621 l2size = this_cpu->legacy_cache_size(c, l2size);
623 /* Allow user to override all this if necessary. */
624 if (cachesize_override != -1)
625 l2size = cachesize_override;
628 return; /* Again, no L2 cache is possible */
631 c->x86_cache_size = l2size;
634 u16 __read_mostly tlb_lli_4k[NR_INFO];
635 u16 __read_mostly tlb_lli_2m[NR_INFO];
636 u16 __read_mostly tlb_lli_4m[NR_INFO];
637 u16 __read_mostly tlb_lld_4k[NR_INFO];
638 u16 __read_mostly tlb_lld_2m[NR_INFO];
639 u16 __read_mostly tlb_lld_4m[NR_INFO];
640 u16 __read_mostly tlb_lld_1g[NR_INFO];
642 static void cpu_detect_tlb(struct cpuinfo_x86 *c)
644 if (this_cpu->c_detect_tlb)
645 this_cpu->c_detect_tlb(c);
647 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
648 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
649 tlb_lli_4m[ENTRIES]);
651 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
652 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
653 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
656 int detect_ht_early(struct cpuinfo_x86 *c)
659 u32 eax, ebx, ecx, edx;
661 if (!cpu_has(c, X86_FEATURE_HT))
664 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
667 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
670 cpuid(1, &eax, &ebx, &ecx, &edx);
672 smp_num_siblings = (ebx & 0xff0000) >> 16;
673 if (smp_num_siblings == 1)
674 pr_info_once("CPU0: Hyper-Threading is disabled\n");
679 void detect_ht(struct cpuinfo_x86 *c)
682 int index_msb, core_bits;
684 if (detect_ht_early(c) < 0)
687 index_msb = get_count_order(smp_num_siblings);
688 c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
690 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
692 index_msb = get_count_order(smp_num_siblings);
694 core_bits = get_count_order(c->x86_max_cores);
696 c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
697 ((1 << core_bits) - 1);
701 static void get_cpu_vendor(struct cpuinfo_x86 *c)
703 char *v = c->x86_vendor_id;
706 for (i = 0; i < X86_VENDOR_NUM; i++) {
710 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
711 (cpu_devs[i]->c_ident[1] &&
712 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
714 this_cpu = cpu_devs[i];
715 c->x86_vendor = this_cpu->c_x86_vendor;
720 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
721 "CPU: Your system may be unstable.\n", v);
723 c->x86_vendor = X86_VENDOR_UNKNOWN;
724 this_cpu = &default_cpu;
727 void cpu_detect(struct cpuinfo_x86 *c)
729 /* Get vendor name */
730 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
731 (unsigned int *)&c->x86_vendor_id[0],
732 (unsigned int *)&c->x86_vendor_id[8],
733 (unsigned int *)&c->x86_vendor_id[4]);
736 /* Intel-defined flags: level 0x00000001 */
737 if (c->cpuid_level >= 0x00000001) {
738 u32 junk, tfms, cap0, misc;
740 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
741 c->x86 = x86_family(tfms);
742 c->x86_model = x86_model(tfms);
743 c->x86_stepping = x86_stepping(tfms);
745 if (cap0 & (1<<19)) {
746 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
747 c->x86_cache_alignment = c->x86_clflush_size;
752 static void apply_forced_caps(struct cpuinfo_x86 *c)
756 for (i = 0; i < NCAPINTS + NBUGINTS; i++) {
757 c->x86_capability[i] &= ~cpu_caps_cleared[i];
758 c->x86_capability[i] |= cpu_caps_set[i];
762 static void init_speculation_control(struct cpuinfo_x86 *c)
765 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support,
766 * and they also have a different bit for STIBP support. Also,
767 * a hypervisor might have set the individual AMD bits even on
768 * Intel CPUs, for finer-grained selection of what's available.
