1 // SPDX-License-Identifier: GPL-2.0-only
2 /* cpu_feature_enabled() cannot be used this early */
3 #define USE_EARLY_PGTABLE_L5
5 #include <linux/memblock.h>
6 #include <linux/linkage.h>
7 #include <linux/bitops.h>
8 #include <linux/kernel.h>
9 #include <linux/export.h>
10 #include <linux/percpu.h>
11 #include <linux/string.h>
12 #include <linux/ctype.h>
13 #include <linux/delay.h>
14 #include <linux/sched/mm.h>
15 #include <linux/sched/clock.h>
16 #include <linux/sched/task.h>
17 #include <linux/sched/smt.h>
18 #include <linux/init.h>
19 #include <linux/kprobes.h>
20 #include <linux/kgdb.h>
21 #include <linux/mem_encrypt.h>
22 #include <linux/smp.h>
23 #include <linux/cpu.h>
25 #include <linux/syscore_ops.h>
26 #include <linux/pgtable.h>
27 #include <linux/stackprotector.h>
28 #include <linux/utsname.h>
30 #include <asm/alternative.h>
31 #include <asm/cmdline.h>
32 #include <asm/perf_event.h>
33 #include <asm/mmu_context.h>
34 #include <asm/doublefault.h>
35 #include <asm/archrandom.h>
36 #include <asm/hypervisor.h>
37 #include <asm/processor.h>
38 #include <asm/tlbflush.h>
39 #include <asm/debugreg.h>
40 #include <asm/sections.h>
41 #include <asm/vsyscall.h>
42 #include <linux/topology.h>
43 #include <linux/cpumask.h>
44 #include <linux/atomic.h>
45 #include <asm/proto.h>
46 #include <asm/setup.h>
49 #include <asm/fpu/api.h>
51 #include <asm/hwcap2.h>
52 #include <linux/numa.h>
59 #include <asm/cacheinfo.h>
60 #include <asm/memtype.h>
61 #include <asm/microcode.h>
62 #include <asm/intel-family.h>
63 #include <asm/cpu_device_id.h>
64 #include <asm/uv/uv.h>
66 #include <asm/set_memory.h>
67 #include <asm/traps.h>
73 u32 elf_hwcap2 __read_mostly;
75 /* Number of siblings per CPU package */
76 unsigned int smp_num_siblings __ro_after_init = 1;
77 EXPORT_SYMBOL(smp_num_siblings);
79 unsigned int __max_dies_per_package __ro_after_init = 1;
80 EXPORT_SYMBOL(__max_dies_per_package);
82 unsigned int __max_logical_packages __ro_after_init = 1;
83 EXPORT_SYMBOL(__max_logical_packages);
85 static struct ppin_info {
90 [X86_VENDOR_INTEL] = {
91 .feature = X86_FEATURE_INTEL_PPIN,
92 .msr_ppin_ctl = MSR_PPIN_CTL,
96 .feature = X86_FEATURE_AMD_PPIN,
97 .msr_ppin_ctl = MSR_AMD_PPIN_CTL,
98 .msr_ppin = MSR_AMD_PPIN
102 static const struct x86_cpu_id ppin_cpuids[] = {
103 X86_MATCH_FEATURE(X86_FEATURE_AMD_PPIN, &ppin_info[X86_VENDOR_AMD]),
104 X86_MATCH_FEATURE(X86_FEATURE_INTEL_PPIN, &ppin_info[X86_VENDOR_INTEL]),
106 /* Legacy models without CPUID enumeration */
107 X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE_X, &ppin_info[X86_VENDOR_INTEL]),
108 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, &ppin_info[X86_VENDOR_INTEL]),
109 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_D, &ppin_info[X86_VENDOR_INTEL]),
110 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, &ppin_info[X86_VENDOR_INTEL]),
111 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_X, &ppin_info[X86_VENDOR_INTEL]),
112 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_X, &ppin_info[X86_VENDOR_INTEL]),
113 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_D, &ppin_info[X86_VENDOR_INTEL]),
114 X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, &ppin_info[X86_VENDOR_INTEL]),
115 X86_MATCH_INTEL_FAM6_MODEL(EMERALDRAPIDS_X, &ppin_info[X86_VENDOR_INTEL]),
116 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNL, &ppin_info[X86_VENDOR_INTEL]),
117 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNM, &ppin_info[X86_VENDOR_INTEL]),
122 static void ppin_init(struct cpuinfo_x86 *c)
124 const struct x86_cpu_id *id;
125 unsigned long long val;
126 struct ppin_info *info;
128 id = x86_match_cpu(ppin_cpuids);
133 * Testing the presence of the MSR is not enough. Need to check
134 * that the PPIN_CTL allows reading of the PPIN.
136 info = (struct ppin_info *)id->driver_data;
138 if (rdmsrl_safe(info->msr_ppin_ctl, &val))
141 if ((val & 3UL) == 1UL) {
142 /* PPIN locked in disabled mode */
146 /* If PPIN is disabled, try to enable */
148 wrmsrl_safe(info->msr_ppin_ctl, val | 2UL);
149 rdmsrl_safe(info->msr_ppin_ctl, &val);
152 /* Is the enable bit set? */
154 c->ppin = __rdmsr(info->msr_ppin);
155 set_cpu_cap(c, info->feature);
160 clear_cpu_cap(c, info->feature);
163 static void default_init(struct cpuinfo_x86 *c)
166 cpu_detect_cache_sizes(c);
168 /* Not much we can do here... */
169 /* Check if at least it has cpuid */
170 if (c->cpuid_level == -1) {
171 /* No cpuid. It must be an ancient CPU */
173 strcpy(c->x86_model_id, "486");
174 else if (c->x86 == 3)
175 strcpy(c->x86_model_id, "386");
180 static const struct cpu_dev default_cpu = {
181 .c_init = default_init,
182 .c_vendor = "Unknown",
183 .c_x86_vendor = X86_VENDOR_UNKNOWN,
186 static const struct cpu_dev *this_cpu = &default_cpu;
188 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
191 * We need valid kernel segments for data and code in long mode too
192 * IRET will check the segment types kkeil 2000/10/28
193 * Also sysret mandates a special GDT layout
195 * TLS descriptors are currently at a different place compared to i386.
196 * Hopefully nobody expects them at a fixed place (Wine?)
198 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(DESC_CODE32, 0, 0xfffff),
199 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(DESC_CODE64, 0, 0xfffff),
200 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(DESC_DATA64, 0, 0xfffff),
201 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(DESC_CODE32 | DESC_USER, 0, 0xfffff),
202 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(DESC_DATA64 | DESC_USER, 0, 0xfffff),
203 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(DESC_CODE64 | DESC_USER, 0, 0xfffff),
205 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(DESC_CODE32, 0, 0xfffff),
206 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff),
207 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(DESC_CODE32 | DESC_USER, 0, 0xfffff),
208 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(DESC_DATA32 | DESC_USER, 0, 0xfffff),
210 * Segments used for calling PnP BIOS have byte granularity.
211 * They code segments and data segments have fixed 64k limits,
212 * the transfer segment sizes are set at run time.
214 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(DESC_CODE32_BIOS, 0, 0xffff),
215 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(DESC_CODE16, 0, 0xffff),
216 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0xffff),
217 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0),
218 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0),
220 * The APM segments have byte granularity and their bases
221 * are set at run time. All have 64k limits.
223 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(DESC_CODE32_BIOS, 0, 0xffff),
224 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(DESC_CODE16, 0, 0xffff),
225 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(DESC_DATA32_BIOS, 0, 0xffff),
227 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff),
228 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff),
231 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
234 static int __init x86_nopcid_setup(char *s)
236 /* nopcid doesn't accept parameters */
240 /* do not emit a message if the feature is not present */
241 if (!boot_cpu_has(X86_FEATURE_PCID))
244 setup_clear_cpu_cap(X86_FEATURE_PCID);
245 pr_info("nopcid: PCID feature disabled\n");
248 early_param("nopcid", x86_nopcid_setup);
251 static int __init x86_noinvpcid_setup(char *s)
253 /* noinvpcid doesn't accept parameters */
257 /* do not emit a message if the feature is not present */
258 if (!boot_cpu_has(X86_FEATURE_INVPCID))
261 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
262 pr_info("noinvpcid: INVPCID feature disabled\n");
265 early_param("noinvpcid", x86_noinvpcid_setup);
268 static int cachesize_override = -1;
269 static int disable_x86_serial_nr = 1;
271 static int __init cachesize_setup(char *str)
273 get_option(&str, &cachesize_override);
276 __setup("cachesize=", cachesize_setup);
278 /* Standard macro to see if a specific flag is changeable */
279 static inline int flag_is_changeable_p(u32 flag)
284 * Cyrix and IDT cpus allow disabling of CPUID
285 * so the code below may return different results
286 * when it is executed before and after enabling
287 * the CPUID. Add "volatile" to not allow gcc to
288 * optimize the subsequent calls to this function.
