Loading...
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/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
357 pr_info_once("x86/cpu: User Mode Instruction Prevention (UMIP) activated\n");
358
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
369static DEFINE_STATIC_KEY_FALSE_RO(cr_pinning);
370static unsigned long cr4_pinned_bits __ro_after_init;
371
372void native_write_cr0(unsigned long val)
373{
374 unsigned long bits_missing = 0;
375
376set_register:
377 asm volatile("mov %0,%%cr0": "+r" (val), "+m" (__force_order));
378
379 if (static_branch_likely(&cr_pinning)) {
380 if (unlikely((val & X86_CR0_WP) != X86_CR0_WP)) {
381 bits_missing = X86_CR0_WP;
382 val |= bits_missing;
383 goto set_register;
384 }
385 /* Warn after we've set the missing bits. */
386 WARN_ONCE(bits_missing, "CR0 WP bit went missing!?\n");
387 }
388}
389EXPORT_SYMBOL(native_write_cr0);
390
391void native_write_cr4(unsigned long val)
392{
393 unsigned long bits_missing = 0;
394
395set_register:
396 asm volatile("mov %0,%%cr4": "+r" (val), "+m" (cr4_pinned_bits));
397
398 if (static_branch_likely(&cr_pinning)) {
399 if (unlikely((val & cr4_pinned_bits) != cr4_pinned_bits)) {
400 bits_missing = ~val & cr4_pinned_bits;
401 val |= bits_missing;
402 goto set_register;
403 }
404 /* Warn after we've set the missing bits. */
405 WARN_ONCE(bits_missing, "CR4 bits went missing: %lx!?\n",
406 bits_missing);
407 }
408}
409EXPORT_SYMBOL(native_write_cr4);
410
411void cr4_init(void)
412{
413 unsigned long cr4 = __read_cr4();
414
415 if (boot_cpu_has(X86_FEATURE_PCID))
416 cr4 |= X86_CR4_PCIDE;
417 if (static_branch_likely(&cr_pinning))
418 cr4 |= cr4_pinned_bits;
419
420 __write_cr4(cr4);
421
422 /* Initialize cr4 shadow for this CPU. */
423 this_cpu_write(cpu_tlbstate.cr4, cr4);
424}
425
426/*
427 * Once CPU feature detection is finished (and boot params have been
428 * parsed), record any of the sensitive CR bits that are set, and
429 * enable CR pinning.
430 */
431static void __init setup_cr_pinning(void)
432{
433 unsigned long mask;
434
435 mask = (X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_UMIP);
436 cr4_pinned_bits = this_cpu_read(cpu_tlbstate.cr4) & mask;
437 static_key_enable(&cr_pinning.key);
438}
439
440/*
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{
447 struct pkru_state *pk;
448
449 /* 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);
459 pk = get_xsave_addr(&init_fpstate.xsave, XFEATURE_PKRU);
460 if (pk)
461 pk->pkru = init_pkru_value;
462 /*
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 */
467 get_cpu_cap(c);
468}
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
587/* 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
875static 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
899void 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;
921
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 }
927 }
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
936 /* 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);
966 init_cqm(c);
967
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
1019#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)
1027
1028#define VULNWL(_vendor, _family, _model, _whitelist) \
1029 { X86_VENDOR_##_vendor, _family, _model, X86_FEATURE_ANY, _whitelist }
1030
1031#define VULNWL_INTEL(model, whitelist) \
1032 VULNWL(INTEL, 6, INTEL_FAM6_##model, whitelist)
1033
1034#define VULNWL_AMD(family, whitelist) \
1035 VULNWL(AMD, family, X86_MODEL_ANY, whitelist)
1036
1037#define VULNWL_HYGON(family, whitelist) \
1038 VULNWL(HYGON, family, X86_MODEL_ANY, whitelist)
1039
1040static const __initconst struct x86_cpu_id cpu_vuln_whitelist[] = {
1041 VULNWL(ANY, 4, X86_MODEL_ANY, NO_SPECULATION),
1042 VULNWL(CENTAUR, 5, X86_MODEL_ANY, NO_SPECULATION),
1043 VULNWL(INTEL, 5, X86_MODEL_ANY, NO_SPECULATION),
1044 VULNWL(NSC, 5, X86_MODEL_ANY, NO_SPECULATION),
1045
1046 /* Intel Family 6 */
1047 VULNWL_INTEL(ATOM_SALTWELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