770 if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) {
771 set_cpu_cap(c, X86_FEATURE_IBRS);
772 set_cpu_cap(c, X86_FEATURE_IBPB);
773 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
776 if (cpu_has(c, X86_FEATURE_INTEL_STIBP))
777 set_cpu_cap(c, X86_FEATURE_STIBP);
779 if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) ||
780 cpu_has(c, X86_FEATURE_VIRT_SSBD))
781 set_cpu_cap(c, X86_FEATURE_SSBD);
783 if (cpu_has(c, X86_FEATURE_AMD_IBRS)) {
784 set_cpu_cap(c, X86_FEATURE_IBRS);
785 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
788 if (cpu_has(c, X86_FEATURE_AMD_IBPB))
789 set_cpu_cap(c, X86_FEATURE_IBPB);
791 if (cpu_has(c, X86_FEATURE_AMD_STIBP)) {
792 set_cpu_cap(c, X86_FEATURE_STIBP);
793 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
796 if (cpu_has(c, X86_FEATURE_AMD_SSBD)) {
797 set_cpu_cap(c, X86_FEATURE_SSBD);
798 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
799 clear_cpu_cap(c, X86_FEATURE_VIRT_SSBD);
803 void get_cpu_cap(struct cpuinfo_x86 *c)
805 u32 eax, ebx, ecx, edx;
807 /* Intel-defined flags: level 0x00000001 */
808 if (c->cpuid_level >= 0x00000001) {
809 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
811 c->x86_capability[CPUID_1_ECX] = ecx;
812 c->x86_capability[CPUID_1_EDX] = edx;
815 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
816 if (c->cpuid_level >= 0x00000006)
817 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
819 /* Additional Intel-defined flags: level 0x00000007 */
820 if (c->cpuid_level >= 0x00000007) {
821 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
822 c->x86_capability[CPUID_7_0_EBX] = ebx;
823 c->x86_capability[CPUID_7_ECX] = ecx;
824 c->x86_capability[CPUID_7_EDX] = edx;
827 /* Extended state features: level 0x0000000d */
828 if (c->cpuid_level >= 0x0000000d) {
829 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
831 c->x86_capability[CPUID_D_1_EAX] = eax;
834 /* Additional Intel-defined flags: level 0x0000000F */
835 if (c->cpuid_level >= 0x0000000F) {
837 /* QoS sub-leaf, EAX=0Fh, ECX=0 */
838 cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
839 c->x86_capability[CPUID_F_0_EDX] = edx;
841 if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
842 /* will be overridden if occupancy monitoring exists */
843 c->x86_cache_max_rmid = ebx;
845 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
846 cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
847 c->x86_capability[CPUID_F_1_EDX] = edx;
849 if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
850 ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
851 (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
852 c->x86_cache_max_rmid = ecx;
853 c->x86_cache_occ_scale = ebx;
856 c->x86_cache_max_rmid = -1;
857 c->x86_cache_occ_scale = -1;
861 /* AMD-defined flags: level 0x80000001 */
862 eax = cpuid_eax(0x80000000);
863 c->extended_cpuid_level = eax;
865 if ((eax & 0xffff0000) == 0x80000000) {
866 if (eax >= 0x80000001) {
867 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
869 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
870 c->x86_capability[CPUID_8000_0001_EDX] = edx;
874 if (c->extended_cpuid_level >= 0x80000007) {
875 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
877 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
881 if (c->extended_cpuid_level >= 0x80000008) {
882 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
883 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
886 if (c->extended_cpuid_level >= 0x8000000a)
887 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
889 init_scattered_cpuid_features(c);
890 init_speculation_control(c);
893 * Clear/Set all flags overridden by options, after probe.
894 * This needs to happen each time we re-probe, which may happen
895 * several times during CPU initialization.