290 asm volatile ("pushfl \n\t"
301 : "=&r" (f1), "=&r" (f2)
304 return ((f1^f2) & flag) != 0;
307 /* Probe for the CPUID instruction */
308 int have_cpuid_p(void)
310 return flag_is_changeable_p(X86_EFLAGS_ID);
313 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
315 unsigned long lo, hi;
317 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
320 /* Disable processor serial number: */
322 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
324 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
326 pr_notice("CPU serial number disabled.\n");
327 clear_cpu_cap(c, X86_FEATURE_PN);
329 /* Disabling the serial number may affect the cpuid level */
330 c->cpuid_level = cpuid_eax(0);
333 static int __init x86_serial_nr_setup(char *s)
335 disable_x86_serial_nr = 0;
338 __setup("serialnumber", x86_serial_nr_setup);
340 static inline int flag_is_changeable_p(u32 flag)
344 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
349 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
351 if (cpu_has(c, X86_FEATURE_SMEP))
352 cr4_set_bits(X86_CR4_SMEP);
355 static __always_inline void setup_smap(struct cpuinfo_x86 *c)
357 unsigned long eflags = native_save_fl();
359 /* This should have been cleared long ago */
360 BUG_ON(eflags & X86_EFLAGS_AC);
362 if (cpu_has(c, X86_FEATURE_SMAP))
363 cr4_set_bits(X86_CR4_SMAP);
366 static __always_inline void setup_umip(struct cpuinfo_x86 *c)
368 /* Check the boot processor, plus build option for UMIP. */
369 if (!cpu_feature_enabled(X86_FEATURE_UMIP))
372 /* Check the current processor's cpuid bits. */
373 if (!cpu_has(c, X86_FEATURE_UMIP))
376 cr4_set_bits(X86_CR4_UMIP);
378 pr_info_once("x86/cpu: User Mode Instruction Prevention (UMIP) activated\n");
384 * Make sure UMIP is disabled in case it was enabled in a
385 * previous boot (e.g., via kexec).
387 cr4_clear_bits(X86_CR4_UMIP);
390 /* These bits should not change their value after CPU init is finished. */
391 static const unsigned long cr4_pinned_mask =
392 X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_UMIP |
393 X86_CR4_FSGSBASE | X86_CR4_CET;
394 static DEFINE_STATIC_KEY_FALSE_RO(cr_pinning);
395 static unsigned long cr4_pinned_bits __ro_after_init;
397 void native_write_cr0(unsigned long val)
399 unsigned long bits_missing = 0;
402 asm volatile("mov %0,%%cr0": "+r" (val) : : "memory");
404 if (static_branch_likely(&cr_pinning)) {
405 if (unlikely((val & X86_CR0_WP) != X86_CR0_WP)) {
406 bits_missing = X86_CR0_WP;
410 /* Warn after we've set the missing bits. */
411 WARN_ONCE(bits_missing, "CR0 WP bit went missing!?\n");
414 EXPORT_SYMBOL(native_write_cr0);
416 void __no_profile native_write_cr4(unsigned long val)
418 unsigned long bits_changed = 0;
421 asm volatile("mov %0,%%cr4": "+r" (val) : : "memory");
423 if (static_branch_likely(&cr_pinning)) {
424 if (unlikely((val & cr4_pinned_mask) != cr4_pinned_bits)) {
425 bits_changed = (val & cr4_pinned_mask) ^ cr4_pinned_bits;
426 val = (val & ~cr4_pinned_mask) | cr4_pinned_bits;
429 /* Warn after we've corrected the changed bits. */
430 WARN_ONCE(bits_changed, "pinned CR4 bits changed: 0x%lx!?\n",
434 #if IS_MODULE(CONFIG_LKDTM)
435 EXPORT_SYMBOL_GPL(native_write_cr4);
438 void cr4_update_irqsoff(unsigned long set, unsigned long clear)
440 unsigned long newval, cr4 = this_cpu_read(cpu_tlbstate.cr4);
442 lockdep_assert_irqs_disabled();
444 newval = (cr4 & ~clear) | set;
446 this_cpu_write(cpu_tlbstate.cr4, newval);
450 EXPORT_SYMBOL(cr4_update_irqsoff);
452 /* Read the CR4 shadow. */
453 unsigned long cr4_read_shadow(void)
455 return this_cpu_read(cpu_tlbstate.cr4);
457 EXPORT_SYMBOL_GPL(cr4_read_shadow);
461 unsigned long cr4 = __read_cr4();
463 if (boot_cpu_has(X86_FEATURE_PCID))
464 cr4 |= X86_CR4_PCIDE;
465 if (static_branch_likely(&cr_pinning))
466 cr4 = (cr4 & ~cr4_pinned_mask) | cr4_pinned_bits;
470 /* Initialize cr4 shadow for this CPU. */
471 this_cpu_write(cpu_tlbstate.cr4, cr4);
475 * Once CPU feature detection is finished (and boot params have been
476 * parsed), record any of the sensitive CR bits that are set, and
479 static void __init setup_cr_pinning(void)
481 cr4_pinned_bits = this_cpu_read(cpu_tlbstate.cr4) & cr4_pinned_mask;
482 static_key_enable(&cr_pinning.key);
485 static __init int x86_nofsgsbase_setup(char *arg)
487 /* Require an exact match without trailing characters. */
491 /* Do not emit a message if the feature is not present. */
492 if (!boot_cpu_has(X86_FEATURE_FSGSBASE))
495 setup_clear_cpu_cap(X86_FEATURE_FSGSBASE);
496 pr_info("FSGSBASE disabled via kernel command line\n");
499 __setup("nofsgsbase", x86_nofsgsbase_setup);
502 * Protection Keys are not available in 32-bit mode.
504 static bool pku_disabled;
506 static __always_inline void setup_pku(struct cpuinfo_x86 *c)
508 if (c == &boot_cpu_data) {
509 if (pku_disabled || !cpu_feature_enabled(X86_FEATURE_PKU))
512 * Setting CR4.PKE will cause the X86_FEATURE_OSPKE cpuid
513 * bit to be set. Enforce it.
515 setup_force_cpu_cap(X86_FEATURE_OSPKE);
517 } else if (!cpu_feature_enabled(X86_FEATURE_OSPKE)) {
521 cr4_set_bits(X86_CR4_PKE);
522 /* Load the default PKRU value */
523 pkru_write_default();
526 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
527 static __init int setup_disable_pku(char *arg)
530 * Do not clear the X86_FEATURE_PKU bit. All of the
531 * runtime checks are against OSPKE so clearing the
534 * This way, we will see "pku" in cpuinfo, but not
535 * "ospke", which is exactly what we want. It shows
536 * that the CPU has PKU, but the OS has not enabled it.
537 * This happens to be exactly how a system would look
538 * if we disabled the config option.
540 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
544 __setup("nopku", setup_disable_pku);
547 #ifdef CONFIG_X86_KERNEL_IBT
549 __noendbr u64 ibt_save(bool disable)
553 if (cpu_feature_enabled(X86_FEATURE_IBT)) {
554 rdmsrl(MSR_IA32_S_CET, msr);
556 wrmsrl(MSR_IA32_S_CET, msr & ~CET_ENDBR_EN);
562 __noendbr void ibt_restore(u64 save)
566 if (cpu_feature_enabled(X86_FEATURE_IBT)) {
567 rdmsrl(MSR_IA32_S_CET, msr);
568 msr &= ~CET_ENDBR_EN;
569 msr |= (save & CET_ENDBR_EN);
570 wrmsrl(MSR_IA32_S_CET, msr);
576 static __always_inline void setup_cet(struct cpuinfo_x86 *c)
578 bool user_shstk, kernel_ibt;
580 if (!IS_ENABLED(CONFIG_X86_CET))
583 kernel_ibt = HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT);
584 user_shstk = cpu_feature_enabled(X86_FEATURE_SHSTK) &&
585 IS_ENABLED(CONFIG_X86_USER_SHADOW_STACK);
587 if (!kernel_ibt && !user_shstk)
591 set_cpu_cap(c, X86_FEATURE_USER_SHSTK);
594 wrmsrl(MSR_IA32_S_CET, CET_ENDBR_EN);
596 wrmsrl(MSR_IA32_S_CET, 0);
598 cr4_set_bits(X86_CR4_CET);
600 if (kernel_ibt && ibt_selftest()) {
601 pr_err("IBT selftest: Failed!\n");
602 wrmsrl(MSR_IA32_S_CET, 0);
603 setup_clear_cpu_cap(X86_FEATURE_IBT);
607 __noendbr void cet_disable(void)
609 if (!(cpu_feature_enabled(X86_FEATURE_IBT) ||
610 cpu_feature_enabled(X86_FEATURE_SHSTK)))
613 wrmsrl(MSR_IA32_S_CET, 0);
614 wrmsrl(MSR_IA32_U_CET, 0);
618 * Some CPU features depend on higher CPUID levels, which may not always
619 * be available due to CPUID level capping or broken virtualization
620 * software. Add those features to this table to auto-disable them.