1048 VULNWL_INTEL(ATOM_SALTWELL_TABLET, NO_SPECULATION | NO_ITLB_MULTIHIT),
1049 VULNWL_INTEL(ATOM_SALTWELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
1050 VULNWL_INTEL(ATOM_BONNELL, NO_SPECULATION | NO_ITLB_MULTIHIT),
1051 VULNWL_INTEL(ATOM_BONNELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT),
1052
1053 VULNWL_INTEL(ATOM_SILVERMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1054 VULNWL_INTEL(ATOM_SILVERMONT_D, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1055 VULNWL_INTEL(ATOM_SILVERMONT_MID, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1056 VULNWL_INTEL(ATOM_AIRMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1057 VULNWL_INTEL(XEON_PHI_KNL, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1058 VULNWL_INTEL(XEON_PHI_KNM, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1059
1060 VULNWL_INTEL(CORE_YONAH, NO_SSB),
1061
1062 VULNWL_INTEL(ATOM_AIRMONT_MID, NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1063 VULNWL_INTEL(ATOM_AIRMONT_NP, NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1064
1065 VULNWL_INTEL(ATOM_GOLDMONT, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1066 VULNWL_INTEL(ATOM_GOLDMONT_D, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1067 VULNWL_INTEL(ATOM_GOLDMONT_PLUS, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1068
1069 /*
1070 * Technically, swapgs isn't serializing on AMD (despite it previously
1071 * being documented as such in the APM). But according to AMD, %gs is
1072 * updated non-speculatively, and the issuing of %gs-relative memory
1073 * operands will be blocked until the %gs update completes, which is
1074 * good enough for our purposes.
1075 */
1076
1077 VULNWL_INTEL(ATOM_TREMONT_D, NO_ITLB_MULTIHIT),
1078
1079 /* AMD Family 0xf - 0x12 */
1080 VULNWL_AMD(0x0f, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1081 VULNWL_AMD(0x10, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1082 VULNWL_AMD(0x11, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1083 VULNWL_AMD(0x12, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1084
1085 /* FAMILY_ANY must be last, otherwise 0x0f - 0x12 matches won't work */
1086 VULNWL_AMD(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1087 VULNWL_HYGON(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1088 {}
1089};
1090
1091static bool __init cpu_matches(unsigned long which)
1092{
1093 const struct x86_cpu_id *m = x86_match_cpu(cpu_vuln_whitelist);
1094
1095 return m && !!(m->driver_data & which);
1096}
1097
1098u64 x86_read_arch_cap_msr(void)
1099{
1100 u64 ia32_cap = 0;
1101
1102 if (boot_cpu_has(X86_FEATURE_ARCH_CAPABILITIES))
1103 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
1104
1105 return ia32_cap;
1106}
1107
1108static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
1109{
1110 u64 ia32_cap = x86_read_arch_cap_msr();
1111
1112 /* Set ITLB_MULTIHIT bug if cpu is not in the whitelist and not mitigated */
1113 if (!cpu_matches(NO_ITLB_MULTIHIT) && !(ia32_cap & ARCH_CAP_PSCHANGE_MC_NO))
1114 setup_force_cpu_bug(X86_BUG_ITLB_MULTIHIT);
1115
1116 if (cpu_matches(NO_SPECULATION))
1117 return;
1118
1119 setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
1120 setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
1121
1122 if (!cpu_matches(NO_SSB) && !(ia32_cap & ARCH_CAP_SSB_NO) &&
1123 !cpu_has(c, X86_FEATURE_AMD_SSB_NO))
1124 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
1125
1126 if (ia32_cap & ARCH_CAP_IBRS_ALL)
1127 setup_force_cpu_cap(X86_FEATURE_IBRS_ENHANCED);
1128
1129 if (!cpu_matches(NO_MDS) && !(ia32_cap & ARCH_CAP_MDS_NO)) {
1130 setup_force_cpu_bug(X86_BUG_MDS);
1131 if (cpu_matches(MSBDS_ONLY))
1132 setup_force_cpu_bug(X86_BUG_MSBDS_ONLY);
1133 }
1134
1135 if (!cpu_matches(NO_SWAPGS))
1136 setup_force_cpu_bug(X86_BUG_SWAPGS);
1137
1138 /*
1139 * When the CPU is not mitigated for TAA (TAA_NO=0) set TAA bug when:
1140 * - TSX is supported or
1141 * - TSX_CTRL is present
1142 *
1143 * TSX_CTRL check is needed for cases when TSX could be disabled before
1144 * the kernel boot e.g. kexec.
