blob: f961a56e9da3ff8c1b1415e63a830cfea046eea6 [file] [log] [blame]
David Brazdil0f672f62019-12-10 10:32:29 +00001// SPDX-License-Identifier: GPL-2.0-only
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00002/* cpu_feature_enabled() cannot be used this early */
3#define USE_EARLY_PGTABLE_L5
4
David Brazdil0f672f62019-12-10 10:32:29 +00005#include <linux/memblock.h>
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00006#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/init.h>
18#include <linux/kprobes.h>
19#include <linux/kgdb.h>
20#include <linux/smp.h>
21#include <linux/io.h>
22#include <linux/syscore_ops.h>
23
24#include <asm/stackprotector.h>
25#include <asm/perf_event.h>
26#include <asm/mmu_context.h>
27#include <asm/archrandom.h>
28#include <asm/hypervisor.h>
29#include <asm/processor.h>
30#include <asm/tlbflush.h>
31#include <asm/debugreg.h>
32#include <asm/sections.h>
33#include <asm/vsyscall.h>
34#include <linux/topology.h>
35#include <linux/cpumask.h>
36#include <asm/pgtable.h>
37#include <linux/atomic.h>
38#include <asm/proto.h>
39#include <asm/setup.h>
40#include <asm/apic.h>
41#include <asm/desc.h>
42#include <asm/fpu/internal.h>
43#include <asm/mtrr.h>
44#include <asm/hwcap2.h>
45#include <linux/numa.h>
46#include <asm/asm.h>
47#include <asm/bugs.h>
48#include <asm/cpu.h>
49#include <asm/mce.h>
50#include <asm/msr.h>
51#include <asm/pat.h>
52#include <asm/microcode.h>
53#include <asm/microcode_intel.h>
54#include <asm/intel-family.h>
55#include <asm/cpu_device_id.h>
56
57#ifdef CONFIG_X86_LOCAL_APIC
58#include <asm/uv/uv.h>
59#endif
60
61#include "cpu.h"
62
63u32 elf_hwcap2 __read_mostly;
64
65/* all of these masks are initialized in setup_cpu_local_masks() */
66cpumask_var_t cpu_initialized_mask;
67cpumask_var_t cpu_callout_mask;
68cpumask_var_t cpu_callin_mask;
69
70/* representing cpus for which sibling maps can be computed */
71cpumask_var_t cpu_sibling_setup_mask;
72
73/* Number of siblings per CPU package */
74int smp_num_siblings = 1;
75EXPORT_SYMBOL(smp_num_siblings);
76
77/* Last level cache ID of each logical CPU */
78DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
79
80/* correctly size the local cpu masks */
81void __init setup_cpu_local_masks(void)
82{
83 alloc_bootmem_cpumask_var(&cpu_initialized_mask);
84 alloc_bootmem_cpumask_var(&cpu_callin_mask);
85 alloc_bootmem_cpumask_var(&cpu_callout_mask);
86 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
87}
88
89static void default_init(struct cpuinfo_x86 *c)
90{
91#ifdef CONFIG_X86_64
92 cpu_detect_cache_sizes(c);
93#else
94 /* Not much we can do here... */
95 /* Check if at least it has cpuid */
96 if (c->cpuid_level == -1) {
97 /* No cpuid. It must be an ancient CPU */
98 if (c->x86 == 4)
99 strcpy(c->x86_model_id, "486");
100 else if (c->x86 == 3)
101 strcpy(c->x86_model_id, "386");
102 }
103#endif
104}
105
106static const struct cpu_dev default_cpu = {
107 .c_init = default_init,
108 .c_vendor = "Unknown",
109 .c_x86_vendor = X86_VENDOR_UNKNOWN,
110};
111
112static const struct cpu_dev *this_cpu = &default_cpu;
113
114DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
115#ifdef CONFIG_X86_64
116 /*
117 * We need valid kernel segments for data and code in long mode too
118 * IRET will check the segment types kkeil 2000/10/28
119 * Also sysret mandates a special GDT layout
120 *
121 * TLS descriptors are currently at a different place compared to i386.
122 * Hopefully nobody expects them at a fixed place (Wine?)
123 */
124 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
125 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
126 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
127 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
128 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
129 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
130#else
131 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
132 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
133 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
134 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
135 /*
136 * Segments used for calling PnP BIOS have byte granularity.
137 * They code segments and data segments have fixed 64k limits,
138 * the transfer segment sizes are set at run time.
139 */
140 /* 32-bit code */
141 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
142 /* 16-bit code */
143 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
144 /* 16-bit data */
145 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(0x0092, 0, 0xffff),
146 /* 16-bit data */
147 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(0x0092, 0, 0),
148 /* 16-bit data */
149 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(0x0092, 0, 0),
150 /*
151 * The APM segments have byte granularity and their bases
152 * are set at run time. All have 64k limits.
153 */
154 /* 32-bit code */
155 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
156 /* 16-bit code */
157 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
158 /* data */
159 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(0x4092, 0, 0xffff),
160
161 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
162 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
163 GDT_STACK_CANARY_INIT
164#endif
165} };
166EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
167
168static int __init x86_mpx_setup(char *s)
169{
170 /* require an exact match without trailing characters */
171 if (strlen(s))
172 return 0;
173
174 /* do not emit a message if the feature is not present */
175 if (!boot_cpu_has(X86_FEATURE_MPX))
176 return 1;
177
178 setup_clear_cpu_cap(X86_FEATURE_MPX);
179 pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
180 return 1;
181}
182__setup("nompx", x86_mpx_setup);
183
184#ifdef CONFIG_X86_64
185static int __init x86_nopcid_setup(char *s)
186{
187 /* nopcid doesn't accept parameters */
188 if (s)
189 return -EINVAL;
190
191 /* do not emit a message if the feature is not present */
192 if (!boot_cpu_has(X86_FEATURE_PCID))
193 return 0;
194
195 setup_clear_cpu_cap(X86_FEATURE_PCID);
196 pr_info("nopcid: PCID feature disabled\n");
197 return 0;
198}
199early_param("nopcid", x86_nopcid_setup);
200#endif
201
202static int __init x86_noinvpcid_setup(char *s)
203{
204 /* noinvpcid doesn't accept parameters */
205 if (s)
206 return -EINVAL;
207
208 /* do not emit a message if the feature is not present */
209 if (!boot_cpu_has(X86_FEATURE_INVPCID))
210 return 0;
211
212 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
213 pr_info("noinvpcid: INVPCID feature disabled\n");
214 return 0;
215}
216early_param("noinvpcid", x86_noinvpcid_setup);
217
218#ifdef CONFIG_X86_32
219static int cachesize_override = -1;
220static int disable_x86_serial_nr = 1;
221
222static int __init cachesize_setup(char *str)
223{
224 get_option(&str, &cachesize_override);
225 return 1;
226}
227__setup("cachesize=", cachesize_setup);
228
229static int __init x86_sep_setup(char *s)
230{
231 setup_clear_cpu_cap(X86_FEATURE_SEP);
232 return 1;
233}
234__setup("nosep", x86_sep_setup);
235
236/* Standard macro to see if a specific flag is changeable */
237static inline int flag_is_changeable_p(u32 flag)
238{
239 u32 f1, f2;
240
241 /*
242 * Cyrix and IDT cpus allow disabling of CPUID
243 * so the code below may return different results
244 * when it is executed before and after enabling
245 * the CPUID. Add "volatile" to not allow gcc to
246 * optimize the subsequent calls to this function.