622 struct cpuid_dependent_feature {
627 static const struct cpuid_dependent_feature
628 cpuid_dependent_features[] = {
629 { X86_FEATURE_MWAIT, 0x00000005 },
630 { X86_FEATURE_DCA, 0x00000009 },
631 { X86_FEATURE_XSAVE, 0x0000000d },
635 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
637 const struct cpuid_dependent_feature *df;
639 for (df = cpuid_dependent_features; df->feature; df++) {
641 if (!cpu_has(c, df->feature))
644 * Note: cpuid_level is set to -1 if unavailable, but
645 * extended_extended_level is set to 0 if unavailable
646 * and the legitimate extended levels are all negative
647 * when signed; hence the weird messing around with
650 if (!((s32)df->level < 0 ?
651 (u32)df->level > (u32)c->extended_cpuid_level :
652 (s32)df->level > (s32)c->cpuid_level))
655 clear_cpu_cap(c, df->feature);
659 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
660 x86_cap_flag(df->feature), df->level);
665 * Naming convention should be: <Name> [(<Codename>)]
666 * This table only is used unless init_<vendor>() below doesn't set it;
667 * in particular, if CPUID levels 0x80000002..4 are supported, this
671 /* Look up CPU names by table lookup. */
672 static const char *table_lookup_model(struct cpuinfo_x86 *c)
675 const struct legacy_cpu_model_info *info;
677 if (c->x86_model >= 16)
678 return NULL; /* Range check */
683 info = this_cpu->legacy_models;
685 while (info->family) {
686 if (info->family == c->x86)
687 return info->model_names[c->x86_model];
691 return NULL; /* Not found */
694 /* Aligned to unsigned long to avoid split lock in atomic bitmap ops */
695 __u32 cpu_caps_cleared[NCAPINTS + NBUGINTS] __aligned(sizeof(unsigned long));
696 __u32 cpu_caps_set[NCAPINTS + NBUGINTS] __aligned(sizeof(unsigned long));
699 /* The 32-bit entry code needs to find cpu_entry_area. */
700 DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area);
703 /* Load the original GDT from the per-cpu structure */
704 void load_direct_gdt(int cpu)
706 struct desc_ptr gdt_descr;
708 gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
709 gdt_descr.size = GDT_SIZE - 1;
710 load_gdt(&gdt_descr);
712 EXPORT_SYMBOL_GPL(load_direct_gdt);
714 /* Load a fixmap remapping of the per-cpu GDT */
715 void load_fixmap_gdt(int cpu)
717 struct desc_ptr gdt_descr;
719 gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
720 gdt_descr.size = GDT_SIZE - 1;
721 load_gdt(&gdt_descr);
723 EXPORT_SYMBOL_GPL(load_fixmap_gdt);
726 * switch_gdt_and_percpu_base - Switch to direct GDT and runtime per CPU base
727 * @cpu: The CPU number for which this is invoked
729 * Invoked during early boot to switch from early GDT and early per CPU to
730 * the direct GDT and the runtime per CPU area. On 32-bit the percpu base
731 * switch is implicit by loading the direct GDT. On 64bit this requires
734 void __init switch_gdt_and_percpu_base(int cpu)
736 load_direct_gdt(cpu);
740 * No need to load %gs. It is already correct.
742 * Writing %gs on 64bit would zero GSBASE which would make any per
743 * CPU operation up to the point of the wrmsrl() fault.
745 * Set GSBASE to the new offset. Until the wrmsrl() happens the
746 * early mapping is still valid. That means the GSBASE update will
747 * lose any prior per CPU data which was not copied over in
748 * setup_per_cpu_areas().
750 * This works even with stackprotector enabled because the
751 * per CPU stack canary is 0 in both per CPU areas.
753 wrmsrl(MSR_GS_BASE, cpu_kernelmode_gs_base(cpu));
756 * %fs is already set to __KERNEL_PERCPU, but after switching GDT
757 * it is required to load FS again so that the 'hidden' part is
758 * updated from the new GDT. Up to this point the early per CPU
759 * translation is active. Any content of the early per CPU data
760 * which was not copied over in setup_per_cpu_areas() is lost.
762 loadsegment(fs, __KERNEL_PERCPU);
766 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
768 static void get_model_name(struct cpuinfo_x86 *c)
773 if (c->extended_cpuid_level < 0x80000004)
776 v = (unsigned int *)c->x86_model_id;
777 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
778 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
779 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
780 c->x86_model_id[48] = 0;
782 /* Trim whitespace */
783 p = q = s = &c->x86_model_id[0];
789 /* Note the last non-whitespace index */
799 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
801 unsigned int n, dummy, ebx, ecx, edx, l2size;
803 n = c->extended_cpuid_level;
805 if (n >= 0x80000005) {
806 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
807 c->x86_cache_size = (ecx>>24) + (edx>>24);
809 /* On K8 L1 TLB is inclusive, so don't count it */
814 if (n < 0x80000006) /* Some chips just has a large L1. */
817 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
821 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
823 /* do processor-specific cache resizing */
824 if (this_cpu->legacy_cache_size)
825 l2size = this_cpu->legacy_cache_size(c, l2size);
827 /* Allow user to override all this if necessary. */
828 if (cachesize_override != -1)
829 l2size = cachesize_override;
832 return; /* Again, no L2 cache is possible */
835 c->x86_cache_size = l2size;
838 u16 __read_mostly tlb_lli_4k[NR_INFO];
839 u16 __read_mostly tlb_lli_2m[NR_INFO];
840 u16 __read_mostly tlb_lli_4m[NR_INFO];
841 u16 __read_mostly tlb_lld_4k[NR_INFO];
842 u16 __read_mostly tlb_lld_2m[NR_INFO];
843 u16 __read_mostly tlb_lld_4m[NR_INFO];
844 u16 __read_mostly tlb_lld_1g[NR_INFO];
846 static void cpu_detect_tlb(struct cpuinfo_x86 *c)
848 if (this_cpu->c_detect_tlb)
849 this_cpu->c_detect_tlb(c);
851 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
852 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
853 tlb_lli_4m[ENTRIES]);
855 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
856 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
857 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
860 static void get_cpu_vendor(struct cpuinfo_x86 *c)
862 char *v = c->x86_vendor_id;
865 for (i = 0; i < X86_VENDOR_NUM; i++) {
869 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
870 (cpu_devs[i]->c_ident[1] &&
871 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
873 this_cpu = cpu_devs[i];
874 c->x86_vendor = this_cpu->c_x86_vendor;
879 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
880 "CPU: Your system may be unstable.\n", v);
882 c->x86_vendor = X86_VENDOR_UNKNOWN;
883 this_cpu = &default_cpu;
886 void cpu_detect(struct cpuinfo_x86 *c)
888 /* Get vendor name */
889 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
890 (unsigned int *)&c->x86_vendor_id[0],
891 (unsigned int *)&c->x86_vendor_id[8],
892 (unsigned int *)&c->x86_vendor_id[4]);
895 /* Intel-defined flags: level 0x00000001 */
896 if (c->cpuid_level >= 0x00000001) {
897 u32 junk, tfms, cap0, misc;
899 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
900 c->x86 = x86_family(tfms);
901 c->x86_model = x86_model(tfms);
902 c->x86_stepping = x86_stepping(tfms);
904 if (cap0 & (1<<19)) {
905 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
906 c->x86_cache_alignment = c->x86_clflush_size;
911 static void apply_forced_caps(struct cpuinfo_x86 *c)
915 for (i = 0; i < NCAPINTS + NBUGINTS; i++) {
916 c->x86_capability[i] &= ~cpu_caps_cleared[i];
917 c->x86_capability[i] |= cpu_caps_set[i];
921 static void init_speculation_control(struct cpuinfo_x86 *c)
924 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support,
925 * and they also have a different bit for STIBP support. Also,
926 * a hypervisor might have set the individual AMD bits even on
927 * Intel CPUs, for finer-grained selection of what's available.