1145 * TSX_CTRL check alone is not sufficient for cases when the microcode
1146 * update is not present or running as guest that don't get TSX_CTRL.
1147 */
1148 if (!(ia32_cap & ARCH_CAP_TAA_NO) &&
1149 (cpu_has(c, X86_FEATURE_RTM) ||
1150 (ia32_cap & ARCH_CAP_TSX_CTRL_MSR)))
1151 setup_force_cpu_bug(X86_BUG_TAA);
1152
1153 if (cpu_matches(NO_MELTDOWN))
1154 return;
1155
1156 /* Rogue Data Cache Load? No! */
1157 if (ia32_cap & ARCH_CAP_RDCL_NO)
1158 return;
1159
1160 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
1161
1162 if (cpu_matches(NO_L1TF))
1163 return;
1164
1165 setup_force_cpu_bug(X86_BUG_L1TF);
1166}
1167
1168/*
1169 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1170 * unfortunately, that's not true in practice because of early VIA
1171 * chips and (more importantly) broken virtualizers that are not easy
1172 * to detect. In the latter case it doesn't even *fail* reliably, so
1173 * probing for it doesn't even work. Disable it completely on 32-bit
1174 * unless we can find a reliable way to detect all the broken cases.
1175 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1176 */
1177static void detect_nopl(void)
1178{
1179#ifdef CONFIG_X86_32
1180 setup_clear_cpu_cap(X86_FEATURE_NOPL);
1181#else
1182 setup_force_cpu_cap(X86_FEATURE_NOPL);
1183#endif
1184}
1185
1186/*
1187 * Do minimum CPU detection early.
1188 * Fields really needed: vendor, cpuid_level, family, model, mask,
1189 * cache alignment.
1190 * The others are not touched to avoid unwanted side effects.
1191 *
1192 * WARNING: this function is only called on the boot CPU. Don't add code
1193 * here that is supposed to run on all CPUs.
1194 */
1195static void __init early_identify_cpu(struct cpuinfo_x86 *c)
1196{
1197#ifdef CONFIG_X86_64
1198 c->x86_clflush_size = 64;
1199 c->x86_phys_bits = 36;
1200 c->x86_virt_bits = 48;
1201#else
1202 c->x86_clflush_size = 32;
1203 c->x86_phys_bits = 32;
1204 c->x86_virt_bits = 32;
1205#endif
1206 c->x86_cache_alignment = c->x86_clflush_size;
1207
1208 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1209 c->extended_cpuid_level = 0;
1210
1211 if (!have_cpuid_p())
1212 identify_cpu_without_cpuid(c);
1213
1214 /* cyrix could have cpuid enabled via c_identify()*/
1215 if (have_cpuid_p()) {
1216 cpu_detect(c);
1217 get_cpu_vendor(c);
1218 get_cpu_cap(c);
1219 get_cpu_address_sizes(c);
1220 setup_force_cpu_cap(X86_FEATURE_CPUID);
1221
1222 if (this_cpu->c_early_init)
1223 this_cpu->c_early_init(c);
1224
1225 c->cpu_index = 0;
1226 filter_cpuid_features(c, false);
1227
1228 if (this_cpu->c_bsp_init)
1229 this_cpu->c_bsp_init(c);
1230 } else {
1231 setup_clear_cpu_cap(X86_FEATURE_CPUID);
1232 }
1233
1234 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1235
1236 cpu_set_bug_bits(c);
1237
1238 fpu__init_system(c);
1239
1240#ifdef CONFIG_X86_32
1241 /*
1242 * Regardless of whether PCID is enumerated, the SDM says
1243 * that it can't be enabled in 32-bit mode.
1244 */
1245 setup_clear_cpu_cap(X86_FEATURE_PCID);
1246#endif
1247
1248 /*
1249 * Later in the boot process pgtable_l5_enabled() relies on
1250 * cpu_feature_enabled(X86_FEATURE_LA57). If 5-level paging is not
1251 * enabled by this point we need to clear the feature bit to avoid
1252 * false-positives at the later stage.