897 apply_forced_caps(c);
900 void get_cpu_address_sizes(struct cpuinfo_x86 *c)
902 u32 eax, ebx, ecx, edx;
904 if (c->extended_cpuid_level >= 0x80000008) {
905 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
907 c->x86_virt_bits = (eax >> 8) & 0xff;
908 c->x86_phys_bits = eax & 0xff;
911 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
912 c->x86_phys_bits = 36;
914 c->x86_cache_bits = c->x86_phys_bits;
917 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
923 * First of all, decide if this is a 486 or higher
924 * It's a 486 if we can modify the AC flag
926 if (flag_is_changeable_p(X86_EFLAGS_AC))
931 for (i = 0; i < X86_VENDOR_NUM; i++)
932 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
933 c->x86_vendor_id[0] = 0;
934 cpu_devs[i]->c_identify(c);
935 if (c->x86_vendor_id[0]) {
943 static const __initconst struct x86_cpu_id cpu_no_speculation[] = {
944 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SALTWELL, X86_FEATURE_ANY },
945 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SALTWELL_TABLET, X86_FEATURE_ANY },
946 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_BONNELL_MID, X86_FEATURE_ANY },
947 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SALTWELL_MID, X86_FEATURE_ANY },
948 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_BONNELL, X86_FEATURE_ANY },
949 { X86_VENDOR_CENTAUR, 5 },
950 { X86_VENDOR_INTEL, 5 },
951 { X86_VENDOR_NSC, 5 },
952 { X86_VENDOR_ANY, 4 },
956 static const __initconst struct x86_cpu_id cpu_no_meltdown[] = {
958 { X86_VENDOR_HYGON },
962 /* Only list CPUs which speculate but are non susceptible to SSB */
963 static const __initconst struct x86_cpu_id cpu_no_spec_store_bypass[] = {
964 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT },
965 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_AIRMONT },
966 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT_X },
967 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT_MID },
968 { X86_VENDOR_INTEL, 6, INTEL_FAM6_CORE_YONAH },
969 { X86_VENDOR_INTEL, 6, INTEL_FAM6_XEON_PHI_KNL },
970 { X86_VENDOR_INTEL, 6, INTEL_FAM6_XEON_PHI_KNM },
971 { X86_VENDOR_AMD, 0x12, },
972 { X86_VENDOR_AMD, 0x11, },
973 { X86_VENDOR_AMD, 0x10, },
974 { X86_VENDOR_AMD, 0xf, },
978 static const __initconst struct x86_cpu_id cpu_no_l1tf[] = {
979 /* in addition to cpu_no_speculation */
980 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT },
981 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT_X },
982 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_AIRMONT },
983 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT_MID },
984 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_AIRMONT_MID },
985 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_GOLDMONT },
986 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_GOLDMONT_X },
987 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_GOLDMONT_PLUS },
988 { X86_VENDOR_INTEL, 6, INTEL_FAM6_XEON_PHI_KNL },
989 { X86_VENDOR_INTEL, 6, INTEL_FAM6_XEON_PHI_KNM },
993 static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
997 if (x86_match_cpu(cpu_no_speculation))
1000 setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
1001 setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
1003 if (cpu_has(c, X86_FEATURE_ARCH_CAPABILITIES))
1004 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
1006 if (!x86_match_cpu(cpu_no_spec_store_bypass) &&
1007 !(ia32_cap & ARCH_CAP_SSB_NO) &&
1008 !cpu_has(c, X86_FEATURE_AMD_SSB_NO))
1009 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
1011 if (ia32_cap & ARCH_CAP_IBRS_ALL)
1012 setup_force_cpu_cap(X86_FEATURE_IBRS_ENHANCED);
1014 if (x86_match_cpu(cpu_no_meltdown))
1017 /* Rogue Data Cache Load? No! */
1018 if (ia32_cap & ARCH_CAP_RDCL_NO)
1021 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
1023 if (x86_match_cpu(cpu_no_l1tf))
1026 setup_force_cpu_bug(X86_BUG_L1TF);
1030 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1031 * unfortunately, that's not true in practice because of early VIA
1032 * chips and (more importantly) broken virtualizers that are not easy
1033 * to detect. In the latter case it doesn't even *fail* reliably, so
1034 * probing for it doesn't even work. Disable it completely on 32-bit
1035 * unless we can find a reliable way to detect all the broken cases.
