1 #include <linux/bootmem.h>
2 #include <linux/linkage.h>
3 #include <linux/bitops.h>
4 #include <linux/kernel.h>
5 #include <linux/export.h>
6 #include <linux/percpu.h>
7 #include <linux/string.h>
8 #include <linux/ctype.h>
9 #include <linux/delay.h>
10 #include <linux/sched/mm.h>
11 #include <linux/sched/clock.h>
12 #include <linux/sched/task.h>
13 #include <linux/init.h>
14 #include <linux/kprobes.h>
15 #include <linux/kgdb.h>
16 #include <linux/smp.h>
18 #include <linux/syscore_ops.h>
20 #include <asm/stackprotector.h>
21 #include <asm/perf_event.h>
22 #include <asm/mmu_context.h>
23 #include <asm/archrandom.h>
24 #include <asm/hypervisor.h>
25 #include <asm/processor.h>
26 #include <asm/tlbflush.h>
27 #include <asm/debugreg.h>
28 #include <asm/sections.h>
29 #include <asm/vsyscall.h>
30 #include <linux/topology.h>
31 #include <linux/cpumask.h>
32 #include <asm/pgtable.h>
33 #include <linux/atomic.h>
34 #include <asm/proto.h>
35 #include <asm/setup.h>
38 #include <asm/fpu/internal.h>
40 #include <asm/hwcap2.h>
41 #include <linux/numa.h>
48 #include <asm/microcode.h>
49 #include <asm/microcode_intel.h>
50 #include <asm/intel-family.h>
51 #include <asm/cpu_device_id.h>
53 #ifdef CONFIG_X86_LOCAL_APIC
54 #include <asm/uv/uv.h>
59 u32 elf_hwcap2 __read_mostly;
61 /* all of these masks are initialized in setup_cpu_local_masks() */
62 cpumask_var_t cpu_initialized_mask;
63 cpumask_var_t cpu_callout_mask;
64 cpumask_var_t cpu_callin_mask;
66 /* representing cpus for which sibling maps can be computed */
67 cpumask_var_t cpu_sibling_setup_mask;
69 /* Number of siblings per CPU package */
70 int smp_num_siblings = 1;
71 EXPORT_SYMBOL(smp_num_siblings);
73 /* Last level cache ID of each logical CPU */
74 DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
76 /* correctly size the local cpu masks */
77 void __init setup_cpu_local_masks(void)
79 alloc_bootmem_cpumask_var(&cpu_initialized_mask);
80 alloc_bootmem_cpumask_var(&cpu_callin_mask);
81 alloc_bootmem_cpumask_var(&cpu_callout_mask);
82 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
85 static void default_init(struct cpuinfo_x86 *c)
88 cpu_detect_cache_sizes(c);
90 /* Not much we can do here... */
91 /* Check if at least it has cpuid */
92 if (c->cpuid_level == -1) {
93 /* No cpuid. It must be an ancient CPU */
95 strcpy(c->x86_model_id, "486");
97 strcpy(c->x86_model_id, "386");
102 static const struct cpu_dev default_cpu = {
103 .c_init = default_init,
104 .c_vendor = "Unknown",
105 .c_x86_vendor = X86_VENDOR_UNKNOWN,
108 static const struct cpu_dev *this_cpu = &default_cpu;
110 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
113 * We need valid kernel segments for data and code in long mode too
114 * IRET will check the segment types kkeil 2000/10/28
115 * Also sysret mandates a special GDT layout
117 * TLS descriptors are currently at a different place compared to i386.
118 * Hopefully nobody expects them at a fixed place (Wine?)
120 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
121 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
122 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
123 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
124 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
125 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
127 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
128 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
129 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
130 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
132 * Segments used for calling PnP BIOS have byte granularity.
133 * They code segments and data segments have fixed 64k limits,
134 * the transfer segment sizes are set at run time.
137 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
139 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
141 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(0x0092, 0, 0xffff),
143 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(0x0092, 0, 0),
145 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(0x0092, 0, 0),
147 * The APM segments have byte granularity and their bases
148 * are set at run time. All have 64k limits.
151 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
153 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
155 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(0x4092, 0, 0xffff),
157 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
158 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
159 GDT_STACK_CANARY_INIT
162 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
164 static int __init x86_mpx_setup(char *s)
166 /* require an exact match without trailing characters */
170 /* do not emit a message if the feature is not present */
171 if (!boot_cpu_has(X86_FEATURE_MPX))
174 setup_clear_cpu_cap(X86_FEATURE_MPX);
175 pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
178 __setup("nompx", x86_mpx_setup);
181 static int __init x86_nopcid_setup(char *s)
183 /* nopcid doesn't accept parameters */
187 /* do not emit a message if the feature is not present */
188 if (!boot_cpu_has(X86_FEATURE_PCID))
191 setup_clear_cpu_cap(X86_FEATURE_PCID);
192 pr_info("nopcid: PCID feature disabled\n");
195 early_param("nopcid", x86_nopcid_setup);
198 static int __init x86_noinvpcid_setup(char *s)
200 /* noinvpcid doesn't accept parameters */
204 /* do not emit a message if the feature is not present */
205 if (!boot_cpu_has(X86_FEATURE_INVPCID))
208 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
209 pr_info("noinvpcid: INVPCID feature disabled\n");
212 early_param("noinvpcid", x86_noinvpcid_setup);
215 static int cachesize_override = -1;
216 static int disable_x86_serial_nr = 1;
218 static int __init cachesize_setup(char *str)
220 get_option(&str, &cachesize_override);
223 __setup("cachesize=", cachesize_setup);
225 static int __init x86_sep_setup(char *s)
227 setup_clear_cpu_cap(X86_FEATURE_SEP);
230 __setup("nosep", x86_sep_setup);
232 /* Standard macro to see if a specific flag is changeable */
233 static inline int flag_is_changeable_p(u32 flag)
238 * Cyrix and IDT cpus allow disabling of CPUID
239 * so the code below may return different results
240 * when it is executed before and after enabling
241 * the CPUID. Add "volatile" to not allow gcc to
242 * optimize the subsequent calls to this function.