1253 *
1254 * pgtable_l5_enabled() can be false here for several reasons:
1255 * - 5-level paging is disabled compile-time;
1256 * - it's 32-bit kernel;
1257 * - machine doesn't support 5-level paging;
1258 * - user specified 'no5lvl' in kernel command line.
1259 */
1260 if (!pgtable_l5_enabled())
1261 setup_clear_cpu_cap(X86_FEATURE_LA57);
1262
1263 detect_nopl();
1264}
1265
1266void __init early_cpu_init(void)
1267{
1268 const struct cpu_dev *const *cdev;
1269 int count = 0;
1270
1271#ifdef CONFIG_PROCESSOR_SELECT
1272 pr_info("KERNEL supported cpus:\n");
1273#endif
1274
1275 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1276 const struct cpu_dev *cpudev = *cdev;
1277
1278 if (count >= X86_VENDOR_NUM)
1279 break;
1280 cpu_devs[count] = cpudev;
1281 count++;
1282
1283#ifdef CONFIG_PROCESSOR_SELECT
1284 {
1285 unsigned int j;
1286
1287 for (j = 0; j < 2; j++) {
1288 if (!cpudev->c_ident[j])
1289 continue;
1290 pr_info(" %s %s\n", cpudev->c_vendor,
1291 cpudev->c_ident[j]);
1292 }
1293 }
1294#endif
1295 }
1296 early_identify_cpu(&boot_cpu_data);
1297}
1298
1299static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
1300{
1301#ifdef CONFIG_X86_64
1302 /*
1303 * Empirically, writing zero to a segment selector on AMD does
1304 * not clear the base, whereas writing zero to a segment
1305 * selector on Intel does clear the base. Intel's behavior
1306 * allows slightly faster context switches in the common case
1307 * where GS is unused by the prev and next threads.
1308 *
1309 * Since neither vendor documents this anywhere that I can see,
1310 * detect it directly instead of hardcoding the choice by
1311 * vendor.
1312 *
1313 * I've designated AMD's behavior as the "bug" because it's
1314 * counterintuitive and less friendly.
1315 */
1316
1317 unsigned long old_base, tmp;
1318 rdmsrl(MSR_FS_BASE, old_base);
1319 wrmsrl(MSR_FS_BASE, 1);
1320 loadsegment(fs, 0);
1321 rdmsrl(MSR_FS_BASE, tmp);
1322 if (tmp != 0)
1323 set_cpu_bug(c, X86_BUG_NULL_SEG);
1324 wrmsrl(MSR_FS_BASE, old_base);
1325#endif
1326}
1327
1328static void generic_identify(struct cpuinfo_x86 *c)
1329{
1330 c->extended_cpuid_level = 0;
1331
1332 if (!have_cpuid_p())
1333 identify_cpu_without_cpuid(c);
1334
1335 /* cyrix could have cpuid enabled via c_identify()*/
1336 if (!have_cpuid_p())
1337 return;
1338
1339 cpu_detect(c);
1340
1341 get_cpu_vendor(c);
1342
1343 get_cpu_cap(c);
1344
1345 get_cpu_address_sizes(c);
1346
1347 if (c->cpuid_level >= 0x00000001) {
1348 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1349#ifdef CONFIG_X86_32
1350# ifdef CONFIG_SMP
1351 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1352# else
1353 c->apicid = c->initial_apicid;
1354# endif
1355#endif
1356 c->phys_proc_id = c->initial_apicid;
1357 }
1358
1359 get_model_name(c); /* Default name */
1360
1361 detect_null_seg_behavior(c);
1362
1363 /*
1364 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
1365 * systems that run Linux at CPL > 0 may or may not have the
1366 * issue, but, even if they have the issue, there's absolutely
1367 * nothing we can do about it because we can't use the real IRET
1368 * instruction.
1369 *
1370 * NB: For the time being, only 32-bit kernels support
1371 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
1372 * whether to apply espfix using paravirt hooks. If any
1373 * non-paravirt system ever shows up that does *not* have the
1374 * ESPFIX issue, we can change this.