1036 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1038 static void detect_nopl(void)
1040 #ifdef CONFIG_X86_32
1041 setup_clear_cpu_cap(X86_FEATURE_NOPL);
1043 setup_force_cpu_cap(X86_FEATURE_NOPL);
1048 * Do minimum CPU detection early.
1049 * Fields really needed: vendor, cpuid_level, family, model, mask,
1051 * The others are not touched to avoid unwanted side effects.
1053 * WARNING: this function is only called on the boot CPU. Don't add code
1054 * here that is supposed to run on all CPUs.
1056 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
1058 #ifdef CONFIG_X86_64
1059 c->x86_clflush_size = 64;
1060 c->x86_phys_bits = 36;
1061 c->x86_virt_bits = 48;
1063 c->x86_clflush_size = 32;
1064 c->x86_phys_bits = 32;
1065 c->x86_virt_bits = 32;
1067 c->x86_cache_alignment = c->x86_clflush_size;
1069 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1070 c->extended_cpuid_level = 0;
1072 if (!have_cpuid_p())
1073 identify_cpu_without_cpuid(c);
1075 /* cyrix could have cpuid enabled via c_identify()*/
1076 if (have_cpuid_p()) {
1080 get_cpu_address_sizes(c);
1081 setup_force_cpu_cap(X86_FEATURE_CPUID);
1083 if (this_cpu->c_early_init)
1084 this_cpu->c_early_init(c);
1087 filter_cpuid_features(c, false);
1089 if (this_cpu->c_bsp_init)
1090 this_cpu->c_bsp_init(c);
1092 setup_clear_cpu_cap(X86_FEATURE_CPUID);
1095 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1097 cpu_set_bug_bits(c);
1099 fpu__init_system(c);
1101 #ifdef CONFIG_X86_32
1103 * Regardless of whether PCID is enumerated, the SDM says
1104 * that it can't be enabled in 32-bit mode.
1106 setup_clear_cpu_cap(X86_FEATURE_PCID);
1110 * Later in the boot process pgtable_l5_enabled() relies on
1111 * cpu_feature_enabled(X86_FEATURE_LA57). If 5-level paging is not
1112 * enabled by this point we need to clear the feature bit to avoid
1113 * false-positives at the later stage.
1115 * pgtable_l5_enabled() can be false here for several reasons:
1116 * - 5-level paging is disabled compile-time;
1117 * - it's 32-bit kernel;
1118 * - machine doesn't support 5-level paging;
1119 * - user specified 'no5lvl' in kernel command line.
1121 if (!pgtable_l5_enabled())
1122 setup_clear_cpu_cap(X86_FEATURE_LA57);
1127 void __init early_cpu_init(void)
1129 const struct cpu_dev *const *cdev;
1132 #ifdef CONFIG_PROCESSOR_SELECT
1133 pr_info("KERNEL supported cpus:\n");
1136 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1137 const struct cpu_dev *cpudev = *cdev;
1139 if (count >= X86_VENDOR_NUM)
1141 cpu_devs[count] = cpudev;
1144 #ifdef CONFIG_PROCESSOR_SELECT
1148 for (j = 0; j < 2; j++) {
1149 if (!cpudev->c_ident[j])
1151 pr_info(" %s %s\n", cpudev->c_vendor,
1152 cpudev->c_ident[j]);
1157 early_identify_cpu(&boot_cpu_data);
1160 static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
1162 #ifdef CONFIG_X86_64
1164 * Empirically, writing zero to a segment selector on AMD does
1165 * not clear the base, whereas writing zero to a segment
1166 * selector on Intel does clear the base. Intel's behavior
1167 * allows slightly faster context switches in the common case
1168 * where GS is unused by the prev and next threads.
1170 * Since neither vendor documents this anywhere that I can see,
1171 * detect it directly instead of hardcoding the choice by
1174 * I've designated AMD's behavior as the "bug" because it's
1175 * counterintuitive and less friendly.
