Loading...
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>
17#include <linux/io.h>
18#include <linux/syscore_ops.h>
19
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>
36#include <asm/apic.h>
37#include <asm/desc.h>
38#include <asm/fpu/internal.h>
39#include <asm/mtrr.h>
40#include <asm/hwcap2.h>
41#include <linux/numa.h>
42#include <asm/asm.h>
43#include <asm/bugs.h>
44#include <asm/cpu.h>
45#include <asm/mce.h>
46#include <asm/msr.h>
47#include <asm/pat.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>
52
53#ifdef CONFIG_X86_LOCAL_APIC
54#include <asm/uv/uv.h>
55#endif
56
57#include "cpu.h"
58
59u32 elf_hwcap2 __read_mostly;
60
61/* all of these masks are initialized in setup_cpu_local_masks() */
62cpumask_var_t cpu_initialized_mask;
63cpumask_var_t cpu_callout_mask;
64cpumask_var_t cpu_callin_mask;
65
66/* representing cpus for which sibling maps can be computed */
67cpumask_var_t cpu_sibling_setup_mask;
68
69/* correctly size the local cpu masks */
70void __init setup_cpu_local_masks(void)
71{
72 alloc_bootmem_cpumask_var(&cpu_initialized_mask);
73 alloc_bootmem_cpumask_var(&cpu_callin_mask);
74 alloc_bootmem_cpumask_var(&cpu_callout_mask);
75 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
76}
77
78static void default_init(struct cpuinfo_x86 *c)
79{
80#ifdef CONFIG_X86_64
81 cpu_detect_cache_sizes(c);
82#else
83 /* Not much we can do here... */
84 /* Check if at least it has cpuid */
85 if (c->cpuid_level == -1) {
86 /* No cpuid. It must be an ancient CPU */
87 if (c->x86 == 4)
88 strcpy(c->x86_model_id, "486");
89 else if (c->x86 == 3)
90 strcpy(c->x86_model_id, "386");
91 }
92#endif
93}
94
95static const struct cpu_dev default_cpu = {
96 .c_init = default_init,
97 .c_vendor = "Unknown",
98 .c_x86_vendor = X86_VENDOR_UNKNOWN,
99};
100
101static const struct cpu_dev *this_cpu = &default_cpu;
102
103DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
104#ifdef CONFIG_X86_64
105 /*
106 * We need valid kernel segments for data and code in long mode too
107 * IRET will check the segment types kkeil 2000/10/28
108 * Also sysret mandates a special GDT layout
109 *
110 * TLS descriptors are currently at a different place compared to i386.
111 * Hopefully nobody expects them at a fixed place (Wine?)
112 */
113 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
114 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
115 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
116 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
117 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
118 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
119#else
120 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
121 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
122 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
123 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
124 /*
125 * Segments used for calling PnP BIOS have byte granularity.
126 * They code segments and data segments have fixed 64k limits,
127 * the transfer segment sizes are set at run time.
128 */
129 /* 32-bit code */
130 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
131 /* 16-bit code */
132 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
133 /* 16-bit data */
134 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(0x0092, 0, 0xffff),
135 /* 16-bit data */
136 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(0x0092, 0, 0),
137 /* 16-bit data */
138 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(0x0092, 0, 0),
139 /*
140 * The APM segments have byte granularity and their bases
141 * are set at run time. All have 64k limits.
142 */
143 /* 32-bit code */
144 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
145 /* 16-bit code */
146 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
147 /* data */
148 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(0x4092, 0, 0xffff),
149
150 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
151 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
152 GDT_STACK_CANARY_INIT
153#endif
154} };
155EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
156
157static int __init x86_mpx_setup(char *s)
158{
159 /* require an exact match without trailing characters */
160 if (strlen(s))
161 return 0;
162
163 /* do not emit a message if the feature is not present */
164 if (!boot_cpu_has(X86_FEATURE_MPX))
165 return 1;
166
167 setup_clear_cpu_cap(X86_FEATURE_MPX);
168 pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
169 return 1;
170}
171__setup("nompx", x86_mpx_setup);
172
173#ifdef CONFIG_X86_64
174static int __init x86_nopcid_setup(char *s)
175{
176 /* nopcid doesn't accept parameters */
177 if (s)
178 return -EINVAL;
179
180 /* do not emit a message if the feature is not present */
181 if (!boot_cpu_has(X86_FEATURE_PCID))
182 return 0;
183
184 setup_clear_cpu_cap(X86_FEATURE_PCID);
185 pr_info("nopcid: PCID feature disabled\n");
186 return 0;
187}
188early_param("nopcid", x86_nopcid_setup);
189#endif
190
191static int __init x86_noinvpcid_setup(char *s)
192{
193 /* noinvpcid doesn't accept parameters */
194 if (s)
195 return -EINVAL;
196
197 /* do not emit a message if the feature is not present */
198 if (!boot_cpu_has(X86_FEATURE_INVPCID))
199 return 0;
200
201 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
202 pr_info("noinvpcid: INVPCID feature disabled\n");
203 return 0;
204}
205early_param("noinvpcid", x86_noinvpcid_setup);
206
207#ifdef CONFIG_X86_32
208static int cachesize_override = -1;
209static int disable_x86_serial_nr = 1;
210
211static int __init cachesize_setup(char *str)
212{
213 get_option(&str, &cachesize_override);
214 return 1;
215}
216__setup("cachesize=", cachesize_setup);
217
218static int __init x86_sep_setup(char *s)
219{
220 setup_clear_cpu_cap(X86_FEATURE_SEP);
221 return 1;
222}
223__setup("nosep", x86_sep_setup);
224
225/* Standard macro to see if a specific flag is changeable */
226static inline int flag_is_changeable_p(u32 flag)
227{
228 u32 f1, f2;
229
230 /*
231 * Cyrix and IDT cpus allow disabling of CPUID
232 * so the code below may return different results
233 * when it is executed before and after enabling
234 * the CPUID. Add "volatile" to not allow gcc to
235 * optimize the subsequent calls to this function.
236 */
237 asm volatile ("pushfl \n\t"
238 "pushfl \n\t"
239 "popl %0 \n\t"
240 "movl %0, %1 \n\t"
241 "xorl %2, %0 \n\t"
242 "pushl %0 \n\t"
243 "popfl \n\t"
244 "pushfl \n\t"
245 "popl %0 \n\t"
246 "popfl \n\t"
247
248 : "=&r" (f1), "=&r" (f2)
249 : "ir" (flag));
250
251 return ((f1^f2) & flag) != 0;
252}
253
254/* Probe for the CPUID instruction */
255int have_cpuid_p(void)
256{
257 return flag_is_changeable_p(X86_EFLAGS_ID);
258}
259
260static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
261{
262 unsigned long lo, hi;
263
264 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
265 return;
266
267 /* Disable processor serial number: */
268
269 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
270 lo |= 0x200000;
271 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
272
273 pr_notice("CPU serial number disabled.\n");
274 clear_cpu_cap(c, X86_FEATURE_PN);
275
276 /* Disabling the serial number may affect the cpuid level */
277 c->cpuid_level = cpuid_eax(0);
278}
279
280static int __init x86_serial_nr_setup(char *s)
281{
282 disable_x86_serial_nr = 0;
283 return 1;
284}
285__setup("serialnumber", x86_serial_nr_setup);
286#else
287static inline int flag_is_changeable_p(u32 flag)
288{
289 return 1;
290}
291static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
292{
293}
294#endif
295
296static __init int setup_disable_smep(char *arg)
297{
298 setup_clear_cpu_cap(X86_FEATURE_SMEP);
299 /* Check for things that depend on SMEP being enabled: */
300 check_mpx_erratum(&boot_cpu_data);
301 return 1;
302}
303__setup("nosmep", setup_disable_smep);
304
305static __always_inline void setup_smep(struct cpuinfo_x86 *c)
306{
307 if (cpu_has(c, X86_FEATURE_SMEP))
308 cr4_set_bits(X86_CR4_SMEP);
309}
310
311static __init int setup_disable_smap(char *arg)
312{
313 setup_clear_cpu_cap(X86_FEATURE_SMAP);
314 return 1;
315}
316__setup("nosmap", setup_disable_smap);
317
318static __always_inline void setup_smap(struct cpuinfo_x86 *c)
319{
320 unsigned long eflags = native_save_fl();
321
322 /* This should have been cleared long ago */
323 BUG_ON(eflags & X86_EFLAGS_AC);
324
325 if (cpu_has(c, X86_FEATURE_SMAP)) {
326#ifdef CONFIG_X86_SMAP
327 cr4_set_bits(X86_CR4_SMAP);
328#else
329 cr4_clear_bits(X86_CR4_SMAP);
330#endif
331 }
332}
333
334static __always_inline void setup_umip(struct cpuinfo_x86 *c)
335{
336 /* Check the boot processor, plus build option for UMIP. */
337 if (!cpu_feature_enabled(X86_FEATURE_UMIP))
338 goto out;
339
340 /* Check the current processor's cpuid bits. */
341 if (!cpu_has(c, X86_FEATURE_UMIP))
342 goto out;
343
344 cr4_set_bits(X86_CR4_UMIP);
345
346 pr_info("x86/cpu: Activated the Intel User Mode Instruction Prevention (UMIP) CPU feature\n");
347
348 return;
349
350out:
351 /*
352 * Make sure UMIP is disabled in case it was enabled in a
353 * previous boot (e.g., via kexec).
354 */
355 cr4_clear_bits(X86_CR4_UMIP);
356}
357
358/*
359 * Protection Keys are not available in 32-bit mode.
360 */
361static bool pku_disabled;
362
363static __always_inline void setup_pku(struct cpuinfo_x86 *c)
364{
365 /* check the boot processor, plus compile options for PKU: */
366 if (!cpu_feature_enabled(X86_FEATURE_PKU))
367 return;
368 /* checks the actual processor's cpuid bits: */
369 if (!cpu_has(c, X86_FEATURE_PKU))
370 return;
371 if (pku_disabled)
372 return;
373
374 cr4_set_bits(X86_CR4_PKE);
375 /*
376 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
377 * cpuid bit to be set. We need to ensure that we
378 * update that bit in this CPU's "cpu_info".
379 */
380 get_cpu_cap(c);
381}
382
383#ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
384static __init int setup_disable_pku(char *arg)
385{
386 /*
387 * Do not clear the X86_FEATURE_PKU bit. All of the
388 * runtime checks are against OSPKE so clearing the
389 * bit does nothing.
390 *
391 * This way, we will see "pku" in cpuinfo, but not
392 * "ospke", which is exactly what we want. It shows
393 * that the CPU has PKU, but the OS has not enabled it.
394 * This happens to be exactly how a system would look
395 * if we disabled the config option.
396 */
397 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
398 pku_disabled = true;
399 return 1;
400}
401__setup("nopku", setup_disable_pku);
402#endif /* CONFIG_X86_64 */
403
404/*
405 * Some CPU features depend on higher CPUID levels, which may not always
406 * be available due to CPUID level capping or broken virtualization
407 * software. Add those features to this table to auto-disable them.
408 */
409struct cpuid_dependent_feature {
410 u32 feature;
411 u32 level;
412};
413
414static const struct cpuid_dependent_feature
415cpuid_dependent_features[] = {
416 { X86_FEATURE_MWAIT, 0x00000005 },
417 { X86_FEATURE_DCA, 0x00000009 },
418 { X86_FEATURE_XSAVE, 0x0000000d },
419 { 0, 0 }
420};
421
422static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
423{
424 const struct cpuid_dependent_feature *df;
425
426 for (df = cpuid_dependent_features; df->feature; df++) {
427
428 if (!cpu_has(c, df->feature))
429 continue;
430 /*
431 * Note: cpuid_level is set to -1 if unavailable, but
432 * extended_extended_level is set to 0 if unavailable
433 * and the legitimate extended levels are all negative
434 * when signed; hence the weird messing around with
435 * signs here...
436 */
437 if (!((s32)df->level < 0 ?
438 (u32)df->level > (u32)c->extended_cpuid_level :
439 (s32)df->level > (s32)c->cpuid_level))
440 continue;
441
442 clear_cpu_cap(c, df->feature);
443 if (!warn)
444 continue;
445
446 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
447 x86_cap_flag(df->feature), df->level);
448 }
449}
450
451/*
452 * Naming convention should be: <Name> [(<Codename>)]
453 * This table only is used unless init_<vendor>() below doesn't set it;
454 * in particular, if CPUID levels 0x80000002..4 are supported, this
455 * isn't used
456 */
457
458/* Look up CPU names by table lookup. */
459static const char *table_lookup_model(struct cpuinfo_x86 *c)
460{
461#ifdef CONFIG_X86_32
462 const struct legacy_cpu_model_info *info;
463
464 if (c->x86_model >= 16)
465 return NULL; /* Range check */
466
467 if (!this_cpu)
468 return NULL;
469
470 info = this_cpu->legacy_models;
471
472 while (info->family) {
473 if (info->family == c->x86)
474 return info->model_names[c->x86_model];
475 info++;
476 }
477#endif
478 return NULL; /* Not found */
479}
480
481__u32 cpu_caps_cleared[NCAPINTS + NBUGINTS];
482__u32 cpu_caps_set[NCAPINTS + NBUGINTS];
483
484void load_percpu_segment(int cpu)
485{
486#ifdef CONFIG_X86_32
487 loadsegment(fs, __KERNEL_PERCPU);
488#else
489 __loadsegment_simple(gs, 0);
490 wrmsrl(MSR_GS_BASE, cpu_kernelmode_gs_base(cpu));
491#endif
492 load_stack_canary_segment();
493}
494
495#ifdef CONFIG_X86_32
496/* The 32-bit entry code needs to find cpu_entry_area. */
497DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area);
498#endif
499
500#ifdef CONFIG_X86_64
501/*
502 * Special IST stacks which the CPU switches to when it calls
503 * an IST-marked descriptor entry. Up to 7 stacks (hardware
504 * limit), all of them are 4K, except the debug stack which
505 * is 8K.
506 */
507static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
508 [0 ... N_EXCEPTION_STACKS - 1] = EXCEPTION_STKSZ,
509 [DEBUG_STACK - 1] = DEBUG_STKSZ
510};
511#endif
512
513/* Load the original GDT from the per-cpu structure */
514void load_direct_gdt(int cpu)
515{
516 struct desc_ptr gdt_descr;
517
518 gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
519 gdt_descr.size = GDT_SIZE - 1;
520 load_gdt(&gdt_descr);
521}
522EXPORT_SYMBOL_GPL(load_direct_gdt);
523
524/* Load a fixmap remapping of the per-cpu GDT */
525void load_fixmap_gdt(int cpu)
526{
527 struct desc_ptr gdt_descr;
528
529 gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
530 gdt_descr.size = GDT_SIZE - 1;
531 load_gdt(&gdt_descr);
532}
533EXPORT_SYMBOL_GPL(load_fixmap_gdt);
534
535/*
536 * Current gdt points %fs at the "master" per-cpu area: after this,
537 * it's on the real one.