929 if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) {
930 set_cpu_cap(c, X86_FEATURE_IBRS);
931 set_cpu_cap(c, X86_FEATURE_IBPB);
932 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
935 if (cpu_has(c, X86_FEATURE_INTEL_STIBP))
936 set_cpu_cap(c, X86_FEATURE_STIBP);
938 if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) ||
939 cpu_has(c, X86_FEATURE_VIRT_SSBD))
940 set_cpu_cap(c, X86_FEATURE_SSBD);
942 if (cpu_has(c, X86_FEATURE_AMD_IBRS)) {
943 set_cpu_cap(c, X86_FEATURE_IBRS);
944 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
947 if (cpu_has(c, X86_FEATURE_AMD_IBPB))
948 set_cpu_cap(c, X86_FEATURE_IBPB);
950 if (cpu_has(c, X86_FEATURE_AMD_STIBP)) {
951 set_cpu_cap(c, X86_FEATURE_STIBP);
952 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
955 if (cpu_has(c, X86_FEATURE_AMD_SSBD)) {
956 set_cpu_cap(c, X86_FEATURE_SSBD);
957 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
958 clear_cpu_cap(c, X86_FEATURE_VIRT_SSBD);
962 void get_cpu_cap(struct cpuinfo_x86 *c)
964 u32 eax, ebx, ecx, edx;
966 /* Intel-defined flags: level 0x00000001 */
967 if (c->cpuid_level >= 0x00000001) {
968 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
970 c->x86_capability[CPUID_1_ECX] = ecx;
971 c->x86_capability[CPUID_1_EDX] = edx;
974 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
975 if (c->cpuid_level >= 0x00000006)
976 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
978 /* Additional Intel-defined flags: level 0x00000007 */
979 if (c->cpuid_level >= 0x00000007) {
980 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
981 c->x86_capability[CPUID_7_0_EBX] = ebx;
982 c->x86_capability[CPUID_7_ECX] = ecx;
983 c->x86_capability[CPUID_7_EDX] = edx;
985 /* Check valid sub-leaf index before accessing it */
987 cpuid_count(0x00000007, 1, &eax, &ebx, &ecx, &edx);
988 c->x86_capability[CPUID_7_1_EAX] = eax;
992 /* Extended state features: level 0x0000000d */
993 if (c->cpuid_level >= 0x0000000d) {
994 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
996 c->x86_capability[CPUID_D_1_EAX] = eax;
999 /* AMD-defined flags: level 0x80000001 */
1000 eax = cpuid_eax(0x80000000);
1001 c->extended_cpuid_level = eax;
1003 if ((eax & 0xffff0000) == 0x80000000) {
1004 if (eax >= 0x80000001) {
1005 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
1007 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
1008 c->x86_capability[CPUID_8000_0001_EDX] = edx;
1012 if (c->extended_cpuid_level >= 0x80000007) {
1013 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1015 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
1019 if (c->extended_cpuid_level >= 0x80000008) {
1020 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
1021 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
1024 if (c->extended_cpuid_level >= 0x8000000a)
1025 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
1027 if (c->extended_cpuid_level >= 0x8000001f)
1028 c->x86_capability[CPUID_8000_001F_EAX] = cpuid_eax(0x8000001f);
1030 if (c->extended_cpuid_level >= 0x80000021)
1031 c->x86_capability[CPUID_8000_0021_EAX] = cpuid_eax(0x80000021);
1033 init_scattered_cpuid_features(c);
1034 init_speculation_control(c);
1037 * Clear/Set all flags overridden by options, after probe.
1038 * This needs to happen each time we re-probe, which may happen
1039 * several times during CPU initialization.
1041 apply_forced_caps(c);
1044 void get_cpu_address_sizes(struct cpuinfo_x86 *c)
1046 u32 eax, ebx, ecx, edx;
1047 bool vp_bits_from_cpuid = true;
1049 if (!cpu_has(c, X86_FEATURE_CPUID) ||
1050 (c->extended_cpuid_level < 0x80000008))
1051 vp_bits_from_cpuid = false;
1053 if (vp_bits_from_cpuid) {
1054 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
1056 c->x86_virt_bits = (eax >> 8) & 0xff;
1057 c->x86_phys_bits = eax & 0xff;
1059 if (IS_ENABLED(CONFIG_X86_64)) {
1060 c->x86_clflush_size = 64;
1061 c->x86_phys_bits = 36;
1062 c->x86_virt_bits = 48;
1064 c->x86_clflush_size = 32;
1065 c->x86_virt_bits = 32;
1066 c->x86_phys_bits = 32;
1068 if (cpu_has(c, X86_FEATURE_PAE) ||
1069 cpu_has(c, X86_FEATURE_PSE36))
1070 c->x86_phys_bits = 36;
1073 c->x86_cache_bits = c->x86_phys_bits;
1074 c->x86_cache_alignment = c->x86_clflush_size;
1077 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
1079 #ifdef CONFIG_X86_32
1083 * First of all, decide if this is a 486 or higher
1084 * It's a 486 if we can modify the AC flag
1086 if (flag_is_changeable_p(X86_EFLAGS_AC))
1091 for (i = 0; i < X86_VENDOR_NUM; i++)
1092 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
1093 c->x86_vendor_id[0] = 0;
1094 cpu_devs[i]->c_identify(c);
1095 if (c->x86_vendor_id[0]) {
1103 #define NO_SPECULATION BIT(0)
1104 #define NO_MELTDOWN BIT(1)
1105 #define NO_SSB BIT(2)
1106 #define NO_L1TF BIT(3)
1107 #define NO_MDS BIT(4)
1108 #define MSBDS_ONLY BIT(5)
1109 #define NO_SWAPGS BIT(6)
1110 #define NO_ITLB_MULTIHIT BIT(7)
1111 #define NO_SPECTRE_V2 BIT(8)
1112 #define NO_MMIO BIT(9)
1113 #define NO_EIBRS_PBRSB BIT(10)
1115 #define VULNWL(vendor, family, model, whitelist) \
1116 X86_MATCH_VENDOR_FAM_MODEL(vendor, family, model, whitelist)
1118 #define VULNWL_INTEL(model, whitelist) \
1119 VULNWL(INTEL, 6, INTEL_FAM6_##model, whitelist)
1121 #define VULNWL_AMD(family, whitelist) \
1122 VULNWL(AMD, family, X86_MODEL_ANY, whitelist)
1124 #define VULNWL_HYGON(family, whitelist) \
1125 VULNWL(HYGON, family, X86_MODEL_ANY, whitelist)
1127 static const __initconst struct x86_cpu_id cpu_vuln_whitelist[] = {
1128 VULNWL(ANY, 4, X86_MODEL_ANY, NO_SPECULATION),
1129 VULNWL(CENTAUR, 5, X86_MODEL_ANY, NO_SPECULATION),
1130 VULNWL(INTEL, 5, X86_MODEL_ANY, NO_SPECULATION),
1131 VULNWL(NSC, 5, X86_MODEL_ANY, NO_SPECULATION),
1132 VULNWL(VORTEX, 5, X86_MODEL_ANY, NO_SPECULATION),
1133 VULNWL(VORTEX, 6, X86_MODEL_ANY, NO_SPECULATION),
1135 /* Intel Family 6 */
1136 VULNWL_INTEL(TIGERLAKE, NO_MMIO),
1137 VULNWL_INTEL(TIGERLAKE_L, NO_MMIO),
1138 VULNWL_INTEL(ALDERLAKE, NO_MMIO),
1139 VULNWL_INTEL(ALDERLAKE_L, NO_MMIO),
1141 VULNWL_INTEL(ATOM_SALTWELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
1142 VULNWL_INTEL(ATOM_SALTWELL_TABLET, NO_SPECULATION | NO_ITLB_MULTIHIT),
1143 VULNWL_INTEL(ATOM_SALTWELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
1144 VULNWL_INTEL(ATOM_BONNELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
1145 VULNWL_INTEL(ATOM_BONNELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
1147 VULNWL_INTEL(ATOM_SILVERMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1148 VULNWL_INTEL(ATOM_SILVERMONT_D, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1149 VULNWL_INTEL(ATOM_SILVERMONT_MID, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1150 VULNWL_INTEL(ATOM_AIRMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1151 VULNWL_INTEL(XEON_PHI_KNL, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1152 VULNWL_INTEL(XEON_PHI_KNM, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1154 VULNWL_INTEL(CORE_YONAH, NO_SSB),
1156 VULNWL_INTEL(ATOM_AIRMONT_MID, NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1157 VULNWL_INTEL(ATOM_AIRMONT_NP, NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1159 VULNWL_INTEL(ATOM_GOLDMONT, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1160 VULNWL_INTEL(ATOM_GOLDMONT_D, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1161 VULNWL_INTEL(ATOM_GOLDMONT_PLUS, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB),
1164 * Technically, swapgs isn't serializing on AMD (despite it previously
1165 * being documented as such in the APM). But according to AMD, %gs is
1166 * updated non-speculatively, and the issuing of %gs-relative memory
1167 * operands will be blocked until the %gs update completes, which is
1168 * good enough for our purposes.