247 */
248 asm volatile ("pushfl \n\t"
249 "pushfl \n\t"
250 "popl %0 \n\t"
251 "movl %0, %1 \n\t"
252 "xorl %2, %0 \n\t"
253 "pushl %0 \n\t"
254 "popfl \n\t"
255 "pushfl \n\t"
256 "popl %0 \n\t"
257 "popfl \n\t"
258
259 : "=&r" (f1), "=&r" (f2)
260 : "ir" (flag));
261
262 return ((f1^f2) & flag) != 0;
263}
264
265/* Probe for the CPUID instruction */
266int have_cpuid_p(void)
267{
268 return flag_is_changeable_p(X86_EFLAGS_ID);
269}
270
271static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
272{
273 unsigned long lo, hi;
274
275 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
276 return;
277
278 /* Disable processor serial number: */
279
280 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
281 lo |= 0x200000;
282 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
283
284 pr_notice("CPU serial number disabled.\n");
285 clear_cpu_cap(c, X86_FEATURE_PN);
286
287 /* Disabling the serial number may affect the cpuid level */
288 c->cpuid_level = cpuid_eax(0);
289}
290
291static int __init x86_serial_nr_setup(char *s)
292{
293 disable_x86_serial_nr = 0;
294 return 1;
295}
296__setup("serialnumber", x86_serial_nr_setup);
297#else
298static inline int flag_is_changeable_p(u32 flag)
299{
300 return 1;
301}
302static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
303{
304}
305#endif
306
307static __init int setup_disable_smep(char *arg)
308{
309 setup_clear_cpu_cap(X86_FEATURE_SMEP);
310 /* Check for things that depend on SMEP being enabled: */
311 check_mpx_erratum(&boot_cpu_data);
312 return 1;
313}
314__setup("nosmep", setup_disable_smep);
315
316static __always_inline void setup_smep(struct cpuinfo_x86 *c)
317{
318 if (cpu_has(c, X86_FEATURE_SMEP))
319 cr4_set_bits(X86_CR4_SMEP);
320}
321
322static __init int setup_disable_smap(char *arg)
323{
324 setup_clear_cpu_cap(X86_FEATURE_SMAP);
325 return 1;
326}
327__setup("nosmap", setup_disable_smap);
328
329static __always_inline void setup_smap(struct cpuinfo_x86 *c)
330{
331 unsigned long eflags = native_save_fl();
332
333 /* This should have been cleared long ago */
334 BUG_ON(eflags & X86_EFLAGS_AC);
335
336 if (cpu_has(c, X86_FEATURE_SMAP)) {
337#ifdef CONFIG_X86_SMAP
338 cr4_set_bits(X86_CR4_SMAP);
339#else
340 cr4_clear_bits(X86_CR4_SMAP);
341#endif
342 }
343}
344
345static __always_inline void setup_umip(struct cpuinfo_x86 *c)
346{
347 /* Check the boot processor, plus build option for UMIP. */
348 if (!cpu_feature_enabled(X86_FEATURE_UMIP))
349 goto out;
350
351 /* Check the current processor's cpuid bits. */
352 if (!cpu_has(c, X86_FEATURE_UMIP))
353 goto out;
354
355 cr4_set_bits(X86_CR4_UMIP);
356
David Brazdil0f672f62019-12-10 10:32:29 +0000357 pr_info_once("x86/cpu: User Mode Instruction Prevention (UMIP) activated\n");
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000358
359 return;
360
361out:
362 /*
363 * Make sure UMIP is disabled in case it was enabled in a
364 * previous boot (e.g., via kexec).
365 */
366 cr4_clear_bits(X86_CR4_UMIP);
367}
368
Olivier Deprez0e641232021-09-23 10:07:05 +0200369/* These bits should not change their value after CPU init is finished. */
370static const unsigned long cr4_pinned_mask =
371 X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_UMIP | X86_CR4_FSGSBASE;
David Brazdil0f672f62019-12-10 10:32:29 +0000372static DEFINE_STATIC_KEY_FALSE_RO(cr_pinning);
373static unsigned long cr4_pinned_bits __ro_after_init;
374
375void native_write_cr0(unsigned long val)
376{
377 unsigned long bits_missing = 0;
378
379set_register:
Olivier Deprez0e641232021-09-23 10:07:05 +0200380 asm volatile("mov %0,%%cr0": "+r" (val) : : "memory");
David Brazdil0f672f62019-12-10 10:32:29 +0000381
382 if (static_branch_likely(&cr_pinning)) {
383 if (unlikely((val & X86_CR0_WP) != X86_CR0_WP)) {
384 bits_missing = X86_CR0_WP;
385 val |= bits_missing;
386 goto set_register;
387 }
388 /* Warn after we've set the missing bits. */
389 WARN_ONCE(bits_missing, "CR0 WP bit went missing!?\n");
390 }
391}
392EXPORT_SYMBOL(native_write_cr0);
393
394void native_write_cr4(unsigned long val)
395{
Olivier Deprez0e641232021-09-23 10:07:05 +0200396 unsigned long bits_changed = 0;
David Brazdil0f672f62019-12-10 10:32:29 +0000397
398set_register:
Olivier Deprez0e641232021-09-23 10:07:05 +0200399 asm volatile("mov %0,%%cr4": "+r" (val) : : "memory");
David Brazdil0f672f62019-12-10 10:32:29 +0000400
401 if (static_branch_likely(&cr_pinning)) {
Olivier Deprez0e641232021-09-23 10:07:05 +0200402 if (unlikely((val & cr4_pinned_mask) != cr4_pinned_bits)) {
403 bits_changed = (val & cr4_pinned_mask) ^ cr4_pinned_bits;
404 val = (val & ~cr4_pinned_mask) | cr4_pinned_bits;
David Brazdil0f672f62019-12-10 10:32:29 +0000405 goto set_register;
406 }
Olivier Deprez0e641232021-09-23 10:07:05 +0200407 /* Warn after we've corrected the changed bits. */
408 WARN_ONCE(bits_changed, "pinned CR4 bits changed: 0x%lx!?\n",
409 bits_changed);
David Brazdil0f672f62019-12-10 10:32:29 +0000410 }
411}
412EXPORT_SYMBOL(native_write_cr4);
413
414void cr4_init(void)
415{
416 unsigned long cr4 = __read_cr4();
417
418 if (boot_cpu_has(X86_FEATURE_PCID))
419 cr4 |= X86_CR4_PCIDE;
420 if (static_branch_likely(&cr_pinning))
Olivier Deprez0e641232021-09-23 10:07:05 +0200421 cr4 = (cr4 & ~cr4_pinned_mask) | cr4_pinned_bits;
David Brazdil0f672f62019-12-10 10:32:29 +0000422
423 __write_cr4(cr4);
424
425 /* Initialize cr4 shadow for this CPU. */
426 this_cpu_write(cpu_tlbstate.cr4, cr4);
427}
428
429/*
430 * Once CPU feature detection is finished (and boot params have been
431 * parsed), record any of the sensitive CR bits that are set, and
432 * enable CR pinning.
433 */
434static void __init setup_cr_pinning(void)
435{
Olivier Deprez0e641232021-09-23 10:07:05 +0200436 cr4_pinned_bits = this_cpu_read(cpu_tlbstate.cr4) & cr4_pinned_mask;
David Brazdil0f672f62019-12-10 10:32:29 +0000437 static_key_enable(&cr_pinning.key);
438}
439
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000440/*
441 * Protection Keys are not available in 32-bit mode.
442 */
443static bool pku_disabled;
444
445static __always_inline void setup_pku(struct cpuinfo_x86 *c)
446{
David Brazdil0f672f62019-12-10 10:32:29 +0000447 struct pkru_state *pk;
448
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000449 /* check the boot processor, plus compile options for PKU: */
450 if (!cpu_feature_enabled(X86_FEATURE_PKU))
451 return;
452 /* checks the actual processor's cpuid bits: */
453 if (!cpu_has(c, X86_FEATURE_PKU))
454 return;
455 if (pku_disabled)
456 return;
457
458 cr4_set_bits(X86_CR4_PKE);
David Brazdil0f672f62019-12-10 10:32:29 +0000459 pk = get_xsave_addr(&init_fpstate.xsave, XFEATURE_PKRU);
460 if (pk)
461 pk->pkru = init_pkru_value;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000462 /*
463 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
464 * cpuid bit to be set. We need to ensure that we
465 * update that bit in this CPU's "cpu_info".
466 */
Olivier Deprez0e641232021-09-23 10:07:05 +0200467 set_cpu_cap(c, X86_FEATURE_OSPKE);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000468}
469
470#ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
471static __init int setup_disable_pku(char *arg)
472{
473 /*
474 * Do not clear the X86_FEATURE_PKU bit. All of the
475 * runtime checks are against OSPKE so clearing the
476 * bit does nothing.
477 *
478 * This way, we will see "pku" in cpuinfo, but not
479 * "ospke", which is exactly what we want. It shows
480 * that the CPU has PKU, but the OS has not enabled it.
481 * This happens to be exactly how a system would look
482 * if we disabled the config option.
483 */
484 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
485 pku_disabled = true;
486 return 1;
487}
488__setup("nopku", setup_disable_pku);
489#endif /* CONFIG_X86_64 */
490
491/*
492 * Some CPU features depend on higher CPUID levels, which may not always
493 * be available due to CPUID level capping or broken virtualization
494 * software. Add those features to this table to auto-disable them.
495 */
496struct cpuid_dependent_feature {
497 u32 feature;
498 u32 level;
499};
500
501static const struct cpuid_dependent_feature
502cpuid_dependent_features[] = {
503 { X86_FEATURE_MWAIT, 0x00000005 },
504 { X86_FEATURE_DCA, 0x00000009 },
505 { X86_FEATURE_XSAVE, 0x0000000d },
506 { 0, 0 }
507};
508
509static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
510{
511 const struct cpuid_dependent_feature *df;
512
513 for (df = cpuid_dependent_features; df->feature; df++) {
514
515 if (!cpu_has(c, df->feature))
516 continue;
517 /*
518 * Note: cpuid_level is set to -1 if unavailable, but
519 * extended_extended_level is set to 0 if unavailable
520 * and the legitimate extended levels are all negative
521 * when signed; hence the weird messing around with
522 * signs here...
523 */
524 if (!((s32)df->level < 0 ?