538 */
539void switch_to_new_gdt(int cpu)
540{
541 /* Load the original GDT */
542 load_direct_gdt(cpu);
543 /* Reload the per-cpu base */
544 load_percpu_segment(cpu);
545}
546
547static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
548
549static void get_model_name(struct cpuinfo_x86 *c)
550{
551 unsigned int *v;
552 char *p, *q, *s;
553
554 if (c->extended_cpuid_level < 0x80000004)
555 return;
556
557 v = (unsigned int *)c->x86_model_id;
558 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
559 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
560 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
561 c->x86_model_id[48] = 0;
562
563 /* Trim whitespace */
564 p = q = s = &c->x86_model_id[0];
565
566 while (*p == ' ')
567 p++;
568
569 while (*p) {
570 /* Note the last non-whitespace index */
571 if (!isspace(*p))
572 s = q;
573
574 *q++ = *p++;
575 }
576
577 *(s + 1) = '\0';
578}
579
580void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
581{
582 unsigned int n, dummy, ebx, ecx, edx, l2size;
583
584 n = c->extended_cpuid_level;
585
586 if (n >= 0x80000005) {
587 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
588 c->x86_cache_size = (ecx>>24) + (edx>>24);
589#ifdef CONFIG_X86_64
590 /* On K8 L1 TLB is inclusive, so don't count it */
591 c->x86_tlbsize = 0;
592#endif
593 }
594
595 if (n < 0x80000006) /* Some chips just has a large L1. */
596 return;
597
598 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
599 l2size = ecx >> 16;
600
601#ifdef CONFIG_X86_64
602 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
603#else
604 /* do processor-specific cache resizing */
605 if (this_cpu->legacy_cache_size)
606 l2size = this_cpu->legacy_cache_size(c, l2size);
607
608 /* Allow user to override all this if necessary. */
609 if (cachesize_override != -1)
610 l2size = cachesize_override;
611
612 if (l2size == 0)
613 return; /* Again, no L2 cache is possible */
614#endif
615
616 c->x86_cache_size = l2size;
617}
618
619u16 __read_mostly tlb_lli_4k[NR_INFO];
620u16 __read_mostly tlb_lli_2m[NR_INFO];
621u16 __read_mostly tlb_lli_4m[NR_INFO];
622u16 __read_mostly tlb_lld_4k[NR_INFO];
623u16 __read_mostly tlb_lld_2m[NR_INFO];
624u16 __read_mostly tlb_lld_4m[NR_INFO];
625u16 __read_mostly tlb_lld_1g[NR_INFO];
626
627static void cpu_detect_tlb(struct cpuinfo_x86 *c)
628{
629 if (this_cpu->c_detect_tlb)
630 this_cpu->c_detect_tlb(c);
631
632 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
633 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
634 tlb_lli_4m[ENTRIES]);
635
636 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
637 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
638 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
639}
640
641void detect_ht(struct cpuinfo_x86 *c)
642{
643#ifdef CONFIG_SMP
644 u32 eax, ebx, ecx, edx;
645 int index_msb, core_bits;
646 static bool printed;
647
648 if (!cpu_has(c, X86_FEATURE_HT))
649 return;
650
651 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
652 goto out;
653
654 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
655 return;
656
657 cpuid(1, &eax, &ebx, &ecx, &edx);
658
659 smp_num_siblings = (ebx & 0xff0000) >> 16;
660
661 if (smp_num_siblings == 1) {
662 pr_info_once("CPU0: Hyper-Threading is disabled\n");
663 goto out;
664 }
665
666 if (smp_num_siblings <= 1)
667 goto out;
668
669 index_msb = get_count_order(smp_num_siblings);
670 c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
671
672 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
673
674 index_msb = get_count_order(smp_num_siblings);
675
676 core_bits = get_count_order(c->x86_max_cores);
677
678 c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
679 ((1 << core_bits) - 1);
680
681out:
682 if (!printed && (c->x86_max_cores * smp_num_siblings) > 1) {
683 pr_info("CPU: Physical Processor ID: %d\n",
684 c->phys_proc_id);
685 pr_info("CPU: Processor Core ID: %d\n",
686 c->cpu_core_id);
687 printed = 1;
688 }
689#endif
690}
691
692static void get_cpu_vendor(struct cpuinfo_x86 *c)
693{
694 char *v = c->x86_vendor_id;
695 int i;
696
697 for (i = 0; i < X86_VENDOR_NUM; i++) {
698 if (!cpu_devs[i])
699 break;
700
701 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
702 (cpu_devs[i]->c_ident[1] &&
703 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
704
705 this_cpu = cpu_devs[i];
706 c->x86_vendor = this_cpu->c_x86_vendor;
707 return;
708 }
709 }
710
711 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
712 "CPU: Your system may be unstable.\n", v);
713
714 c->x86_vendor = X86_VENDOR_UNKNOWN;
715 this_cpu = &default_cpu;
716}
717
718void cpu_detect(struct cpuinfo_x86 *c)
719{
720 /* Get vendor name */
721 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
722 (unsigned int *)&c->x86_vendor_id[0],
723 (unsigned int *)&c->x86_vendor_id[8],
724 (unsigned int *)&c->x86_vendor_id[4]);
725
726 c->x86 = 4;
727 /* Intel-defined flags: level 0x00000001 */
728 if (c->cpuid_level >= 0x00000001) {
729 u32 junk, tfms, cap0, misc;
730
731 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
732 c->x86 = x86_family(tfms);
733 c->x86_model = x86_model(tfms);
734 c->x86_stepping = x86_stepping(tfms);
735
736 if (cap0 & (1<<19)) {
737 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
738 c->x86_cache_alignment = c->x86_clflush_size;
739 }
740 }
741}
742
743static void apply_forced_caps(struct cpuinfo_x86 *c)
744{
745 int i;
746
747 for (i = 0; i < NCAPINTS + NBUGINTS; i++) {
748 c->x86_capability[i] &= ~cpu_caps_cleared[i];
749 c->x86_capability[i] |= cpu_caps_set[i];
750 }
751}
752
753static void init_speculation_control(struct cpuinfo_x86 *c)
754{
755 /*
756 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support,
757 * and they also have a different bit for STIBP support. Also,
758 * a hypervisor might have set the individual AMD bits even on
759 * Intel CPUs, for finer-grained selection of what's available.
760 */
761 if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) {
762 set_cpu_cap(c, X86_FEATURE_IBRS);
763 set_cpu_cap(c, X86_FEATURE_IBPB);
764 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
765 }
766
767 if (cpu_has(c, X86_FEATURE_INTEL_STIBP))
768 set_cpu_cap(c, X86_FEATURE_STIBP);
769
770 if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) ||
771 cpu_has(c, X86_FEATURE_VIRT_SSBD))
772 set_cpu_cap(c, X86_FEATURE_SSBD);
773
774 if (cpu_has(c, X86_FEATURE_AMD_IBRS)) {
775 set_cpu_cap(c, X86_FEATURE_IBRS);
776 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
777 }
778
779 if (cpu_has(c, X86_FEATURE_AMD_IBPB))
780 set_cpu_cap(c, X86_FEATURE_IBPB);
781
782 if (cpu_has(c, X86_FEATURE_AMD_STIBP)) {
783 set_cpu_cap(c, X86_FEATURE_STIBP);
784 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
785 }
786}
787
788void get_cpu_cap(struct cpuinfo_x86 *c)
789{
790 u32 eax, ebx, ecx, edx;
791
792 /* Intel-defined flags: level 0x00000001 */
793 if (c->cpuid_level >= 0x00000001) {
794 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
795
796 c->x86_capability[CPUID_1_ECX] = ecx;
797 c->x86_capability[CPUID_1_EDX] = edx;
798 }
799
800 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
801 if (c->cpuid_level >= 0x00000006)
802 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
803
804 /* Additional Intel-defined flags: level 0x00000007 */
805 if (c->cpuid_level >= 0x00000007) {
806 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
807 c->x86_capability[CPUID_7_0_EBX] = ebx;
808 c->x86_capability[CPUID_7_ECX] = ecx;
809 c->x86_capability[CPUID_7_EDX] = edx;
810 }
811
812 /* Extended state features: level 0x0000000d */
813 if (c->cpuid_level >= 0x0000000d) {
814 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
815
816 c->x86_capability[CPUID_D_1_EAX] = eax;
817 }
818
819 /* Additional Intel-defined flags: level 0x0000000F */
820 if (c->cpuid_level >= 0x0000000F) {
821
822 /* QoS sub-leaf, EAX=0Fh, ECX=0 */
823 cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
824 c->x86_capability[CPUID_F_0_EDX] = edx;
825
826 if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
827 /* will be overridden if occupancy monitoring exists */
828 c->x86_cache_max_rmid = ebx;
829
830 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
831 cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
832 c->x86_capability[CPUID_F_1_EDX] = edx;
833
834 if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
835 ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
836 (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
837 c->x86_cache_max_rmid = ecx;
838 c->x86_cache_occ_scale = ebx;
839 }
840 } else {
841 c->x86_cache_max_rmid = -1;
842 c->x86_cache_occ_scale = -1;
843 }
844 }
845
846 /* AMD-defined flags: level 0x80000001 */
847 eax = cpuid_eax(0x80000000);
848 c->extended_cpuid_level = eax;
849
850 if ((eax & 0xffff0000) == 0x80000000) {
851 if (eax >= 0x80000001) {
852 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
853
854 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
855 c->x86_capability[CPUID_8000_0001_EDX] = edx;
856 }
857 }
858
859 if (c->extended_cpuid_level >= 0x80000007) {
860 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
861
862 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
863 c->x86_power = edx;
864 }
865
866 if (c->extended_cpuid_level >= 0x80000008) {
867 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
868 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
869 }
870
871 if (c->extended_cpuid_level >= 0x8000000a)
872 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
873
874 init_scattered_cpuid_features(c);
875 init_speculation_control(c);
876
877 /*
878 * Clear/Set all flags overridden by options, after probe.
879 * This needs to happen each time we re-probe, which may happen
880 * several times during CPU initialization.
881 */
882 apply_forced_caps(c);
883}
884
885static void get_cpu_address_sizes(struct cpuinfo_x86 *c)
886{
887 u32 eax, ebx, ecx, edx;
888
889 if (c->extended_cpuid_level >= 0x80000008) {
890 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
891
892 c->x86_virt_bits = (eax >> 8) & 0xff;
893 c->x86_phys_bits = eax & 0xff;
894 }
895#ifdef CONFIG_X86_32
896 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
897 c->x86_phys_bits = 36;
898#endif
899}
900
901static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
902{
903#ifdef CONFIG_X86_32
904 int i;
905
906 /*
907 * First of all, decide if this is a 486 or higher
908 * It's a 486 if we can modify the AC flag
909 */
910 if (flag_is_changeable_p(X86_EFLAGS_AC))
911 c->x86 = 4;
912 else
913 c->x86 = 3;
914
915 for (i = 0; i < X86_VENDOR_NUM; i++)
916 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
917 c->x86_vendor_id[0] = 0;
918 cpu_devs[i]->c_identify(c);
919 if (c->x86_vendor_id[0]) {
920 get_cpu_vendor(c);
921 break;
922 }
923 }
924#endif
925}
926
927static const __initconst struct x86_cpu_id cpu_no_speculation[] = {
928 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_CEDARVIEW, X86_FEATURE_ANY },
929 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_CLOVERVIEW, X86_FEATURE_ANY },
930 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_LINCROFT, X86_FEATURE_ANY },
931 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_PENWELL, X86_FEATURE_ANY },
932 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_PINEVIEW, X86_FEATURE_ANY },
933 { X86_VENDOR_CENTAUR, 5 },
934 { X86_VENDOR_INTEL, 5 },
935 { X86_VENDOR_NSC, 5 },
936 { X86_VENDOR_ANY, 4 },
937 {}
938};
939
940static const __initconst struct x86_cpu_id cpu_no_meltdown[] = {
941 { X86_VENDOR_AMD },
942 {}
943};
944
945/* Only list CPUs which speculate but are non susceptible to SSB */
946static const __initconst struct x86_cpu_id cpu_no_spec_store_bypass[] = {
947 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT1 },
948 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_AIRMONT },
949 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_SILVERMONT2 },
950 { X86_VENDOR_INTEL, 6, INTEL_FAM6_ATOM_MERRIFIELD },
951 { X86_VENDOR_INTEL, 6, INTEL_FAM6_CORE_YONAH },
952 { X86_VENDOR_INTEL, 6, INTEL_FAM6_XEON_PHI_KNL },
953 { X86_VENDOR_INTEL, 6, INTEL_FAM6_XEON_PHI_KNM },
954 { X86_VENDOR_AMD, 0x12, },
955 { X86_VENDOR_AMD, 0x11, },
956 { X86_VENDOR_AMD, 0x10, },
957 { X86_VENDOR_AMD, 0xf, },
958 {}
959};
960
961static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
962{
963 u64 ia32_cap = 0;
964
965 if (x86_match_cpu(cpu_no_speculation))
966 return;
967
968 setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
969 setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
970
971 if (cpu_has(c, X86_FEATURE_ARCH_CAPABILITIES))
972 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
973
974 if (!x86_match_cpu(cpu_no_spec_store_bypass) &&
975 !(ia32_cap & ARCH_CAP_SSB_NO))
976 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
977
978 if (x86_match_cpu(cpu_no_meltdown))
979 return;
980
981 /* Rogue Data Cache Load? No! */
982 if (ia32_cap & ARCH_CAP_RDCL_NO)
983 return;
984
985 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
986}
987
988/*
989 * Do minimum CPU detection early.
990 * Fields really needed: vendor, cpuid_level, family, model, mask,
991 * cache alignment.
992 * The others are not touched to avoid unwanted side effects.
993 *
994 * WARNING: this function is only called on the boot CPU. Don't add code
995 * here that is supposed to run on all CPUs.
996 */
997static void __init early_identify_cpu(struct cpuinfo_x86 *c)
998{
999#ifdef CONFIG_X86_64
1000 c->x86_clflush_size = 64;
1001 c->x86_phys_bits = 36;
1002 c->x86_virt_bits = 48;
1003#else
1004 c->x86_clflush_size = 32;
1005 c->x86_phys_bits = 32;
1006 c->x86_virt_bits = 32;
1007#endif
1008 c->x86_cache_alignment = c->x86_clflush_size;
1009
1010 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1011 c->extended_cpuid_level = 0;
1012
1013 /* cyrix could have cpuid enabled via c_identify()*/
1014 if (have_cpuid_p()) {
1015 cpu_detect(c);
1016 get_cpu_vendor(c);
1017 get_cpu_cap(c);
1018 get_cpu_address_sizes(c);
1019 setup_force_cpu_cap(X86_FEATURE_CPUID);
1020
1021 if (this_cpu->c_early_init)
1022 this_cpu->c_early_init(c);
1023
1024 c->cpu_index = 0;
1025 filter_cpuid_features(c, false);
1026
1027 if (this_cpu->c_bsp_init)
1028 this_cpu->c_bsp_init(c);
1029 } else {
1030 identify_cpu_without_cpuid(c);
1031 setup_clear_cpu_cap(X86_FEATURE_CPUID);
1032 }
1033
1034 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1035
1036 cpu_set_bug_bits(c);
1037
1038 fpu__init_system(c);
1039
1040#ifdef CONFIG_X86_32
1041 /*
1042 * Regardless of whether PCID is enumerated, the SDM says
1043 * that it can't be enabled in 32-bit mode.
1044 */
1045 setup_clear_cpu_cap(X86_FEATURE_PCID);
1046#endif
1047}
1048
1049void __init early_cpu_init(void)
1050{
1051 const struct cpu_dev *const *cdev;
1052 int count = 0;
1053
1054#ifdef CONFIG_PROCESSOR_SELECT
1055 pr_info("KERNEL supported cpus:\n");
1056#endif
1057
1058 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1059 const struct cpu_dev *cpudev = *cdev;
1060
1061 if (count >= X86_VENDOR_NUM)
1062 break;
1063 cpu_devs[count] = cpudev;
1064 count++;
1065
1066#ifdef CONFIG_PROCESSOR_SELECT
1067 {
1068 unsigned int j;
1069
1070 for (j = 0; j < 2; j++) {
1071 if (!cpudev->c_ident[j])
1072 continue;
1073 pr_info(" %s %s\n", cpudev->c_vendor,
1074 cpudev->c_ident[j]);
1075 }
1076 }
1077#endif
1078 }
1079 early_identify_cpu(&boot_cpu_data);
1080}
1081
1082/*
1083 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1084 * unfortunately, that's not true in practice because of early VIA
1085 * chips and (more importantly) broken virtualizers that are not easy
1086 * to detect. In the latter case it doesn't even *fail* reliably, so
1087 * probing for it doesn't even work. Disable it completely on 32-bit
1088 * unless we can find a reliable way to detect all the broken cases.
1089 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1090 */
1091static void detect_nopl(struct cpuinfo_x86 *c)
1092{
1093#ifdef CONFIG_X86_32
1094 clear_cpu_cap(c, X86_FEATURE_NOPL);
1095#else
1096 set_cpu_cap(c, X86_FEATURE_NOPL);
1097#endif
1098}
1099
1100static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
1101{
1102#ifdef CONFIG_X86_64
1103 /*
1104 * Empirically, writing zero to a segment selector on AMD does
1105 * not clear the base, whereas writing zero to a segment
1106 * selector on Intel does clear the base. Intel's behavior
1107 * allows slightly faster context switches in the common case
1108 * where GS is unused by the prev and next threads.
1109 *
1110 * Since neither vendor documents this anywhere that I can see,
1111 * detect it directly instead of hardcoding the choice by
1112 * vendor.
1113 *
1114 * I've designated AMD's behavior as the "bug" because it's
1115 * counterintuitive and less friendly.
1116 */
1117
1118 unsigned long old_base, tmp;
1119 rdmsrl(MSR_FS_BASE, old_base);
1120 wrmsrl(MSR_FS_BASE, 1);
1121 loadsegment(fs, 0);
1122 rdmsrl(MSR_FS_BASE, tmp);
1123 if (tmp != 0)
1124 set_cpu_bug(c, X86_BUG_NULL_SEG);
1125 wrmsrl(MSR_FS_BASE, old_base);
1126#endif
1127}
1128
1129static void generic_identify(struct cpuinfo_x86 *c)
1130{
1131 c->extended_cpuid_level = 0;
1132
1133 if (!have_cpuid_p())
1134 identify_cpu_without_cpuid(c);
1135
1136 /* cyrix could have cpuid enabled via c_identify()*/
1137 if (!have_cpuid_p())
1138 return;
1139
1140 cpu_detect(c);
1141
1142 get_cpu_vendor(c);
1143
1144 get_cpu_cap(c);
1145
1146 get_cpu_address_sizes(c);
1147
1148 if (c->cpuid_level >= 0x00000001) {
1149 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1150#ifdef CONFIG_X86_32
1151# ifdef CONFIG_SMP
1152 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1153# else
1154 c->apicid = c->initial_apicid;
1155# endif
1156#endif
1157 c->phys_proc_id = c->initial_apicid;
1158 }
1159
1160 get_model_name(c); /* Default name */
1161
1162 detect_nopl(c);
1163
1164 detect_null_seg_behavior(c);
1165
1166 /*
1167 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1168 * systems that run Linux at CPL > 0 may or may not have the
1169 * issue, but, even if they have the issue, there's absolutely
1170 * nothing we can do about it because we can't use the real IRET
1171 * instruction.
1172 *
1173 * NB: For the time being, only 32-bit kernels support
1174 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1175 * whether to apply espfix using paravirt hooks. If any
1176 * non-paravirt system ever shows up that does *not* have the
1177 * ESPFIX issue, we can change this.
1178 */
1179#ifdef CONFIG_X86_32
1180# ifdef CONFIG_PARAVIRT
1181 do {
1182 extern void native_iret(void);
1183 if (pv_cpu_ops.iret == native_iret)
1184 set_cpu_bug(c, X86_BUG_ESPFIX);
1185 } while (0);
1186# else
1187 set_cpu_bug(c, X86_BUG_ESPFIX);
1188# endif
1189#endif
1190}
1191
1192static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1193{
1194 /*
1195 * The heavy lifting of max_rmid and cache_occ_scale are handled
1196 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
1197 * in case CQM bits really aren't there in this CPU.