1375 */
1376#ifdef CONFIG_X86_32
1377# ifdef CONFIG_PARAVIRT_XXL
1378 do {
1379 extern void native_iret(void);
1380 if (pv_ops.cpu.iret == native_iret)
1381 set_cpu_bug(c, X86_BUG_ESPFIX);
1382 } while (0);
1383# else
1384 set_cpu_bug(c, X86_BUG_ESPFIX);
1385# endif
1386#endif
1387}
1388
1389static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1390{
1391 /*
1392 * The heavy lifting of max_rmid and cache_occ_scale are handled
1393 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
1394 * in case CQM bits really aren't there in this CPU.
1395 */
1396 if (c != &boot_cpu_data) {
1397 boot_cpu_data.x86_cache_max_rmid =
1398 min(boot_cpu_data.x86_cache_max_rmid,
1399 c->x86_cache_max_rmid);
1400 }
1401}
1402
1403/*
1404 * Validate that ACPI/mptables have the same information about the
1405 * effective APIC id and update the package map.
1406 */
1407static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1408{
1409#ifdef CONFIG_SMP
1410 unsigned int apicid, cpu = smp_processor_id();
1411
1412 apicid = apic->cpu_present_to_apicid(cpu);
1413
1414 if (apicid != c->apicid) {
1415 pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1416 cpu, apicid, c->initial_apicid);
1417 }
1418 BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
1419 BUG_ON(topology_update_die_map(c->cpu_die_id, cpu));
1420#else
1421 c->logical_proc_id = 0;
1422#endif
1423}
1424
1425/*
1426 * This does the hard work of actually picking apart the CPU stuff...
1427 */
1428static void identify_cpu(struct cpuinfo_x86 *c)
1429{
1430 int i;
1431
1432 c->loops_per_jiffy = loops_per_jiffy;
1433 c->x86_cache_size = 0;
1434 c->x86_vendor = X86_VENDOR_UNKNOWN;
1435 c->x86_model = c->x86_stepping = 0; /* So far unknown... */
1436 c->x86_vendor_id[0] = '\0'; /* Unset */
1437 c->x86_model_id[0] = '\0'; /* Unset */
1438 c->x86_max_cores = 1;
1439 c->x86_coreid_bits = 0;
1440 c->cu_id = 0xff;
1441#ifdef CONFIG_X86_64
1442 c->x86_clflush_size = 64;
1443 c->x86_phys_bits = 36;
1444 c->x86_virt_bits = 48;
1445#else
1446 c->cpuid_level = -1; /* CPUID not detected */
1447 c->x86_clflush_size = 32;
1448 c->x86_phys_bits = 32;
1449 c->x86_virt_bits = 32;
1450#endif
1451 c->x86_cache_alignment = c->x86_clflush_size;
1452 memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1453
1454 generic_identify(c);
1455
1456 if (this_cpu->c_identify)
1457 this_cpu->c_identify(c);
1458
1459 /* Clear/Set all flags overridden by options, after probe */
1460 apply_forced_caps(c);
1461
1462#ifdef CONFIG_X86_64
1463 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1464#endif
1465
1466 /*
1467 * Vendor-specific initialization. In this section we
1468 * canonicalize the feature flags, meaning if there are
1469 * features a certain CPU supports which CPUID doesn't
1470 * tell us, CPUID claiming incorrect flags, or other bugs,
1471 * we handle them here.
1472 *
1473 * At the end of this section, c->x86_capability better
1474 * indicate the features this CPU genuinely supports!
1475 */
1476 if (this_cpu->c_init)
1477 this_cpu->c_init(c);
1478
1479 /* Disable the PN if appropriate */
1480 squash_the_stupid_serial_number(c);
1481
1482 /* Set up SMEP/SMAP/UMIP */
1483 setup_smep(c);
1484 setup_smap(c);
1485 setup_umip(c);
1486
1487 /*
1488 * The vendor-specific functions might have changed features.
1489 * Now we do "generic changes."
1490 */
1491
1492 /* Filter out anything that depends on CPUID levels we don't have */
1493 filter_cpuid_features(c, true);
1494
1495 /* If the model name is still unset, do table lookup. */
1496 if (!c->x86_model_id[0]) {
1497 const char *p;
1498 p = table_lookup_model(c);
1499 if (p)
1500 strcpy(c->x86_model_id, p);
1501 else
1502 /* Last resort... */
1503 sprintf(c->x86_model_id, "%02x/%02x",
1504 c->x86, c->x86_model);
1505 }
1506
1507#ifdef CONFIG_X86_64
1508 detect_ht(c);
1509#endif
1510
1511 x86_init_rdrand(c);
1512 x86_init_cache_qos(c);
1513 setup_pku(c);
1514
1515 /*
1516 * Clear/Set all flags overridden by options, need do it
1517 * before following smp all cpus cap AND.