525 (u32)df->level > (u32)c->extended_cpuid_level :
526 (s32)df->level > (s32)c->cpuid_level))
527 continue;
528
529 clear_cpu_cap(c, df->feature);
530 if (!warn)
531 continue;
532
533 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
534 x86_cap_flag(df->feature), df->level);
535 }
536}
537
538/*
539 * Naming convention should be: <Name> [(<Codename>)]
540 * This table only is used unless init_<vendor>() below doesn't set it;
541 * in particular, if CPUID levels 0x80000002..4 are supported, this
542 * isn't used
543 */
544
545/* Look up CPU names by table lookup. */
546static const char *table_lookup_model(struct cpuinfo_x86 *c)
547{
548#ifdef CONFIG_X86_32
549 const struct legacy_cpu_model_info *info;
550
551 if (c->x86_model >= 16)
552 return NULL; /* Range check */
553
554 if (!this_cpu)
555 return NULL;
556
557 info = this_cpu->legacy_models;
558
559 while (info->family) {
560 if (info->family == c->x86)
561 return info->model_names[c->x86_model];
562 info++;
563 }
564#endif
565 return NULL; /* Not found */
566}
567
568__u32 cpu_caps_cleared[NCAPINTS + NBUGINTS];
569__u32 cpu_caps_set[NCAPINTS + NBUGINTS];
570
571void load_percpu_segment(int cpu)
572{
573#ifdef CONFIG_X86_32
574 loadsegment(fs, __KERNEL_PERCPU);
575#else
576 __loadsegment_simple(gs, 0);
577 wrmsrl(MSR_GS_BASE, cpu_kernelmode_gs_base(cpu));
578#endif
579 load_stack_canary_segment();
580}
581
582#ifdef CONFIG_X86_32
583/* The 32-bit entry code needs to find cpu_entry_area. */
584DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area);
585#endif
586
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000587/* Load the original GDT from the per-cpu structure */
588void load_direct_gdt(int cpu)
589{
590 struct desc_ptr gdt_descr;
591
592 gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
593 gdt_descr.size = GDT_SIZE - 1;
594 load_gdt(&gdt_descr);
595}
596EXPORT_SYMBOL_GPL(load_direct_gdt);
597
598/* Load a fixmap remapping of the per-cpu GDT */
599void load_fixmap_gdt(int cpu)
600{
601 struct desc_ptr gdt_descr;
602
603 gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
604 gdt_descr.size = GDT_SIZE - 1;
605 load_gdt(&gdt_descr);
606}
607EXPORT_SYMBOL_GPL(load_fixmap_gdt);
608
609/*
610 * Current gdt points %fs at the "master" per-cpu area: after this,
611 * it's on the real one.
612 */
613void switch_to_new_gdt(int cpu)
614{
615 /* Load the original GDT */
616 load_direct_gdt(cpu);
617 /* Reload the per-cpu base */
618 load_percpu_segment(cpu);
619}
620
621static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
622
623static void get_model_name(struct cpuinfo_x86 *c)
624{
625 unsigned int *v;
626 char *p, *q, *s;
627
628 if (c->extended_cpuid_level < 0x80000004)
629 return;
630
631 v = (unsigned int *)c->x86_model_id;
632 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
633 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
634 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
635 c->x86_model_id[48] = 0;
636
637 /* Trim whitespace */
638 p = q = s = &c->x86_model_id[0];
639
640 while (*p == ' ')
641 p++;
642
643 while (*p) {
644 /* Note the last non-whitespace index */
645 if (!isspace(*p))
646 s = q;
647
648 *q++ = *p++;
649 }
650
651 *(s + 1) = '\0';
652}
653
654void detect_num_cpu_cores(struct cpuinfo_x86 *c)
655{
656 unsigned int eax, ebx, ecx, edx;
657
658 c->x86_max_cores = 1;
659 if (!IS_ENABLED(CONFIG_SMP) || c->cpuid_level < 4)
660 return;
661
662 cpuid_count(4, 0, &eax, &ebx, &ecx, &edx);
663 if (eax & 0x1f)
664 c->x86_max_cores = (eax >> 26) + 1;
665}
666
667void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
668{
669 unsigned int n, dummy, ebx, ecx, edx, l2size;
670
671 n = c->extended_cpuid_level;
672
673 if (n >= 0x80000005) {
674 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
675 c->x86_cache_size = (ecx>>24) + (edx>>24);
676#ifdef CONFIG_X86_64
677 /* On K8 L1 TLB is inclusive, so don't count it */
678 c->x86_tlbsize = 0;
679#endif
680 }
681
682 if (n < 0x80000006) /* Some chips just has a large L1. */
683 return;
684
685 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
686 l2size = ecx >> 16;
687
688#ifdef CONFIG_X86_64
689 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
690#else
691 /* do processor-specific cache resizing */
692 if (this_cpu->legacy_cache_size)
693 l2size = this_cpu->legacy_cache_size(c, l2size);
694
695 /* Allow user to override all this if necessary. */
696 if (cachesize_override != -1)
697 l2size = cachesize_override;
698
699 if (l2size == 0)
700 return; /* Again, no L2 cache is possible */
701#endif
702
703 c->x86_cache_size = l2size;
704}
705
706u16 __read_mostly tlb_lli_4k[NR_INFO];
707u16 __read_mostly tlb_lli_2m[NR_INFO];
708u16 __read_mostly tlb_lli_4m[NR_INFO];
709u16 __read_mostly tlb_lld_4k[NR_INFO];
710u16 __read_mostly tlb_lld_2m[NR_INFO];
711u16 __read_mostly tlb_lld_4m[NR_INFO];
712u16 __read_mostly tlb_lld_1g[NR_INFO];
713
714static void cpu_detect_tlb(struct cpuinfo_x86 *c)
715{
716 if (this_cpu->c_detect_tlb)
717 this_cpu->c_detect_tlb(c);
718
719 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
720 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
721 tlb_lli_4m[ENTRIES]);
722
723 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
724 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
725 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
726}
727
728int detect_ht_early(struct cpuinfo_x86 *c)
729{
730#ifdef CONFIG_SMP
731 u32 eax, ebx, ecx, edx;
732
733 if (!cpu_has(c, X86_FEATURE_HT))
734 return -1;
735
736 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
737 return -1;
738
739 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
740 return -1;
741
742 cpuid(1, &eax, &ebx, &ecx, &edx);
743
744 smp_num_siblings = (ebx & 0xff0000) >> 16;
745 if (smp_num_siblings == 1)
746 pr_info_once("CPU0: Hyper-Threading is disabled\n");
747#endif
748 return 0;
749}
750
751void detect_ht(struct cpuinfo_x86 *c)
752{
753#ifdef CONFIG_SMP
754 int index_msb, core_bits;
755
756 if (detect_ht_early(c) < 0)
757 return;
758
759 index_msb = get_count_order(smp_num_siblings);
760 c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
761
762 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
763
764 index_msb = get_count_order(smp_num_siblings);
765
766 core_bits = get_count_order(c->x86_max_cores);
767
768 c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
769 ((1 << core_bits) - 1);
770#endif
771}
772
773static void get_cpu_vendor(struct cpuinfo_x86 *c)
774{
775 char *v = c->x86_vendor_id;
776 int i;
777
778 for (i = 0; i < X86_VENDOR_NUM; i++) {
779 if (!cpu_devs[i])
780 break;
781
782 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
783 (cpu_devs[i]->c_ident[1] &&
784 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
785
786 this_cpu = cpu_devs[i];
787 c->x86_vendor = this_cpu->c_x86_vendor;
788 return;
789 }
790 }
791
792 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
793 "CPU: Your system may be unstable.\n", v);
794
795 c->x86_vendor = X86_VENDOR_UNKNOWN;
796 this_cpu = &default_cpu;
797}
798
799void cpu_detect(struct cpuinfo_x86 *c)
800{
801 /* Get vendor name */
802 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
803 (unsigned int *)&c->x86_vendor_id[0],
804 (unsigned int *)&c->x86_vendor_id[8],
805 (unsigned int *)&c->x86_vendor_id[4]);
806
807 c->x86 = 4;
808 /* Intel-defined flags: level 0x00000001 */
809 if (c->cpuid_level >= 0x00000001) {
810 u32 junk, tfms, cap0, misc;
811
812 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
813 c->x86 = x86_family(tfms);
814 c->x86_model = x86_model(tfms);
815 c->x86_stepping = x86_stepping(tfms);
816
817 if (cap0 & (1<<19)) {
818 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
819 c->x86_cache_alignment = c->x86_clflush_size;
820 }
821 }
822}
823
824static void apply_forced_caps(struct cpuinfo_x86 *c)
825{
826 int i;
827
828 for (i = 0; i < NCAPINTS + NBUGINTS; i++) {
829 c->x86_capability[i] &= ~cpu_caps_cleared[i];
830 c->x86_capability[i] |= cpu_caps_set[i];
831 }
832}
833
834static void init_speculation_control(struct cpuinfo_x86 *c)
835{
836 /*
837 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support,
838 * and they also have a different bit for STIBP support. Also,
839 * a hypervisor might have set the individual AMD bits even on
840 * Intel CPUs, for finer-grained selection of what's available.