1198 */
1199 if (c != &boot_cpu_data) {
1200 boot_cpu_data.x86_cache_max_rmid =
1201 min(boot_cpu_data.x86_cache_max_rmid,
1202 c->x86_cache_max_rmid);
1203 }
1204}
1205
1206/*
1207 * Validate that ACPI/mptables have the same information about the
1208 * effective APIC id and update the package map.
1209 */
1210static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1211{
1212#ifdef CONFIG_SMP
1213 unsigned int apicid, cpu = smp_processor_id();
1214
1215 apicid = apic->cpu_present_to_apicid(cpu);
1216
1217 if (apicid != c->apicid) {
1218 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1219 cpu, apicid, c->initial_apicid);
1220 }
1221 BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
1222#else
1223 c->logical_proc_id = 0;
1224#endif
1225}
1226
1227/*
1228 * This does the hard work of actually picking apart the CPU stuff...
1229 */
1230static void identify_cpu(struct cpuinfo_x86 *c)
1231{
1232 int i;
1233
1234 c->loops_per_jiffy = loops_per_jiffy;
1235 c->x86_cache_size = 0;
1236 c->x86_vendor = X86_VENDOR_UNKNOWN;
1237 c->x86_model = c->x86_stepping = 0; /* So far unknown... */
1238 c->x86_vendor_id[0] = '\0'; /* Unset */
1239 c->x86_model_id[0] = '\0'; /* Unset */
1240 c->x86_max_cores = 1;
1241 c->x86_coreid_bits = 0;
1242 c->cu_id = 0xff;
1243#ifdef CONFIG_X86_64
1244 c->x86_clflush_size = 64;
1245 c->x86_phys_bits = 36;
1246 c->x86_virt_bits = 48;
1247#else
1248 c->cpuid_level = -1; /* CPUID not detected */
1249 c->x86_clflush_size = 32;
1250 c->x86_phys_bits = 32;
1251 c->x86_virt_bits = 32;
1252#endif
1253 c->x86_cache_alignment = c->x86_clflush_size;
1254 memset(&c->x86_capability, 0, sizeof c->x86_capability);
1255
1256 generic_identify(c);
1257
1258 if (this_cpu->c_identify)
1259 this_cpu->c_identify(c);
1260
1261 /* Clear/Set all flags overridden by options, after probe */
1262 apply_forced_caps(c);
1263
1264#ifdef CONFIG_X86_64
1265 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1266#endif
1267
1268 /*
1269 * Vendor-specific initialization. In this section we
1270 * canonicalize the feature flags, meaning if there are
1271 * features a certain CPU supports which CPUID doesn't
1272 * tell us, CPUID claiming incorrect flags, or other bugs,
1273 * we handle them here.
1274 *
1275 * At the end of this section, c->x86_capability better
1276 * indicate the features this CPU genuinely supports!
1277 */
1278 if (this_cpu->c_init)
1279 this_cpu->c_init(c);
1280
1281 /* Disable the PN if appropriate */
1282 squash_the_stupid_serial_number(c);
1283
1284 /* Set up SMEP/SMAP/UMIP */
1285 setup_smep(c);
1286 setup_smap(c);
1287 setup_umip(c);
1288
1289 /*
1290 * The vendor-specific functions might have changed features.
1291 * Now we do "generic changes."
1292 */
1293
1294 /* Filter out anything that depends on CPUID levels we don't have */
1295 filter_cpuid_features(c, true);
1296
1297 /* If the model name is still unset, do table lookup. */
1298 if (!c->x86_model_id[0]) {
1299 const char *p;
1300 p = table_lookup_model(c);
1301 if (p)
1302 strcpy(c->x86_model_id, p);
1303 else
1304 /* Last resort... */
1305 sprintf(c->x86_model_id, "%02x/%02x",
1306 c->x86, c->x86_model);
1307 }
1308
1309#ifdef CONFIG_X86_64
1310 detect_ht(c);
1311#endif
1312
1313 x86_init_rdrand(c);
1314 x86_init_cache_qos(c);
1315 setup_pku(c);
1316
1317 /*
1318 * Clear/Set all flags overridden by options, need do it
1319 * before following smp all cpus cap AND.
1320 */
1321 apply_forced_caps(c);
1322
1323 /*
1324 * On SMP, boot_cpu_data holds the common feature set between
1325 * all CPUs; so make sure that we indicate which features are
1326 * common between the CPUs. The first time this routine gets
1327 * executed, c == &boot_cpu_data.
1328 */
1329 if (c != &boot_cpu_data) {
1330 /* AND the already accumulated flags with these */
1331 for (i = 0; i < NCAPINTS; i++)
1332 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1333
1334 /* OR, i.e. replicate the bug flags */
1335 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1336 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1337 }
1338
1339 /* Init Machine Check Exception if available. */
1340 mcheck_cpu_init(c);
1341
1342 select_idle_routine(c);
1343
1344#ifdef CONFIG_NUMA
1345 numa_add_cpu(smp_processor_id());
1346#endif
1347}
1348
1349/*
1350 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1351 * on 32-bit kernels:
1352 */
1353#ifdef CONFIG_X86_32
1354void enable_sep_cpu(void)
1355{
1356 struct tss_struct *tss;
1357 int cpu;
1358
1359 if (!boot_cpu_has(X86_FEATURE_SEP))
1360 return;
1361
1362 cpu = get_cpu();
1363 tss = &per_cpu(cpu_tss_rw, cpu);
1364
1365 /*
1366 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1367 * see the big comment in struct x86_hw_tss's definition.
1368 */
1369
1370 tss->x86_tss.ss1 = __KERNEL_CS;
1371 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1372 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
1373 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1374
1375 put_cpu();
1376}
1377#endif
1378
1379void __init identify_boot_cpu(void)
1380{
1381 identify_cpu(&boot_cpu_data);
1382#ifdef CONFIG_X86_32
1383 sysenter_setup();
1384 enable_sep_cpu();
1385#endif
1386 cpu_detect_tlb(&boot_cpu_data);
1387}
1388
1389void identify_secondary_cpu(struct cpuinfo_x86 *c)
1390{
1391 BUG_ON(c == &boot_cpu_data);
1392 identify_cpu(c);
1393#ifdef CONFIG_X86_32
1394 enable_sep_cpu();
1395#endif
1396 mtrr_ap_init();
1397 validate_apic_and_package_id(c);
1398 x86_spec_ctrl_setup_ap();
1399}
1400
1401static __init int setup_noclflush(char *arg)
1402{
1403 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1404 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1405 return 1;
1406}
1407__setup("noclflush", setup_noclflush);
1408
1409void print_cpu_info(struct cpuinfo_x86 *c)
1410{
1411 const char *vendor = NULL;
1412
1413 if (c->x86_vendor < X86_VENDOR_NUM) {
1414 vendor = this_cpu->c_vendor;
1415 } else {
1416 if (c->cpuid_level >= 0)
1417 vendor = c->x86_vendor_id;
1418 }
1419
1420 if (vendor && !strstr(c->x86_model_id, vendor))
1421 pr_cont("%s ", vendor);
1422
1423 if (c->x86_model_id[0])
1424 pr_cont("%s", c->x86_model_id);
1425 else
1426 pr_cont("%d86", c->x86);
1427
1428 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1429
1430 if (c->x86_stepping || c->cpuid_level >= 0)
1431 pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
1432 else
1433 pr_cont(")\n");
1434}
1435
1436/*
1437 * clearcpuid= was already parsed in fpu__init_parse_early_param.
1438 * But we need to keep a dummy __setup around otherwise it would
1439 * show up as an environment variable for init.
1440 */
1441static __init int setup_clearcpuid(char *arg)
1442{
1443 return 1;
1444}
1445__setup("clearcpuid=", setup_clearcpuid);
1446
1447#ifdef CONFIG_X86_64
1448DEFINE_PER_CPU_FIRST(union irq_stack_union,
1449 irq_stack_union) __aligned(PAGE_SIZE) __visible;
1450EXPORT_PER_CPU_SYMBOL_GPL(irq_stack_union);
1451
1452/*
1453 * The following percpu variables are hot. Align current_task to
1454 * cacheline size such that they fall in the same cacheline.
1455 */
1456DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1457 &init_task;
1458EXPORT_PER_CPU_SYMBOL(current_task);
1459
1460DEFINE_PER_CPU(char *, irq_stack_ptr) =
1461 init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE;
1462
1463DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1464
1465DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1466EXPORT_PER_CPU_SYMBOL(__preempt_count);
1467
1468/* May not be marked __init: used by software suspend */
1469void syscall_init(void)
1470{
1471 extern char _entry_trampoline[];
1472 extern char entry_SYSCALL_64_trampoline[];
1473
1474 int cpu = smp_processor_id();
1475 unsigned long SYSCALL64_entry_trampoline =
1476 (unsigned long)get_cpu_entry_area(cpu)->entry_trampoline +
1477 (entry_SYSCALL_64_trampoline - _entry_trampoline);
1478
1479 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1480 if (static_cpu_has(X86_FEATURE_PTI))
1481 wrmsrl(MSR_LSTAR, SYSCALL64_entry_trampoline);
1482 else
1483 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1484
1485#ifdef CONFIG_IA32_EMULATION
1486 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1487 /*
1488 * This only works on Intel CPUs.
1489 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1490 * This does not cause SYSENTER to jump to the wrong location, because
1491 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1492 */
1493 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1494 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1));
1495 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1496#else
1497 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1498 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1499 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1500 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1501#endif
1502
1503 /* Flags to clear on syscall */
1504 wrmsrl(MSR_SYSCALL_MASK,
1505 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1506 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1507}
1508
1509/*
1510 * Copies of the original ist values from the tss are only accessed during
1511 * debugging, no special alignment required.
1512 */
1513DEFINE_PER_CPU(struct orig_ist, orig_ist);
1514
1515static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1516DEFINE_PER_CPU(int, debug_stack_usage);
1517
1518int is_debug_stack(unsigned long addr)
1519{
1520 return __this_cpu_read(debug_stack_usage) ||
1521 (addr <= __this_cpu_read(debug_stack_addr) &&
1522 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1523}
1524NOKPROBE_SYMBOL(is_debug_stack);
1525
1526DEFINE_PER_CPU(u32, debug_idt_ctr);
1527
1528void debug_stack_set_zero(void)
1529{
1530 this_cpu_inc(debug_idt_ctr);
1531 load_current_idt();
1532}
1533NOKPROBE_SYMBOL(debug_stack_set_zero);
1534
1535void debug_stack_reset(void)
1536{
1537 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1538 return;
1539 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1540 load_current_idt();
1541}
1542NOKPROBE_SYMBOL(debug_stack_reset);
1543
1544#else /* CONFIG_X86_64 */
1545
1546DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1547EXPORT_PER_CPU_SYMBOL(current_task);
1548DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1549EXPORT_PER_CPU_SYMBOL(__preempt_count);
1550
1551/*
1552 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1553 * the top of the kernel stack. Use an extra percpu variable to track the
1554 * top of the kernel stack directly.
1555 */
1556DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1557 (unsigned long)&init_thread_union + THREAD_SIZE;
1558EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1559
1560#ifdef CONFIG_CC_STACKPROTECTOR
1561DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1562#endif
1563
1564#endif /* CONFIG_X86_64 */
1565
1566/*
1567 * Clear all 6 debug registers:
1568 */
1569static void clear_all_debug_regs(void)
1570{
1571 int i;
1572
1573 for (i = 0; i < 8; i++) {
1574 /* Ignore db4, db5 */
1575 if ((i == 4) || (i == 5))
1576 continue;
1577
1578 set_debugreg(0, i);
1579 }
1580}
1581
1582#ifdef CONFIG_KGDB
1583/*
1584 * Restore debug regs if using kgdbwait and you have a kernel debugger
1585 * connection established.
1586 */
1587static void dbg_restore_debug_regs(void)
1588{
1589 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1590 arch_kgdb_ops.correct_hw_break();
1591}
1592#else /* ! CONFIG_KGDB */
1593#define dbg_restore_debug_regs()
1594#endif /* ! CONFIG_KGDB */
1595
1596static void wait_for_master_cpu(int cpu)
1597{
1598#ifdef CONFIG_SMP
1599 /*
1600 * wait for ACK from master CPU before continuing
1601 * with AP initialization
1602 */
1603 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1604 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1605 cpu_relax();
1606#endif
1607}
1608
1609/*
1610 * cpu_init() initializes state that is per-CPU. Some data is already
1611 * initialized (naturally) in the bootstrap process, such as the GDT
1612 * and IDT. We reload them nevertheless, this function acts as a
1613 * 'CPU state barrier', nothing should get across.
1614 * A lot of state is already set up in PDA init for 64 bit
1615 */
1616#ifdef CONFIG_X86_64
1617
1618void cpu_init(void)
1619{
1620 struct orig_ist *oist;
1621 struct task_struct *me;
1622 struct tss_struct *t;
1623 unsigned long v;
1624 int cpu = raw_smp_processor_id();
1625 int i;
1626
1627 wait_for_master_cpu(cpu);
1628
1629 /*
1630 * Initialize the CR4 shadow before doing anything that could
1631 * try to read it.
1632 */
1633 cr4_init_shadow();
1634
1635 if (cpu)
1636 load_ucode_ap();
1637
1638 t = &per_cpu(cpu_tss_rw, cpu);
1639 oist = &per_cpu(orig_ist, cpu);
1640
1641#ifdef CONFIG_NUMA
1642 if (this_cpu_read(numa_node) == 0 &&
1643 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1644 set_numa_node(early_cpu_to_node(cpu));
1645#endif
1646
1647 me = current;
1648
1649 pr_debug("Initializing CPU#%d\n", cpu);
1650
1651 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1652
1653 /*
1654 * Initialize the per-CPU GDT with the boot GDT,
1655 * and set up the GDT descriptor:
1656 */
1657
1658 switch_to_new_gdt(cpu);
1659 loadsegment(fs, 0);
1660
1661 load_current_idt();
1662
1663 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1664 syscall_init();
1665
1666 wrmsrl(MSR_FS_BASE, 0);
1667 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1668 barrier();
1669
1670 x86_configure_nx();
1671 x2apic_setup();
1672
1673 /*
1674 * set up and load the per-CPU TSS
1675 */
1676 if (!oist->ist[0]) {
1677 char *estacks = get_cpu_entry_area(cpu)->exception_stacks;
1678
1679 for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1680 estacks += exception_stack_sizes[v];
1681 oist->ist[v] = t->x86_tss.ist[v] =
1682 (unsigned long)estacks;
1683 if (v == DEBUG_STACK-1)
1684 per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1685 }
1686 }
1687
1688 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1689
1690 /*
1691 * <= is required because the CPU will access up to
1692 * 8 bits beyond the end of the IO permission bitmap.
1693 */
1694 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1695 t->io_bitmap[i] = ~0UL;
1696
1697 mmgrab(&init_mm);
1698 me->active_mm = &init_mm;
1699 BUG_ON(me->mm);
1700 initialize_tlbstate_and_flush();
1701 enter_lazy_tlb(&init_mm, me);
1702
1703 /*
1704 * Initialize the TSS. sp0 points to the entry trampoline stack
1705 * regardless of what task is running.
1706 */
1707 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1708 load_TR_desc();
1709 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1710
1711 load_mm_ldt(&init_mm);
1712
1713 clear_all_debug_regs();
1714 dbg_restore_debug_regs();
1715
1716 fpu__init_cpu();
1717
1718 if (is_uv_system())
1719 uv_cpu_init();
1720
1721 load_fixmap_gdt(cpu);
1722}
1723
1724#else
1725
1726void cpu_init(void)
1727{
1728 int cpu = smp_processor_id();
1729 struct task_struct *curr = current;
1730 struct tss_struct *t = &per_cpu(cpu_tss_rw, cpu);
1731
1732 wait_for_master_cpu(cpu);
1733
1734 /*
1735 * Initialize the CR4 shadow before doing anything that could
1736 * try to read it.
1737 */
1738 cr4_init_shadow();
1739
1740 show_ucode_info_early();
1741
1742 pr_info("Initializing CPU#%d\n", cpu);
1743
1744 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1745 boot_cpu_has(X86_FEATURE_TSC) ||
1746 boot_cpu_has(X86_FEATURE_DE))
1747 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1748
1749 load_current_idt();
1750 switch_to_new_gdt(cpu);
1751
1752 /*
1753 * Set up and load the per-CPU TSS and LDT
1754 */
1755 mmgrab(&init_mm);
1756 curr->active_mm = &init_mm;
1757 BUG_ON(curr->mm);
1758 initialize_tlbstate_and_flush();
1759 enter_lazy_tlb(&init_mm, curr);
1760
1761 /*
1762 * Initialize the TSS. Don't bother initializing sp0, as the initial
1763 * task never enters user mode.
1764 */
1765 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1766 load_TR_desc();
1767
1768 load_mm_ldt(&init_mm);
1769
1770 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1771
1772#ifdef CONFIG_DOUBLEFAULT
1773 /* Set up doublefault TSS pointer in the GDT */
1774 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1775#endif
1776
1777 clear_all_debug_regs();
1778 dbg_restore_debug_regs();
1779
1780 fpu__init_cpu();
1781
1782 load_fixmap_gdt(cpu);
1783}
1784#endif
1785
1786static void bsp_resume(void)
1787{
1788 if (this_cpu->c_bsp_resume)
1789 this_cpu->c_bsp_resume(&boot_cpu_data);
1790}
1791
1792static struct syscore_ops cpu_syscore_ops = {
1793 .resume = bsp_resume,
1794};
1795
1796static int __init init_cpu_syscore(void)
1797{
1798 register_syscore_ops(&cpu_syscore_ops);
1799 return 0;
1800}
1801core_initcall(init_cpu_syscore);
1802
1803/*
1804 * The microcode loader calls this upon late microcode load to recheck features,
1805 * only when microcode has been updated. Caller holds microcode_mutex and CPU
1806 * hotplug lock.