244 asm volatile ("pushfl \n\t"
255 : "=&r" (f1), "=&r" (f2)
258 return ((f1^f2) & flag) != 0;
261 /* Probe for the CPUID instruction */
262 int have_cpuid_p(void)
264 return flag_is_changeable_p(X86_EFLAGS_ID);
267 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
269 unsigned long lo, hi;
271 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
274 /* Disable processor serial number: */
276 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
278 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
280 pr_notice("CPU serial number disabled.\n");
281 clear_cpu_cap(c, X86_FEATURE_PN);
283 /* Disabling the serial number may affect the cpuid level */
284 c->cpuid_level = cpuid_eax(0);
287 static int __init x86_serial_nr_setup(char *s)
289 disable_x86_serial_nr = 0;
292 __setup("serialnumber", x86_serial_nr_setup);
294 static inline int flag_is_changeable_p(u32 flag)
298 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
303 static __init int setup_disable_smep(char *arg)
305 setup_clear_cpu_cap(X86_FEATURE_SMEP);
306 /* Check for things that depend on SMEP being enabled: */
307 check_mpx_erratum(&boot_cpu_data);
310 __setup("nosmep", setup_disable_smep);
312 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
314 if (cpu_has(c, X86_FEATURE_SMEP))
315 cr4_set_bits(X86_CR4_SMEP);
318 static __init int setup_disable_smap(char *arg)
320 setup_clear_cpu_cap(X86_FEATURE_SMAP);
323 __setup("nosmap", setup_disable_smap);
325 static __always_inline void setup_smap(struct cpuinfo_x86 *c)
327 unsigned long eflags = native_save_fl();
329 /* This should have been cleared long ago */
330 BUG_ON(eflags & X86_EFLAGS_AC);
332 if (cpu_has(c, X86_FEATURE_SMAP)) {
333 #ifdef CONFIG_X86_SMAP
334 cr4_set_bits(X86_CR4_SMAP);
336 cr4_clear_bits(X86_CR4_SMAP);
341 static __always_inline void setup_umip(struct cpuinfo_x86 *c)
343 /* Check the boot processor, plus build option for UMIP. */
344 if (!cpu_feature_enabled(X86_FEATURE_UMIP))
347 /* Check the current processor's cpuid bits. */
348 if (!cpu_has(c, X86_FEATURE_UMIP))
351 cr4_set_bits(X86_CR4_UMIP);
353 pr_info("x86/cpu: Activated the Intel User Mode Instruction Prevention (UMIP) CPU feature\n");
359 * Make sure UMIP is disabled in case it was enabled in a
360 * previous boot (e.g., via kexec).
362 cr4_clear_bits(X86_CR4_UMIP);
366 * Protection Keys are not available in 32-bit mode.
368 static bool pku_disabled;
370 static __always_inline void setup_pku(struct cpuinfo_x86 *c)
372 /* check the boot processor, plus compile options for PKU: */
373 if (!cpu_feature_enabled(X86_FEATURE_PKU))
375 /* checks the actual processor's cpuid bits: */
376 if (!cpu_has(c, X86_FEATURE_PKU))
381 cr4_set_bits(X86_CR4_PKE);
383 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
384 * cpuid bit to be set. We need to ensure that we
385 * update that bit in this CPU's "cpu_info".
390 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
391 static __init int setup_disable_pku(char *arg)
394 * Do not clear the X86_FEATURE_PKU bit. All of the
395 * runtime checks are against OSPKE so clearing the
398 * This way, we will see "pku" in cpuinfo, but not
399 * "ospke", which is exactly what we want. It shows
400 * that the CPU has PKU, but the OS has not enabled it.
401 * This happens to be exactly how a system would look
402 * if we disabled the config option.
404 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
408 __setup("nopku", setup_disable_pku);
409 #endif /* CONFIG_X86_64 */
412 * Some CPU features depend on higher CPUID levels, which may not always
413 * be available due to CPUID level capping or broken virtualization
414 * software. Add those features to this table to auto-disable them.
416 struct cpuid_dependent_feature {
421 static const struct cpuid_dependent_feature
422 cpuid_dependent_features[] = {
423 { X86_FEATURE_MWAIT, 0x00000005 },
424 { X86_FEATURE_DCA, 0x00000009 },
425 { X86_FEATURE_XSAVE, 0x0000000d },
429 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
431 const struct cpuid_dependent_feature *df;
433 for (df = cpuid_dependent_features; df->feature; df++) {
435 if (!cpu_has(c, df->feature))
438 * Note: cpuid_level is set to -1 if unavailable, but
439 * extended_extended_level is set to 0 if unavailable
440 * and the legitimate extended levels are all negative
441 * when signed; hence the weird messing around with
444 if (!((s32)df->level < 0 ?