841 */
842 if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) {
843 set_cpu_cap(c, X86_FEATURE_IBRS);
844 set_cpu_cap(c, X86_FEATURE_IBPB);
845 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
846 }
847
848 if (cpu_has(c, X86_FEATURE_INTEL_STIBP))
849 set_cpu_cap(c, X86_FEATURE_STIBP);
850
851 if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) ||
852 cpu_has(c, X86_FEATURE_VIRT_SSBD))
853 set_cpu_cap(c, X86_FEATURE_SSBD);
854
855 if (cpu_has(c, X86_FEATURE_AMD_IBRS)) {
856 set_cpu_cap(c, X86_FEATURE_IBRS);
857 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
858 }
859
860 if (cpu_has(c, X86_FEATURE_AMD_IBPB))
861 set_cpu_cap(c, X86_FEATURE_IBPB);
862
863 if (cpu_has(c, X86_FEATURE_AMD_STIBP)) {
864 set_cpu_cap(c, X86_FEATURE_STIBP);
865 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
866 }
867
868 if (cpu_has(c, X86_FEATURE_AMD_SSBD)) {
869 set_cpu_cap(c, X86_FEATURE_SSBD);
870 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
871 clear_cpu_cap(c, X86_FEATURE_VIRT_SSBD);
872 }
873}
874
David Brazdil0f672f62019-12-10 10:32:29 +0000875static void init_cqm(struct cpuinfo_x86 *c)
876{
877 if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
878 c->x86_cache_max_rmid = -1;
879 c->x86_cache_occ_scale = -1;
880 return;
881 }
882
883 /* will be overridden if occupancy monitoring exists */
884 c->x86_cache_max_rmid = cpuid_ebx(0xf);
885
886 if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
887 cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
888 cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
889 u32 eax, ebx, ecx, edx;
890
891 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
892 cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
893
894 c->x86_cache_max_rmid = ecx;
895 c->x86_cache_occ_scale = ebx;
896 }
897}
898
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000899void get_cpu_cap(struct cpuinfo_x86 *c)
900{
901 u32 eax, ebx, ecx, edx;
902
903 /* Intel-defined flags: level 0x00000001 */
904 if (c->cpuid_level >= 0x00000001) {
905 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
906
907 c->x86_capability[CPUID_1_ECX] = ecx;
908 c->x86_capability[CPUID_1_EDX] = edx;
909 }
910
911 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */
912 if (c->cpuid_level >= 0x00000006)
913 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
914
915 /* Additional Intel-defined flags: level 0x00000007 */
916 if (c->cpuid_level >= 0x00000007) {
917 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
918 c->x86_capability[CPUID_7_0_EBX] = ebx;
919 c->x86_capability[CPUID_7_ECX] = ecx;
920 c->x86_capability[CPUID_7_EDX] = edx;
David Brazdil0f672f62019-12-10 10:32:29 +0000921
922 /* Check valid sub-leaf index before accessing it */
923 if (eax >= 1) {
924 cpuid_count(0x00000007, 1, &eax, &ebx, &ecx, &edx);
925 c->x86_capability[CPUID_7_1_EAX] = eax;
926 }
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000927 }
928
929 /* Extended state features: level 0x0000000d */
930 if (c->cpuid_level >= 0x0000000d) {
931 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
932
933 c->x86_capability[CPUID_D_1_EAX] = eax;
934 }
935
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000936 /* AMD-defined flags: level 0x80000001 */
937 eax = cpuid_eax(0x80000000);
938 c->extended_cpuid_level = eax;
939
940 if ((eax & 0xffff0000) == 0x80000000) {
941 if (eax >= 0x80000001) {
942 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
943
944 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
945 c->x86_capability[CPUID_8000_0001_EDX] = edx;
946 }
947 }
948
949 if (c->extended_cpuid_level >= 0x80000007) {
950 cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
951
952 c->x86_capability[CPUID_8000_0007_EBX] = ebx;
953 c->x86_power = edx;
954 }
955
956 if (c->extended_cpuid_level >= 0x80000008) {
957 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
958 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
959 }
960
961 if (c->extended_cpuid_level >= 0x8000000a)
962 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
963
964 init_scattered_cpuid_features(c);
965 init_speculation_control(c);
David Brazdil0f672f62019-12-10 10:32:29 +0000966 init_cqm(c);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +0000967
968 /*
969 * Clear/Set all flags overridden by options, after probe.
970 * This needs to happen each time we re-probe, which may happen
971 * several times during CPU initialization.
972 */
973 apply_forced_caps(c);
974}
975
976void get_cpu_address_sizes(struct cpuinfo_x86 *c)
977{
978 u32 eax, ebx, ecx, edx;
979
980 if (c->extended_cpuid_level >= 0x80000008) {
981 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
982
983 c->x86_virt_bits = (eax >> 8) & 0xff;
984 c->x86_phys_bits = eax & 0xff;
985 }
986#ifdef CONFIG_X86_32
987 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
988 c->x86_phys_bits = 36;
989#endif
990 c->x86_cache_bits = c->x86_phys_bits;
991}
992
993static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
994{
995#ifdef CONFIG_X86_32
996 int i;
997
998 /*
999 * First of all, decide if this is a 486 or higher
1000 * It's a 486 if we can modify the AC flag
1001 */
1002 if (flag_is_changeable_p(X86_EFLAGS_AC))
1003 c->x86 = 4;
1004 else
1005 c->x86 = 3;
1006
1007 for (i = 0; i < X86_VENDOR_NUM; i++)
1008 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
1009 c->x86_vendor_id[0] = 0;
1010 cpu_devs[i]->c_identify(c);
1011 if (c->x86_vendor_id[0]) {
1012 get_cpu_vendor(c);
1013 break;
1014 }
1015 }
1016#endif
1017}
1018
David Brazdil0f672f62019-12-10 10:32:29 +00001019#define NO_SPECULATION BIT(0)
1020#define NO_MELTDOWN BIT(1)
1021#define NO_SSB BIT(2)
1022#define NO_L1TF BIT(3)
1023#define NO_MDS BIT(4)
1024#define MSBDS_ONLY BIT(5)
1025#define NO_SWAPGS BIT(6)
1026#define NO_ITLB_MULTIHIT BIT(7)
Olivier Deprez0e641232021-09-23 10:07:05 +02001027#define NO_SPECTRE_V2 BIT(8)
David Brazdil0f672f62019-12-10 10:32:29 +00001028
1029#define VULNWL(_vendor, _family, _model, _whitelist) \
1030 { X86_VENDOR_##_vendor, _family, _model, X86_FEATURE_ANY, _whitelist }
1031
1032#define VULNWL_INTEL(model, whitelist) \
1033 VULNWL(INTEL, 6, INTEL_FAM6_##model, whitelist)
1034
1035#define VULNWL_AMD(family, whitelist) \
1036 VULNWL(AMD, family, X86_MODEL_ANY, whitelist)
1037
1038#define VULNWL_HYGON(family, whitelist) \
1039 VULNWL(HYGON, family, X86_MODEL_ANY, whitelist)
1040
1041static const __initconst struct x86_cpu_id cpu_vuln_whitelist[] = {
1042 VULNWL(ANY, 4, X86_MODEL_ANY, NO_SPECULATION),
1043 VULNWL(CENTAUR, 5, X86_MODEL_ANY, NO_SPECULATION),
1044 VULNWL(INTEL, 5, X86_MODEL_ANY, NO_SPECULATION),
1045 VULNWL(NSC, 5, X86_MODEL_ANY, NO_SPECULATION),
1046
1047 /* Intel Family 6 */
1048 VULNWL_INTEL(ATOM_SALTWELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
1049 VULNWL_INTEL(ATOM_SALTWELL_TABLET, NO_SPECULATION | NO_ITLB_MULTIHIT),
1050 VULNWL_INTEL(ATOM_SALTWELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
1051 VULNWL_INTEL(ATOM_BONNELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
1052 VULNWL_INTEL(ATOM_BONNELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
1053
1054 VULNWL_INTEL(ATOM_SILVERMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1055 VULNWL_INTEL(ATOM_SILVERMONT_D, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1056 VULNWL_INTEL(ATOM_SILVERMONT_MID, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1057 VULNWL_INTEL(ATOM_AIRMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1058 VULNWL_INTEL(XEON_PHI_KNL, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1059 VULNWL_INTEL(XEON_PHI_KNM, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1060
1061 VULNWL_INTEL(CORE_YONAH, NO_SSB),
1062
1063 VULNWL_INTEL(ATOM_AIRMONT_MID, NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1064 VULNWL_INTEL(ATOM_AIRMONT_NP, NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1065
1066 VULNWL_INTEL(ATOM_GOLDMONT, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1067 VULNWL_INTEL(ATOM_GOLDMONT_D, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1068 VULNWL_INTEL(ATOM_GOLDMONT_PLUS, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1069
1070 /*
1071 * Technically, swapgs isn't serializing on AMD (despite it previously
1072 * being documented as such in the APM). But according to AMD, %gs is
1073 * updated non-speculatively, and the issuing of %gs-relative memory
1074 * operands will be blocked until the %gs update completes, which is
1075 * good enough for our purposes.