1807 */
1808void microcode_check(void)
1809{
1810 struct cpuinfo_x86 info;
1811
1812 perf_check_microcode();
1813
1814 /* Reload CPUID max function as it might've changed. */
1815 info.cpuid_level = cpuid_eax(0);
1816
1817 /*
1818 * Copy all capability leafs to pick up the synthetic ones so that
1819 * memcmp() below doesn't fail on that. The ones coming from CPUID will
1820 * get overwritten in get_cpu_cap().
1821 */
1822 memcpy(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability));
1823
1824 get_cpu_cap(&info);
1825
1826 if (!memcmp(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability)))
1827 return;
1828
1829 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
1830 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");
1831}
1// SPDX-License-Identifier: GPL-2.0-only
2/* cpu_feature_enabled() cannot be used this early */
3#define USE_EARLY_PGTABLE_L5
4
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>
24#include <linux/io.h>
25#include <linux/syscore_ops.h>
26#include <linux/pgtable.h>
27#include <linux/stackprotector.h>
28#include <linux/utsname.h>
29
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>
47#include <asm/apic.h>
48#include <asm/desc.h>
49#include <asm/fpu/api.h>
50#include <asm/mtrr.h>
51#include <asm/hwcap2.h>
52#include <linux/numa.h>
53#include <asm/numa.h>
54#include <asm/asm.h>
55#include <asm/bugs.h>
56#include <asm/cpu.h>
57#include <asm/mce.h>
58#include <asm/msr.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>
65#include <asm/ia32.h>
66#include <asm/set_memory.h>
67#include <asm/traps.h>
68#include <asm/sev.h>
69#include <asm/tdx.h>
70
71#include "cpu.h"
72
73u32 elf_hwcap2 __read_mostly;
74
75/* Number of siblings per CPU package */
76int smp_num_siblings = 1;
77EXPORT_SYMBOL(smp_num_siblings);
78
79static struct ppin_info {
80 int feature;
81 int msr_ppin_ctl;
82 int msr_ppin;
83} ppin_info[] = {
84 [X86_VENDOR_INTEL] = {
85 .feature = X86_FEATURE_INTEL_PPIN,
86 .msr_ppin_ctl = MSR_PPIN_CTL,
87 .msr_ppin = MSR_PPIN
88 },
89 [X86_VENDOR_AMD] = {
90 .feature = X86_FEATURE_AMD_PPIN,
91 .msr_ppin_ctl = MSR_AMD_PPIN_CTL,
92 .msr_ppin = MSR_AMD_PPIN
93 },
94};
95
96static const struct x86_cpu_id ppin_cpuids[] = {
97 X86_MATCH_FEATURE(X86_FEATURE_AMD_PPIN, &ppin_info[X86_VENDOR_AMD]),
98 X86_MATCH_FEATURE(X86_FEATURE_INTEL_PPIN, &ppin_info[X86_VENDOR_INTEL]),
99
100 /* Legacy models without CPUID enumeration */
101 X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE_X, &ppin_info[X86_VENDOR_INTEL]),
102 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, &ppin_info[X86_VENDOR_INTEL]),
103 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_D, &ppin_info[X86_VENDOR_INTEL]),
104 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, &ppin_info[X86_VENDOR_INTEL]),
105 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_X, &ppin_info[X86_VENDOR_INTEL]),
106 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_X, &ppin_info[X86_VENDOR_INTEL]),
107 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_D, &ppin_info[X86_VENDOR_INTEL]),
108 X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, &ppin_info[X86_VENDOR_INTEL]),
109 X86_MATCH_INTEL_FAM6_MODEL(EMERALDRAPIDS_X, &ppin_info[X86_VENDOR_INTEL]),
110 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNL, &ppin_info[X86_VENDOR_INTEL]),
111 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNM, &ppin_info[X86_VENDOR_INTEL]),
112
113 {}
114};
115
116static void ppin_init(struct cpuinfo_x86 *c)
117{
118 const struct x86_cpu_id *id;
119 unsigned long long val;
120 struct ppin_info *info;
121
122 id = x86_match_cpu(ppin_cpuids);
123 if (!id)
124 return;
125
126 /*
127 * Testing the presence of the MSR is not enough. Need to check
128 * that the PPIN_CTL allows reading of the PPIN.
129 */
130 info = (struct ppin_info *)id->driver_data;
131
132 if (rdmsrl_safe(info->msr_ppin_ctl, &val))
133 goto clear_ppin;
134
135 if ((val & 3UL) == 1UL) {
136 /* PPIN locked in disabled mode */
137 goto clear_ppin;
138 }
139
140 /* If PPIN is disabled, try to enable */
141 if (!(val & 2UL)) {
142 wrmsrl_safe(info->msr_ppin_ctl, val | 2UL);
143 rdmsrl_safe(info->msr_ppin_ctl, &val);
144 }
145
146 /* Is the enable bit set? */
147 if (val & 2UL) {
148 c->ppin = __rdmsr(info->msr_ppin);
149 set_cpu_cap(c, info->feature);
150 return;
151 }
152
153clear_ppin:
154 clear_cpu_cap(c, info->feature);
155}
156
157static void default_init(struct cpuinfo_x86 *c)
158{
159#ifdef CONFIG_X86_64
160 cpu_detect_cache_sizes(c);
161#else
162 /* Not much we can do here... */
163 /* Check if at least it has cpuid */
164 if (c->cpuid_level == -1) {
165 /* No cpuid. It must be an ancient CPU */
166 if (c->x86 == 4)
167 strcpy(c->x86_model_id, "486");
168 else if (c->x86 == 3)
169 strcpy(c->x86_model_id, "386");
170 }
171#endif
172}
173
174static const struct cpu_dev default_cpu = {
175 .c_init = default_init,
176 .c_vendor = "Unknown",
177 .c_x86_vendor = X86_VENDOR_UNKNOWN,
178};
179
180static const struct cpu_dev *this_cpu = &default_cpu;
181
182DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
183#ifdef CONFIG_X86_64
184 /*
185 * We need valid kernel segments for data and code in long mode too
186 * IRET will check the segment types kkeil 2000/10/28
187 * Also sysret mandates a special GDT layout
188 *
189 * TLS descriptors are currently at a different place compared to i386.
190 * Hopefully nobody expects them at a fixed place (Wine?)
191 */
192 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(DESC_CODE32, 0, 0xfffff),
193 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(DESC_CODE64, 0, 0xfffff),
194 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(DESC_DATA64, 0, 0xfffff),
195 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(DESC_CODE32 | DESC_USER, 0, 0xfffff),
196 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(DESC_DATA64 | DESC_USER, 0, 0xfffff),
197 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(DESC_CODE64 | DESC_USER, 0, 0xfffff),
198#else
199 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(DESC_CODE32, 0, 0xfffff),
200 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff),
201 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(DESC_CODE32 | DESC_USER, 0, 0xfffff),
202 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(DESC_DATA32 | DESC_USER, 0, 0xfffff),
203 /*
204 * Segments used for calling PnP BIOS have byte granularity.
205 * They code segments and data segments have fixed 64k limits,
206 * the transfer segment sizes are set at run time.
207 */
208 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(DESC_CODE32_BIOS, 0, 0xffff),
209 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(DESC_CODE16, 0, 0xffff),
210 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0xffff),
211 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0),
212 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0),
213 /*
214 * The APM segments have byte granularity and their bases
215 * are set at run time. All have 64k limits.
216 */
217 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(DESC_CODE32_BIOS, 0, 0xffff),
218 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(DESC_CODE16, 0, 0xffff),
219 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(DESC_DATA32_BIOS, 0, 0xffff),
220
221 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff),
222 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff),
223#endif
224} };
225EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
226
227#ifdef CONFIG_X86_64
228static int __init x86_nopcid_setup(char *s)
229{
230 /* nopcid doesn't accept parameters */
231 if (s)
232 return -EINVAL;
233
234 /* do not emit a message if the feature is not present */
235 if (!boot_cpu_has(X86_FEATURE_PCID))
236 return 0;
237
238 setup_clear_cpu_cap(X86_FEATURE_PCID);
239 pr_info("nopcid: PCID feature disabled\n");
240 return 0;
241}
242early_param("nopcid", x86_nopcid_setup);
243#endif
244
245static int __init x86_noinvpcid_setup(char *s)
246{
247 /* noinvpcid doesn't accept parameters */
248 if (s)
249 return -EINVAL;
250
251 /* do not emit a message if the feature is not present */
252 if (!boot_cpu_has(X86_FEATURE_INVPCID))
253 return 0;
254
255 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
256 pr_info("noinvpcid: INVPCID feature disabled\n");
257 return 0;
258}
259early_param("noinvpcid", x86_noinvpcid_setup);
260
261#ifdef CONFIG_X86_32
262static int cachesize_override = -1;
263static int disable_x86_serial_nr = 1;
264
265static int __init cachesize_setup(char *str)
266{
267 get_option(&str, &cachesize_override);
268 return 1;
269}
270__setup("cachesize=", cachesize_setup);
271
272/* Standard macro to see if a specific flag is changeable */
273static inline int flag_is_changeable_p(u32 flag)
274{
275 u32 f1, f2;
276
277 /*
278 * Cyrix and IDT cpus allow disabling of CPUID
279 * so the code below may return different results
280 * when it is executed before and after enabling
281 * the CPUID. Add "volatile" to not allow gcc to
282 * optimize the subsequent calls to this function.
283 */
284 asm volatile ("pushfl \n\t"
285 "pushfl \n\t"
286 "popl %0 \n\t"
287 "movl %0, %1 \n\t"
288 "xorl %2, %0 \n\t"
289 "pushl %0 \n\t"
290 "popfl \n\t"
291 "pushfl \n\t"
292 "popl %0 \n\t"
293 "popfl \n\t"
294
295 : "=&r" (f1), "=&r" (f2)
296 : "ir" (flag));
297
298 return ((f1^f2) & flag) != 0;
299}
300
301/* Probe for the CPUID instruction */
302int have_cpuid_p(void)
303{
304 return flag_is_changeable_p(X86_EFLAGS_ID);
305}
306
307static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
308{
309 unsigned long lo, hi;
310
311 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
312 return;
313
314 /* Disable processor serial number: */
315
316 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
317 lo |= 0x200000;
318 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
319
320 pr_notice("CPU serial number disabled.\n");
321 clear_cpu_cap(c, X86_FEATURE_PN);
322
323 /* Disabling the serial number may affect the cpuid level */
324 c->cpuid_level = cpuid_eax(0);
325}
326
327static int __init x86_serial_nr_setup(char *s)
328{
329 disable_x86_serial_nr = 0;
330 return 1;
331}
332__setup("serialnumber", x86_serial_nr_setup);
333#else
334static inline int flag_is_changeable_p(u32 flag)
335{
336 return 1;
337}
338static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
339{
340}
341#endif
342
343static __always_inline void setup_smep(struct cpuinfo_x86 *c)
344{
345 if (cpu_has(c, X86_FEATURE_SMEP))
346 cr4_set_bits(X86_CR4_SMEP);
347}
348
349static __always_inline void setup_smap(struct cpuinfo_x86 *c)
350{
351 unsigned long eflags = native_save_fl();
352
353 /* This should have been cleared long ago */
354 BUG_ON(eflags & X86_EFLAGS_AC);
355
356 if (cpu_has(c, X86_FEATURE_SMAP))
357 cr4_set_bits(X86_CR4_SMAP);
358}
359
360static __always_inline void setup_umip(struct cpuinfo_x86 *c)
361{
362 /* Check the boot processor, plus build option for UMIP. */
363 if (!cpu_feature_enabled(X86_FEATURE_UMIP))
364 goto out;
365
366 /* Check the current processor's cpuid bits. */
367 if (!cpu_has(c, X86_FEATURE_UMIP))
368 goto out;
369
370 cr4_set_bits(X86_CR4_UMIP);
371
372 pr_info_once("x86/cpu: User Mode Instruction Prevention (UMIP) activated\n");
373
374 return;
375
376out:
377 /*
378 * Make sure UMIP is disabled in case it was enabled in a
379 * previous boot (e.g., via kexec).
380 */
381 cr4_clear_bits(X86_CR4_UMIP);
382}
383
384/* These bits should not change their value after CPU init is finished. */
385static const unsigned long cr4_pinned_mask =
386 X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_UMIP |
387 X86_CR4_FSGSBASE | X86_CR4_CET;
388static DEFINE_STATIC_KEY_FALSE_RO(cr_pinning);
389static unsigned long cr4_pinned_bits __ro_after_init;
390
391void native_write_cr0(unsigned long val)
392{
393 unsigned long bits_missing = 0;
394
395set_register:
396 asm volatile("mov %0,%%cr0": "+r" (val) : : "memory");
397
398 if (static_branch_likely(&cr_pinning)) {
399 if (unlikely((val & X86_CR0_WP) != X86_CR0_WP)) {
400 bits_missing = X86_CR0_WP;
401 val |= bits_missing;
402 goto set_register;
403 }
404 /* Warn after we've set the missing bits. */
405 WARN_ONCE(bits_missing, "CR0 WP bit went missing!?\n");
406 }
407}
408EXPORT_SYMBOL(native_write_cr0);
409
410void __no_profile native_write_cr4(unsigned long val)
411{
412 unsigned long bits_changed = 0;
413
414set_register:
415 asm volatile("mov %0,%%cr4": "+r" (val) : : "memory");
416
417 if (static_branch_likely(&cr_pinning)) {
418 if (unlikely((val & cr4_pinned_mask) != cr4_pinned_bits)) {
419 bits_changed = (val & cr4_pinned_mask) ^ cr4_pinned_bits;
420 val = (val & ~cr4_pinned_mask) | cr4_pinned_bits;
421 goto set_register;
422 }
423 /* Warn after we've corrected the changed bits. */
424 WARN_ONCE(bits_changed, "pinned CR4 bits changed: 0x%lx!?\n",
425 bits_changed);
426 }
427}
428#if IS_MODULE(CONFIG_LKDTM)
429EXPORT_SYMBOL_GPL(native_write_cr4);
430#endif
431
432void cr4_update_irqsoff(unsigned long set, unsigned long clear)
433{
434 unsigned long newval, cr4 = this_cpu_read(cpu_tlbstate.cr4);
435
436 lockdep_assert_irqs_disabled();
437
438 newval = (cr4 & ~clear) | set;
439 if (newval != cr4) {
440 this_cpu_write(cpu_tlbstate.cr4, newval);
441 __write_cr4(newval);
442 }
443}
444EXPORT_SYMBOL(cr4_update_irqsoff);
445
446/* Read the CR4 shadow. */
447unsigned long cr4_read_shadow(void)
448{
449 return this_cpu_read(cpu_tlbstate.cr4);
450}
451EXPORT_SYMBOL_GPL(cr4_read_shadow);
452
453void cr4_init(void)
454{
455 unsigned long cr4 = __read_cr4();
456
457 if (boot_cpu_has(X86_FEATURE_PCID))
458 cr4 |= X86_CR4_PCIDE;
459 if (static_branch_likely(&cr_pinning))
460 cr4 = (cr4 & ~cr4_pinned_mask) | cr4_pinned_bits;
461
462 __write_cr4(cr4);
463
464 /* Initialize cr4 shadow for this CPU. */
465 this_cpu_write(cpu_tlbstate.cr4, cr4);
466}
467
468/*
469 * Once CPU feature detection is finished (and boot params have been
470 * parsed), record any of the sensitive CR bits that are set, and
471 * enable CR pinning.
472 */
473static void __init setup_cr_pinning(void)
474{
475 cr4_pinned_bits = this_cpu_read(cpu_tlbstate.cr4) & cr4_pinned_mask;
476 static_key_enable(&cr_pinning.key);
477}
478
479static __init int x86_nofsgsbase_setup(char *arg)
480{
481 /* Require an exact match without trailing characters. */
482 if (strlen(arg))
483 return 0;
484
485 /* Do not emit a message if the feature is not present. */
486 if (!boot_cpu_has(X86_FEATURE_FSGSBASE))
487 return 1;
488
489 setup_clear_cpu_cap(X86_FEATURE_FSGSBASE);
490 pr_info("FSGSBASE disabled via kernel command line\n");
491 return 1;
492}
493__setup("nofsgsbase", x86_nofsgsbase_setup);
494
495/*
496 * Protection Keys are not available in 32-bit mode.
497 */
498static bool pku_disabled;
499
500static __always_inline void setup_pku(struct cpuinfo_x86 *c)
501{
502 if (c == &boot_cpu_data) {
503 if (pku_disabled || !cpu_feature_enabled(X86_FEATURE_PKU))
504 return;
505 /*
506 * Setting CR4.PKE will cause the X86_FEATURE_OSPKE cpuid
507 * bit to be set. Enforce it.
508 */
509 setup_force_cpu_cap(X86_FEATURE_OSPKE);
510
511 } else if (!cpu_feature_enabled(X86_FEATURE_OSPKE)) {
512 return;
513 }
514
515 cr4_set_bits(X86_CR4_PKE);
516 /* Load the default PKRU value */
517 pkru_write_default();
518}
519
520#ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
521static __init int setup_disable_pku(char *arg)
522{
523 /*
524 * Do not clear the X86_FEATURE_PKU bit. All of the
525 * runtime checks are against OSPKE so clearing the
526 * bit does nothing.