1171 VULNWL_INTEL(ATOM_TREMONT, NO_EIBRS_PBRSB),
1172 VULNWL_INTEL(ATOM_TREMONT_L, NO_EIBRS_PBRSB),
1173 VULNWL_INTEL(ATOM_TREMONT_D, NO_ITLB_MULTIHIT | NO_EIBRS_PBRSB),
1175 /* AMD Family 0xf - 0x12 */
1176 VULNWL_AMD(0x0f, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1177 VULNWL_AMD(0x10, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1178 VULNWL_AMD(0x11, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1179 VULNWL_AMD(0x12, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1181 /* FAMILY_ANY must be last, otherwise 0x0f - 0x12 matches won't work */
1182 VULNWL_AMD(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB),
1183 VULNWL_HYGON(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB),
1185 /* Zhaoxin Family 7 */
1186 VULNWL(CENTAUR, 7, X86_MODEL_ANY, NO_SPECTRE_V2 | NO_SWAPGS | NO_MMIO),
1187 VULNWL(ZHAOXIN, 7, X86_MODEL_ANY, NO_SPECTRE_V2 | NO_SWAPGS | NO_MMIO),
1191 #define VULNBL(vendor, family, model, blacklist) \
1192 X86_MATCH_VENDOR_FAM_MODEL(vendor, family, model, blacklist)
1194 #define VULNBL_INTEL_STEPPINGS(model, steppings, issues) \
1195 X86_MATCH_VENDOR_FAM_MODEL_STEPPINGS_FEATURE(INTEL, 6, \
1196 INTEL_FAM6_##model, steppings, \
1197 X86_FEATURE_ANY, issues)
1199 #define VULNBL_AMD(family, blacklist) \
1200 VULNBL(AMD, family, X86_MODEL_ANY, blacklist)
1202 #define VULNBL_HYGON(family, blacklist) \
1203 VULNBL(HYGON, family, X86_MODEL_ANY, blacklist)
1205 #define SRBDS BIT(0)
1206 /* CPU is affected by X86_BUG_MMIO_STALE_DATA */
1208 /* CPU is affected by Shared Buffers Data Sampling (SBDS), a variant of X86_BUG_MMIO_STALE_DATA */
1209 #define MMIO_SBDS BIT(2)
1210 /* CPU is affected by RETbleed, speculating where you would not expect it */
1211 #define RETBLEED BIT(3)
1212 /* CPU is affected by SMT (cross-thread) return predictions */
1213 #define SMT_RSB BIT(4)
1214 /* CPU is affected by SRSO */
1216 /* CPU is affected by GDS */
1219 static const struct x86_cpu_id cpu_vuln_blacklist[] __initconst = {
1220 VULNBL_INTEL_STEPPINGS(IVYBRIDGE, X86_STEPPING_ANY, SRBDS),
1221 VULNBL_INTEL_STEPPINGS(HASWELL, X86_STEPPING_ANY, SRBDS),
1222 VULNBL_INTEL_STEPPINGS(HASWELL_L, X86_STEPPING_ANY, SRBDS),
1223 VULNBL_INTEL_STEPPINGS(HASWELL_G, X86_STEPPING_ANY, SRBDS),
1224 VULNBL_INTEL_STEPPINGS(HASWELL_X, X86_STEPPING_ANY, MMIO),
1225 VULNBL_INTEL_STEPPINGS(BROADWELL_D, X86_STEPPING_ANY, MMIO),
1226 VULNBL_INTEL_STEPPINGS(BROADWELL_G, X86_STEPPING_ANY, SRBDS),
1227 VULNBL_INTEL_STEPPINGS(BROADWELL_X, X86_STEPPING_ANY, MMIO),
1228 VULNBL_INTEL_STEPPINGS(BROADWELL, X86_STEPPING_ANY, SRBDS),
1229 VULNBL_INTEL_STEPPINGS(SKYLAKE_X, X86_STEPPING_ANY, MMIO | RETBLEED | GDS),
1230 VULNBL_INTEL_STEPPINGS(SKYLAKE_L, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS),
1231 VULNBL_INTEL_STEPPINGS(SKYLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS),
1232 VULNBL_INTEL_STEPPINGS(KABYLAKE_L, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS),
1233 VULNBL_INTEL_STEPPINGS(KABYLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS),
1234 VULNBL_INTEL_STEPPINGS(CANNONLAKE_L, X86_STEPPING_ANY, RETBLEED),
1235 VULNBL_INTEL_STEPPINGS(ICELAKE_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS),
1236 VULNBL_INTEL_STEPPINGS(ICELAKE_D, X86_STEPPING_ANY, MMIO | GDS),
1237 VULNBL_INTEL_STEPPINGS(ICELAKE_X, X86_STEPPING_ANY, MMIO | GDS),
1238 VULNBL_INTEL_STEPPINGS(COMETLAKE, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS),
1239 VULNBL_INTEL_STEPPINGS(COMETLAKE_L, X86_STEPPINGS(0x0, 0x0), MMIO | RETBLEED),
1240 VULNBL_INTEL_STEPPINGS(COMETLAKE_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS),
1241 VULNBL_INTEL_STEPPINGS(TIGERLAKE_L, X86_STEPPING_ANY, GDS),
1242 VULNBL_INTEL_STEPPINGS(TIGERLAKE, X86_STEPPING_ANY, GDS),
1243 VULNBL_INTEL_STEPPINGS(LAKEFIELD, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED),
1244 VULNBL_INTEL_STEPPINGS(ROCKETLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS),
1245 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT, X86_STEPPING_ANY, MMIO | MMIO_SBDS),
1246 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT_D, X86_STEPPING_ANY, MMIO),
1247 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS),
1249 VULNBL_AMD(0x15, RETBLEED),
1250 VULNBL_AMD(0x16, RETBLEED),
1251 VULNBL_AMD(0x17, RETBLEED | SMT_RSB | SRSO),
1252 VULNBL_HYGON(0x18, RETBLEED | SMT_RSB | SRSO),
1253 VULNBL_AMD(0x19, SRSO),
1257 static bool __init cpu_matches(const struct x86_cpu_id *table, unsigned long which)
1259 const struct x86_cpu_id *m = x86_match_cpu(table);
1261 return m && !!(m->driver_data & which);
1264 u64 x86_read_arch_cap_msr(void)
1268 if (boot_cpu_has(X86_FEATURE_ARCH_CAPABILITIES))
1269 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
1274 static bool arch_cap_mmio_immune(u64 ia32_cap)
1276 return (ia32_cap & ARCH_CAP_FBSDP_NO &&
1277 ia32_cap & ARCH_CAP_PSDP_NO &&
1278 ia32_cap & ARCH_CAP_SBDR_SSDP_NO);
1281 static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
1283 u64 ia32_cap = x86_read_arch_cap_msr();
1285 /* Set ITLB_MULTIHIT bug if cpu is not in the whitelist and not mitigated */
1286 if (!cpu_matches(cpu_vuln_whitelist, NO_ITLB_MULTIHIT) &&
1287 !(ia32_cap & ARCH_CAP_PSCHANGE_MC_NO))
1288 setup_force_cpu_bug(X86_BUG_ITLB_MULTIHIT);
1290 if (cpu_matches(cpu_vuln_whitelist, NO_SPECULATION))
1293 setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
1295 if (!cpu_matches(cpu_vuln_whitelist, NO_SPECTRE_V2))
1296 setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
1298 if (!cpu_matches(cpu_vuln_whitelist, NO_SSB) &&
1299 !(ia32_cap & ARCH_CAP_SSB_NO) &&
1300 !cpu_has(c, X86_FEATURE_AMD_SSB_NO))
1301 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
1304 * AMD's AutoIBRS is equivalent to Intel's eIBRS - use the Intel feature
1305 * flag and protect from vendor-specific bugs via the whitelist.
1307 if ((ia32_cap & ARCH_CAP_IBRS_ALL) || cpu_has(c, X86_FEATURE_AUTOIBRS)) {
1308 setup_force_cpu_cap(X86_FEATURE_IBRS_ENHANCED);
1309 if (!cpu_matches(cpu_vuln_whitelist, NO_EIBRS_PBRSB) &&
1310 !(ia32_cap & ARCH_CAP_PBRSB_NO))
1311 setup_force_cpu_bug(X86_BUG_EIBRS_PBRSB);
1314 if (!cpu_matches(cpu_vuln_whitelist, NO_MDS) &&
1315 !(ia32_cap & ARCH_CAP_MDS_NO)) {
1316 setup_force_cpu_bug(X86_BUG_MDS);
1317 if (cpu_matches(cpu_vuln_whitelist, MSBDS_ONLY))
1318 setup_force_cpu_bug(X86_BUG_MSBDS_ONLY);
1321 if (!cpu_matches(cpu_vuln_whitelist, NO_SWAPGS))
1322 setup_force_cpu_bug(X86_BUG_SWAPGS);
1325 * When the CPU is not mitigated for TAA (TAA_NO=0) set TAA bug when:
1326 * - TSX is supported or
1327 * - TSX_CTRL is present
1329 * TSX_CTRL check is needed for cases when TSX could be disabled before
1330 * the kernel boot e.g. kexec.
1331 * TSX_CTRL check alone is not sufficient for cases when the microcode
1332 * update is not present or running as guest that don't get TSX_CTRL.
1334 if (!(ia32_cap & ARCH_CAP_TAA_NO) &&
1335 (cpu_has(c, X86_FEATURE_RTM) ||
1336 (ia32_cap & ARCH_CAP_TSX_CTRL_MSR)))
1337 setup_force_cpu_bug(X86_BUG_TAA);
1340 * SRBDS affects CPUs which support RDRAND or RDSEED and are listed
1341 * in the vulnerability blacklist.
1343 * Some of the implications and mitigation of Shared Buffers Data
1344 * Sampling (SBDS) are similar to SRBDS. Give SBDS same treatment as
1347 if ((cpu_has(c, X86_FEATURE_RDRAND) ||
1348 cpu_has(c, X86_FEATURE_RDSEED)) &&
1349 cpu_matches(cpu_vuln_blacklist, SRBDS | MMIO_SBDS))
1350 setup_force_cpu_bug(X86_BUG_SRBDS);
1353 * Processor MMIO Stale Data bug enumeration
1355 * Affected CPU list is generally enough to enumerate the vulnerability,
1356 * but for virtualization case check for ARCH_CAP MSR bits also, VMM may
1357 * not want the guest to enumerate the bug.
1359 * Set X86_BUG_MMIO_UNKNOWN for CPUs that are neither in the blacklist,
1360 * nor in the whitelist and also don't enumerate MSR ARCH_CAP MMIO bits.