1178 unsigned long old_base, tmp;
1179 rdmsrl(MSR_FS_BASE, old_base);
1180 wrmsrl(MSR_FS_BASE, 1);
1182 rdmsrl(MSR_FS_BASE, tmp);
1184 set_cpu_bug(c, X86_BUG_NULL_SEG);
1185 wrmsrl(MSR_FS_BASE, old_base);
1189 static void generic_identify(struct cpuinfo_x86 *c)
1191 c->extended_cpuid_level = 0;
1193 if (!have_cpuid_p())
1194 identify_cpu_without_cpuid(c);
1196 /* cyrix could have cpuid enabled via c_identify()*/
1197 if (!have_cpuid_p())
1206 get_cpu_address_sizes(c);
1208 if (c->cpuid_level >= 0x00000001) {
1209 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1210 #ifdef CONFIG_X86_32
1212 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1214 c->apicid = c->initial_apicid;
1217 c->phys_proc_id = c->initial_apicid;
1220 get_model_name(c); /* Default name */
1222 detect_null_seg_behavior(c);
1225 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1226 * systems that run Linux at CPL > 0 may or may not have the
1227 * issue, but, even if they have the issue, there's absolutely
1228 * nothing we can do about it because we can't use the real IRET
1231 * NB: For the time being, only 32-bit kernels support
1232 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1233 * whether to apply espfix using paravirt hooks. If any
1234 * non-paravirt system ever shows up that does *not* have the
1235 * ESPFIX issue, we can change this.
1237 #ifdef CONFIG_X86_32
1238 # ifdef CONFIG_PARAVIRT_XXL
1240 extern void native_iret(void);
1241 if (pv_ops.cpu.iret == native_iret)
1242 set_cpu_bug(c, X86_BUG_ESPFIX);
1245 set_cpu_bug(c, X86_BUG_ESPFIX);
1250 static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1253 * The heavy lifting of max_rmid and cache_occ_scale are handled
1254 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
1255 * in case CQM bits really aren't there in this CPU.
1257 if (c != &boot_cpu_data) {
1258 boot_cpu_data.x86_cache_max_rmid =
1259 min(boot_cpu_data.x86_cache_max_rmid,
1260 c->x86_cache_max_rmid);
1265 * Validate that ACPI/mptables have the same information about the
1266 * effective APIC id and update the package map.
1268 static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1271 unsigned int apicid, cpu = smp_processor_id();
1273 apicid = apic->cpu_present_to_apicid(cpu);
1275 if (apicid != c->apicid) {
1276 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1277 cpu, apicid, c->initial_apicid);
1279 BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
1281 c->logical_proc_id = 0;
1286 * This does the hard work of actually picking apart the CPU stuff...
1288 static void identify_cpu(struct cpuinfo_x86 *c)
1292 c->loops_per_jiffy = loops_per_jiffy;
1293 c->x86_cache_size = 0;
1294 c->x86_vendor = X86_VENDOR_UNKNOWN;
1295 c->x86_model = c->x86_stepping = 0; /* So far unknown... */
1296 c->x86_vendor_id[0] = '\0'; /* Unset */
1297 c->x86_model_id[0] = '\0'; /* Unset */
1298 c->x86_max_cores = 1;
1299 c->x86_coreid_bits = 0;
1301 #ifdef CONFIG_X86_64
1302 c->x86_clflush_size = 64;
1303 c->x86_phys_bits = 36;
1304 c->x86_virt_bits = 48;
1306 c->cpuid_level = -1; /* CPUID not detected */
1307 c->x86_clflush_size = 32;
1308 c->x86_phys_bits = 32;
1309 c->x86_virt_bits = 32;
1311 c->x86_cache_alignment = c->x86_clflush_size;
1312 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1314 generic_identify(c);
1316 if (this_cpu->c_identify)
1317 this_cpu->c_identify(c);
1319 /* Clear/Set all flags overridden by options, after probe */
1320 apply_forced_caps(c);
1322 #ifdef CONFIG_X86_64
1323 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1327 * Vendor-specific initialization. In this section we
1328 * canonicalize the feature flags, meaning if there are
1329 * features a certain CPU supports which CPUID doesn't
1330 * tell us, CPUID claiming incorrect flags, or other bugs,
1331 * we handle them here.