527 *
528 * This way, we will see "pku" in cpuinfo, but not
529 * "ospke", which is exactly what we want. It shows
530 * that the CPU has PKU, but the OS has not enabled it.
531 * This happens to be exactly how a system would look
532 * if we disabled the config option.
533 */
534 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
535 pku_disabled = true;
536 return 1;
537}
538__setup("nopku", setup_disable_pku);
539#endif
540
541#ifdef CONFIG_X86_KERNEL_IBT
542
543__noendbr u64 ibt_save(bool disable)
544{
545 u64 msr = 0;
546
547 if (cpu_feature_enabled(X86_FEATURE_IBT)) {
548 rdmsrl(MSR_IA32_S_CET, msr);
549 if (disable)
550 wrmsrl(MSR_IA32_S_CET, msr & ~CET_ENDBR_EN);
551 }
552
553 return msr;
554}
555
556__noendbr void ibt_restore(u64 save)
557{
558 u64 msr;
559
560 if (cpu_feature_enabled(X86_FEATURE_IBT)) {
561 rdmsrl(MSR_IA32_S_CET, msr);
562 msr &= ~CET_ENDBR_EN;
563 msr |= (save & CET_ENDBR_EN);
564 wrmsrl(MSR_IA32_S_CET, msr);
565 }
566}
567
568#endif
569
570static __always_inline void setup_cet(struct cpuinfo_x86 *c)
571{
572 bool user_shstk, kernel_ibt;
573
574 if (!IS_ENABLED(CONFIG_X86_CET))
575 return;
576
577 kernel_ibt = HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT);
578 user_shstk = cpu_feature_enabled(X86_FEATURE_SHSTK) &&
579 IS_ENABLED(CONFIG_X86_USER_SHADOW_STACK);
580
581 if (!kernel_ibt && !user_shstk)
582 return;
583
584 if (user_shstk)
585 set_cpu_cap(c, X86_FEATURE_USER_SHSTK);
586
587 if (kernel_ibt)
588 wrmsrl(MSR_IA32_S_CET, CET_ENDBR_EN);
589 else
590 wrmsrl(MSR_IA32_S_CET, 0);
591
592 cr4_set_bits(X86_CR4_CET);
593
594 if (kernel_ibt && ibt_selftest()) {
595 pr_err("IBT selftest: Failed!\n");
596 wrmsrl(MSR_IA32_S_CET, 0);
597 setup_clear_cpu_cap(X86_FEATURE_IBT);
598 }
599}
600
601__noendbr void cet_disable(void)
602{
603 if (!(cpu_feature_enabled(X86_FEATURE_IBT) ||
604 cpu_feature_enabled(X86_FEATURE_SHSTK)))
605 return;
606
607 wrmsrl(MSR_IA32_S_CET, 0);
608 wrmsrl(MSR_IA32_U_CET, 0);
609}
610
611/*
612 * Some CPU features depend on higher CPUID levels, which may not always
613 * be available due to CPUID level capping or broken virtualization
614 * software. Add those features to this table to auto-disable them.
615 */
616struct cpuid_dependent_feature {
617 u32 feature;
618 u32 level;
619};
620
621static const struct cpuid_dependent_feature
622cpuid_dependent_features[] = {
623 { X86_FEATURE_MWAIT, 0x00000005 },
624 { X86_FEATURE_DCA, 0x00000009 },
625 { X86_FEATURE_XSAVE, 0x0000000d },
626 { 0, 0 }
627};
628
629static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
630{
631 const struct cpuid_dependent_feature *df;
632
633 for (df = cpuid_dependent_features; df->feature; df++) {
634
635 if (!cpu_has(c, df->feature))
636 continue;
637 /*
638 * Note: cpuid_level is set to -1 if unavailable, but
639 * extended_extended_level is set to 0 if unavailable
640 * and the legitimate extended levels are all negative
641 * when signed; hence the weird messing around with
642 * signs here...
643 */
644 if (!((s32)df->level < 0 ?
645 (u32)df->level > (u32)c->extended_cpuid_level :
646 (s32)df->level > (s32)c->cpuid_level))
647 continue;
648
649 clear_cpu_cap(c, df->feature);
650 if (!warn)
651 continue;
652
653 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
654 x86_cap_flag(df->feature), df->level);
655 }
656}
657
658/*
659 * Naming convention should be: <Name> [(<Codename>)]
660 * This table only is used unless init_<vendor>() below doesn't set it;
661 * in particular, if CPUID levels 0x80000002..4 are supported, this
662 * isn't used
663 */
664
665/* Look up CPU names by table lookup. */
666static const char *table_lookup_model(struct cpuinfo_x86 *c)
667{
668#ifdef CONFIG_X86_32
669 const struct legacy_cpu_model_info *info;
670
671 if (c->x86_model >= 16)
672 return NULL; /* Range check */
673
674 if (!this_cpu)
675 return NULL;
676
677 info = this_cpu->legacy_models;
678
679 while (info->family) {
680 if (info->family == c->x86)
681 return info->model_names[c->x86_model];
682 info++;
683 }
684#endif
685 return NULL; /* Not found */
686}
687
688/* Aligned to unsigned long to avoid split lock in atomic bitmap ops */
689__u32 cpu_caps_cleared[NCAPINTS + NBUGINTS] __aligned(sizeof(unsigned long));
690__u32 cpu_caps_set[NCAPINTS + NBUGINTS] __aligned(sizeof(unsigned long));
691
692#ifdef CONFIG_X86_32
693/* The 32-bit entry code needs to find cpu_entry_area. */
694DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area);
695#endif
696
697/* Load the original GDT from the per-cpu structure */
698void load_direct_gdt(int cpu)
699{
700 struct desc_ptr gdt_descr;
701
702 gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
703 gdt_descr.size = GDT_SIZE - 1;
704 load_gdt(&gdt_descr);
705}
706EXPORT_SYMBOL_GPL(load_direct_gdt);
707
708/* Load a fixmap remapping of the per-cpu GDT */
709void load_fixmap_gdt(int cpu)
710{
711 struct desc_ptr gdt_descr;
712
713 gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
714 gdt_descr.size = GDT_SIZE - 1;
715 load_gdt(&gdt_descr);
716}
717EXPORT_SYMBOL_GPL(load_fixmap_gdt);
718
719/**
720 * switch_gdt_and_percpu_base - Switch to direct GDT and runtime per CPU base
721 * @cpu: The CPU number for which this is invoked
722 *
723 * Invoked during early boot to switch from early GDT and early per CPU to
724 * the direct GDT and the runtime per CPU area. On 32-bit the percpu base
725 * switch is implicit by loading the direct GDT. On 64bit this requires
726 * to update GSBASE.
727 */
728void __init switch_gdt_and_percpu_base(int cpu)
729{
730 load_direct_gdt(cpu);
731
732#ifdef CONFIG_X86_64
733 /*
734 * No need to load %gs. It is already correct.
735 *
736 * Writing %gs on 64bit would zero GSBASE which would make any per
737 * CPU operation up to the point of the wrmsrl() fault.
738 *
739 * Set GSBASE to the new offset. Until the wrmsrl() happens the
740 * early mapping is still valid. That means the GSBASE update will
741 * lose any prior per CPU data which was not copied over in
742 * setup_per_cpu_areas().
743 *
744 * This works even with stackprotector enabled because the
745 * per CPU stack canary is 0 in both per CPU areas.
746 */
747 wrmsrl(MSR_GS_BASE, cpu_kernelmode_gs_base(cpu));
748#else
749 /*
750 * %fs is already set to __KERNEL_PERCPU, but after switching GDT
751 * it is required to load FS again so that the 'hidden' part is
752 * updated from the new GDT. Up to this point the early per CPU
753 * translation is active. Any content of the early per CPU data
754 * which was not copied over in setup_per_cpu_areas() is lost.
755 */
756 loadsegment(fs, __KERNEL_PERCPU);
757#endif
758}
759
760static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
761
762static void get_model_name(struct cpuinfo_x86 *c)
763{
764 unsigned int *v;
765 char *p, *q, *s;
766
767 if (c->extended_cpuid_level < 0x80000004)
768 return;
769
770 v = (unsigned int *)c->x86_model_id;
771 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
772 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
773 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
774 c->x86_model_id[48] = 0;
775
776 /* Trim whitespace */
777 p = q = s = &c->x86_model_id[0];
778
779 while (*p == ' ')
780 p++;
781
782 while (*p) {
783 /* Note the last non-whitespace index */
784 if (!isspace(*p))
785 s = q;
786
787 *q++ = *p++;
788 }
789
790 *(s + 1) = '\0';
791}
792
793void detect_num_cpu_cores(struct cpuinfo_x86 *c)
794{
795 unsigned int eax, ebx, ecx, edx;
796
797 c->x86_max_cores = 1;
798 if (!IS_ENABLED(CONFIG_SMP) || c->cpuid_level < 4)
799 return;
800
801 cpuid_count(4, 0, &eax, &ebx, &ecx, &edx);
802 if (eax & 0x1f)
803 c->x86_max_cores = (eax >> 26) + 1;
804}
805
806void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
807{
808 unsigned int n, dummy, ebx, ecx, edx, l2size;
809
810 n = c->extended_cpuid_level;
811
812 if (n >= 0x80000005) {
813 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
814 c->x86_cache_size = (ecx>>24) + (edx>>24);
815#ifdef CONFIG_X86_64
816 /* On K8 L1 TLB is inclusive, so don't count it */
817 c->x86_tlbsize = 0;
818#endif
819 }
820
821 if (n < 0x80000006) /* Some chips just has a large L1. */
822 return;
823
824 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
825 l2size = ecx >> 16;
826
827#ifdef CONFIG_X86_64
828 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
829#else
830 /* do processor-specific cache resizing */
831 if (this_cpu->legacy_cache_size)
832 l2size = this_cpu->legacy_cache_size(c, l2size);
833
834 /* Allow user to override all this if necessary. */
835 if (cachesize_override != -1)
836 l2size = cachesize_override;
837
838 if (l2size == 0)
839 return; /* Again, no L2 cache is possible */
840#endif
841
842 c->x86_cache_size = l2size;
843}
844
845u16 __read_mostly tlb_lli_4k[NR_INFO];
846u16 __read_mostly tlb_lli_2m[NR_INFO];
847u16 __read_mostly tlb_lli_4m[NR_INFO];
848u16 __read_mostly tlb_lld_4k[NR_INFO];
849u16 __read_mostly tlb_lld_2m[NR_INFO];
850u16 __read_mostly tlb_lld_4m[NR_INFO];
851u16 __read_mostly tlb_lld_1g[NR_INFO];
852
853static void cpu_detect_tlb(struct cpuinfo_x86 *c)
854{
855 if (this_cpu->c_detect_tlb)
856 this_cpu->c_detect_tlb(c);
857
858 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
859 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
860 tlb_lli_4m[ENTRIES]);
861
862 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
863 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
864 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
865}
866
867int detect_ht_early(struct cpuinfo_x86 *c)
868{
869#ifdef CONFIG_SMP
870 u32 eax, ebx, ecx, edx;
871
872 if (!cpu_has(c, X86_FEATURE_HT))
873 return -1;
874
875 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
876 return -1;
877
878 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
879 return -1;
880
881 cpuid(1, &eax, &ebx, &ecx, &edx);
882
883 smp_num_siblings = (ebx & 0xff0000) >> 16;
884 if (smp_num_siblings == 1)
885 pr_info_once("CPU0: Hyper-Threading is disabled\n");
886#endif
887 return 0;
888}
889
890void detect_ht(struct cpuinfo_x86 *c)
891{
892#ifdef CONFIG_SMP
893 int index_msb, core_bits;
894
895 if (detect_ht_early(c) < 0)
896 return;
897
898 index_msb = get_count_order(smp_num_siblings);
899 c->topo.pkg_id = apic->phys_pkg_id(c->topo.initial_apicid, index_msb);
900
901 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
902
903 index_msb = get_count_order(smp_num_siblings);
904
905 core_bits = get_count_order(c->x86_max_cores);
906
907 c->topo.core_id = apic->phys_pkg_id(c->topo.initial_apicid, index_msb) &
908 ((1 << core_bits) - 1);
909#endif
910}
911
912static void get_cpu_vendor(struct cpuinfo_x86 *c)
913{
914 char *v = c->x86_vendor_id;
915 int i;
916
917 for (i = 0; i < X86_VENDOR_NUM; i++) {
918 if (!cpu_devs[i])
919 break;
920
921 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
922 (cpu_devs[i]->c_ident[1] &&
923 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
924
925 this_cpu = cpu_devs[i];
926 c->x86_vendor = this_cpu->c_x86_vendor;
927 return;
928 }
929 }
930
931 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
932 "CPU: Your system may be unstable.\n", v);
933
934 c->x86_vendor = X86_VENDOR_UNKNOWN;
935 this_cpu = &default_cpu;
936}
937
938void cpu_detect(struct cpuinfo_x86 *c)
939{
940 /* Get vendor name */
941 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
942 (unsigned int *)&c->x86_vendor_id[0],
943 (unsigned int *)&c->x86_vendor_id[8],
944 (unsigned int *)&c->x86_vendor_id[4]);
945
946 c->x86 = 4;
947 /* Intel-defined flags: level 0x00000001 */
948 if (c->cpuid_level >= 0x00000001) {
949 u32 junk, tfms, cap0, misc;
950
951 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
952 c->x86 = x86_family(tfms);
953 c->x86_model = x86_model(tfms);
954 c->x86_stepping = x86_stepping(tfms);
955
956 if (cap0 & (1<<19)) {
957 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
958 c->x86_cache_alignment = c->x86_clflush_size;
959 }
960 }
961}
962
963static void apply_forced_caps(struct cpuinfo_x86 *c)
964{
965 int i;
966
967 for (i = 0; i < NCAPINTS + NBUGINTS; i++) {
968 c->x86_capability[i] &= ~cpu_caps_cleared[i];
969 c->x86_capability[i] |= cpu_caps_set[i];
970 }
971}
972
973static void init_speculation_control(struct cpuinfo_x86 *c)
974{
975 /*
976 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support,
977 * and they also have a different bit for STIBP support. Also,
978 * a hypervisor might have set the individual AMD bits even on
979 * Intel CPUs, for finer-grained selection of what's available.
980 */
981 if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) {
982 set_cpu_cap(c, X86_FEATURE_IBRS);
983 set_cpu_cap(c, X86_FEATURE_IBPB);
984 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
985 }
986
987 if (cpu_has(c, X86_FEATURE_INTEL_STIBP))
988 set_cpu_cap(c, X86_FEATURE_STIBP);
989
990 if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) ||
991 cpu_has(c, X86_FEATURE_VIRT_SSBD))
992 set_cpu_cap(c, X86_FEATURE_SSBD);
993
994 if (cpu_has(c, X86_FEATURE_AMD_IBRS)) {
995 set_cpu_cap(c, X86_FEATURE_IBRS);
996 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
997 }
998
999 if (cpu_has(c, X86_FEATURE_AMD_IBPB))
1000 set_cpu_cap(c, X86_FEATURE_IBPB);
1001
1002 if (cpu_has(c, X86_FEATURE_AMD_STIBP)) {
1003 set_cpu_cap(c, X86_FEATURE_STIBP);
1004 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
1005 }
1006
1007 if (cpu_has(c, X86_FEATURE_AMD_SSBD)) {
1008 set_cpu_cap(c, X86_FEATURE_SSBD);
1009 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
1010 clear_cpu_cap(c, X86_FEATURE_VIRT_SSBD);
1011 }
1012}
1013
1014void get_cpu_cap(struct cpuinfo_x86 *c)
1015{
1016 u32 eax, ebx, ecx, edx;
1017
1018 /* Intel-defined flags: level 0x00000001 */
1019 if (c->cpuid_level >= 0x00000001) {
1020 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
1021
1022 c->x86_capability[CPUID_1_ECX] = ecx;
1023 c->x86_capability[CPUID_1_EDX] = edx;
1024 }
1025
1026 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
1027 if (c->cpuid_level >= 0x00000006)
1028 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
1029
1030 /* Additional Intel-defined flags: level 0x00000007 */
1031 if (c->cpuid_level >= 0x00000007) {
1032 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
1033 c->x86_capability[CPUID_7_0_EBX] = ebx;
1034 c->x86_capability[CPUID_7_ECX] = ecx;
1035 c->x86_capability[CPUID_7_EDX] = edx;
1036
1037 /* Check valid sub-leaf index before accessing it */
1038 if (eax >= 1) {
1039 cpuid_count(0x00000007, 1, &eax, &ebx, &ecx, &edx);
1040 c->x86_capability[CPUID_7_1_EAX] = eax;
1041 }
1042 }
1043
1044 /* Extended state features: level 0x0000000d */
1045 if (c->cpuid_level >= 0x0000000d) {
1046 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
1047
1048 c->x86_capability[CPUID_D_1_EAX] = eax;
1049 }
1050
1051 /* AMD-defined flags: level 0x80000001 */
1052 eax = cpuid_eax(0x80000000);
1053 c->extended_cpuid_level = eax;
1054
1055 if ((eax & 0xffff0000) == 0x80000000) {
1056 if (eax >= 0x80000001) {
1057 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
1058
1059 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
1060 c->x86_capability[CPUID_8000_0001_EDX] = edx;
1061 }
1062 }
1063
1064 if (c->extended_cpuid_level >= 0x80000007) {
1065 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1066
1067 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
1068 c->x86_power = edx;
1069 }
1070
1071 if (c->extended_cpuid_level >= 0x80000008) {
1072 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
1073 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
1074 }
1075
1076 if (c->extended_cpuid_level >= 0x8000000a)
1077 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
1078
1079 if (c->extended_cpuid_level >= 0x8000001f)
1080 c->x86_capability[CPUID_8000_001F_EAX] = cpuid_eax(0x8000001f);
1081
1082 if (c->extended_cpuid_level >= 0x80000021)
1083 c->x86_capability[CPUID_8000_0021_EAX] = cpuid_eax(0x80000021);
1084
1085 init_scattered_cpuid_features(c);
1086 init_speculation_control(c);
1087
1088 /*
1089 * Clear/Set all flags overridden by options, after probe.