445 (u32)df->level > (u32)c->extended_cpuid_level :
446 (s32)df->level > (s32)c->cpuid_level))
449 clear_cpu_cap(c, df->feature);
453 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
454 x86_cap_flag(df->feature), df->level);
459 * Naming convention should be: <Name> [(<Codename>)]
460 * This table only is used unless init_<vendor>() below doesn't set it;
461 * in particular, if CPUID levels 0x80000002..4 are supported, this
465 /* Look up CPU names by table lookup. */
466 static const char *table_lookup_model(struct cpuinfo_x86 *c)
469 const struct legacy_cpu_model_info *info;
471 if (c->x86_model >= 16)
472 return NULL; /* Range check */
477 info = this_cpu->legacy_models;
479 while (info->family) {
480 if (info->family == c->x86)
481 return info->model_names[c->x86_model];
485 return NULL; /* Not found */
488 __u32 cpu_caps_cleared[NCAPINTS + NBUGINTS];
489 __u32 cpu_caps_set[NCAPINTS + NBUGINTS];
491 void load_percpu_segment(int cpu)
494 loadsegment(fs, __KERNEL_PERCPU);
496 __loadsegment_simple(gs, 0);
497 wrmsrl(MSR_GS_BASE, cpu_kernelmode_gs_base(cpu));
499 load_stack_canary_segment();
503 /* The 32-bit entry code needs to find cpu_entry_area. */
504 DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area);
509 * Special IST stacks which the CPU switches to when it calls
510 * an IST-marked descriptor entry. Up to 7 stacks (hardware
511 * limit), all of them are 4K, except the debug stack which
514 static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
515 [0 ... N_EXCEPTION_STACKS - 1] = EXCEPTION_STKSZ,
516 [DEBUG_STACK - 1] = DEBUG_STKSZ
520 /* Load the original GDT from the per-cpu structure */
521 void load_direct_gdt(int cpu)
523 struct desc_ptr gdt_descr;
525 gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
526 gdt_descr.size = GDT_SIZE - 1;
527 load_gdt(&gdt_descr);
529 EXPORT_SYMBOL_GPL(load_direct_gdt);
531 /* Load a fixmap remapping of the per-cpu GDT */
532 void load_fixmap_gdt(int cpu)
534 struct desc_ptr gdt_descr;
536 gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
537 gdt_descr.size = GDT_SIZE - 1;
538 load_gdt(&gdt_descr);
540 EXPORT_SYMBOL_GPL(load_fixmap_gdt);
543 * Current gdt points %fs at the "master" per-cpu area: after this,
544 * it's on the real one.
546 void switch_to_new_gdt(int cpu)
548 /* Load the original GDT */
549 load_direct_gdt(cpu);
550 /* Reload the per-cpu base */
551 load_percpu_segment(cpu);
554 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
556 static void get_model_name(struct cpuinfo_x86 *c)
561 if (c->extended_cpuid_level < 0x80000004)
564 v = (unsigned int *)c->x86_model_id;
565 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
566 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
567 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
568 c->x86_model_id[48] = 0;
570 /* Trim whitespace */
571 p = q = s = &c->x86_model_id[0];
577 /* Note the last non-whitespace index */
587 void detect_num_cpu_cores(struct cpuinfo_x86 *c)
589 unsigned int eax, ebx, ecx, edx;
591 c->x86_max_cores = 1;
592 if (!IS_ENABLED(CONFIG_SMP) || c->cpuid_level < 4)
595 cpuid_count(4, 0, &eax, &ebx, &ecx, &edx);
597 c->x86_max_cores = (eax >> 26) + 1;
600 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
602 unsigned int n, dummy, ebx, ecx, edx, l2size;
604 n = c->extended_cpuid_level;
606 if (n >= 0x80000005) {
607 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
608 c->x86_cache_size = (ecx>>24) + (edx>>24);
610 /* On K8 L1 TLB is inclusive, so don't count it */
615 if (n < 0x80000006) /* Some chips just has a large L1. */
618 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
622 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
624 /* do processor-specific cache resizing */
625 if (this_cpu->legacy_cache_size)
626 l2size = this_cpu->legacy_cache_size(c, l2size);
628 /* Allow user to override all this if necessary. */
629 if (cachesize_override != -1)
630 l2size = cachesize_override;
633 return; /* Again, no L2 cache is possible */
636 c->x86_cache_size = l2size;
639 u16 __read_mostly tlb_lli_4k[NR_INFO];
640 u16 __read_mostly tlb_lli_2m[NR_INFO];
641 u16 __read_mostly tlb_lli_4m[NR_INFO];
642 u16 __read_mostly tlb_lld_4k[NR_INFO];
643 u16 __read_mostly tlb_lld_2m[NR_INFO];
644 u16 __read_mostly tlb_lld_4m[NR_INFO];
645 u16 __read_mostly tlb_lld_1g[NR_INFO];
647 static void cpu_detect_tlb(struct cpuinfo_x86 *c)
649 if (this_cpu->c_detect_tlb)
650 this_cpu->c_detect_tlb(c);
652 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
653 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
654 tlb_lli_4m[ENTRIES]);
656 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
657 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
658 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
661 void detect_ht(struct cpuinfo_x86 *c)
664 u32 eax, ebx, ecx, edx;
665 int index_msb, core_bits;
668 if (!cpu_has(c, X86_FEATURE_HT))
671 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
674 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
677 cpuid(1, &eax, &ebx, &ecx, &edx);
679 smp_num_siblings = (ebx & 0xff0000) >> 16;
681 if (smp_num_siblings == 1) {
682 pr_info_once("CPU0: Hyper-Threading is disabled\n");
686 if (smp_num_siblings <= 1)
689 index_msb = get_count_order(smp_num_siblings);
690 c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
692 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
694 index_msb = get_count_order(smp_num_siblings);
696 core_bits = get_count_order(c->x86_max_cores);
698 c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
699 ((1 << core_bits) - 1);
702 if (!printed && (c->x86_max_cores * smp_num_siblings) > 1) {
703 pr_info("CPU: Physical Processor ID: %d\n",
705 pr_info("CPU: Processor Core ID: %d\n",
712 static void get_cpu_vendor(struct cpuinfo_x86 *c)
714 char *v = c->x86_vendor_id;
717 for (i = 0; i < X86_VENDOR_NUM; i++) {
721 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
722 (cpu_devs[i]->c_ident[1] &&
723 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
725 this_cpu = cpu_devs[i];
726 c->x86_vendor = this_cpu->c_x86_vendor;
731 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
732 "CPU: Your system may be unstable.\n", v);
734 c->x86_vendor = X86_VENDOR_UNKNOWN;
735 this_cpu = &default_cpu;
738 void cpu_detect(struct cpuinfo_x86 *c)
740 /* Get vendor name */
741 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
742 (unsigned int *)&c->x86_vendor_id[0],
743 (unsigned int *)&c->x86_vendor_id[8],
744 (unsigned int *)&c->x86_vendor_id[4]);
747 /* Intel-defined flags: level 0x00000001 */
748 if (c->cpuid_level >= 0x00000001) {
749 u32 junk, tfms, cap0, misc;
751 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
752 c->x86 = x86_family(tfms);
753 c->x86_model = x86_model(tfms);
754 c->x86_stepping = x86_stepping(tfms);
756 if (cap0 & (1<<19)) {
757 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
758 c->x86_cache_alignment = c->x86_clflush_size;
763 static void apply_forced_caps(struct cpuinfo_x86 *c)
767 for (i = 0; i < NCAPINTS + NBUGINTS; i++) {
768 c->x86_capability[i] &= ~cpu_caps_cleared[i];
769 c->x86_capability[i] |= cpu_caps_set[i];
773 static void init_speculation_control(struct cpuinfo_x86 *c)
776 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support,
777 * and they also have a different bit for STIBP support. Also,
778 * a hypervisor might have set the individual AMD bits even on
779 * Intel CPUs, for finer-grained selection of what's available.