1518 */
1519 apply_forced_caps(c);
1520
1521 /*
1522 * On SMP, boot_cpu_data holds the common feature set between
1523 * all CPUs; so make sure that we indicate which features are
1524 * common between the CPUs. The first time this routine gets
1525 * executed, c == &boot_cpu_data.
1526 */
1527 if (c != &boot_cpu_data) {
1528 /* AND the already accumulated flags with these */
1529 for (i = 0; i < NCAPINTS; i++)
1530 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1531
1532 /* OR, i.e. replicate the bug flags */
1533 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1534 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1535 }
1536
1537 /* Init Machine Check Exception if available. */
1538 mcheck_cpu_init(c);
1539
1540 select_idle_routine(c);
1541
1542#ifdef CONFIG_NUMA
1543 numa_add_cpu(smp_processor_id());
1544#endif
1545}
1546
1547/*
1548 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1549 * on 32-bit kernels:
1550 */
1551#ifdef CONFIG_X86_32
1552void enable_sep_cpu(void)
1553{
1554 struct tss_struct *tss;
1555 int cpu;
1556
1557 if (!boot_cpu_has(X86_FEATURE_SEP))
1558 return;
1559
1560 cpu = get_cpu();
1561 tss = &per_cpu(cpu_tss_rw, cpu);
1562
1563 /*
1564 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1565 * see the big comment in struct x86_hw_tss's definition.
1566 */
1567
1568 tss->x86_tss.ss1 = __KERNEL_CS;
1569 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1570 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
1571 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1572
1573 put_cpu();
1574}
1575#endif
1576
1577void __init identify_boot_cpu(void)
1578{
1579 identify_cpu(&boot_cpu_data);
1580#ifdef CONFIG_X86_32
1581 sysenter_setup();
1582 enable_sep_cpu();
1583#endif
1584 cpu_detect_tlb(&boot_cpu_data);
1585 setup_cr_pinning();
1586
1587 tsx_init();
1588}
1589
1590void identify_secondary_cpu(struct cpuinfo_x86 *c)
1591{
1592 BUG_ON(c == &boot_cpu_data);
1593 identify_cpu(c);
1594#ifdef CONFIG_X86_32
1595 enable_sep_cpu();
1596#endif
1597 mtrr_ap_init();
1598 validate_apic_and_package_id(c);
1599 x86_spec_ctrl_setup_ap();
1600}
1601
1602static __init int setup_noclflush(char *arg)
1603{
1604 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1605 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1606 return 1;
1607}
1608__setup("noclflush", setup_noclflush);
1609
1610void print_cpu_info(struct cpuinfo_x86 *c)
1611{
1612 const char *vendor = NULL;
1613
1614 if (c->x86_vendor < X86_VENDOR_NUM) {
1615 vendor = this_cpu->c_vendor;
1616 } else {
1617 if (c->cpuid_level >= 0)
1618 vendor = c->x86_vendor_id;
1619 }
1620
1621 if (vendor && !strstr(c->x86_model_id, vendor))
1622 pr_cont("%s ", vendor);
1623
1624 if (c->x86_model_id[0])
1625 pr_cont("%s", c->x86_model_id);
1626 else
1627 pr_cont("%d86", c->x86);
1628
1629 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1630
1631 if (c->x86_stepping || c->cpuid_level >= 0)
1632 pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
1633 else
1634 pr_cont(")\n");
1635}
1636
1637/*
1638 * clearcpuid= was already parsed in fpu__init_parse_early_param.
1639 * But we need to keep a dummy __setup around otherwise it would
1640 * show up as an environment variable for init.
1641 */
1642static __init int setup_clearcpuid(char *arg)
1643{
1644 return 1;
1645}
1646__setup("clearcpuid=", setup_clearcpuid);
1647
1648#ifdef CONFIG_X86_64
1649DEFINE_PER_CPU_FIRST(struct fixed_percpu_data,
1650 fixed_percpu_data) __aligned(PAGE_SIZE) __visible;
1651EXPORT_PER_CPU_SYMBOL_GPL(fixed_percpu_data);
1652
1653/*
1654 * The following percpu variables are hot. Align current_task to
1655 * cacheline size such that they fall in the same cacheline.