1362 if (!arch_cap_mmio_immune(ia32_cap)) {
1363 if (cpu_matches(cpu_vuln_blacklist, MMIO))
1364 setup_force_cpu_bug(X86_BUG_MMIO_STALE_DATA);
1365 else if (!cpu_matches(cpu_vuln_whitelist, NO_MMIO))
1366 setup_force_cpu_bug(X86_BUG_MMIO_UNKNOWN);
1369 if (!cpu_has(c, X86_FEATURE_BTC_NO)) {
1370 if (cpu_matches(cpu_vuln_blacklist, RETBLEED) || (ia32_cap & ARCH_CAP_RSBA))
1371 setup_force_cpu_bug(X86_BUG_RETBLEED);
1374 if (cpu_matches(cpu_vuln_blacklist, SMT_RSB))
1375 setup_force_cpu_bug(X86_BUG_SMT_RSB);
1377 if (!cpu_has(c, X86_FEATURE_SRSO_NO)) {
1378 if (cpu_matches(cpu_vuln_blacklist, SRSO))
1379 setup_force_cpu_bug(X86_BUG_SRSO);
1383 * Check if CPU is vulnerable to GDS. If running in a virtual machine on
1384 * an affected processor, the VMM may have disabled the use of GATHER by
1385 * disabling AVX2. The only way to do this in HW is to clear XCR0[2],
1386 * which means that AVX will be disabled.
1388 if (cpu_matches(cpu_vuln_blacklist, GDS) && !(ia32_cap & ARCH_CAP_GDS_NO) &&
1389 boot_cpu_has(X86_FEATURE_AVX))
1390 setup_force_cpu_bug(X86_BUG_GDS);
1392 if (cpu_matches(cpu_vuln_whitelist, NO_MELTDOWN))
1395 /* Rogue Data Cache Load? No! */
1396 if (ia32_cap & ARCH_CAP_RDCL_NO)
1399 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
1401 if (cpu_matches(cpu_vuln_whitelist, NO_L1TF))
1404 setup_force_cpu_bug(X86_BUG_L1TF);
1408 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1409 * unfortunately, that's not true in practice because of early VIA
1410 * chips and (more importantly) broken virtualizers that are not easy
1411 * to detect. In the latter case it doesn't even *fail* reliably, so
1412 * probing for it doesn't even work. Disable it completely on 32-bit
1413 * unless we can find a reliable way to detect all the broken cases.
1414 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1416 static void detect_nopl(void)
1418 #ifdef CONFIG_X86_32
1419 setup_clear_cpu_cap(X86_FEATURE_NOPL);
1421 setup_force_cpu_cap(X86_FEATURE_NOPL);
1426 * We parse cpu parameters early because fpu__init_system() is executed
1427 * before parse_early_param().
1429 static void __init cpu_parse_early_param(void)
1432 char *argptr = arg, *opt;
1433 int arglen, taint = 0;
1435 #ifdef CONFIG_X86_32
1436 if (cmdline_find_option_bool(boot_command_line, "no387"))
1437 #ifdef CONFIG_MATH_EMULATION
1438 setup_clear_cpu_cap(X86_FEATURE_FPU);
1440 pr_err("Option 'no387' required CONFIG_MATH_EMULATION enabled.\n");
1443 if (cmdline_find_option_bool(boot_command_line, "nofxsr"))
1444 setup_clear_cpu_cap(X86_FEATURE_FXSR);
1447 if (cmdline_find_option_bool(boot_command_line, "noxsave"))
1448 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
1450 if (cmdline_find_option_bool(boot_command_line, "noxsaveopt"))
1451 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
1453 if (cmdline_find_option_bool(boot_command_line, "noxsaves"))
1454 setup_clear_cpu_cap(X86_FEATURE_XSAVES);
1456 if (cmdline_find_option_bool(boot_command_line, "nousershstk"))
1457 setup_clear_cpu_cap(X86_FEATURE_USER_SHSTK);
1459 arglen = cmdline_find_option(boot_command_line, "clearcpuid", arg, sizeof(arg));
1463 pr_info("Clearing CPUID bits:");
1466 bool found __maybe_unused = false;
1469 opt = strsep(&argptr, ",");
1472 * Handle naked numbers first for feature flags which don't
1475 if (!kstrtouint(opt, 10, &bit)) {
1476 if (bit < NCAPINTS * 32) {
1478 /* empty-string, i.e., ""-defined feature flags */
1479 if (!x86_cap_flags[bit])
1480 pr_cont(" " X86_CAP_FMT_NUM, x86_cap_flag_num(bit));
1482 pr_cont(" " X86_CAP_FMT, x86_cap_flag(bit));
1484 setup_clear_cpu_cap(bit);
1488 * The assumption is that there are no feature names with only
1489 * numbers in the name thus go to the next argument.
1494 for (bit = 0; bit < 32 * NCAPINTS; bit++) {
1495 if (!x86_cap_flag(bit))
1498 if (strcmp(x86_cap_flag(bit), opt))
1501 pr_cont(" %s", opt);
1502 setup_clear_cpu_cap(bit);
1509 pr_cont(" (unknown: %s)", opt);
1514 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1518 * Do minimum CPU detection early.
1519 * Fields really needed: vendor, cpuid_level, family, model, mask,
1521 * The others are not touched to avoid unwanted side effects.
1523 * WARNING: this function is only called on the boot CPU. Don't add code
1524 * here that is supposed to run on all CPUs.
1526 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
1528 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1529 c->extended_cpuid_level = 0;
1531 if (!have_cpuid_p())
1532 identify_cpu_without_cpuid(c);
1534 /* cyrix could have cpuid enabled via c_identify()*/
1535 if (have_cpuid_p()) {
1539 setup_force_cpu_cap(X86_FEATURE_CPUID);
1540 cpu_parse_early_param();
1542 cpu_init_topology(c);
1544 if (this_cpu->c_early_init)
1545 this_cpu->c_early_init(c);
1548 filter_cpuid_features(c, false);
1550 if (this_cpu->c_bsp_init)
1551 this_cpu->c_bsp_init(c);
1553 setup_clear_cpu_cap(X86_FEATURE_CPUID);
1554 cpu_init_topology(c);
1557 get_cpu_address_sizes(c);
1559 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1561 cpu_set_bug_bits(c);
1565 #ifdef CONFIG_X86_32
1567 * Regardless of whether PCID is enumerated, the SDM says
1568 * that it can't be enabled in 32-bit mode.
1570 setup_clear_cpu_cap(X86_FEATURE_PCID);
1574 * Later in the boot process pgtable_l5_enabled() relies on
1575 * cpu_feature_enabled(X86_FEATURE_LA57). If 5-level paging is not
1576 * enabled by this point we need to clear the feature bit to avoid
1577 * false-positives at the later stage.
1579 * pgtable_l5_enabled() can be false here for several reasons:
1580 * - 5-level paging is disabled compile-time;
1581 * - it's 32-bit kernel;
1582 * - machine doesn't support 5-level paging;
1583 * - user specified 'no5lvl' in kernel command line.
1585 if (!pgtable_l5_enabled())
1586 setup_clear_cpu_cap(X86_FEATURE_LA57);
1591 void __init early_cpu_init(void)
1593 const struct cpu_dev *const *cdev;
1596 #ifdef CONFIG_PROCESSOR_SELECT
1597 pr_info("KERNEL supported cpus:\n");
1600 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1601 const struct cpu_dev *cpudev = *cdev;
1603 if (count >= X86_VENDOR_NUM)
1605 cpu_devs[count] = cpudev;
1608 #ifdef CONFIG_PROCESSOR_SELECT
1612 for (j = 0; j < 2; j++) {
1613 if (!cpudev->c_ident[j])
1615 pr_info(" %s %s\n", cpudev->c_vendor,
1616 cpudev->c_ident[j]);
1621 early_identify_cpu(&boot_cpu_data);
1624 static bool detect_null_seg_behavior(void)
1627 * Empirically, writing zero to a segment selector on AMD does
1628 * not clear the base, whereas writing zero to a segment
1629 * selector on Intel does clear the base. Intel's behavior
1630 * allows slightly faster context switches in the common case
1631 * where GS is unused by the prev and next threads.
1633 * Since neither vendor documents this anywhere that I can see,
1634 * detect it directly instead of hard-coding the choice by
1637 * I've designated AMD's behavior as the "bug" because it's
1638 * counterintuitive and less friendly.
1641 unsigned long old_base, tmp;
1642 rdmsrl(MSR_FS_BASE, old_base);
1643 wrmsrl(MSR_FS_BASE, 1);
1645 rdmsrl(MSR_FS_BASE, tmp);
1646 wrmsrl(MSR_FS_BASE, old_base);
1650 void check_null_seg_clears_base(struct cpuinfo_x86 *c)
1652 /* BUG_NULL_SEG is only relevant with 64bit userspace */
1653 if (!IS_ENABLED(CONFIG_X86_64))
1656 if (cpu_has(c, X86_FEATURE_NULL_SEL_CLR_BASE))
1660 * CPUID bit above wasn't set. If this kernel is still running
1661 * as a HV guest, then the HV has decided not to advertize
1662 * that CPUID bit for whatever reason. For example, one
1663 * member of the migration pool might be vulnerable. Which
1664 * means, the bug is present: set the BUG flag and return.