781 if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) {
782 set_cpu_cap(c, X86_FEATURE_IBRS);
783 set_cpu_cap(c, X86_FEATURE_IBPB);
784 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
787 if (cpu_has(c, X86_FEATURE_INTEL_STIBP))
788 set_cpu_cap(c, X86_FEATURE_STIBP);
790 if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) ||
791 cpu_has(c, X86_FEATURE_VIRT_SSBD))
792 set_cpu_cap(c, X86_FEATURE_SSBD);
794 if (cpu_has(c, X86_FEATURE_AMD_IBRS)) {
795 set_cpu_cap(c, X86_FEATURE_IBRS);
796 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
799 if (cpu_has(c, X86_FEATURE_AMD_IBPB))
800 set_cpu_cap(c, X86_FEATURE_IBPB);
802 if (cpu_has(c, X86_FEATURE_AMD_STIBP)) {
803 set_cpu_cap(c, X86_FEATURE_STIBP);
804 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
807 if (cpu_has(c, X86_FEATURE_AMD_SSBD)) {
808 set_cpu_cap(c, X86_FEATURE_SSBD);
809 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
810 clear_cpu_cap(c, X86_FEATURE_VIRT_SSBD);
814 void get_cpu_cap(struct cpuinfo_x86 *c)
816 u32 eax, ebx, ecx, edx;
818 /* Intel-defined flags: level 0x00000001 */
819 if (c->cpuid_level >= 0x00000001) {
820 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
822 c->x86_capability[CPUID_1_ECX] = ecx;
823 c->x86_capability[CPUID_1_EDX] = edx;
826 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
827 if (c->cpuid_level >= 0x00000006)
828 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
830 /* Additional Intel-defined flags: level 0x00000007 */
831 if (c->cpuid_level >= 0x00000007) {
832 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
833 c->x86_capability[CPUID_7_0_EBX] = ebx;
834 c->x86_capability[CPUID_7_ECX] = ecx;
835 c->x86_capability[CPUID_7_EDX] = edx;
838 /* Extended state features: level 0x0000000d */
839 if (c->cpuid_level >= 0x0000000d) {
840 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
842 c->x86_capability[CPUID_D_1_EAX] = eax;
845 /* Additional Intel-defined flags: level 0x0000000F */
846 if (c->cpuid_level >= 0x0000000F) {
848 /* QoS sub-leaf, EAX=0Fh, ECX=0 */
849 cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
850 c->x86_capability[CPUID_F_0_EDX] = edx;
852 if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
853 /* will be overridden if occupancy monitoring exists */
854 c->x86_cache_max_rmid = ebx;
856 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
857 cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
858 c->x86_capability[CPUID_F_1_EDX] = edx;
860 if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
861 ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
862 (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
863 c->x86_cache_max_rmid = ecx;
864 c->x86_cache_occ_scale = ebx;
867 c->x86_cache_max_rmid = -1;
868 c->x86_cache_occ_scale = -1;
872 /* AMD-defined flags: level 0x80000001 */
873 eax = cpuid_eax(0x80000000);
874 c->extended_cpuid_level = eax;
876 if ((eax & 0xffff0000) == 0x80000000) {
877 if (eax >= 0x80000001) {
878 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
880 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
881 c->x86_capability[CPUID_8000_0001_EDX] = edx;
885 if (c->extended_cpuid_level >= 0x80000007) {
886 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
888 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
892 if (c->extended_cpuid_level >= 0x80000008) {
893 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
894 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
897 if (c->extended_cpuid_level >= 0x8000000a)
898 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
900 init_scattered_cpuid_features(c);
901 init_speculation_control(c);
904 * Clear/Set all flags overridden by options, after probe.
905 * This needs to happen each time we re-probe, which may happen
906 * several times during CPU initialization.