1656 */
1657DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1658 &init_task;
1659EXPORT_PER_CPU_SYMBOL(current_task);
1660
1661DEFINE_PER_CPU(struct irq_stack *, hardirq_stack_ptr);
1662DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1663
1664DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1665EXPORT_PER_CPU_SYMBOL(__preempt_count);
1666
1667/* May not be marked __init: used by software suspend */
1668void syscall_init(void)
1669{
1670 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1671 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1672
1673#ifdef CONFIG_IA32_EMULATION
1674 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1675 /*
1676 * This only works on Intel CPUs.
1677 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1678 * This does not cause SYSENTER to jump to the wrong location, because
1679 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1680 */
1681 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1682 wrmsrl_safe(MSR_IA32_SYSENTER_ESP,
1683 (unsigned long)(cpu_entry_stack(smp_processor_id()) + 1));
1684 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1685#else
1686 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1687 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1688 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1689 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1690#endif
1691
1692 /* Flags to clear on syscall */
1693 wrmsrl(MSR_SYSCALL_MASK,
1694 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1695 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1696}
1697
1698DEFINE_PER_CPU(int, debug_stack_usage);
1699DEFINE_PER_CPU(u32, debug_idt_ctr);
1700
1701void debug_stack_set_zero(void)
1702{
1703 this_cpu_inc(debug_idt_ctr);
1704 load_current_idt();
1705}
1706NOKPROBE_SYMBOL(debug_stack_set_zero);
1707
1708void debug_stack_reset(void)
1709{
1710 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1711 return;
1712 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1713 load_current_idt();
1714}
1715NOKPROBE_SYMBOL(debug_stack_reset);
1716
1717#else /* CONFIG_X86_64 */
1718
1719DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1720EXPORT_PER_CPU_SYMBOL(current_task);
1721DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1722EXPORT_PER_CPU_SYMBOL(__preempt_count);
1723
1724/*
1725 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1726 * the top of the kernel stack. Use an extra percpu variable to track the
1727 * top of the kernel stack directly.
1728 */
1729DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1730 (unsigned long)&init_thread_union + THREAD_SIZE;
1731EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1732
1733#ifdef CONFIG_STACKPROTECTOR
1734DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1735#endif
1736
1737#endif /* CONFIG_X86_64 */
1738
1739/*
1740 * Clear all 6 debug registers:
1741 */
1742static void clear_all_debug_regs(void)
1743{
1744 int i;
1745
1746 for (i = 0; i < 8; i++) {
1747 /* Ignore db4, db5 */
1748 if ((i == 4) || (i == 5))
1749 continue;
1750
1751 set_debugreg(0, i);
1752 }
1753}
1754
1755#ifdef CONFIG_KGDB
1756/*
1757 * Restore debug regs if using kgdbwait and you have a kernel debugger
1758 * connection established.
1759 */
1760static void dbg_restore_debug_regs(void)
1761{
1762 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1763 arch_kgdb_ops.correct_hw_break();
1764}
1765#else /* ! CONFIG_KGDB */
1766#define dbg_restore_debug_regs()
1767#endif /* ! CONFIG_KGDB */
1768
1769static void wait_for_master_cpu(int cpu)
1770{
1771#ifdef CONFIG_SMP
1772 /*
1773 * wait for ACK from master CPU before continuing
1774 * with AP initialization
1775 */
1776 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1777 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1778 cpu_relax();
1779#endif
1780}
1781
1782#ifdef CONFIG_X86_64
1783static void setup_getcpu(int cpu)
1784{
1785 unsigned long cpudata = vdso_encode_cpunode(cpu, early_cpu_to_node(cpu));
1786 struct desc_struct d = { };
1787
1788 if (boot_cpu_has(X86_FEATURE_RDTSCP))
1789 write_rdtscp_aux(cpudata);
1790
1791 /* Store CPU and node number in limit. */
1792 d.limit0 = cpudata;
1793 d.limit1 = cpudata >> 16;
1794
1795 d.type = 5; /* RO data, expand down, accessed */
1796 d.dpl = 3; /* Visible to user code */
1797 d.s = 1; /* Not a system segment */
1798 d.p = 1; /* Present */
1799 d.d = 1; /* 32-bit */
1800
1801 write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_CPUNODE, &d, DESCTYPE_S);
1802}
1803#endif
1804
1805/*
1806 * cpu_init() initializes state that is per-CPU. Some data is already
1807 * initialized (naturally) in the bootstrap process, such as the GDT
1808 * and IDT. We reload them nevertheless, this function acts as a
1809 * 'CPU state barrier', nothing should get across.