1076 */
1077
1078 VULNWL_INTEL(ATOM_TREMONT_D, NO_ITLB_MULTIHIT),
1079
1080 /* AMD Family 0xf - 0x12 */
1081 VULNWL_AMD(0x0f, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1082 VULNWL_AMD(0x10, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1083 VULNWL_AMD(0x11, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1084 VULNWL_AMD(0x12, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1085
1086 /* FAMILY_ANY must be last, otherwise 0x0f - 0x12 matches won't work */
1087 VULNWL_AMD(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1088 VULNWL_HYGON(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
Olivier Deprez0e641232021-09-23 10:07:05 +02001089
1090 /* Zhaoxin Family 7 */
1091 VULNWL(CENTAUR, 7, X86_MODEL_ANY, NO_SPECTRE_V2),
1092 VULNWL(ZHAOXIN, 7, X86_MODEL_ANY, NO_SPECTRE_V2),
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001093 {}
1094};
1095
Olivier Deprez0e641232021-09-23 10:07:05 +02001096#define VULNBL_INTEL_STEPPINGS(model, steppings, issues) \
1097 X86_MATCH_VENDOR_FAM_MODEL_STEPPINGS_FEATURE(INTEL, 6, \
1098 INTEL_FAM6_##model, steppings, \
1099 X86_FEATURE_ANY, issues)
1100
1101#define SRBDS BIT(0)
1102
1103static const struct x86_cpu_id cpu_vuln_blacklist[] __initconst = {
1104 VULNBL_INTEL_STEPPINGS(IVYBRIDGE, X86_STEPPING_ANY, SRBDS),
1105 VULNBL_INTEL_STEPPINGS(HASWELL, X86_STEPPING_ANY, SRBDS),
1106 VULNBL_INTEL_STEPPINGS(HASWELL_L, X86_STEPPING_ANY, SRBDS),
1107 VULNBL_INTEL_STEPPINGS(HASWELL_G, X86_STEPPING_ANY, SRBDS),
1108 VULNBL_INTEL_STEPPINGS(BROADWELL_G, X86_STEPPING_ANY, SRBDS),
1109 VULNBL_INTEL_STEPPINGS(BROADWELL, X86_STEPPING_ANY, SRBDS),
1110 VULNBL_INTEL_STEPPINGS(SKYLAKE_L, X86_STEPPING_ANY, SRBDS),
1111 VULNBL_INTEL_STEPPINGS(SKYLAKE, X86_STEPPING_ANY, SRBDS),
1112 VULNBL_INTEL_STEPPINGS(KABYLAKE_L, X86_STEPPINGS(0x0, 0xC), SRBDS),
1113 VULNBL_INTEL_STEPPINGS(KABYLAKE, X86_STEPPINGS(0x0, 0xD), SRBDS),
1114 {}
1115};
1116
1117static bool __init cpu_matches(const struct x86_cpu_id *table, unsigned long which)
David Brazdil0f672f62019-12-10 10:32:29 +00001118{
Olivier Deprez0e641232021-09-23 10:07:05 +02001119 const struct x86_cpu_id *m = x86_match_cpu(table);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001120
David Brazdil0f672f62019-12-10 10:32:29 +00001121 return m && !!(m->driver_data & which);
1122}
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001123
David Brazdil0f672f62019-12-10 10:32:29 +00001124u64 x86_read_arch_cap_msr(void)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001125{
1126 u64 ia32_cap = 0;
1127
David Brazdil0f672f62019-12-10 10:32:29 +00001128 if (boot_cpu_has(X86_FEATURE_ARCH_CAPABILITIES))
1129 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
1130
1131 return ia32_cap;
1132}
1133
1134static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
1135{
1136 u64 ia32_cap = x86_read_arch_cap_msr();
1137
1138 /* Set ITLB_MULTIHIT bug if cpu is not in the whitelist and not mitigated */
Olivier Deprez0e641232021-09-23 10:07:05 +02001139 if (!cpu_matches(cpu_vuln_whitelist, NO_ITLB_MULTIHIT) &&
1140 !(ia32_cap & ARCH_CAP_PSCHANGE_MC_NO))
David Brazdil0f672f62019-12-10 10:32:29 +00001141 setup_force_cpu_bug(X86_BUG_ITLB_MULTIHIT);
1142
Olivier Deprez0e641232021-09-23 10:07:05 +02001143 if (cpu_matches(cpu_vuln_whitelist, NO_SPECULATION))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001144 return;
1145
1146 setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001147
Olivier Deprez0e641232021-09-23 10:07:05 +02001148 if (!cpu_matches(cpu_vuln_whitelist, NO_SPECTRE_V2))
1149 setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
1150
1151 if (!cpu_matches(cpu_vuln_whitelist, NO_SSB) &&
1152 !(ia32_cap & ARCH_CAP_SSB_NO) &&
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001153 !cpu_has(c, X86_FEATURE_AMD_SSB_NO))
1154 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
1155
1156 if (ia32_cap & ARCH_CAP_IBRS_ALL)
1157 setup_force_cpu_cap(X86_FEATURE_IBRS_ENHANCED);
1158
Olivier Deprez0e641232021-09-23 10:07:05 +02001159 if (!cpu_matches(cpu_vuln_whitelist, NO_MDS) &&
1160 !(ia32_cap & ARCH_CAP_MDS_NO)) {
David Brazdil0f672f62019-12-10 10:32:29 +00001161 setup_force_cpu_bug(X86_BUG_MDS);
Olivier Deprez0e641232021-09-23 10:07:05 +02001162 if (cpu_matches(cpu_vuln_whitelist, MSBDS_ONLY))
David Brazdil0f672f62019-12-10 10:32:29 +00001163 setup_force_cpu_bug(X86_BUG_MSBDS_ONLY);
1164 }
1165
Olivier Deprez0e641232021-09-23 10:07:05 +02001166 if (!cpu_matches(cpu_vuln_whitelist, NO_SWAPGS))
David Brazdil0f672f62019-12-10 10:32:29 +00001167 setup_force_cpu_bug(X86_BUG_SWAPGS);
1168
1169 /*
1170 * When the CPU is not mitigated for TAA (TAA_NO=0) set TAA bug when:
1171 * - TSX is supported or
1172 * - TSX_CTRL is present
1173 *
1174 * TSX_CTRL check is needed for cases when TSX could be disabled before
1175 * the kernel boot e.g. kexec.
1176 * TSX_CTRL check alone is not sufficient for cases when the microcode
1177 * update is not present or running as guest that don't get TSX_CTRL.
1178 */
1179 if (!(ia32_cap & ARCH_CAP_TAA_NO) &&
1180 (cpu_has(c, X86_FEATURE_RTM) ||
1181 (ia32_cap & ARCH_CAP_TSX_CTRL_MSR)))
1182 setup_force_cpu_bug(X86_BUG_TAA);
1183
Olivier Deprez0e641232021-09-23 10:07:05 +02001184 /*
1185 * SRBDS affects CPUs which support RDRAND or RDSEED and are listed
1186 * in the vulnerability blacklist.
1187 */
1188 if ((cpu_has(c, X86_FEATURE_RDRAND) ||
1189 cpu_has(c, X86_FEATURE_RDSEED)) &&
1190 cpu_matches(cpu_vuln_blacklist, SRBDS))
1191 setup_force_cpu_bug(X86_BUG_SRBDS);
1192
1193 if (cpu_matches(cpu_vuln_whitelist, NO_MELTDOWN))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001194 return;
1195
1196 /* Rogue Data Cache Load? No! */
1197 if (ia32_cap & ARCH_CAP_RDCL_NO)
1198 return;
1199
1200 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
1201
Olivier Deprez0e641232021-09-23 10:07:05 +02001202 if (cpu_matches(cpu_vuln_whitelist, NO_L1TF))
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001203 return;
1204
1205 setup_force_cpu_bug(X86_BUG_L1TF);
1206}
1207
1208/*
1209 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1210 * unfortunately, that's not true in practice because of early VIA
1211 * chips and (more importantly) broken virtualizers that are not easy
1212 * to detect. In the latter case it doesn't even *fail* reliably, so
1213 * probing for it doesn't even work. Disable it completely on 32-bit
1214 * unless we can find a reliable way to detect all the broken cases.
1215 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1216 */
1217static void detect_nopl(void)
1218{
1219#ifdef CONFIG_X86_32
1220 setup_clear_cpu_cap(X86_FEATURE_NOPL);
1221#else
1222 setup_force_cpu_cap(X86_FEATURE_NOPL);
1223#endif
1224}
1225
1226/*
1227 * Do minimum CPU detection early.