1333 * At the end of this section, c->x86_capability better
1334 * indicate the features this CPU genuinely supports!
1336 if (this_cpu->c_init)
1337 this_cpu->c_init(c);
1339 /* Disable the PN if appropriate */
1340 squash_the_stupid_serial_number(c);
1342 /* Set up SMEP/SMAP/UMIP */
1348 * The vendor-specific functions might have changed features.
1349 * Now we do "generic changes."
1352 /* Filter out anything that depends on CPUID levels we don't have */
1353 filter_cpuid_features(c, true);
1355 /* If the model name is still unset, do table lookup. */
1356 if (!c->x86_model_id[0]) {
1358 p = table_lookup_model(c);
1360 strcpy(c->x86_model_id, p);
1362 /* Last resort... */
1363 sprintf(c->x86_model_id, "%02x/%02x",
1364 c->x86, c->x86_model);
1367 #ifdef CONFIG_X86_64
1372 x86_init_cache_qos(c);
1376 * Clear/Set all flags overridden by options, need do it
1377 * before following smp all cpus cap AND.
1379 apply_forced_caps(c);
1382 * On SMP, boot_cpu_data holds the common feature set between
1383 * all CPUs; so make sure that we indicate which features are
1384 * common between the CPUs. The first time this routine gets
1385 * executed, c == &boot_cpu_data.
1387 if (c != &boot_cpu_data) {
1388 /* AND the already accumulated flags with these */
1389 for (i = 0; i < NCAPINTS; i++)
1390 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1392 /* OR, i.e. replicate the bug flags */
1393 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1394 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1397 /* Init Machine Check Exception if available. */
1400 select_idle_routine(c);
1403 numa_add_cpu(smp_processor_id());
1408 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1409 * on 32-bit kernels:
1411 #ifdef CONFIG_X86_32
1412 void enable_sep_cpu(void)
1414 struct tss_struct *tss;
1417 if (!boot_cpu_has(X86_FEATURE_SEP))
1421 tss = &per_cpu(cpu_tss_rw, cpu);
1424 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1425 * see the big comment in struct x86_hw_tss's definition.
1428 tss->x86_tss.ss1 = __KERNEL_CS;
1429 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1430 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
1431 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1437 void __init identify_boot_cpu(void)
1439 identify_cpu(&boot_cpu_data);
1440 #ifdef CONFIG_X86_32
1444 cpu_detect_tlb(&boot_cpu_data);
1447 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1449 BUG_ON(c == &boot_cpu_data);
1451 #ifdef CONFIG_X86_32
1455 validate_apic_and_package_id(c);
1456 x86_spec_ctrl_setup_ap();
1459 static __init int setup_noclflush(char *arg)
1461 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1462 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1465 __setup("noclflush", setup_noclflush);
1467 void print_cpu_info(struct cpuinfo_x86 *c)
1469 const char *vendor = NULL;
1471 if (c->x86_vendor < X86_VENDOR_NUM) {
1472 vendor = this_cpu->c_vendor;
1474 if (c->cpuid_level >= 0)
1475 vendor = c->x86_vendor_id;
1478 if (vendor && !strstr(c->x86_model_id, vendor))
1479 pr_cont("%s ", vendor);
1481 if (c->x86_model_id[0])
1482 pr_cont("%s", c->x86_model_id);
1484 pr_cont("%d86", c->x86);
1486 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1488 if (c->x86_stepping || c->cpuid_level >= 0)
1489 pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
1495 * clearcpuid= was already parsed in fpu__init_parse_early_param.
1496 * But we need to keep a dummy __setup around otherwise it would
1497 * show up as an environment variable for init.