1090 * This needs to happen each time we re-probe, which may happen
1091 * several times during CPU initialization.
1092 */
1093 apply_forced_caps(c);
1094}
1095
1096void get_cpu_address_sizes(struct cpuinfo_x86 *c)
1097{
1098 u32 eax, ebx, ecx, edx;
1099 bool vp_bits_from_cpuid = true;
1100
1101 if (!cpu_has(c, X86_FEATURE_CPUID) ||
1102 (c->extended_cpuid_level < 0x80000008))
1103 vp_bits_from_cpuid = false;
1104
1105 if (vp_bits_from_cpuid) {
1106 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
1107
1108 c->x86_virt_bits = (eax >> 8) & 0xff;
1109 c->x86_phys_bits = eax & 0xff;
1110 } else {
1111 if (IS_ENABLED(CONFIG_X86_64)) {
1112 c->x86_clflush_size = 64;
1113 c->x86_phys_bits = 36;
1114 c->x86_virt_bits = 48;
1115 } else {
1116 c->x86_clflush_size = 32;
1117 c->x86_virt_bits = 32;
1118 c->x86_phys_bits = 32;
1119
1120 if (cpu_has(c, X86_FEATURE_PAE) ||
1121 cpu_has(c, X86_FEATURE_PSE36))
1122 c->x86_phys_bits = 36;
1123 }
1124 }
1125 c->x86_cache_bits = c->x86_phys_bits;
1126 c->x86_cache_alignment = c->x86_clflush_size;
1127}
1128
1129static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
1130{
1131#ifdef CONFIG_X86_32
1132 int i;
1133
1134 /*
1135 * First of all, decide if this is a 486 or higher
1136 * It's a 486 if we can modify the AC flag
1137 */
1138 if (flag_is_changeable_p(X86_EFLAGS_AC))
1139 c->x86 = 4;
1140 else
1141 c->x86 = 3;
1142
1143 for (i = 0; i < X86_VENDOR_NUM; i++)
1144 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
1145 c->x86_vendor_id[0] = 0;
1146 cpu_devs[i]->c_identify(c);
1147 if (c->x86_vendor_id[0]) {
1148 get_cpu_vendor(c);
1149 break;
1150 }
1151 }
1152#endif
1153}
1154
1155#define NO_SPECULATION BIT(0)
1156#define NO_MELTDOWN BIT(1)
1157#define NO_SSB BIT(2)
1158#define NO_L1TF BIT(3)
1159#define NO_MDS BIT(4)
1160#define MSBDS_ONLY BIT(5)
1161#define NO_SWAPGS BIT(6)
1162#define NO_ITLB_MULTIHIT BIT(7)
1163#define NO_SPECTRE_V2 BIT(8)
1164#define NO_MMIO BIT(9)
1165#define NO_EIBRS_PBRSB BIT(10)
1166
1167#define VULNWL(vendor, family, model, whitelist) \
1168 X86_MATCH_VENDOR_FAM_MODEL(vendor, family, model, whitelist)
1169
1170#define VULNWL_INTEL(model, whitelist) \
1171 VULNWL(INTEL, 6, INTEL_FAM6_##model, whitelist)
1172
1173#define VULNWL_AMD(family, whitelist) \
1174 VULNWL(AMD, family, X86_MODEL_ANY, whitelist)
1175
1176#define VULNWL_HYGON(family, whitelist) \
1177 VULNWL(HYGON, family, X86_MODEL_ANY, whitelist)
1178
1179static const __initconst struct x86_cpu_id cpu_vuln_whitelist[] = {
1180 VULNWL(ANY, 4, X86_MODEL_ANY, NO_SPECULATION),
1181 VULNWL(CENTAUR, 5, X86_MODEL_ANY, NO_SPECULATION),
1182 VULNWL(INTEL, 5, X86_MODEL_ANY, NO_SPECULATION),
1183 VULNWL(NSC, 5, X86_MODEL_ANY, NO_SPECULATION),
1184 VULNWL(VORTEX, 5, X86_MODEL_ANY, NO_SPECULATION),
1185 VULNWL(VORTEX, 6, X86_MODEL_ANY, NO_SPECULATION),
1186
1187 /* Intel Family 6 */
1188 VULNWL_INTEL(TIGERLAKE, NO_MMIO),
1189 VULNWL_INTEL(TIGERLAKE_L, NO_MMIO),
1190 VULNWL_INTEL(ALDERLAKE, NO_MMIO),
1191 VULNWL_INTEL(ALDERLAKE_L, NO_MMIO),
1192
1193 VULNWL_INTEL(ATOM_SALTWELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
1194 VULNWL_INTEL(ATOM_SALTWELL_TABLET, NO_SPECULATION | NO_ITLB_MULTIHIT),
1195 VULNWL_INTEL(ATOM_SALTWELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
1196 VULNWL_INTEL(ATOM_BONNELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
1197 VULNWL_INTEL(ATOM_BONNELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
1198
1199 VULNWL_INTEL(ATOM_SILVERMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1200 VULNWL_INTEL(ATOM_SILVERMONT_D, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1201 VULNWL_INTEL(ATOM_SILVERMONT_MID, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1202 VULNWL_INTEL(ATOM_AIRMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1203 VULNWL_INTEL(XEON_PHI_KNL, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1204 VULNWL_INTEL(XEON_PHI_KNM, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1205
1206 VULNWL_INTEL(CORE_YONAH, NO_SSB),
1207
1208 VULNWL_INTEL(ATOM_AIRMONT_MID, NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1209 VULNWL_INTEL(ATOM_AIRMONT_NP, NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1210
1211 VULNWL_INTEL(ATOM_GOLDMONT, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1212 VULNWL_INTEL(ATOM_GOLDMONT_D, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1213 VULNWL_INTEL(ATOM_GOLDMONT_PLUS, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB),
1214
1215 /*
1216 * Technically, swapgs isn't serializing on AMD (despite it previously
1217 * being documented as such in the APM). But according to AMD, %gs is
1218 * updated non-speculatively, and the issuing of %gs-relative memory
1219 * operands will be blocked until the %gs update completes, which is
1220 * good enough for our purposes.
1221 */
1222
1223 VULNWL_INTEL(ATOM_TREMONT, NO_EIBRS_PBRSB),
1224 VULNWL_INTEL(ATOM_TREMONT_L, NO_EIBRS_PBRSB),
1225 VULNWL_INTEL(ATOM_TREMONT_D, NO_ITLB_MULTIHIT | NO_EIBRS_PBRSB),
1226
1227 /* AMD Family 0xf - 0x12 */
1228 VULNWL_AMD(0x0f, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1229 VULNWL_AMD(0x10, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1230 VULNWL_AMD(0x11, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1231 VULNWL_AMD(0x12, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO),
1232
1233 /* FAMILY_ANY must be last, otherwise 0x0f - 0x12 matches won't work */
1234 VULNWL_AMD(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB),
1235 VULNWL_HYGON(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB),
1236
1237 /* Zhaoxin Family 7 */
1238 VULNWL(CENTAUR, 7, X86_MODEL_ANY, NO_SPECTRE_V2 | NO_SWAPGS | NO_MMIO),
1239 VULNWL(ZHAOXIN, 7, X86_MODEL_ANY, NO_SPECTRE_V2 | NO_SWAPGS | NO_MMIO),
1240 {}
1241};
1242
1243#define VULNBL(vendor, family, model, blacklist) \
1244 X86_MATCH_VENDOR_FAM_MODEL(vendor, family, model, blacklist)
1245
1246#define VULNBL_INTEL_STEPPINGS(model, steppings, issues) \
1247 X86_MATCH_VENDOR_FAM_MODEL_STEPPINGS_FEATURE(INTEL, 6, \
1248 INTEL_FAM6_##model, steppings, \
1249 X86_FEATURE_ANY, issues)
1250
1251#define VULNBL_AMD(family, blacklist) \
1252 VULNBL(AMD, family, X86_MODEL_ANY, blacklist)
1253
1254#define VULNBL_HYGON(family, blacklist) \
1255 VULNBL(HYGON, family, X86_MODEL_ANY, blacklist)
1256
1257#define SRBDS BIT(0)
1258/* CPU is affected by X86_BUG_MMIO_STALE_DATA */
1259#define MMIO BIT(1)
1260/* CPU is affected by Shared Buffers Data Sampling (SBDS), a variant of X86_BUG_MMIO_STALE_DATA */
1261#define MMIO_SBDS BIT(2)
1262/* CPU is affected by RETbleed, speculating where you would not expect it */
1263#define RETBLEED BIT(3)
1264/* CPU is affected by SMT (cross-thread) return predictions */
1265#define SMT_RSB BIT(4)
1266/* CPU is affected by SRSO */
1267#define SRSO BIT(5)
1268/* CPU is affected by GDS */
1269#define GDS BIT(6)
1270
1271static const struct x86_cpu_id cpu_vuln_blacklist[] __initconst = {
1272 VULNBL_INTEL_STEPPINGS(IVYBRIDGE, X86_STEPPING_ANY, SRBDS),
1273 VULNBL_INTEL_STEPPINGS(HASWELL, X86_STEPPING_ANY, SRBDS),
1274 VULNBL_INTEL_STEPPINGS(HASWELL_L, X86_STEPPING_ANY, SRBDS),
1275 VULNBL_INTEL_STEPPINGS(HASWELL_G, X86_STEPPING_ANY, SRBDS),
1276 VULNBL_INTEL_STEPPINGS(HASWELL_X, X86_STEPPING_ANY, MMIO),
1277 VULNBL_INTEL_STEPPINGS(BROADWELL_D, X86_STEPPING_ANY, MMIO),
1278 VULNBL_INTEL_STEPPINGS(BROADWELL_G, X86_STEPPING_ANY, SRBDS),
1279 VULNBL_INTEL_STEPPINGS(BROADWELL_X, X86_STEPPING_ANY, MMIO),
1280 VULNBL_INTEL_STEPPINGS(BROADWELL, X86_STEPPING_ANY, SRBDS),
1281 VULNBL_INTEL_STEPPINGS(SKYLAKE_X, X86_STEPPING_ANY, MMIO | RETBLEED | GDS),
1282 VULNBL_INTEL_STEPPINGS(SKYLAKE_L, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS),
1283 VULNBL_INTEL_STEPPINGS(SKYLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS),
1284 VULNBL_INTEL_STEPPINGS(KABYLAKE_L, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS),
1285 VULNBL_INTEL_STEPPINGS(KABYLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS),
1286 VULNBL_INTEL_STEPPINGS(CANNONLAKE_L, X86_STEPPING_ANY, RETBLEED),
1287 VULNBL_INTEL_STEPPINGS(ICELAKE_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS),
1288 VULNBL_INTEL_STEPPINGS(ICELAKE_D, X86_STEPPING_ANY, MMIO | GDS),
1289 VULNBL_INTEL_STEPPINGS(ICELAKE_X, X86_STEPPING_ANY, MMIO | GDS),
1290 VULNBL_INTEL_STEPPINGS(COMETLAKE, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS),
1291 VULNBL_INTEL_STEPPINGS(COMETLAKE_L, X86_STEPPINGS(0x0, 0x0), MMIO | RETBLEED),
1292 VULNBL_INTEL_STEPPINGS(COMETLAKE_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS),
1293 VULNBL_INTEL_STEPPINGS(TIGERLAKE_L, X86_STEPPING_ANY, GDS),
1294 VULNBL_INTEL_STEPPINGS(TIGERLAKE, X86_STEPPING_ANY, GDS),
1295 VULNBL_INTEL_STEPPINGS(LAKEFIELD, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED),
1296 VULNBL_INTEL_STEPPINGS(ROCKETLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS),
1297 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT, X86_STEPPING_ANY, MMIO | MMIO_SBDS),
1298 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT_D, X86_STEPPING_ANY, MMIO),
1299 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS),
1300
1301 VULNBL_AMD(0x15, RETBLEED),
1302 VULNBL_AMD(0x16, RETBLEED),
1303 VULNBL_AMD(0x17, RETBLEED | SMT_RSB | SRSO),
1304 VULNBL_HYGON(0x18, RETBLEED | SMT_RSB | SRSO),
1305 VULNBL_AMD(0x19, SRSO),
1306 {}
1307};
1308
1309static bool __init cpu_matches(const struct x86_cpu_id *table, unsigned long which)
1310{
1311 const struct x86_cpu_id *m = x86_match_cpu(table);
1312
1313 return m && !!(m->driver_data & which);
1314}
1315
1316u64 x86_read_arch_cap_msr(void)
1317{
1318 u64 ia32_cap = 0;
1319
1320 if (boot_cpu_has(X86_FEATURE_ARCH_CAPABILITIES))
1321 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
1322
1323 return ia32_cap;
1324}
1325
1326static bool arch_cap_mmio_immune(u64 ia32_cap)
1327{
1328 return (ia32_cap & ARCH_CAP_FBSDP_NO &&
1329 ia32_cap & ARCH_CAP_PSDP_NO &&
1330 ia32_cap & ARCH_CAP_SBDR_SSDP_NO);
1331}
1332
1333static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
1334{
1335 u64 ia32_cap = x86_read_arch_cap_msr();
1336
1337 /* Set ITLB_MULTIHIT bug if cpu is not in the whitelist and not mitigated */
1338 if (!cpu_matches(cpu_vuln_whitelist, NO_ITLB_MULTIHIT) &&
1339 !(ia32_cap & ARCH_CAP_PSCHANGE_MC_NO))
1340 setup_force_cpu_bug(X86_BUG_ITLB_MULTIHIT);
1341
1342 if (cpu_matches(cpu_vuln_whitelist, NO_SPECULATION))
1343 return;
1344
1345 setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
1346
1347 if (!cpu_matches(cpu_vuln_whitelist, NO_SPECTRE_V2))
1348 setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
1349
1350 if (!cpu_matches(cpu_vuln_whitelist, NO_SSB) &&
1351 !(ia32_cap & ARCH_CAP_SSB_NO) &&
1352 !cpu_has(c, X86_FEATURE_AMD_SSB_NO))
1353 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
1354
1355 /*
1356 * AMD's AutoIBRS is equivalent to Intel's eIBRS - use the Intel feature
1357 * flag and protect from vendor-specific bugs via the whitelist.
1358 */
1359 if ((ia32_cap & ARCH_CAP_IBRS_ALL) || cpu_has(c, X86_FEATURE_AUTOIBRS)) {
1360 setup_force_cpu_cap(X86_FEATURE_IBRS_ENHANCED);
1361 if (!cpu_matches(cpu_vuln_whitelist, NO_EIBRS_PBRSB) &&
1362 !(ia32_cap & ARCH_CAP_PBRSB_NO))
1363 setup_force_cpu_bug(X86_BUG_EIBRS_PBRSB);
1364 }
1365
1366 if (!cpu_matches(cpu_vuln_whitelist, NO_MDS) &&
1367 !(ia32_cap & ARCH_CAP_MDS_NO)) {
1368 setup_force_cpu_bug(X86_BUG_MDS);
1369 if (cpu_matches(cpu_vuln_whitelist, MSBDS_ONLY))
1370 setup_force_cpu_bug(X86_BUG_MSBDS_ONLY);
1371 }
1372
1373 if (!cpu_matches(cpu_vuln_whitelist, NO_SWAPGS))
1374 setup_force_cpu_bug(X86_BUG_SWAPGS);
1375
1376 /*
1377 * When the CPU is not mitigated for TAA (TAA_NO=0) set TAA bug when:
1378 * - TSX is supported or
1379 * - TSX_CTRL is present
1380 *
1381 * TSX_CTRL check is needed for cases when TSX could be disabled before
1382 * the kernel boot e.g. kexec.
1383 * TSX_CTRL check alone is not sufficient for cases when the microcode
1384 * update is not present or running as guest that don't get TSX_CTRL.
1385 */
1386 if (!(ia32_cap & ARCH_CAP_TAA_NO) &&
1387 (cpu_has(c, X86_FEATURE_RTM) ||
1388 (ia32_cap & ARCH_CAP_TSX_CTRL_MSR)))
1389 setup_force_cpu_bug(X86_BUG_TAA);
1390
1391 /*
1392 * SRBDS affects CPUs which support RDRAND or RDSEED and are listed
1393 * in the vulnerability blacklist.
1394 *
1395 * Some of the implications and mitigation of Shared Buffers Data
1396 * Sampling (SBDS) are similar to SRBDS. Give SBDS same treatment as
1397 * SRBDS.
1398 */
1399 if ((cpu_has(c, X86_FEATURE_RDRAND) ||
1400 cpu_has(c, X86_FEATURE_RDSEED)) &&
1401 cpu_matches(cpu_vuln_blacklist, SRBDS | MMIO_SBDS))
1402 setup_force_cpu_bug(X86_BUG_SRBDS);
1403
1404 /*
1405 * Processor MMIO Stale Data bug enumeration
1406 *
1407 * Affected CPU list is generally enough to enumerate the vulnerability,
1408 * but for virtualization case check for ARCH_CAP MSR bits also, VMM may
1409 * not want the guest to enumerate the bug.
1410 *
1411 * Set X86_BUG_MMIO_UNKNOWN for CPUs that are neither in the blacklist,
1412 * nor in the whitelist and also don't enumerate MSR ARCH_CAP MMIO bits.