908 apply_forced_caps(c);
911 static void get_cpu_address_sizes(struct cpuinfo_x86 *c)
913 u32 eax, ebx, ecx, edx;
915 if (c->extended_cpuid_level >= 0x80000008) {
916 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
918 c->x86_virt_bits = (eax >> 8) & 0xff;
919 c->x86_phys_bits = eax & 0xff;
922 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
923 c->x86_phys_bits = 36;
927 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
933 * First of all, decide if this is a 486 or higher
934 * It's a 486 if we can modify the AC flag
936 if (flag_is_changeable_p(X86_EFLAGS_AC))
941 for (i = 0; i < X86_VENDOR_NUM; i++)
942 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
943 c->x86_vendor_id[0] = 0;
944 cpu_devs[i]->c_identify(c);
945 if (c->x86_vendor_id[0]) {
953 static const __initconst struct x86_cpu_id cpu_no_speculation[] = {
954 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_CEDARVIEW, X86_FEATURE_ANY },
955 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_CLOVERVIEW, X86_FEATURE_ANY },
956 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_LINCROFT, X86_FEATURE_ANY },
957 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_PENWELL, X86_FEATURE_ANY },
958 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_PINEVIEW, X86_FEATURE_ANY },
959 { X86_VENDOR_CENTAUR, 5 },
960 { X86_VENDOR_INTEL, 5 },
961 { X86_VENDOR_NSC, 5 },
962 { X86_VENDOR_ANY, 4 },
966 static const __initconst struct x86_cpu_id cpu_no_meltdown[] = {
971 /* Only list CPUs which speculate but are non susceptible to SSB */
972 static const __initconst struct x86_cpu_id cpu_no_spec_store_bypass[] = {
973 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT1 },
974 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_AIRMONT },
975 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT2 },
976 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_MERRIFIELD },
977 { X86_VENDOR_INTEL, 6, INTEL_FAM6_CORE_YONAH },
978 { X86_VENDOR_INTEL, 6, INTEL_FAM6_XEON_PHI_KNL },
979 { X86_VENDOR_INTEL, 6, INTEL_FAM6_XEON_PHI_KNM },
980 { X86_VENDOR_AMD, 0x12, },
981 { X86_VENDOR_AMD, 0x11, },
982 { X86_VENDOR_AMD, 0x10, },
983 { X86_VENDOR_AMD, 0xf, },
987 static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
991 if (x86_match_cpu(cpu_no_speculation))
994 setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
995 setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
997 if (cpu_has(c, X86_FEATURE_ARCH_CAPABILITIES))
998 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
1000 if (!x86_match_cpu(cpu_no_spec_store_bypass) &&
1001 !(ia32_cap & ARCH_CAP_SSB_NO) &&
1002 !cpu_has(c, X86_FEATURE_AMD_SSB_NO))
1003 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
1005 if (x86_match_cpu(cpu_no_meltdown))
1008 /* Rogue Data Cache Load? No! */
1009 if (ia32_cap & ARCH_CAP_RDCL_NO)
1012 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
1016 * Do minimum CPU detection early.
1017 * Fields really needed: vendor, cpuid_level, family, model, mask,
1019 * The others are not touched to avoid unwanted side effects.
1021 * WARNING: this function is only called on the boot CPU. Don't add code
1022 * here that is supposed to run on all CPUs.
1024 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
1026 #ifdef CONFIG_X86_64
1027 c->x86_clflush_size = 64;
1028 c->x86_phys_bits = 36;
1029 c->x86_virt_bits = 48;
1031 c->x86_clflush_size = 32;
1032 c->x86_phys_bits = 32;
1033 c->x86_virt_bits = 32;
1035 c->x86_cache_alignment = c->x86_clflush_size;
1037 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1038 c->extended_cpuid_level = 0;
1040 /* cyrix could have cpuid enabled via c_identify()*/
1041 if (have_cpuid_p()) {
1045 get_cpu_address_sizes(c);
1046 setup_force_cpu_cap(X86_FEATURE_CPUID);
1048 if (this_cpu->c_early_init)
1049 this_cpu->c_early_init(c);
1052 filter_cpuid_features(c, false);
1054 if (this_cpu->c_bsp_init)
1055 this_cpu->c_bsp_init(c);
1057 identify_cpu_without_cpuid(c);
1058 setup_clear_cpu_cap(X86_FEATURE_CPUID);
1061 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1063 cpu_set_bug_bits(c);
1065 fpu__init_system(c);
1067 #ifdef CONFIG_X86_32
1069 * Regardless of whether PCID is enumerated, the SDM says
1070 * that it can't be enabled in 32-bit mode.
1072 setup_clear_cpu_cap(X86_FEATURE_PCID);
1076 * Later in the boot process pgtable_l5_enabled() relies on
1077 * cpu_feature_enabled(X86_FEATURE_LA57). If 5-level paging is not
1078 * enabled by this point we need to clear the feature bit to avoid
1079 * false-positives at the later stage.
1081 * pgtable_l5_enabled() can be false here for several reasons:
1082 * - 5-level paging is disabled compile-time;
1083 * - it's 32-bit kernel;
1084 * - machine doesn't support 5-level paging;
1085 * - user specified 'no5lvl' in kernel command line.
1087 if (!pgtable_l5_enabled())
1088 setup_clear_cpu_cap(X86_FEATURE_LA57);
1091 void __init early_cpu_init(void)
1093 const struct cpu_dev *const *cdev;
1096 #ifdef CONFIG_PROCESSOR_SELECT
1097 pr_info("KERNEL supported cpus:\n");
1100 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1101 const struct cpu_dev *cpudev = *cdev;
1103 if (count >= X86_VENDOR_NUM)
1105 cpu_devs[count] = cpudev;
1108 #ifdef CONFIG_PROCESSOR_SELECT
1112 for (j = 0; j < 2; j++) {
1113 if (!cpudev->c_ident[j])
1115 pr_info(" %s %s\n", cpudev->c_vendor,
1116 cpudev->c_ident[j]);
1121 early_identify_cpu(&boot_cpu_data);
1125 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1126 * unfortunately, that's not true in practice because of early VIA
1127 * chips and (more importantly) broken virtualizers that are not easy
1128 * to detect. In the latter case it doesn't even *fail* reliably, so
1129 * probing for it doesn't even work. Disable it completely on 32-bit
1130 * unless we can find a reliable way to detect all the broken cases.