1499 static __init int setup_clearcpuid(char *arg)
1503 __setup("clearcpuid=", setup_clearcpuid);
1505 #ifdef CONFIG_X86_64
1506 DEFINE_PER_CPU_FIRST(struct fixed_percpu_data,
1507 fixed_percpu_data) __aligned(PAGE_SIZE) __visible;
1508 EXPORT_PER_CPU_SYMBOL_GPL(fixed_percpu_data);
1511 * The following percpu variables are hot. Align current_task to
1512 * cacheline size such that they fall in the same cacheline.
1514 DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1516 EXPORT_PER_CPU_SYMBOL(current_task);
1518 DEFINE_PER_CPU(struct irq_stack *, hardirq_stack_ptr);
1519 DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1521 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1522 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1524 /* May not be marked __init: used by software suspend */
1525 void syscall_init(void)
1527 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1528 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1530 #ifdef CONFIG_IA32_EMULATION
1531 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1533 * This only works on Intel CPUs.
1534 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1535 * This does not cause SYSENTER to jump to the wrong location, because
1536 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1538 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1539 wrmsrl_safe(MSR_IA32_SYSENTER_ESP,
1540 (unsigned long)(cpu_entry_stack(smp_processor_id()) + 1));
1541 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1543 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1544 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1545 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1546 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1549 /* Flags to clear on syscall */
1550 wrmsrl(MSR_SYSCALL_MASK,
1551 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1552 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1555 DEFINE_PER_CPU(int, debug_stack_usage);
1556 DEFINE_PER_CPU(u32, debug_idt_ctr);
1558 void debug_stack_set_zero(void)
1560 this_cpu_inc(debug_idt_ctr);
1563 NOKPROBE_SYMBOL(debug_stack_set_zero);
1565 void debug_stack_reset(void)
1567 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1569 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1572 NOKPROBE_SYMBOL(debug_stack_reset);
1574 #else /* CONFIG_X86_64 */
1576 DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1577 EXPORT_PER_CPU_SYMBOL(current_task);
1578 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1579 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1582 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1583 * the top of the kernel stack. Use an extra percpu variable to track the
1584 * top of the kernel stack directly.
1586 DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1587 (unsigned long)&init_thread_union + THREAD_SIZE;
1588 EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1590 #ifdef CONFIG_STACKPROTECTOR
1591 DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1594 #endif /* CONFIG_X86_64 */
1597 * Clear all 6 debug registers:
1599 static void clear_all_debug_regs(void)
1603 for (i = 0; i < 8; i++) {
1604 /* Ignore db4, db5 */
1605 if ((i == 4) || (i == 5))
1614 * Restore debug regs if using kgdbwait and you have a kernel debugger
1615 * connection established.
1617 static void dbg_restore_debug_regs(void)
1619 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1620 arch_kgdb_ops.correct_hw_break();
1622 #else /* ! CONFIG_KGDB */
1623 #define dbg_restore_debug_regs()
1624 #endif /* ! CONFIG_KGDB */
1626 static void wait_for_master_cpu(int cpu)
1630 * wait for ACK from master CPU before continuing
1631 * with AP initialization
1633 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1634 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1639 #ifdef CONFIG_X86_64
1640 static void setup_getcpu(int cpu)
1642 unsigned long cpudata = vdso_encode_cpunode(cpu, early_cpu_to_node(cpu));
1643 struct desc_struct d = { };
1645 if (boot_cpu_has(X86_FEATURE_RDTSCP))
1646 write_rdtscp_aux(cpudata);
1648 /* Store CPU and node number in limit. */
1650 d.limit1 = cpudata >> 16;
1652 d.type = 5; /* RO data, expand down, accessed */
1653 d.dpl = 3; /* Visible to user code */
1654 d.s = 1; /* Not a system segment */
1655 d.p = 1; /* Present */
1656 d.d = 1; /* 32-bit */
1658 write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_CPUNODE, &d, DESCTYPE_S);
1663 * cpu_init() initializes state that is per-CPU. Some data is already
1664 * initialized (naturally) in the bootstrap process, such as the GDT
1665 * and IDT. We reload them nevertheless, this function acts as a
1666 * 'CPU state barrier', nothing should get across.