1413 */
1414 if (!arch_cap_mmio_immune(ia32_cap)) {
1415 if (cpu_matches(cpu_vuln_blacklist, MMIO))
1416 setup_force_cpu_bug(X86_BUG_MMIO_STALE_DATA);
1417 else if (!cpu_matches(cpu_vuln_whitelist, NO_MMIO))
1418 setup_force_cpu_bug(X86_BUG_MMIO_UNKNOWN);
1419 }
1420
1421 if (!cpu_has(c, X86_FEATURE_BTC_NO)) {
1422 if (cpu_matches(cpu_vuln_blacklist, RETBLEED) || (ia32_cap & ARCH_CAP_RSBA))
1423 setup_force_cpu_bug(X86_BUG_RETBLEED);
1424 }
1425
1426 if (cpu_matches(cpu_vuln_blacklist, SMT_RSB))
1427 setup_force_cpu_bug(X86_BUG_SMT_RSB);
1428
1429 if (!cpu_has(c, X86_FEATURE_SRSO_NO)) {
1430 if (cpu_matches(cpu_vuln_blacklist, SRSO))
1431 setup_force_cpu_bug(X86_BUG_SRSO);
1432 }
1433
1434 /*
1435 * Check if CPU is vulnerable to GDS. If running in a virtual machine on
1436 * an affected processor, the VMM may have disabled the use of GATHER by
1437 * disabling AVX2. The only way to do this in HW is to clear XCR0[2],
1438 * which means that AVX will be disabled.
1439 */
1440 if (cpu_matches(cpu_vuln_blacklist, GDS) && !(ia32_cap & ARCH_CAP_GDS_NO) &&
1441 boot_cpu_has(X86_FEATURE_AVX))
1442 setup_force_cpu_bug(X86_BUG_GDS);
1443
1444 if (cpu_matches(cpu_vuln_whitelist, NO_MELTDOWN))
1445 return;
1446
1447 /* Rogue Data Cache Load? No! */
1448 if (ia32_cap & ARCH_CAP_RDCL_NO)
1449 return;
1450
1451 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
1452
1453 if (cpu_matches(cpu_vuln_whitelist, NO_L1TF))
1454 return;
1455
1456 setup_force_cpu_bug(X86_BUG_L1TF);
1457}
1458
1459/*
1460 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1461 * unfortunately, that's not true in practice because of early VIA
1462 * chips and (more importantly) broken virtualizers that are not easy
1463 * to detect. In the latter case it doesn't even *fail* reliably, so
1464 * probing for it doesn't even work. Disable it completely on 32-bit
1465 * unless we can find a reliable way to detect all the broken cases.
1466 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1467 */
1468static void detect_nopl(void)
1469{
1470#ifdef CONFIG_X86_32
1471 setup_clear_cpu_cap(X86_FEATURE_NOPL);
1472#else
1473 setup_force_cpu_cap(X86_FEATURE_NOPL);
1474#endif
1475}
1476
1477/*
1478 * We parse cpu parameters early because fpu__init_system() is executed
1479 * before parse_early_param().
1480 */
1481static void __init cpu_parse_early_param(void)
1482{
1483 char arg[128];
1484 char *argptr = arg, *opt;
1485 int arglen, taint = 0;
1486
1487#ifdef CONFIG_X86_32
1488 if (cmdline_find_option_bool(boot_command_line, "no387"))
1489#ifdef CONFIG_MATH_EMULATION
1490 setup_clear_cpu_cap(X86_FEATURE_FPU);
1491#else
1492 pr_err("Option 'no387' required CONFIG_MATH_EMULATION enabled.\n");
1493#endif
1494
1495 if (cmdline_find_option_bool(boot_command_line, "nofxsr"))
1496 setup_clear_cpu_cap(X86_FEATURE_FXSR);
1497#endif
1498
1499 if (cmdline_find_option_bool(boot_command_line, "noxsave"))
1500 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
1501
1502 if (cmdline_find_option_bool(boot_command_line, "noxsaveopt"))
1503 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
1504
1505 if (cmdline_find_option_bool(boot_command_line, "noxsaves"))
1506 setup_clear_cpu_cap(X86_FEATURE_XSAVES);
1507
1508 if (cmdline_find_option_bool(boot_command_line, "nousershstk"))
1509 setup_clear_cpu_cap(X86_FEATURE_USER_SHSTK);
1510
1511 arglen = cmdline_find_option(boot_command_line, "clearcpuid", arg, sizeof(arg));
1512 if (arglen <= 0)
1513 return;
1514
1515 pr_info("Clearing CPUID bits:");
1516
1517 while (argptr) {
1518 bool found __maybe_unused = false;
1519 unsigned int bit;
1520
1521 opt = strsep(&argptr, ",");
1522
1523 /*
1524 * Handle naked numbers first for feature flags which don't
1525 * have names.
1526 */
1527 if (!kstrtouint(opt, 10, &bit)) {
1528 if (bit < NCAPINTS * 32) {
1529
1530 /* empty-string, i.e., ""-defined feature flags */
1531 if (!x86_cap_flags[bit])
1532 pr_cont(" " X86_CAP_FMT_NUM, x86_cap_flag_num(bit));
1533 else
1534 pr_cont(" " X86_CAP_FMT, x86_cap_flag(bit));
1535
1536 setup_clear_cpu_cap(bit);
1537 taint++;
1538 }
1539 /*
1540 * The assumption is that there are no feature names with only
1541 * numbers in the name thus go to the next argument.
1542 */
1543 continue;
1544 }
1545
1546 for (bit = 0; bit < 32 * NCAPINTS; bit++) {
1547 if (!x86_cap_flag(bit))
1548 continue;
1549
1550 if (strcmp(x86_cap_flag(bit), opt))
1551 continue;
1552
1553 pr_cont(" %s", opt);
1554 setup_clear_cpu_cap(bit);
1555 taint++;
1556 found = true;
1557 break;
1558 }
1559
1560 if (!found)
1561 pr_cont(" (unknown: %s)", opt);
1562 }
1563 pr_cont("\n");
1564
1565 if (taint)
1566 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
1567}
1568
1569/*
1570 * Do minimum CPU detection early.
1571 * Fields really needed: vendor, cpuid_level, family, model, mask,
1572 * cache alignment.
1573 * The others are not touched to avoid unwanted side effects.
1574 *
1575 * WARNING: this function is only called on the boot CPU. Don't add code
1576 * here that is supposed to run on all CPUs.
1577 */
1578static void __init early_identify_cpu(struct cpuinfo_x86 *c)
1579{
1580 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1581 c->extended_cpuid_level = 0;
1582
1583 if (!have_cpuid_p())
1584 identify_cpu_without_cpuid(c);
1585
1586 /* cyrix could have cpuid enabled via c_identify()*/
1587 if (have_cpuid_p()) {
1588 cpu_detect(c);
1589 get_cpu_vendor(c);
1590 get_cpu_cap(c);
1591 setup_force_cpu_cap(X86_FEATURE_CPUID);
1592 get_cpu_address_sizes(c);
1593 cpu_parse_early_param();
1594
1595 if (this_cpu->c_early_init)
1596 this_cpu->c_early_init(c);
1597
1598 c->cpu_index = 0;
1599 filter_cpuid_features(c, false);
1600
1601 if (this_cpu->c_bsp_init)
1602 this_cpu->c_bsp_init(c);
1603 } else {
1604 setup_clear_cpu_cap(X86_FEATURE_CPUID);
1605 get_cpu_address_sizes(c);
1606 }
1607
1608 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1609
1610 cpu_set_bug_bits(c);
1611
1612 sld_setup(c);
1613
1614#ifdef CONFIG_X86_32
1615 /*
1616 * Regardless of whether PCID is enumerated, the SDM says
1617 * that it can't be enabled in 32-bit mode.
1618 */
1619 setup_clear_cpu_cap(X86_FEATURE_PCID);
1620#endif
1621
1622 /*
1623 * Later in the boot process pgtable_l5_enabled() relies on
1624 * cpu_feature_enabled(X86_FEATURE_LA57). If 5-level paging is not
1625 * enabled by this point we need to clear the feature bit to avoid
1626 * false-positives at the later stage.
1627 *
1628 * pgtable_l5_enabled() can be false here for several reasons:
1629 * - 5-level paging is disabled compile-time;
1630 * - it's 32-bit kernel;
1631 * - machine doesn't support 5-level paging;
1632 * - user specified 'no5lvl' in kernel command line.
1633 */
1634 if (!pgtable_l5_enabled())
1635 setup_clear_cpu_cap(X86_FEATURE_LA57);
1636
1637 detect_nopl();
1638}
1639
1640void __init early_cpu_init(void)
1641{
1642 const struct cpu_dev *const *cdev;
1643 int count = 0;
1644
1645#ifdef CONFIG_PROCESSOR_SELECT
1646 pr_info("KERNEL supported cpus:\n");
1647#endif
1648
1649 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1650 const struct cpu_dev *cpudev = *cdev;
1651
1652 if (count >= X86_VENDOR_NUM)
1653 break;
1654 cpu_devs[count] = cpudev;
1655 count++;
1656
1657#ifdef CONFIG_PROCESSOR_SELECT
1658 {
1659 unsigned int j;
1660
1661 for (j = 0; j < 2; j++) {
1662 if (!cpudev->c_ident[j])
1663 continue;
1664 pr_info(" %s %s\n", cpudev->c_vendor,
1665 cpudev->c_ident[j]);
1666 }
1667 }
1668#endif
1669 }
1670 early_identify_cpu(&boot_cpu_data);
1671}
1672
1673static bool detect_null_seg_behavior(void)
1674{
1675 /*
1676 * Empirically, writing zero to a segment selector on AMD does
1677 * not clear the base, whereas writing zero to a segment
1678 * selector on Intel does clear the base. Intel's behavior
1679 * allows slightly faster context switches in the common case
1680 * where GS is unused by the prev and next threads.
1681 *
1682 * Since neither vendor documents this anywhere that I can see,
1683 * detect it directly instead of hard-coding the choice by
1684 * vendor.
1685 *
1686 * I've designated AMD's behavior as the "bug" because it's
1687 * counterintuitive and less friendly.
1688 */
1689
1690 unsigned long old_base, tmp;
1691 rdmsrl(MSR_FS_BASE, old_base);
1692 wrmsrl(MSR_FS_BASE, 1);
1693 loadsegment(fs, 0);
1694 rdmsrl(MSR_FS_BASE, tmp);
1695 wrmsrl(MSR_FS_BASE, old_base);
1696 return tmp == 0;
1697}
1698
1699void check_null_seg_clears_base(struct cpuinfo_x86 *c)
1700{
1701 /* BUG_NULL_SEG is only relevant with 64bit userspace */
1702 if (!IS_ENABLED(CONFIG_X86_64))
1703 return;
1704
1705 if (cpu_has(c, X86_FEATURE_NULL_SEL_CLR_BASE))
1706 return;
1707
1708 /*
1709 * CPUID bit above wasn't set. If this kernel is still running
1710 * as a HV guest, then the HV has decided not to advertize
1711 * that CPUID bit for whatever reason. For example, one
1712 * member of the migration pool might be vulnerable. Which
1713 * means, the bug is present: set the BUG flag and return.
1714 */
1715 if (cpu_has(c, X86_FEATURE_HYPERVISOR)) {
1716 set_cpu_bug(c, X86_BUG_NULL_SEG);
1717 return;
1718 }
1719
1720 /*
1721 * Zen2 CPUs also have this behaviour, but no CPUID bit.
1722 * 0x18 is the respective family for Hygon.
1723 */
1724 if ((c->x86 == 0x17 || c->x86 == 0x18) &&
1725 detect_null_seg_behavior())
1726 return;
1727
1728 /* All the remaining ones are affected */
1729 set_cpu_bug(c, X86_BUG_NULL_SEG);
1730}
1731
1732static void generic_identify(struct cpuinfo_x86 *c)
1733{
1734 c->extended_cpuid_level = 0;
1735
1736 if (!have_cpuid_p())
1737 identify_cpu_without_cpuid(c);
1738
1739 /* cyrix could have cpuid enabled via c_identify()*/
1740 if (!have_cpuid_p())
1741 return;
1742
1743 cpu_detect(c);
1744
1745 get_cpu_vendor(c);
1746
1747 get_cpu_cap(c);
1748
1749 get_cpu_address_sizes(c);
1750
1751 if (c->cpuid_level >= 0x00000001) {
1752 c->topo.initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1753#ifdef CONFIG_X86_32
1754# ifdef CONFIG_SMP
1755 c->topo.apicid = apic->phys_pkg_id(c->topo.initial_apicid, 0);
1756# else
1757 c->topo.apicid = c->topo.initial_apicid;
1758# endif
1759#endif
1760 c->topo.pkg_id = c->topo.initial_apicid;
1761 }
1762
1763 get_model_name(c); /* Default name */
1764
1765 /*
1766 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1767 * systems that run Linux at CPL > 0 may or may not have the
1768 * issue, but, even if they have the issue, there's absolutely
1769 * nothing we can do about it because we can't use the real IRET
1770 * instruction.
1771 *
1772 * NB: For the time being, only 32-bit kernels support
1773 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1774 * whether to apply espfix using paravirt hooks. If any
1775 * non-paravirt system ever shows up that does *not* have the
1776 * ESPFIX issue, we can change this.
1777 */
1778#ifdef CONFIG_X86_32
1779 set_cpu_bug(c, X86_BUG_ESPFIX);
1780#endif
1781}
1782
1783/*
1784 * Validate that ACPI/mptables have the same information about the
1785 * effective APIC id and update the package map.
1786 */
1787static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1788{
1789#ifdef CONFIG_SMP
1790 unsigned int cpu = smp_processor_id();
1791 u32 apicid;
1792
1793 apicid = apic->cpu_present_to_apicid(cpu);
1794
1795 if (apicid != c->topo.apicid) {
1796 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1797 cpu, apicid, c->topo.initial_apicid);
1798 }
1799 BUG_ON(topology_update_package_map(c->topo.pkg_id, cpu));
1800 BUG_ON(topology_update_die_map(c->topo.die_id, cpu));
1801#else
1802 c->topo.logical_pkg_id = 0;
1803#endif
1804}
1805
1806/*
1807 * This does the hard work of actually picking apart the CPU stuff...
1808 */
1809static void identify_cpu(struct cpuinfo_x86 *c)
1810{
1811 int i;
1812
1813 c->loops_per_jiffy = loops_per_jiffy;
1814 c->x86_cache_size = 0;
1815 c->x86_vendor = X86_VENDOR_UNKNOWN;
1816 c->x86_model = c->x86_stepping = 0; /* So far unknown... */
1817 c->x86_vendor_id[0] = '\0'; /* Unset */
1818 c->x86_model_id[0] = '\0'; /* Unset */
1819 c->x86_max_cores = 1;
1820 c->x86_coreid_bits = 0;
1821 c->topo.cu_id = 0xff;
1822 c->topo.llc_id = BAD_APICID;
1823 c->topo.l2c_id = BAD_APICID;
1824#ifdef CONFIG_X86_64
1825 c->x86_clflush_size = 64;
1826 c->x86_phys_bits = 36;
1827 c->x86_virt_bits = 48;
1828#else
1829 c->cpuid_level = -1; /* CPUID not detected */
1830 c->x86_clflush_size = 32;
1831 c->x86_phys_bits = 32;
1832 c->x86_virt_bits = 32;
1833#endif
1834 c->x86_cache_alignment = c->x86_clflush_size;
1835 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1836#ifdef CONFIG_X86_VMX_FEATURE_NAMES
1837 memset(&c->vmx_capability, 0, sizeof(c->vmx_capability));
1838#endif
1839
1840 generic_identify(c);
1841
1842 if (this_cpu->c_identify)
1843 this_cpu->c_identify(c);
1844
1845 /* Clear/Set all flags overridden by options, after probe */
1846 apply_forced_caps(c);
1847
1848#ifdef CONFIG_X86_64
1849 c->topo.apicid = apic->phys_pkg_id(c->topo.initial_apicid, 0);
1850#endif
1851
1852
1853 /*
1854 * Set default APIC and TSC_DEADLINE MSR fencing flag. AMD and
1855 * Hygon will clear it in ->c_init() below.
1856 */
1857 set_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE);
1858
1859 /*
1860 * Vendor-specific initialization. In this section we
1861 * canonicalize the feature flags, meaning if there are
1862 * features a certain CPU supports which CPUID doesn't
1863 * tell us, CPUID claiming incorrect flags, or other bugs,
1864 * we handle them here.
1865 *
1866 * At the end of this section, c->x86_capability better
1867 * indicate the features this CPU genuinely supports!
1868 */
1869 if (this_cpu->c_init)
1870 this_cpu->c_init(c);
1871
1872 /* Disable the PN if appropriate */
1873 squash_the_stupid_serial_number(c);
1874
1875 /* Set up SMEP/SMAP/UMIP */
1876 setup_smep(c);
1877 setup_smap(c);
1878 setup_umip(c);
1879
1880 /* Enable FSGSBASE instructions if available. */
1881 if (cpu_has(c, X86_FEATURE_FSGSBASE)) {
1882 cr4_set_bits(X86_CR4_FSGSBASE);
1883 elf_hwcap2 |= HWCAP2_FSGSBASE;
1884 }
1885
1886 /*
1887 * The vendor-specific functions might have changed features.
1888 * Now we do "generic changes."