1131 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1133 static void detect_nopl(struct cpuinfo_x86 *c)
1135 #ifdef CONFIG_X86_32
1136 clear_cpu_cap(c, X86_FEATURE_NOPL);
1138 set_cpu_cap(c, X86_FEATURE_NOPL);
1142 static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
1144 #ifdef CONFIG_X86_64
1146 * Empirically, writing zero to a segment selector on AMD does
1147 * not clear the base, whereas writing zero to a segment
1148 * selector on Intel does clear the base. Intel's behavior
1149 * allows slightly faster context switches in the common case
1150 * where GS is unused by the prev and next threads.
1152 * Since neither vendor documents this anywhere that I can see,
1153 * detect it directly instead of hardcoding the choice by
1156 * I've designated AMD's behavior as the "bug" because it's
1157 * counterintuitive and less friendly.
1160 unsigned long old_base, tmp;
1161 rdmsrl(MSR_FS_BASE, old_base);
1162 wrmsrl(MSR_FS_BASE, 1);
1164 rdmsrl(MSR_FS_BASE, tmp);
1166 set_cpu_bug(c, X86_BUG_NULL_SEG);
1167 wrmsrl(MSR_FS_BASE, old_base);
1171 static void generic_identify(struct cpuinfo_x86 *c)
1173 c->extended_cpuid_level = 0;
1175 if (!have_cpuid_p())
1176 identify_cpu_without_cpuid(c);
1178 /* cyrix could have cpuid enabled via c_identify()*/
1179 if (!have_cpuid_p())
1188 get_cpu_address_sizes(c);
1190 if (c->cpuid_level >= 0x00000001) {
1191 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1192 #ifdef CONFIG_X86_32
1194 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1196 c->apicid = c->initial_apicid;
1199 c->phys_proc_id = c->initial_apicid;
1202 get_model_name(c); /* Default name */
1206 detect_null_seg_behavior(c);
1209 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1210 * systems that run Linux at CPL > 0 may or may not have the
1211 * issue, but, even if they have the issue, there's absolutely
1212 * nothing we can do about it because we can't use the real IRET
1215 * NB: For the time being, only 32-bit kernels support
1216 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1217 * whether to apply espfix using paravirt hooks. If any
1218 * non-paravirt system ever shows up that does *not* have the
1219 * ESPFIX issue, we can change this.
1221 #ifdef CONFIG_X86_32
1222 # ifdef CONFIG_PARAVIRT
1224 extern void native_iret(void);
1225 if (pv_cpu_ops.iret == native_iret)
1226 set_cpu_bug(c, X86_BUG_ESPFIX);
1229 set_cpu_bug(c, X86_BUG_ESPFIX);
1234 static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1237 * The heavy lifting of max_rmid and cache_occ_scale are handled
1238 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
1239 * in case CQM bits really aren't there in this CPU.
1241 if (c != &boot_cpu_data) {
1242 boot_cpu_data.x86_cache_max_rmid =
1243 min(boot_cpu_data.x86_cache_max_rmid,
1244 c->x86_cache_max_rmid);
1249 * Validate that ACPI/mptables have the same information about the
1250 * effective APIC id and update the package map.
1252 static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1255 unsigned int apicid, cpu = smp_processor_id();
1257 apicid = apic->cpu_present_to_apicid(cpu);
1259 if (apicid != c->apicid) {
1260 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1261 cpu, apicid, c->initial_apicid);
1263 BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
1265 c->logical_proc_id = 0;
1270 * This does the hard work of actually picking apart the CPU stuff...
1272 static void identify_cpu(struct cpuinfo_x86 *c)
1276 c->loops_per_jiffy = loops_per_jiffy;
1277 c->x86_cache_size = 0;
1278 c->x86_vendor = X86_VENDOR_UNKNOWN;
1279 c->x86_model = c->x86_stepping = 0; /* So far unknown... */
1280 c->x86_vendor_id[0] = '\0'; /* Unset */
1281 c->x86_model_id[0] = '\0'; /* Unset */
1282 c->x86_max_cores = 1;
1283 c->x86_coreid_bits = 0;
1285 #ifdef CONFIG_X86_64
1286 c->x86_clflush_size = 64;
1287 c->x86_phys_bits = 36;
1288 c->x86_virt_bits = 48;
1290 c->cpuid_level = -1; /* CPUID not detected */
1291 c->x86_clflush_size = 32;
1292 c->x86_phys_bits = 32;
1293 c->x86_virt_bits = 32;
1295 c->x86_cache_alignment = c->x86_clflush_size;
1296 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1298 generic_identify(c);
1300 if (this_cpu->c_identify)
1301 this_cpu->c_identify(c);
1303 /* Clear/Set all flags overridden by options, after probe */
1304 apply_forced_caps(c);
1306 #ifdef CONFIG_X86_64
1307 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1311 * Vendor-specific initialization. In this section we
1312 * canonicalize the feature flags, meaning if there are
1313 * features a certain CPU supports which CPUID doesn't
1314 * tell us, CPUID claiming incorrect flags, or other bugs,
1315 * we handle them here.
1317 * At the end of this section, c->x86_capability better
1318 * indicate the features this CPU genuinely supports!
1320 if (this_cpu->c_init)
1321 this_cpu->c_init(c);
1323 /* Disable the PN if appropriate */
1324 squash_the_stupid_serial_number(c);
1326 /* Set up SMEP/SMAP/UMIP */
1332 * The vendor-specific functions might have changed features.
1333 * Now we do "generic changes."