1668 #ifdef CONFIG_X86_64
1672 int cpu = raw_smp_processor_id();
1673 struct task_struct *me;
1674 struct tss_struct *t;
1677 wait_for_master_cpu(cpu);
1680 * Initialize the CR4 shadow before doing anything that could
1688 t = &per_cpu(cpu_tss_rw, cpu);
1691 if (this_cpu_read(numa_node) == 0 &&
1692 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1693 set_numa_node(early_cpu_to_node(cpu));
1699 pr_debug("Initializing CPU#%d\n", cpu);
1701 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1704 * Initialize the per-CPU GDT with the boot GDT,
1705 * and set up the GDT descriptor:
1708 switch_to_new_gdt(cpu);
1713 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1716 wrmsrl(MSR_FS_BASE, 0);
1717 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1724 * set up and load the per-CPU TSS
1726 if (!t->x86_tss.ist[0]) {
1727 t->x86_tss.ist[IST_INDEX_DF] = __this_cpu_ist_top_va(DF);
1728 t->x86_tss.ist[IST_INDEX_NMI] = __this_cpu_ist_top_va(NMI);
1729 t->x86_tss.ist[IST_INDEX_DB] = __this_cpu_ist_top_va(DB);
1730 t->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
1733 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1736 * <= is required because the CPU will access up to
1737 * 8 bits beyond the end of the IO permission bitmap.
1739 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1740 t->io_bitmap[i] = ~0UL;
1743 me->active_mm = &init_mm;
1745 initialize_tlbstate_and_flush();
1746 enter_lazy_tlb(&init_mm, me);
1749 * Initialize the TSS. sp0 points to the entry trampoline stack
1750 * regardless of what task is running.
1752 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1754 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1756 load_mm_ldt(&init_mm);
1758 clear_all_debug_regs();
1759 dbg_restore_debug_regs();
1766 load_fixmap_gdt(cpu);
1773 int cpu = smp_processor_id();
1774 struct task_struct *curr = current;
1775 struct tss_struct *t = &per_cpu(cpu_tss_rw, cpu);
1777 wait_for_master_cpu(cpu);
1780 * Initialize the CR4 shadow before doing anything that could
1785 show_ucode_info_early();
1787 pr_info("Initializing CPU#%d\n", cpu);
1789 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1790 boot_cpu_has(X86_FEATURE_TSC) ||
1791 boot_cpu_has(X86_FEATURE_DE))
1792 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1795 switch_to_new_gdt(cpu);
1798 * Set up and load the per-CPU TSS and LDT
1801 curr->active_mm = &init_mm;
1803 initialize_tlbstate_and_flush();
1804 enter_lazy_tlb(&init_mm, curr);
1807 * Initialize the TSS. sp0 points to the entry trampoline stack
1808 * regardless of what task is running.
1810 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1812 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1814 load_mm_ldt(&init_mm);
1816 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1818 #ifdef CONFIG_DOUBLEFAULT
1819 /* Set up doublefault TSS pointer in the GDT */
1820 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1823 clear_all_debug_regs();
1824 dbg_restore_debug_regs();
1828 load_fixmap_gdt(cpu);
1833 * The microcode loader calls this upon late microcode load to recheck features,
1834 * only when microcode has been updated. Caller holds microcode_mutex and CPU
1837 void microcode_check(void)
1839 struct cpuinfo_x86 info;
1841 perf_check_microcode();
1843 /* Reload CPUID max function as it might've changed. */
1844 info.cpuid_level = cpuid_eax(0);
1847 * Copy all capability leafs to pick up the synthetic ones so that
1848 * memcmp() below doesn't fail on that. The ones coming from CPUID will
1849 * get overwritten in get_cpu_cap().
1851 memcpy(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability));
1855 if (!memcmp(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability)))
1858 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
1859 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");