1666 if (cpu_has(c, X86_FEATURE_HYPERVISOR)) {
1667 set_cpu_bug(c, X86_BUG_NULL_SEG);
1672 * Zen2 CPUs also have this behaviour, but no CPUID bit.
1673 * 0x18 is the respective family for Hygon.
1675 if ((c->x86 == 0x17 || c->x86 == 0x18) &&
1676 detect_null_seg_behavior())
1679 /* All the remaining ones are affected */
1680 set_cpu_bug(c, X86_BUG_NULL_SEG);
1683 static void generic_identify(struct cpuinfo_x86 *c)
1685 c->extended_cpuid_level = 0;
1687 if (!have_cpuid_p())
1688 identify_cpu_without_cpuid(c);
1690 /* cyrix could have cpuid enabled via c_identify()*/
1691 if (!have_cpuid_p())
1700 get_cpu_address_sizes(c);
1702 get_model_name(c); /* Default name */
1705 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1706 * systems that run Linux at CPL > 0 may or may not have the
1707 * issue, but, even if they have the issue, there's absolutely
1708 * nothing we can do about it because we can't use the real IRET
1711 * NB: For the time being, only 32-bit kernels support
1712 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1713 * whether to apply espfix using paravirt hooks. If any
1714 * non-paravirt system ever shows up that does *not* have the
1715 * ESPFIX issue, we can change this.
1717 #ifdef CONFIG_X86_32
1718 set_cpu_bug(c, X86_BUG_ESPFIX);
1723 * This does the hard work of actually picking apart the CPU stuff...
1725 static void identify_cpu(struct cpuinfo_x86 *c)
1729 c->loops_per_jiffy = loops_per_jiffy;
1730 c->x86_cache_size = 0;
1731 c->x86_vendor = X86_VENDOR_UNKNOWN;
1732 c->x86_model = c->x86_stepping = 0; /* So far unknown... */
1733 c->x86_vendor_id[0] = '\0'; /* Unset */
1734 c->x86_model_id[0] = '\0'; /* Unset */
1735 c->x86_max_cores = 1;
1736 #ifdef CONFIG_X86_64
1737 c->x86_clflush_size = 64;
1738 c->x86_phys_bits = 36;
1739 c->x86_virt_bits = 48;
1741 c->cpuid_level = -1; /* CPUID not detected */
1742 c->x86_clflush_size = 32;
1743 c->x86_phys_bits = 32;
1744 c->x86_virt_bits = 32;
1746 c->x86_cache_alignment = c->x86_clflush_size;
1747 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1748 #ifdef CONFIG_X86_VMX_FEATURE_NAMES
1749 memset(&c->vmx_capability, 0, sizeof(c->vmx_capability));
1752 generic_identify(c);
1754 cpu_parse_topology(c);
1756 if (this_cpu->c_identify)
1757 this_cpu->c_identify(c);
1759 /* Clear/Set all flags overridden by options, after probe */
1760 apply_forced_caps(c);
1763 * Set default APIC and TSC_DEADLINE MSR fencing flag. AMD and
1764 * Hygon will clear it in ->c_init() below.
1766 set_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE);
1769 * Vendor-specific initialization. In this section we
1770 * canonicalize the feature flags, meaning if there are
1771 * features a certain CPU supports which CPUID doesn't
1772 * tell us, CPUID claiming incorrect flags, or other bugs,
1773 * we handle them here.
1775 * At the end of this section, c->x86_capability better
1776 * indicate the features this CPU genuinely supports!
1778 if (this_cpu->c_init)
1779 this_cpu->c_init(c);
1781 /* Disable the PN if appropriate */
1782 squash_the_stupid_serial_number(c);
1784 /* Set up SMEP/SMAP/UMIP */
1789 /* Enable FSGSBASE instructions if available. */
1790 if (cpu_has(c, X86_FEATURE_FSGSBASE)) {
1791 cr4_set_bits(X86_CR4_FSGSBASE);
1792 elf_hwcap2 |= HWCAP2_FSGSBASE;
1796 * The vendor-specific functions might have changed features.
1797 * Now we do "generic changes."
1800 /* Filter out anything that depends on CPUID levels we don't have */
1801 filter_cpuid_features(c, true);
1803 /* If the model name is still unset, do table lookup. */
1804 if (!c->x86_model_id[0]) {
1806 p = table_lookup_model(c);
1808 strcpy(c->x86_model_id, p);
1810 /* Last resort... */
1811 sprintf(c->x86_model_id, "%02x/%02x",
1812 c->x86, c->x86_model);
1820 * Clear/Set all flags overridden by options, need do it
1821 * before following smp all cpus cap AND.
1823 apply_forced_caps(c);
1826 * On SMP, boot_cpu_data holds the common feature set between
1827 * all CPUs; so make sure that we indicate which features are
1828 * common between the CPUs. The first time this routine gets
1829 * executed, c == &boot_cpu_data.
1831 if (c != &boot_cpu_data) {
1832 /* AND the already accumulated flags with these */
1833 for (i = 0; i < NCAPINTS; i++)
1834 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1836 /* OR, i.e. replicate the bug flags */
1837 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1838 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1843 /* Init Machine Check Exception if available. */
1846 select_idle_routine(c);
1849 numa_add_cpu(smp_processor_id());
1854 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1855 * on 32-bit kernels:
1857 #ifdef CONFIG_X86_32
1858 void enable_sep_cpu(void)
1860 struct tss_struct *tss;
1863 if (!boot_cpu_has(X86_FEATURE_SEP))
1867 tss = &per_cpu(cpu_tss_rw, cpu);
1870 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1871 * see the big comment in struct x86_hw_tss's definition.
1874 tss->x86_tss.ss1 = __KERNEL_CS;
1875 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1876 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
1877 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1883 static __init void identify_boot_cpu(void)
1885 identify_cpu(&boot_cpu_data);
1886 if (HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT))
1887 pr_info("CET detected: Indirect Branch Tracking enabled\n");
1888 #ifdef CONFIG_X86_32
1891 cpu_detect_tlb(&boot_cpu_data);
1899 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1901 BUG_ON(c == &boot_cpu_data);
1903 #ifdef CONFIG_X86_32
1906 x86_spec_ctrl_setup_ap();
1908 if (boot_cpu_has_bug(X86_BUG_GDS))
1914 void print_cpu_info(struct cpuinfo_x86 *c)
1916 const char *vendor = NULL;
1918 if (c->x86_vendor < X86_VENDOR_NUM) {
1919 vendor = this_cpu->c_vendor;
1921 if (c->cpuid_level >= 0)
1922 vendor = c->x86_vendor_id;
1925 if (vendor && !strstr(c->x86_model_id, vendor))
1926 pr_cont("%s ", vendor);
1928 if (c->x86_model_id[0])
1929 pr_cont("%s", c->x86_model_id);
1931 pr_cont("%d86", c->x86);
1933 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1935 if (c->x86_stepping || c->cpuid_level >= 0)
1936 pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
1942 * clearcpuid= was already parsed in cpu_parse_early_param(). This dummy
1943 * function prevents it from becoming an environment variable for init.
1945 static __init int setup_clearcpuid(char *arg)
1949 __setup("clearcpuid=", setup_clearcpuid);
1951 DEFINE_PER_CPU_ALIGNED(struct pcpu_hot, pcpu_hot) = {
1952 .current_task = &init_task,
1953 .preempt_count = INIT_PREEMPT_COUNT,
1954 .top_of_stack = TOP_OF_INIT_STACK,
1956 EXPORT_PER_CPU_SYMBOL(pcpu_hot);
1958 #ifdef CONFIG_X86_64
1959 DEFINE_PER_CPU_FIRST(struct fixed_percpu_data,
1960 fixed_percpu_data) __aligned(PAGE_SIZE) __visible;
1961 EXPORT_PER_CPU_SYMBOL_GPL(fixed_percpu_data);
1963 static void wrmsrl_cstar(unsigned long val)
1966 * Intel CPUs do not support 32-bit SYSCALL. Writing to MSR_CSTAR
1967 * is so far ignored by the CPU, but raises a #VE trap in a TDX
1968 * guest. Avoid the pointless write on all Intel CPUs.
1970 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
1971 wrmsrl(MSR_CSTAR, val);
1974 /* May not be marked __init: used by software suspend */
1975 void syscall_init(void)
1977 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1978 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1980 if (ia32_enabled()) {
1981 wrmsrl_cstar((unsigned long)entry_SYSCALL_compat);
1983 * This only works on Intel CPUs.
1984 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1985 * This does not cause SYSENTER to jump to the wrong location, because
1986 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1988 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1989 wrmsrl_safe(MSR_IA32_SYSENTER_ESP,
1990 (unsigned long)(cpu_entry_stack(smp_processor_id()) + 1));
1991 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1993 wrmsrl_cstar((unsigned long)entry_SYSCALL32_ignore);
1994 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1995 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1996 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
2000 * Flags to clear on syscall; clear as much as possible
2001 * to minimize user space-kernel interference.