1336 /* Filter out anything that depends on CPUID levels we don't have */
1337 filter_cpuid_features(c, true);
1339 /* If the model name is still unset, do table lookup. */
1340 if (!c->x86_model_id[0]) {
1342 p = table_lookup_model(c);
1344 strcpy(c->x86_model_id, p);
1346 /* Last resort... */
1347 sprintf(c->x86_model_id, "%02x/%02x",
1348 c->x86, c->x86_model);
1351 #ifdef CONFIG_X86_64
1356 x86_init_cache_qos(c);
1360 * Clear/Set all flags overridden by options, need do it
1361 * before following smp all cpus cap AND.
1363 apply_forced_caps(c);
1366 * On SMP, boot_cpu_data holds the common feature set between
1367 * all CPUs; so make sure that we indicate which features are
1368 * common between the CPUs. The first time this routine gets
1369 * executed, c == &boot_cpu_data.
1371 if (c != &boot_cpu_data) {
1372 /* AND the already accumulated flags with these */
1373 for (i = 0; i < NCAPINTS; i++)
1374 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1376 /* OR, i.e. replicate the bug flags */
1377 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1378 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1381 /* Init Machine Check Exception if available. */
1384 select_idle_routine(c);
1387 numa_add_cpu(smp_processor_id());
1392 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1393 * on 32-bit kernels:
1395 #ifdef CONFIG_X86_32
1396 void enable_sep_cpu(void)
1398 struct tss_struct *tss;
1401 if (!boot_cpu_has(X86_FEATURE_SEP))
1405 tss = &per_cpu(cpu_tss_rw, cpu);
1408 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1409 * see the big comment in struct x86_hw_tss's definition.
1412 tss->x86_tss.ss1 = __KERNEL_CS;
1413 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1414 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
1415 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1421 void __init identify_boot_cpu(void)
1423 identify_cpu(&boot_cpu_data);
1424 #ifdef CONFIG_X86_32
1428 cpu_detect_tlb(&boot_cpu_data);
1431 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1433 BUG_ON(c == &boot_cpu_data);
1435 #ifdef CONFIG_X86_32
1439 validate_apic_and_package_id(c);
1440 x86_spec_ctrl_setup_ap();
1443 static __init int setup_noclflush(char *arg)
1445 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1446 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1449 __setup("noclflush", setup_noclflush);
1451 void print_cpu_info(struct cpuinfo_x86 *c)
1453 const char *vendor = NULL;
1455 if (c->x86_vendor < X86_VENDOR_NUM) {
1456 vendor = this_cpu->c_vendor;
1458 if (c->cpuid_level >= 0)
1459 vendor = c->x86_vendor_id;
1462 if (vendor && !strstr(c->x86_model_id, vendor))
1463 pr_cont("%s ", vendor);
1465 if (c->x86_model_id[0])
1466 pr_cont("%s", c->x86_model_id);
1468 pr_cont("%d86", c->x86);
1470 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1472 if (c->x86_stepping || c->cpuid_level >= 0)
1473 pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
1479 * clearcpuid= was already parsed in fpu__init_parse_early_param.
1480 * But we need to keep a dummy __setup around otherwise it would
1481 * show up as an environment variable for init.
1483 static __init int setup_clearcpuid(char *arg)
1487 __setup("clearcpuid=", setup_clearcpuid);
1489 #ifdef CONFIG_X86_64
1490 DEFINE_PER_CPU_FIRST(union irq_stack_union,
1491 irq_stack_union) __aligned(PAGE_SIZE) __visible;
1492 EXPORT_PER_CPU_SYMBOL_GPL(irq_stack_union);
1495 * The following percpu variables are hot. Align current_task to
1496 * cacheline size such that they fall in the same cacheline.
1498 DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1500 EXPORT_PER_CPU_SYMBOL(current_task);
1502 DEFINE_PER_CPU(char *, irq_stack_ptr) =
1503 init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE;
1505 DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1507 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1508 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1510 /* May not be marked __init: used by software suspend */
1511 void syscall_init(void)
1513 extern char _entry_trampoline[];
1514 extern char entry_SYSCALL_64_trampoline[];
1516 int cpu = smp_processor_id();
1517 unsigned long SYSCALL64_entry_trampoline =
1518 (unsigned long)get_cpu_entry_area(cpu)->entry_trampoline +
1519 (entry_SYSCALL_64_trampoline - _entry_trampoline);
1521 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1522 if (static_cpu_has(X86_FEATURE_PTI))
1523 wrmsrl(MSR_LSTAR, SYSCALL64_entry_trampoline);
1525 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1527 #ifdef CONFIG_IA32_EMULATION
1528 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1530 * This only works on Intel CPUs.
1531 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1532 * This does not cause SYSENTER to jump to the wrong location, because
1533 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1535 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1536 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1));
1537 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1539 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1540 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1541 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1542 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1545 /* Flags to clear on syscall */
1546 wrmsrl(MSR_SYSCALL_MASK,
1547 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1548 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1552 * Copies of the original ist values from the tss are only accessed during
1553 * debugging, no special alignment required.
1555 DEFINE_PER_CPU(struct orig_ist, orig_ist);
1557 static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1558 DEFINE_PER_CPU(int, debug_stack_usage);
1560 int is_debug_stack(unsigned long addr)
1562 return __this_cpu_read(debug_stack_usage) ||
1563 (addr <= __this_cpu_read(debug_stack_addr) &&
1564 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1566 NOKPROBE_SYMBOL(is_debug_stack);
1568 DEFINE_PER_CPU(u32, debug_idt_ctr);
1570 void debug_stack_set_zero(void)
1572 this_cpu_inc(debug_idt_ctr);
1575 NOKPROBE_SYMBOL(debug_stack_set_zero);
1577 void debug_stack_reset(void)
1579 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1581 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1584 NOKPROBE_SYMBOL(debug_stack_reset);
1586 #else /* CONFIG_X86_64 */
1588 DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1589 EXPORT_PER_CPU_SYMBOL(current_task);
1590 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1591 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1594 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1595 * the top of the kernel stack. Use an extra percpu variable to track the
1596 * top of the kernel stack directly.