1889 */
1890
1891 /* Filter out anything that depends on CPUID levels we don't have */
1892 filter_cpuid_features(c, true);
1893
1894 /* If the model name is still unset, do table lookup. */
1895 if (!c->x86_model_id[0]) {
1896 const char *p;
1897 p = table_lookup_model(c);
1898 if (p)
1899 strcpy(c->x86_model_id, p);
1900 else
1901 /* Last resort... */
1902 sprintf(c->x86_model_id, "%02x/%02x",
1903 c->x86, c->x86_model);
1904 }
1905
1906#ifdef CONFIG_X86_64
1907 detect_ht(c);
1908#endif
1909
1910 x86_init_rdrand(c);
1911 setup_pku(c);
1912 setup_cet(c);
1913
1914 /*
1915 * Clear/Set all flags overridden by options, need do it
1916 * before following smp all cpus cap AND.
1917 */
1918 apply_forced_caps(c);
1919
1920 /*
1921 * On SMP, boot_cpu_data holds the common feature set between
1922 * all CPUs; so make sure that we indicate which features are
1923 * common between the CPUs. The first time this routine gets
1924 * executed, c == &boot_cpu_data.
1925 */
1926 if (c != &boot_cpu_data) {
1927 /* AND the already accumulated flags with these */
1928 for (i = 0; i < NCAPINTS; i++)
1929 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1930
1931 /* OR, i.e. replicate the bug flags */
1932 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1933 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1934 }
1935
1936 ppin_init(c);
1937
1938 /* Init Machine Check Exception if available. */
1939 mcheck_cpu_init(c);
1940
1941 select_idle_routine(c);
1942
1943#ifdef CONFIG_NUMA
1944 numa_add_cpu(smp_processor_id());
1945#endif
1946}
1947
1948/*
1949 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1950 * on 32-bit kernels:
1951 */
1952#ifdef CONFIG_X86_32
1953void enable_sep_cpu(void)
1954{
1955 struct tss_struct *tss;
1956 int cpu;
1957
1958 if (!boot_cpu_has(X86_FEATURE_SEP))
1959 return;
1960
1961 cpu = get_cpu();
1962 tss = &per_cpu(cpu_tss_rw, cpu);
1963
1964 /*
1965 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1966 * see the big comment in struct x86_hw_tss's definition.
1967 */
1968
1969 tss->x86_tss.ss1 = __KERNEL_CS;
1970 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1971 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
1972 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1973
1974 put_cpu();
1975}
1976#endif
1977
1978static __init void identify_boot_cpu(void)
1979{
1980 identify_cpu(&boot_cpu_data);
1981 if (HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT))
1982 pr_info("CET detected: Indirect Branch Tracking enabled\n");
1983#ifdef CONFIG_X86_32
1984 enable_sep_cpu();
1985#endif
1986 cpu_detect_tlb(&boot_cpu_data);
1987 setup_cr_pinning();
1988
1989 tsx_init();
1990 tdx_init();
1991 lkgs_init();
1992}
1993
1994void identify_secondary_cpu(struct cpuinfo_x86 *c)
1995{
1996 BUG_ON(c == &boot_cpu_data);
1997 identify_cpu(c);
1998#ifdef CONFIG_X86_32
1999 enable_sep_cpu();
2000#endif
2001 validate_apic_and_package_id(c);
2002 x86_spec_ctrl_setup_ap();
2003 update_srbds_msr();
2004 if (boot_cpu_has_bug(X86_BUG_GDS))
2005 update_gds_msr();
2006
2007 tsx_ap_init();
2008}
2009
2010void print_cpu_info(struct cpuinfo_x86 *c)
2011{
2012 const char *vendor = NULL;
2013
2014 if (c->x86_vendor < X86_VENDOR_NUM) {
2015 vendor = this_cpu->c_vendor;
2016 } else {
2017 if (c->cpuid_level >= 0)
2018 vendor = c->x86_vendor_id;
2019 }
2020
2021 if (vendor && !strstr(c->x86_model_id, vendor))
2022 pr_cont("%s ", vendor);
2023
2024 if (c->x86_model_id[0])
2025 pr_cont("%s", c->x86_model_id);
2026 else
2027 pr_cont("%d86", c->x86);
2028
2029 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
2030
2031 if (c->x86_stepping || c->cpuid_level >= 0)
2032 pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
2033 else
2034 pr_cont(")\n");
2035}
2036
2037/*
2038 * clearcpuid= was already parsed in cpu_parse_early_param(). This dummy
2039 * function prevents it from becoming an environment variable for init.
2040 */
2041static __init int setup_clearcpuid(char *arg)
2042{
2043 return 1;
2044}
2045__setup("clearcpuid=", setup_clearcpuid);
2046
2047DEFINE_PER_CPU_ALIGNED(struct pcpu_hot, pcpu_hot) = {
2048 .current_task = &init_task,
2049 .preempt_count = INIT_PREEMPT_COUNT,
2050 .top_of_stack = TOP_OF_INIT_STACK,
2051};
2052EXPORT_PER_CPU_SYMBOL(pcpu_hot);
2053
2054#ifdef CONFIG_X86_64
2055DEFINE_PER_CPU_FIRST(struct fixed_percpu_data,
2056 fixed_percpu_data) __aligned(PAGE_SIZE) __visible;
2057EXPORT_PER_CPU_SYMBOL_GPL(fixed_percpu_data);
2058
2059static void wrmsrl_cstar(unsigned long val)
2060{
2061 /*
2062 * Intel CPUs do not support 32-bit SYSCALL. Writing to MSR_CSTAR
2063 * is so far ignored by the CPU, but raises a #VE trap in a TDX
2064 * guest. Avoid the pointless write on all Intel CPUs.
2065 */
2066 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
2067 wrmsrl(MSR_CSTAR, val);
2068}
2069
2070/* May not be marked __init: used by software suspend */
2071void syscall_init(void)
2072{
2073 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
2074 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
2075
2076 if (ia32_enabled()) {
2077 wrmsrl_cstar((unsigned long)entry_SYSCALL_compat);
2078 /*
2079 * This only works on Intel CPUs.
2080 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
2081 * This does not cause SYSENTER to jump to the wrong location, because
2082 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
2083 */
2084 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
2085 wrmsrl_safe(MSR_IA32_SYSENTER_ESP,
2086 (unsigned long)(cpu_entry_stack(smp_processor_id()) + 1));
2087 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
2088 } else {
2089 wrmsrl_cstar((unsigned long)entry_SYSCALL32_ignore);
2090 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
2091 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
2092 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
2093 }
2094
2095 /*
2096 * Flags to clear on syscall; clear as much as possible
2097 * to minimize user space-kernel interference.
2098 */
2099 wrmsrl(MSR_SYSCALL_MASK,
2100 X86_EFLAGS_CF|X86_EFLAGS_PF|X86_EFLAGS_AF|
2101 X86_EFLAGS_ZF|X86_EFLAGS_SF|X86_EFLAGS_TF|
2102 X86_EFLAGS_IF|X86_EFLAGS_DF|X86_EFLAGS_OF|
2103 X86_EFLAGS_IOPL|X86_EFLAGS_NT|X86_EFLAGS_RF|
2104 X86_EFLAGS_AC|X86_EFLAGS_ID);
2105}
2106
2107#else /* CONFIG_X86_64 */
2108
2109#ifdef CONFIG_STACKPROTECTOR
2110DEFINE_PER_CPU(unsigned long, __stack_chk_guard);
2111EXPORT_PER_CPU_SYMBOL(__stack_chk_guard);
2112#endif
2113
2114#endif /* CONFIG_X86_64 */
2115
2116/*
2117 * Clear all 6 debug registers:
2118 */
2119static void clear_all_debug_regs(void)
2120{
2121 int i;
2122
2123 for (i = 0; i < 8; i++) {
2124 /* Ignore db4, db5 */
2125 if ((i == 4) || (i == 5))
2126 continue;
2127
2128 set_debugreg(0, i);
2129 }
2130}
2131
2132#ifdef CONFIG_KGDB
2133/*
2134 * Restore debug regs if using kgdbwait and you have a kernel debugger
2135 * connection established.
2136 */
2137static void dbg_restore_debug_regs(void)
2138{
2139 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
2140 arch_kgdb_ops.correct_hw_break();
2141}
2142#else /* ! CONFIG_KGDB */
2143#define dbg_restore_debug_regs()
2144#endif /* ! CONFIG_KGDB */
2145
2146static inline void setup_getcpu(int cpu)
2147{
2148 unsigned long cpudata = vdso_encode_cpunode(cpu, early_cpu_to_node(cpu));
2149 struct desc_struct d = { };
2150
2151 if (boot_cpu_has(X86_FEATURE_RDTSCP) || boot_cpu_has(X86_FEATURE_RDPID))
2152 wrmsr(MSR_TSC_AUX, cpudata, 0);
2153
2154 /* Store CPU and node number in limit. */
2155 d.limit0 = cpudata;
2156 d.limit1 = cpudata >> 16;
2157
2158 d.type = 5; /* RO data, expand down, accessed */
2159 d.dpl = 3; /* Visible to user code */
2160 d.s = 1; /* Not a system segment */
2161 d.p = 1; /* Present */
2162 d.d = 1; /* 32-bit */
2163
2164 write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_CPUNODE, &d, DESCTYPE_S);
2165}
2166
2167#ifdef CONFIG_X86_64
2168static inline void tss_setup_ist(struct tss_struct *tss)
2169{
2170 /* Set up the per-CPU TSS IST stacks */
2171 tss->x86_tss.ist[IST_INDEX_DF] = __this_cpu_ist_top_va(DF);
2172 tss->x86_tss.ist[IST_INDEX_NMI] = __this_cpu_ist_top_va(NMI);
2173 tss->x86_tss.ist[IST_INDEX_DB] = __this_cpu_ist_top_va(DB);
2174 tss->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
2175 /* Only mapped when SEV-ES is active */
2176 tss->x86_tss.ist[IST_INDEX_VC] = __this_cpu_ist_top_va(VC);
2177}
2178#else /* CONFIG_X86_64 */
2179static inline void tss_setup_ist(struct tss_struct *tss) { }
2180#endif /* !CONFIG_X86_64 */
2181
2182static inline void tss_setup_io_bitmap(struct tss_struct *tss)
2183{
2184 tss->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET_INVALID;
2185
2186#ifdef CONFIG_X86_IOPL_IOPERM
2187 tss->io_bitmap.prev_max = 0;
2188 tss->io_bitmap.prev_sequence = 0;
2189 memset(tss->io_bitmap.bitmap, 0xff, sizeof(tss->io_bitmap.bitmap));
2190 /*
2191 * Invalidate the extra array entry past the end of the all
2192 * permission bitmap as required by the hardware.
2193 */
2194 tss->io_bitmap.mapall[IO_BITMAP_LONGS] = ~0UL;
2195#endif
2196}
2197
2198/*
2199 * Setup everything needed to handle exceptions from the IDT, including the IST
2200 * exceptions which use paranoid_entry().
2201 */
2202void cpu_init_exception_handling(void)
2203{
2204 struct tss_struct *tss = this_cpu_ptr(&cpu_tss_rw);
2205 int cpu = raw_smp_processor_id();
2206
2207 /* paranoid_entry() gets the CPU number from the GDT */
2208 setup_getcpu(cpu);
2209
2210 /* IST vectors need TSS to be set up. */
2211 tss_setup_ist(tss);
2212 tss_setup_io_bitmap(tss);
2213 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
2214
2215 load_TR_desc();
2216
2217 /* GHCB needs to be setup to handle #VC. */
2218 setup_ghcb();
2219
2220 /* Finally load the IDT */
2221 load_current_idt();
2222}
2223
2224/*
2225 * cpu_init() initializes state that is per-CPU. Some data is already
2226 * initialized (naturally) in the bootstrap process, such as the GDT. We
2227 * reload it nevertheless, this function acts as a 'CPU state barrier',
2228 * nothing should get across.
2229 */
2230void cpu_init(void)
2231{
2232 struct task_struct *cur = current;
2233 int cpu = raw_smp_processor_id();
2234
2235#ifdef CONFIG_NUMA
2236 if (this_cpu_read(numa_node) == 0 &&
2237 early_cpu_to_node(cpu) != NUMA_NO_NODE)
2238 set_numa_node(early_cpu_to_node(cpu));
2239#endif
2240 pr_debug("Initializing CPU#%d\n", cpu);
2241
2242 if (IS_ENABLED(CONFIG_X86_64) || cpu_feature_enabled(X86_FEATURE_VME) ||
2243 boot_cpu_has(X86_FEATURE_TSC) || boot_cpu_has(X86_FEATURE_DE))
2244 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
2245
2246 if (IS_ENABLED(CONFIG_X86_64)) {
2247 loadsegment(fs, 0);
2248 memset(cur->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
2249 syscall_init();
2250
2251 wrmsrl(MSR_FS_BASE, 0);
2252 wrmsrl(MSR_KERNEL_GS_BASE, 0);
2253 barrier();
2254
2255 x2apic_setup();
2256 }
2257
2258 mmgrab(&init_mm);
2259 cur->active_mm = &init_mm;
2260 BUG_ON(cur->mm);
2261 initialize_tlbstate_and_flush();
2262 enter_lazy_tlb(&init_mm, cur);
2263
2264 /*
2265 * sp0 points to the entry trampoline stack regardless of what task
2266 * is running.
2267 */
2268 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
2269
2270 load_mm_ldt(&init_mm);
2271
2272 clear_all_debug_regs();
2273 dbg_restore_debug_regs();
2274
2275 doublefault_init_cpu_tss();
2276
2277 if (is_uv_system())
2278 uv_cpu_init();
2279
2280 load_fixmap_gdt(cpu);
2281}
2282
2283#ifdef CONFIG_MICROCODE_LATE_LOADING
2284/**
2285 * store_cpu_caps() - Store a snapshot of CPU capabilities
2286 * @curr_info: Pointer where to store it
2287 *
2288 * Returns: None
2289 */
2290void store_cpu_caps(struct cpuinfo_x86 *curr_info)
2291{
2292 /* Reload CPUID max function as it might've changed. */
2293 curr_info->cpuid_level = cpuid_eax(0);
2294
2295 /* Copy all capability leafs and pick up the synthetic ones. */
2296 memcpy(&curr_info->x86_capability, &boot_cpu_data.x86_capability,
2297 sizeof(curr_info->x86_capability));
2298
2299 /* Get the hardware CPUID leafs */
2300 get_cpu_cap(curr_info);
2301}
2302
2303/**
2304 * microcode_check() - Check if any CPU capabilities changed after an update.
2305 * @prev_info: CPU capabilities stored before an update.
2306 *
2307 * The microcode loader calls this upon late microcode load to recheck features,
2308 * only when microcode has been updated. Caller holds and CPU hotplug lock.
2309 *
2310 * Return: None
2311 */
2312void microcode_check(struct cpuinfo_x86 *prev_info)
2313{
2314 struct cpuinfo_x86 curr_info;
2315
2316 perf_check_microcode();
2317
2318 amd_check_microcode();
2319
2320 store_cpu_caps(&curr_info);
2321
2322 if (!memcmp(&prev_info->x86_capability, &curr_info.x86_capability,
2323 sizeof(prev_info->x86_capability)))
2324 return;
2325
2326 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
2327 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");
2328}
2329#endif
2330
2331/*
2332 * Invoked from core CPU hotplug code after hotplug operations
2333 */
2334void arch_smt_update(void)
2335{
2336 /* Handle the speculative execution misfeatures */
2337 cpu_bugs_smt_update();
2338 /* Check whether IPI broadcasting can be enabled */
2339 apic_smt_update();
2340}
2341
2342void __init arch_cpu_finalize_init(void)
2343{
2344 identify_boot_cpu();
2345
2346 /*
2347 * identify_boot_cpu() initialized SMT support information, let the
2348 * core code know.
2349 */
2350 cpu_smt_set_num_threads(smp_num_siblings, smp_num_siblings);
2351
2352 if (!IS_ENABLED(CONFIG_SMP)) {
2353 pr_info("CPU: ");
2354 print_cpu_info(&boot_cpu_data);
2355 }
2356
2357 cpu_select_mitigations();
2358
2359 arch_smt_update();
2360
2361 if (IS_ENABLED(CONFIG_X86_32)) {
2362 /*
2363 * Check whether this is a real i386 which is not longer
2364 * supported and fixup the utsname.
2365 */
2366 if (boot_cpu_data.x86 < 4)
2367 panic("Kernel requires i486+ for 'invlpg' and other features");
2368
2369 init_utsname()->machine[1] =
2370 '0' + (boot_cpu_data.x86 > 6 ? 6 : boot_cpu_data.x86);
2371 }
2372
2373 /*
2374 * Must be before alternatives because it might set or clear
2375 * feature bits.
2376 */
2377 fpu__init_system();
2378 fpu__init_cpu();
2379
2380 alternative_instructions();
2381
2382 if (IS_ENABLED(CONFIG_X86_64)) {
2383 /*
2384 * Make sure the first 2MB area is not mapped by huge pages
2385 * There are typically fixed size MTRRs in there and overlapping
2386 * MTRRs into large pages causes slow downs.
2387 *
2388 * Right now we don't do that with gbpages because there seems
2389 * very little benefit for that case.
2390 */
2391 if (!direct_gbpages)
2392 set_memory_4k((unsigned long)__va(0), 1);
2393 } else {
2394 fpu__init_check_bugs();
2395 }
2396
2397 /*
2398 * This needs to be called before any devices perform DMA
2399 * operations that might use the SWIOTLB bounce buffers. It will
2400 * mark the bounce buffers as decrypted so that their usage will
2401 * not cause "plain-text" data to be decrypted when accessed. It
2402 * must be called after late_time_init() so that Hyper-V x86/x64
2403 * hypercalls work when the SWIOTLB bounce buffers are decrypted.
2404 */
2405 mem_encrypt_init();
2406}