1228 * Fields really needed: vendor, cpuid_level, family, model, mask,
1229 * cache alignment.
1230 * The others are not touched to avoid unwanted side effects.
1231 *
1232 * WARNING: this function is only called on the boot CPU. Don't add code
1233 * here that is supposed to run on all CPUs.
1234 */
1235static void __init early_identify_cpu(struct cpuinfo_x86 *c)
1236{
1237#ifdef CONFIG_X86_64
1238 c->x86_clflush_size = 64;
1239 c->x86_phys_bits = 36;
1240 c->x86_virt_bits = 48;
1241#else
1242 c->x86_clflush_size = 32;
1243 c->x86_phys_bits = 32;
1244 c->x86_virt_bits = 32;
1245#endif
1246 c->x86_cache_alignment = c->x86_clflush_size;
1247
David Brazdil0f672f62019-12-10 10:32:29 +00001248 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001249 c->extended_cpuid_level = 0;
1250
David Brazdil0f672f62019-12-10 10:32:29 +00001251 if (!have_cpuid_p())
1252 identify_cpu_without_cpuid(c);
1253
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001254 /* cyrix could have cpuid enabled via c_identify()*/
1255 if (have_cpuid_p()) {
1256 cpu_detect(c);
1257 get_cpu_vendor(c);
1258 get_cpu_cap(c);
1259 get_cpu_address_sizes(c);
1260 setup_force_cpu_cap(X86_FEATURE_CPUID);
1261
1262 if (this_cpu->c_early_init)
1263 this_cpu->c_early_init(c);
1264
1265 c->cpu_index = 0;
1266 filter_cpuid_features(c, false);
1267
1268 if (this_cpu->c_bsp_init)
1269 this_cpu->c_bsp_init(c);
1270 } else {
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001271 setup_clear_cpu_cap(X86_FEATURE_CPUID);
1272 }
1273
1274 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1275
1276 cpu_set_bug_bits(c);
1277
1278 fpu__init_system(c);
1279
1280#ifdef CONFIG_X86_32
1281 /*
1282 * Regardless of whether PCID is enumerated, the SDM says
1283 * that it can't be enabled in 32-bit mode.
1284 */
1285 setup_clear_cpu_cap(X86_FEATURE_PCID);
1286#endif
1287
1288 /*
1289 * Later in the boot process pgtable_l5_enabled() relies on
1290 * cpu_feature_enabled(X86_FEATURE_LA57). If 5-level paging is not
1291 * enabled by this point we need to clear the feature bit to avoid
1292 * false-positives at the later stage.
1293 *
1294 * pgtable_l5_enabled() can be false here for several reasons:
1295 * - 5-level paging is disabled compile-time;
1296 * - it's 32-bit kernel;
1297 * - machine doesn't support 5-level paging;
1298 * - user specified 'no5lvl' in kernel command line.
1299 */
1300 if (!pgtable_l5_enabled())
1301 setup_clear_cpu_cap(X86_FEATURE_LA57);
1302
1303 detect_nopl();
1304}
1305
1306void __init early_cpu_init(void)
1307{
1308 const struct cpu_dev *const *cdev;
1309 int count = 0;
1310
1311#ifdef CONFIG_PROCESSOR_SELECT
1312 pr_info("KERNEL supported cpus:\n");
1313#endif
1314
1315 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1316 const struct cpu_dev *cpudev = *cdev;
1317
1318 if (count >= X86_VENDOR_NUM)
1319 break;
1320 cpu_devs[count] = cpudev;
1321 count++;
1322
1323#ifdef CONFIG_PROCESSOR_SELECT
1324 {
1325 unsigned int j;
1326
1327 for (j = 0; j < 2; j++) {
1328 if (!cpudev->c_ident[j])
1329 continue;
1330 pr_info(" %s %s\n", cpudev->c_vendor,
1331 cpudev->c_ident[j]);
1332 }
1333 }
1334#endif
1335 }
1336 early_identify_cpu(&boot_cpu_data);
1337}
1338
1339static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
1340{
1341#ifdef CONFIG_X86_64
1342 /*
1343 * Empirically, writing zero to a segment selector on AMD does
1344 * not clear the base, whereas writing zero to a segment
1345 * selector on Intel does clear the base. Intel's behavior
1346 * allows slightly faster context switches in the common case
1347 * where GS is unused by the prev and next threads.
1348 *
1349 * Since neither vendor documents this anywhere that I can see,
1350 * detect it directly instead of hardcoding the choice by
1351 * vendor.
1352 *
1353 * I've designated AMD's behavior as the "bug" because it's
1354 * counterintuitive and less friendly.
1355 */
1356
1357 unsigned long old_base, tmp;
1358 rdmsrl(MSR_FS_BASE, old_base);
1359 wrmsrl(MSR_FS_BASE, 1);
1360 loadsegment(fs, 0);
1361 rdmsrl(MSR_FS_BASE, tmp);
1362 if (tmp != 0)
1363 set_cpu_bug(c, X86_BUG_NULL_SEG);
1364 wrmsrl(MSR_FS_BASE, old_base);
1365#endif
1366}
1367
1368static void generic_identify(struct cpuinfo_x86 *c)
1369{
1370 c->extended_cpuid_level = 0;
1371
1372 if (!have_cpuid_p())
1373 identify_cpu_without_cpuid(c);
1374
1375 /* cyrix could have cpuid enabled via c_identify()*/
1376 if (!have_cpuid_p())
1377 return;
1378
1379 cpu_detect(c);
1380
1381 get_cpu_vendor(c);
1382
1383 get_cpu_cap(c);
1384
1385 get_cpu_address_sizes(c);
1386
1387 if (c->cpuid_level >= 0x00000001) {
1388 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1389#ifdef CONFIG_X86_32
1390# ifdef CONFIG_SMP
1391 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1392# else
1393 c->apicid = c->initial_apicid;
1394# endif
1395#endif
1396 c->phys_proc_id = c->initial_apicid;
1397 }
1398
1399 get_model_name(c); /* Default name */
1400
1401 detect_null_seg_behavior(c);
1402
1403 /*
1404 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1405 * systems that run Linux at CPL > 0 may or may not have the
1406 * issue, but, even if they have the issue, there's absolutely
1407 * nothing we can do about it because we can't use the real IRET
1408 * instruction.
1409 *
1410 * NB: For the time being, only 32-bit kernels support
1411 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1412 * whether to apply espfix using paravirt hooks. If any
1413 * non-paravirt system ever shows up that does *not* have the
1414 * ESPFIX issue, we can change this.
1415 */
1416#ifdef CONFIG_X86_32
David Brazdil0f672f62019-12-10 10:32:29 +00001417# ifdef CONFIG_PARAVIRT_XXL
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001418 do {
1419 extern void native_iret(void);
David Brazdil0f672f62019-12-10 10:32:29 +00001420 if (pv_ops.cpu.iret == native_iret)
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001421 set_cpu_bug(c, X86_BUG_ESPFIX);
1422 } while (0);
1423# else
1424 set_cpu_bug(c, X86_BUG_ESPFIX);
1425# endif
1426#endif
1427}
1428
1429static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1430{
1431 /*
1432 * The heavy lifting of max_rmid and cache_occ_scale are handled
1433 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
1434 * in case CQM bits really aren't there in this CPU.
1435 */
1436 if (c != &boot_cpu_data) {
1437 boot_cpu_data.x86_cache_max_rmid =
1438 min(boot_cpu_data.x86_cache_max_rmid,
1439 c->x86_cache_max_rmid);
1440 }
1441}
1442
1443/*
1444 * Validate that ACPI/mptables have the same information about the
1445 * effective APIC id and update the package map.
1446 */
1447static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1448{
1449#ifdef CONFIG_SMP
1450 unsigned int apicid, cpu = smp_processor_id();
1451
1452 apicid = apic->cpu_present_to_apicid(cpu);
1453
1454 if (apicid != c->apicid) {
1455 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1456 cpu, apicid, c->initial_apicid);
1457 }
1458 BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
David Brazdil0f672f62019-12-10 10:32:29 +00001459 BUG_ON(topology_update_die_map(c->cpu_die_id, cpu));
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001460#else
1461 c->logical_proc_id = 0;
1462#endif
1463}
1464
1465/*
1466 * This does the hard work of actually picking apart the CPU stuff...