2003 wrmsrl(MSR_SYSCALL_MASK,
2004 X86_EFLAGS_CF|X86_EFLAGS_PF|X86_EFLAGS_AF|
2005 X86_EFLAGS_ZF|X86_EFLAGS_SF|X86_EFLAGS_TF|
2006 X86_EFLAGS_IF|X86_EFLAGS_DF|X86_EFLAGS_OF|
2007 X86_EFLAGS_IOPL|X86_EFLAGS_NT|X86_EFLAGS_RF|
2008 X86_EFLAGS_AC|X86_EFLAGS_ID);
2011 #else /* CONFIG_X86_64 */
2013 #ifdef CONFIG_STACKPROTECTOR
2014 DEFINE_PER_CPU(unsigned long, __stack_chk_guard);
2015 EXPORT_PER_CPU_SYMBOL(__stack_chk_guard);
2018 #endif /* CONFIG_X86_64 */
2021 * Clear all 6 debug registers:
2023 static void clear_all_debug_regs(void)
2027 for (i = 0; i < 8; i++) {
2028 /* Ignore db4, db5 */
2029 if ((i == 4) || (i == 5))
2038 * Restore debug regs if using kgdbwait and you have a kernel debugger
2039 * connection established.
2041 static void dbg_restore_debug_regs(void)
2043 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
2044 arch_kgdb_ops.correct_hw_break();
2046 #else /* ! CONFIG_KGDB */
2047 #define dbg_restore_debug_regs()
2048 #endif /* ! CONFIG_KGDB */
2050 static inline void setup_getcpu(int cpu)
2052 unsigned long cpudata = vdso_encode_cpunode(cpu, early_cpu_to_node(cpu));
2053 struct desc_struct d = { };
2055 if (boot_cpu_has(X86_FEATURE_RDTSCP) || boot_cpu_has(X86_FEATURE_RDPID))
2056 wrmsr(MSR_TSC_AUX, cpudata, 0);
2058 /* Store CPU and node number in limit. */
2060 d.limit1 = cpudata >> 16;
2062 d.type = 5; /* RO data, expand down, accessed */
2063 d.dpl = 3; /* Visible to user code */
2064 d.s = 1; /* Not a system segment */
2065 d.p = 1; /* Present */
2066 d.d = 1; /* 32-bit */
2068 write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_CPUNODE, &d, DESCTYPE_S);
2071 #ifdef CONFIG_X86_64
2072 static inline void tss_setup_ist(struct tss_struct *tss)
2074 /* Set up the per-CPU TSS IST stacks */
2075 tss->x86_tss.ist[IST_INDEX_DF] = __this_cpu_ist_top_va(DF);
2076 tss->x86_tss.ist[IST_INDEX_NMI] = __this_cpu_ist_top_va(NMI);
2077 tss->x86_tss.ist[IST_INDEX_DB] = __this_cpu_ist_top_va(DB);
2078 tss->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
2079 /* Only mapped when SEV-ES is active */
2080 tss->x86_tss.ist[IST_INDEX_VC] = __this_cpu_ist_top_va(VC);
2082 #else /* CONFIG_X86_64 */
2083 static inline void tss_setup_ist(struct tss_struct *tss) { }
2084 #endif /* !CONFIG_X86_64 */
2086 static inline void tss_setup_io_bitmap(struct tss_struct *tss)
2088 tss->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET_INVALID;
2090 #ifdef CONFIG_X86_IOPL_IOPERM
2091 tss->io_bitmap.prev_max = 0;
2092 tss->io_bitmap.prev_sequence = 0;
2093 memset(tss->io_bitmap.bitmap, 0xff, sizeof(tss->io_bitmap.bitmap));
2095 * Invalidate the extra array entry past the end of the all
2096 * permission bitmap as required by the hardware.
2098 tss->io_bitmap.mapall[IO_BITMAP_LONGS] = ~0UL;
2103 * Setup everything needed to handle exceptions from the IDT, including the IST
2104 * exceptions which use paranoid_entry().
2106 void cpu_init_exception_handling(void)
2108 struct tss_struct *tss = this_cpu_ptr(&cpu_tss_rw);
2109 int cpu = raw_smp_processor_id();
2111 /* paranoid_entry() gets the CPU number from the GDT */
2114 /* IST vectors need TSS to be set up. */
2116 tss_setup_io_bitmap(tss);
2117 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
2121 /* GHCB needs to be setup to handle #VC. */
2124 /* Finally load the IDT */
2129 * cpu_init() initializes state that is per-CPU. Some data is already
2130 * initialized (naturally) in the bootstrap process, such as the GDT. We
2131 * reload it nevertheless, this function acts as a 'CPU state barrier',
2132 * nothing should get across.
2136 struct task_struct *cur = current;
2137 int cpu = raw_smp_processor_id();
2140 if (this_cpu_read(numa_node) == 0 &&
2141 early_cpu_to_node(cpu) != NUMA_NO_NODE)
2142 set_numa_node(early_cpu_to_node(cpu));
2144 pr_debug("Initializing CPU#%d\n", cpu);
2146 if (IS_ENABLED(CONFIG_X86_64) || cpu_feature_enabled(X86_FEATURE_VME) ||
2147 boot_cpu_has(X86_FEATURE_TSC) || boot_cpu_has(X86_FEATURE_DE))
2148 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
2150 if (IS_ENABLED(CONFIG_X86_64)) {
2152 memset(cur->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
2155 wrmsrl(MSR_FS_BASE, 0);
2156 wrmsrl(MSR_KERNEL_GS_BASE, 0);
2163 cur->active_mm = &init_mm;
2165 initialize_tlbstate_and_flush();
2166 enter_lazy_tlb(&init_mm, cur);
2169 * sp0 points to the entry trampoline stack regardless of what task
2172 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
2174 load_mm_ldt(&init_mm);
2176 clear_all_debug_regs();
2177 dbg_restore_debug_regs();
2179 doublefault_init_cpu_tss();
2184 load_fixmap_gdt(cpu);
2187 #ifdef CONFIG_MICROCODE_LATE_LOADING
2189 * store_cpu_caps() - Store a snapshot of CPU capabilities
2190 * @curr_info: Pointer where to store it
2194 void store_cpu_caps(struct cpuinfo_x86 *curr_info)
2196 /* Reload CPUID max function as it might've changed. */
2197 curr_info->cpuid_level = cpuid_eax(0);
2199 /* Copy all capability leafs and pick up the synthetic ones. */
2200 memcpy(&curr_info->x86_capability, &boot_cpu_data.x86_capability,
2201 sizeof(curr_info->x86_capability));
2203 /* Get the hardware CPUID leafs */
2204 get_cpu_cap(curr_info);
2208 * microcode_check() - Check if any CPU capabilities changed after an update.
2209 * @prev_info: CPU capabilities stored before an update.
2211 * The microcode loader calls this upon late microcode load to recheck features,
2212 * only when microcode has been updated. Caller holds and CPU hotplug lock.
2216 void microcode_check(struct cpuinfo_x86 *prev_info)
2218 struct cpuinfo_x86 curr_info;
2220 perf_check_microcode();
2222 amd_check_microcode();
2224 store_cpu_caps(&curr_info);
2226 if (!memcmp(&prev_info->x86_capability, &curr_info.x86_capability,
2227 sizeof(prev_info->x86_capability)))
2230 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
2231 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");
2236 * Invoked from core CPU hotplug code after hotplug operations
2238 void arch_smt_update(void)
2240 /* Handle the speculative execution misfeatures */
2241 cpu_bugs_smt_update();
2242 /* Check whether IPI broadcasting can be enabled */
2246 void __init arch_cpu_finalize_init(void)
2248 identify_boot_cpu();
2251 * identify_boot_cpu() initialized SMT support information, let the
2254 cpu_smt_set_num_threads(smp_num_siblings, smp_num_siblings);
2256 if (!IS_ENABLED(CONFIG_SMP)) {
2258 print_cpu_info(&boot_cpu_data);
2261 cpu_select_mitigations();
2265 if (IS_ENABLED(CONFIG_X86_32)) {
2267 * Check whether this is a real i386 which is not longer
2268 * supported and fixup the utsname.
2270 if (boot_cpu_data.x86 < 4)
2271 panic("Kernel requires i486+ for 'invlpg' and other features");
2273 init_utsname()->machine[1] =
2274 '0' + (boot_cpu_data.x86 > 6 ? 6 : boot_cpu_data.x86);
2278 * Must be before alternatives because it might set or clear
2284 alternative_instructions();
2286 if (IS_ENABLED(CONFIG_X86_64)) {
2288 * Make sure the first 2MB area is not mapped by huge pages
2289 * There are typically fixed size MTRRs in there and overlapping
2290 * MTRRs into large pages causes slow downs.
2292 * Right now we don't do that with gbpages because there seems
2293 * very little benefit for that case.
2295 if (!direct_gbpages)
2296 set_memory_4k((unsigned long)__va(0), 1);
2298 fpu__init_check_bugs();
2302 * This needs to be called before any devices perform DMA
2303 * operations that might use the SWIOTLB bounce buffers. It will
2304 * mark the bounce buffers as decrypted so that their usage will
2305 * not cause "plain-text" data to be decrypted when accessed. It
2306 * must be called after late_time_init() so that Hyper-V x86/x64
2307 * hypercalls work when the SWIOTLB bounce buffers are decrypted.