1598 DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1599 (unsigned long)&init_thread_union + THREAD_SIZE;
1600 EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1602 #ifdef CONFIG_STACKPROTECTOR
1603 DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1606 #endif /* CONFIG_X86_64 */
1609 * Clear all 6 debug registers:
1611 static void clear_all_debug_regs(void)
1615 for (i = 0; i < 8; i++) {
1616 /* Ignore db4, db5 */
1617 if ((i == 4) || (i == 5))
1626 * Restore debug regs if using kgdbwait and you have a kernel debugger
1627 * connection established.
1629 static void dbg_restore_debug_regs(void)
1631 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1632 arch_kgdb_ops.correct_hw_break();
1634 #else /* ! CONFIG_KGDB */
1635 #define dbg_restore_debug_regs()
1636 #endif /* ! CONFIG_KGDB */
1638 static void wait_for_master_cpu(int cpu)
1642 * wait for ACK from master CPU before continuing
1643 * with AP initialization
1645 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1646 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1652 * cpu_init() initializes state that is per-CPU. Some data is already
1653 * initialized (naturally) in the bootstrap process, such as the GDT
1654 * and IDT. We reload them nevertheless, this function acts as a
1655 * 'CPU state barrier', nothing should get across.
1656 * A lot of state is already set up in PDA init for 64 bit
1658 #ifdef CONFIG_X86_64
1662 struct orig_ist *oist;
1663 struct task_struct *me;
1664 struct tss_struct *t;
1666 int cpu = raw_smp_processor_id();
1669 wait_for_master_cpu(cpu);
1672 * Initialize the CR4 shadow before doing anything that could
1680 t = &per_cpu(cpu_tss_rw, cpu);
1681 oist = &per_cpu(orig_ist, cpu);
1684 if (this_cpu_read(numa_node) == 0 &&
1685 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1686 set_numa_node(early_cpu_to_node(cpu));
1691 pr_debug("Initializing CPU#%d\n", cpu);
1693 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1696 * Initialize the per-CPU GDT with the boot GDT,
1697 * and set up the GDT descriptor:
1700 switch_to_new_gdt(cpu);
1705 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1708 wrmsrl(MSR_FS_BASE, 0);
1709 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1716 * set up and load the per-CPU TSS
1718 if (!oist->ist[0]) {
1719 char *estacks = get_cpu_entry_area(cpu)->exception_stacks;
1721 for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1722 estacks += exception_stack_sizes[v];
1723 oist->ist[v] = t->x86_tss.ist[v] =
1724 (unsigned long)estacks;
1725 if (v == DEBUG_STACK-1)
1726 per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1730 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1733 * <= is required because the CPU will access up to
1734 * 8 bits beyond the end of the IO permission bitmap.
1736 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1737 t->io_bitmap[i] = ~0UL;
1740 me->active_mm = &init_mm;
1742 initialize_tlbstate_and_flush();
1743 enter_lazy_tlb(&init_mm, me);
1746 * Initialize the TSS. sp0 points to the entry trampoline stack
1747 * regardless of what task is running.
1749 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1751 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1753 load_mm_ldt(&init_mm);
1755 clear_all_debug_regs();
1756 dbg_restore_debug_regs();
1763 load_fixmap_gdt(cpu);
1770 int cpu = smp_processor_id();
1771 struct task_struct *curr = current;
1772 struct tss_struct *t = &per_cpu(cpu_tss_rw, cpu);
1774 wait_for_master_cpu(cpu);
1777 * Initialize the CR4 shadow before doing anything that could
1782 show_ucode_info_early();
1784 pr_info("Initializing CPU#%d\n", cpu);
1786 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1787 boot_cpu_has(X86_FEATURE_TSC) ||
1788 boot_cpu_has(X86_FEATURE_DE))
1789 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1792 switch_to_new_gdt(cpu);
1795 * Set up and load the per-CPU TSS and LDT
1798 curr->active_mm = &init_mm;
1800 initialize_tlbstate_and_flush();
1801 enter_lazy_tlb(&init_mm, curr);
1804 * Initialize the TSS. Don't bother initializing sp0, as the initial
1805 * task never enters user mode.
1807 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1810 load_mm_ldt(&init_mm);
1812 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1814 #ifdef CONFIG_DOUBLEFAULT
1815 /* Set up doublefault TSS pointer in the GDT */
1816 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1819 clear_all_debug_regs();
1820 dbg_restore_debug_regs();
1824 load_fixmap_gdt(cpu);
1828 static void bsp_resume(void)
1830 if (this_cpu->c_bsp_resume)
1831 this_cpu->c_bsp_resume(&boot_cpu_data);
1834 static struct syscore_ops cpu_syscore_ops = {
1835 .resume = bsp_resume,
1838 static int __init init_cpu_syscore(void)
1840 register_syscore_ops(&cpu_syscore_ops);
1843 core_initcall(init_cpu_syscore);
1846 * The microcode loader calls this upon late microcode load to recheck features,
1847 * only when microcode has been updated. Caller holds microcode_mutex and CPU
1850 void microcode_check(void)
1852 struct cpuinfo_x86 info;
1854 perf_check_microcode();
1856 /* Reload CPUID max function as it might've changed. */
1857 info.cpuid_level = cpuid_eax(0);
1860 * Copy all capability leafs to pick up the synthetic ones so that
1861 * memcmp() below doesn't fail on that. The ones coming from CPUID will
1862 * get overwritten in get_cpu_cap().
1864 memcpy(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability));
1868 if (!memcmp(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability)))
1871 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
1872 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");