1467 */
1468static void identify_cpu(struct cpuinfo_x86 *c)
1469{
1470 int i;
1471
1472 c->loops_per_jiffy = loops_per_jiffy;
1473 c->x86_cache_size = 0;
1474 c->x86_vendor = X86_VENDOR_UNKNOWN;
1475 c->x86_model = c->x86_stepping = 0; /* So far unknown... */
1476 c->x86_vendor_id[0] = '\0'; /* Unset */
1477 c->x86_model_id[0] = '\0'; /* Unset */
1478 c->x86_max_cores = 1;
1479 c->x86_coreid_bits = 0;
1480 c->cu_id = 0xff;
1481#ifdef CONFIG_X86_64
1482 c->x86_clflush_size = 64;
1483 c->x86_phys_bits = 36;
1484 c->x86_virt_bits = 48;
1485#else
1486 c->cpuid_level = -1; /* CPUID not detected */
1487 c->x86_clflush_size = 32;
1488 c->x86_phys_bits = 32;
1489 c->x86_virt_bits = 32;
1490#endif
1491 c->x86_cache_alignment = c->x86_clflush_size;
David Brazdil0f672f62019-12-10 10:32:29 +00001492 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001493
1494 generic_identify(c);
1495
1496 if (this_cpu->c_identify)
1497 this_cpu->c_identify(c);
1498
1499 /* Clear/Set all flags overridden by options, after probe */
1500 apply_forced_caps(c);
1501
1502#ifdef CONFIG_X86_64
1503 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1504#endif
1505
1506 /*
1507 * Vendor-specific initialization. In this section we
1508 * canonicalize the feature flags, meaning if there are
1509 * features a certain CPU supports which CPUID doesn't
1510 * tell us, CPUID claiming incorrect flags, or other bugs,
1511 * we handle them here.
1512 *
1513 * At the end of this section, c->x86_capability better
1514 * indicate the features this CPU genuinely supports!
1515 */
1516 if (this_cpu->c_init)
1517 this_cpu->c_init(c);
1518
1519 /* Disable the PN if appropriate */
1520 squash_the_stupid_serial_number(c);
1521
1522 /* Set up SMEP/SMAP/UMIP */
1523 setup_smep(c);
1524 setup_smap(c);
1525 setup_umip(c);
1526
1527 /*
1528 * The vendor-specific functions might have changed features.
1529 * Now we do "generic changes."
1530 */
1531
1532 /* Filter out anything that depends on CPUID levels we don't have */
1533 filter_cpuid_features(c, true);
1534
1535 /* If the model name is still unset, do table lookup. */
1536 if (!c->x86_model_id[0]) {
1537 const char *p;
1538 p = table_lookup_model(c);
1539 if (p)
1540 strcpy(c->x86_model_id, p);
1541 else
1542 /* Last resort... */
1543 sprintf(c->x86_model_id, "%02x/%02x",
1544 c->x86, c->x86_model);
1545 }
1546
1547#ifdef CONFIG_X86_64
1548 detect_ht(c);
1549#endif
1550
1551 x86_init_rdrand(c);
1552 x86_init_cache_qos(c);
1553 setup_pku(c);
1554
1555 /*
1556 * Clear/Set all flags overridden by options, need do it
1557 * before following smp all cpus cap AND.
1558 */
1559 apply_forced_caps(c);
1560
1561 /*
1562 * On SMP, boot_cpu_data holds the common feature set between
1563 * all CPUs; so make sure that we indicate which features are
1564 * common between the CPUs. The first time this routine gets
1565 * executed, c == &boot_cpu_data.
1566 */
1567 if (c != &boot_cpu_data) {
1568 /* AND the already accumulated flags with these */
1569 for (i = 0; i < NCAPINTS; i++)
1570 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1571
1572 /* OR, i.e. replicate the bug flags */
1573 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1574 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1575 }
1576
1577 /* Init Machine Check Exception if available. */
1578 mcheck_cpu_init(c);
1579
1580 select_idle_routine(c);
1581
1582#ifdef CONFIG_NUMA
1583 numa_add_cpu(smp_processor_id());
1584#endif
1585}
1586
1587/*
1588 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1589 * on 32-bit kernels:
1590 */
1591#ifdef CONFIG_X86_32
1592void enable_sep_cpu(void)
1593{
1594 struct tss_struct *tss;
1595 int cpu;
1596
1597 if (!boot_cpu_has(X86_FEATURE_SEP))
1598 return;
1599
1600 cpu = get_cpu();
1601 tss = &per_cpu(cpu_tss_rw, cpu);
1602
1603 /*
1604 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1605 * see the big comment in struct x86_hw_tss's definition.
1606 */
1607
1608 tss->x86_tss.ss1 = __KERNEL_CS;
1609 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1610 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
1611 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1612
1613 put_cpu();
1614}
1615#endif
1616
1617void __init identify_boot_cpu(void)
1618{
1619 identify_cpu(&boot_cpu_data);
1620#ifdef CONFIG_X86_32
1621 sysenter_setup();
1622 enable_sep_cpu();
1623#endif
1624 cpu_detect_tlb(&boot_cpu_data);
David Brazdil0f672f62019-12-10 10:32:29 +00001625 setup_cr_pinning();
1626
1627 tsx_init();
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001628}
1629
1630void identify_secondary_cpu(struct cpuinfo_x86 *c)
1631{
1632 BUG_ON(c == &boot_cpu_data);
1633 identify_cpu(c);
1634#ifdef CONFIG_X86_32
1635 enable_sep_cpu();
1636#endif
1637 mtrr_ap_init();
1638 validate_apic_and_package_id(c);
1639 x86_spec_ctrl_setup_ap();
Olivier Deprez0e641232021-09-23 10:07:05 +02001640 update_srbds_msr();
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001641}
1642
1643static __init int setup_noclflush(char *arg)
1644{
1645 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1646 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1647 return 1;
1648}
1649__setup("noclflush", setup_noclflush);
1650
1651void print_cpu_info(struct cpuinfo_x86 *c)
1652{
1653 const char *vendor = NULL;
1654
1655 if (c->x86_vendor < X86_VENDOR_NUM) {
1656 vendor = this_cpu->c_vendor;
1657 } else {
1658 if (c->cpuid_level >= 0)
1659 vendor = c->x86_vendor_id;
1660 }
1661
1662 if (vendor && !strstr(c->x86_model_id, vendor))
1663 pr_cont("%s ", vendor);
1664
1665 if (c->x86_model_id[0])
1666 pr_cont("%s", c->x86_model_id);
1667 else
1668 pr_cont("%d86", c->x86);
1669
1670 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1671
1672 if (c->x86_stepping || c->cpuid_level >= 0)
1673 pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
1674 else
1675 pr_cont(")\n");
1676}
1677
1678/*
1679 * clearcpuid= was already parsed in fpu__init_parse_early_param.
1680 * But we need to keep a dummy __setup around otherwise it would
1681 * show up as an environment variable for init.
1682 */
1683static __init int setup_clearcpuid(char *arg)
1684{
1685 return 1;
1686}
1687__setup("clearcpuid=", setup_clearcpuid);
1688
1689#ifdef CONFIG_X86_64
David Brazdil0f672f62019-12-10 10:32:29 +00001690DEFINE_PER_CPU_FIRST(struct fixed_percpu_data,
1691 fixed_percpu_data) __aligned(PAGE_SIZE) __visible;
1692EXPORT_PER_CPU_SYMBOL_GPL(fixed_percpu_data);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001693
1694/*
1695 * The following percpu variables are hot. Align current_task to
1696 * cacheline size such that they fall in the same cacheline.
1697 */
1698DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1699 &init_task;
1700EXPORT_PER_CPU_SYMBOL(current_task);
1701
David Brazdil0f672f62019-12-10 10:32:29 +00001702DEFINE_PER_CPU(struct irq_stack *, hardirq_stack_ptr);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001703DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1704
1705DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1706EXPORT_PER_CPU_SYMBOL(__preempt_count);
1707
1708/* May not be marked __init: used by software suspend */
1709void syscall_init(void)
1710{
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001711 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
David Brazdil0f672f62019-12-10 10:32:29 +00001712 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001713
1714#ifdef CONFIG_IA32_EMULATION
1715 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1716 /*
1717 * This only works on Intel CPUs.
1718 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1719 * This does not cause SYSENTER to jump to the wrong location, because
1720 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1721 */
1722 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
David Brazdil0f672f62019-12-10 10:32:29 +00001723 wrmsrl_safe(MSR_IA32_SYSENTER_ESP,
1724 (unsigned long)(cpu_entry_stack(smp_processor_id()) + 1));
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001725 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1726#else
1727 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1728 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1729 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1730 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1731#endif
1732
1733 /* Flags to clear on syscall */
1734 wrmsrl(MSR_SYSCALL_MASK,
1735 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1736 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1737}
1738
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001739DEFINE_PER_CPU(int, debug_stack_usage);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001740DEFINE_PER_CPU(u32, debug_idt_ctr);
1741
1742void debug_stack_set_zero(void)
1743{
1744 this_cpu_inc(debug_idt_ctr);
1745 load_current_idt();
1746}
1747NOKPROBE_SYMBOL(debug_stack_set_zero);
1748
1749void debug_stack_reset(void)
1750{
1751 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1752 return;
1753 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1754 load_current_idt();
1755}
1756NOKPROBE_SYMBOL(debug_stack_reset);
1757
1758#else /* CONFIG_X86_64 */
1759
1760DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1761EXPORT_PER_CPU_SYMBOL(current_task);
1762DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1763EXPORT_PER_CPU_SYMBOL(__preempt_count);
1764
1765/*
1766 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1767 * the top of the kernel stack. Use an extra percpu variable to track the
1768 * top of the kernel stack directly.
1769 */
1770DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1771 (unsigned long)&init_thread_union + THREAD_SIZE;
1772EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1773
1774#ifdef CONFIG_STACKPROTECTOR
1775DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1776#endif
1777
1778#endif /* CONFIG_X86_64 */
1779
1780/*
1781 * Clear all 6 debug registers:
1782 */
1783static void clear_all_debug_regs(void)
1784{
1785 int i;
1786
1787 for (i = 0; i < 8; i++) {
1788 /* Ignore db4, db5 */
1789 if ((i == 4) || (i == 5))
1790 continue;
1791
1792 set_debugreg(0, i);
1793 }
1794}
1795
1796#ifdef CONFIG_KGDB
1797/*
1798 * Restore debug regs if using kgdbwait and you have a kernel debugger
1799 * connection established.
1800 */
1801static void dbg_restore_debug_regs(void)
1802{
1803 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1804 arch_kgdb_ops.correct_hw_break();
1805}
1806#else /* ! CONFIG_KGDB */
1807#define dbg_restore_debug_regs()
1808#endif /* ! CONFIG_KGDB */
1809
1810static void wait_for_master_cpu(int cpu)
1811{
1812#ifdef CONFIG_SMP
1813 /*
1814 * wait for ACK from master CPU before continuing
1815 * with AP initialization
1816 */
1817 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1818 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1819 cpu_relax();
1820#endif
1821}
1822
David Brazdil0f672f62019-12-10 10:32:29 +00001823#ifdef CONFIG_X86_64
1824static void setup_getcpu(int cpu)
1825{
1826 unsigned long cpudata = vdso_encode_cpunode(cpu, early_cpu_to_node(cpu));
1827 struct desc_struct d = { };
1828
Olivier Deprez0e641232021-09-23 10:07:05 +02001829 if (boot_cpu_has(X86_FEATURE_RDTSCP) || boot_cpu_has(X86_FEATURE_RDPID))
David Brazdil0f672f62019-12-10 10:32:29 +00001830 write_rdtscp_aux(cpudata);
1831
1832 /* Store CPU and node number in limit. */
1833 d.limit0 = cpudata;
1834 d.limit1 = cpudata >> 16;
1835
1836 d.type = 5; /* RO data, expand down, accessed */
1837 d.dpl = 3; /* Visible to user code */
1838 d.s = 1; /* Not a system segment */
1839 d.p = 1; /* Present */
1840 d.d = 1; /* 32-bit */
1841
1842 write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_CPUNODE, &d, DESCTYPE_S);
1843}
1844#endif
1845
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001846/*
1847 * cpu_init() initializes state that is per-CPU. Some data is already
1848 * initialized (naturally) in the bootstrap process, such as the GDT
1849 * and IDT. We reload them nevertheless, this function acts as a
1850 * 'CPU state barrier', nothing should get across.
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001851 */
1852#ifdef CONFIG_X86_64
1853
1854void cpu_init(void)
1855{
David Brazdil0f672f62019-12-10 10:32:29 +00001856 int cpu = raw_smp_processor_id();
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001857 struct task_struct *me;
1858 struct tss_struct *t;
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001859 int i;
1860
1861 wait_for_master_cpu(cpu);
1862
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001863 if (cpu)
1864 load_ucode_ap();
1865
1866 t = &per_cpu(cpu_tss_rw, cpu);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001867
1868#ifdef CONFIG_NUMA
1869 if (this_cpu_read(numa_node) == 0 &&
1870 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1871 set_numa_node(early_cpu_to_node(cpu));
1872#endif
David Brazdil0f672f62019-12-10 10:32:29 +00001873 setup_getcpu(cpu);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001874
1875 me = current;
1876
1877 pr_debug("Initializing CPU#%d\n", cpu);
1878
1879 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1880
1881 /*
1882 * Initialize the per-CPU GDT with the boot GDT,
1883 * and set up the GDT descriptor:
1884 */
1885
1886 switch_to_new_gdt(cpu);
1887 loadsegment(fs, 0);
1888
1889 load_current_idt();
1890
1891 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1892 syscall_init();
1893
1894 wrmsrl(MSR_FS_BASE, 0);
1895 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1896 barrier();
1897
1898 x86_configure_nx();
1899 x2apic_setup();
1900
1901 /*
1902 * set up and load the per-CPU TSS
1903 */
David Brazdil0f672f62019-12-10 10:32:29 +00001904 if (!t->x86_tss.ist[0]) {
1905 t->x86_tss.ist[IST_INDEX_DF] = __this_cpu_ist_top_va(DF);
1906 t->x86_tss.ist[IST_INDEX_NMI] = __this_cpu_ist_top_va(NMI);
1907 t->x86_tss.ist[IST_INDEX_DB] = __this_cpu_ist_top_va(DB);
1908 t->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001909 }
1910
1911 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1912
1913 /*
1914 * <= is required because the CPU will access up to
1915 * 8 bits beyond the end of the IO permission bitmap.
1916 */
1917 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1918 t->io_bitmap[i] = ~0UL;
1919
1920 mmgrab(&init_mm);
1921 me->active_mm = &init_mm;
1922 BUG_ON(me->mm);
1923 initialize_tlbstate_and_flush();
1924 enter_lazy_tlb(&init_mm, me);
1925
1926 /*
1927 * Initialize the TSS. sp0 points to the entry trampoline stack
1928 * regardless of what task is running.
1929 */
1930 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1931 load_TR_desc();
1932 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1933
1934 load_mm_ldt(&init_mm);
1935
1936 clear_all_debug_regs();
1937 dbg_restore_debug_regs();
1938
1939 fpu__init_cpu();
1940
1941 if (is_uv_system())
1942 uv_cpu_init();
1943
1944 load_fixmap_gdt(cpu);
1945}
1946
1947#else
1948
1949void cpu_init(void)
1950{
1951 int cpu = smp_processor_id();
1952 struct task_struct *curr = current;
1953 struct tss_struct *t = &per_cpu(cpu_tss_rw, cpu);
1954
1955 wait_for_master_cpu(cpu);
1956
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00001957 show_ucode_info_early();
1958
1959 pr_info("Initializing CPU#%d\n", cpu);
1960
1961 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1962 boot_cpu_has(X86_FEATURE_TSC) ||
1963 boot_cpu_has(X86_FEATURE_DE))
1964 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1965
1966 load_current_idt();
1967 switch_to_new_gdt(cpu);
1968
1969 /*
1970 * Set up and load the per-CPU TSS and LDT
1971 */
1972 mmgrab(&init_mm);
1973 curr->active_mm = &init_mm;
1974 BUG_ON(curr->mm);
1975 initialize_tlbstate_and_flush();
1976 enter_lazy_tlb(&init_mm, curr);
1977
1978 /*
1979 * Initialize the TSS. sp0 points to the entry trampoline stack
1980 * regardless of what task is running.
1981 */
1982 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1983 load_TR_desc();
1984 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1985
1986 load_mm_ldt(&init_mm);
1987
1988 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1989
1990#ifdef CONFIG_DOUBLEFAULT
1991 /* Set up doublefault TSS pointer in the GDT */
1992 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1993#endif
1994
1995 clear_all_debug_regs();
1996 dbg_restore_debug_regs();
1997
1998 fpu__init_cpu();
1999
2000 load_fixmap_gdt(cpu);
2001}
2002#endif
2003
Andrew Scullb4b6d4a2019-01-02 15:54:55 +00002004/*
2005 * The microcode loader calls this upon late microcode load to recheck features,
2006 * only when microcode has been updated. Caller holds microcode_mutex and CPU
2007 * hotplug lock.
2008 */
2009void microcode_check(void)
2010{
2011 struct cpuinfo_x86 info;
2012
2013 perf_check_microcode();
2014
2015 /* Reload CPUID max function as it might've changed. */
2016 info.cpuid_level = cpuid_eax(0);
2017
2018 /*
2019 * Copy all capability leafs to pick up the synthetic ones so that
2020 * memcmp() below doesn't fail on that. The ones coming from CPUID will
2021 * get overwritten in get_cpu_cap().
2022 */
2023 memcpy(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability));
2024
2025 get_cpu_cap(&info);
2026
2027 if (!memcmp(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability)))
2028 return;
2029
2030 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
2031 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");
2032}
David Brazdil0f672f62019-12-10 10:32:29 +00002033
2034/*
2035 * Invoked from core CPU hotplug code after hotplug operations
2036 */
2037void arch_smt_update(void)
2038{
2039 /* Handle the speculative execution misfeatures */
2040 cpu_bugs_smt_update();
2041 /* Check whether IPI broadcasting can be enabled */
2042 apic_smt_update();
2043}