1810 */
1811#ifdef CONFIG_X86_64
1812
1813void cpu_init(void)
1814{
1815 int cpu = raw_smp_processor_id();
1816 struct task_struct *me;
1817 struct tss_struct *t;
1818 int i;
1819
1820 wait_for_master_cpu(cpu);
1821
1822 if (cpu)
1823 load_ucode_ap();
1824
1825 t = &per_cpu(cpu_tss_rw, cpu);
1826
1827#ifdef CONFIG_NUMA
1828 if (this_cpu_read(numa_node) == 0 &&
1829 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1830 set_numa_node(early_cpu_to_node(cpu));
1831#endif
1832 setup_getcpu(cpu);
1833
1834 me = current;
1835
1836 pr_debug("Initializing CPU#%d\n", cpu);
1837
1838 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1839
1840 /*
1841 * Initialize the per-CPU GDT with the boot GDT,
1842 * and set up the GDT descriptor:
1843 */
1844
1845 switch_to_new_gdt(cpu);
1846 loadsegment(fs, 0);
1847
1848 load_current_idt();
1849
1850 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1851 syscall_init();
1852
1853 wrmsrl(MSR_FS_BASE, 0);
1854 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1855 barrier();
1856
1857 x86_configure_nx();
1858 x2apic_setup();
1859
1860 /*
1861 * set up and load the per-CPU TSS
1862 */
1863 if (!t->x86_tss.ist[0]) {
1864 t->x86_tss.ist[IST_INDEX_DF] = __this_cpu_ist_top_va(DF);
1865 t->x86_tss.ist[IST_INDEX_NMI] = __this_cpu_ist_top_va(NMI);
1866 t->x86_tss.ist[IST_INDEX_DB] = __this_cpu_ist_top_va(DB);
1867 t->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
1868 }
1869
1870 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1871
1872 /*
1873 * <= is required because the CPU will access up to
1874 * 8 bits beyond the end of the IO permission bitmap.
1875 */
1876 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1877 t->io_bitmap[i] = ~0UL;
1878
1879 mmgrab(&init_mm);
1880 me->active_mm = &init_mm;
1881 BUG_ON(me->mm);
1882 initialize_tlbstate_and_flush();
1883 enter_lazy_tlb(&init_mm, me);
1884
1885 /*
1886 * Initialize the TSS. sp0 points to the entry trampoline stack
1887 * regardless of what task is running.
1888 */
1889 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1890 load_TR_desc();
1891 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1892
1893 load_mm_ldt(&init_mm);
1894
1895 clear_all_debug_regs();
1896 dbg_restore_debug_regs();
1897
1898 fpu__init_cpu();
1899
1900 if (is_uv_system())
1901 uv_cpu_init();
1902
1903 load_fixmap_gdt(cpu);
1904}
1905
1906#else
1907
1908void cpu_init(void)
1909{
1910 int cpu = smp_processor_id();
1911 struct task_struct *curr = current;
1912 struct tss_struct *t = &per_cpu(cpu_tss_rw, cpu);
1913
1914 wait_for_master_cpu(cpu);
1915
1916 show_ucode_info_early();
1917
1918 pr_info("Initializing CPU#%d\n", cpu);
1919
1920 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1921 boot_cpu_has(X86_FEATURE_TSC) ||
1922 boot_cpu_has(X86_FEATURE_DE))
1923 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1924
1925 load_current_idt();
1926 switch_to_new_gdt(cpu);
1927
1928 /*
1929 * Set up and load the per-CPU TSS and LDT
1930 */
1931 mmgrab(&init_mm);
1932 curr->active_mm = &init_mm;
1933 BUG_ON(curr->mm);
1934 initialize_tlbstate_and_flush();
1935 enter_lazy_tlb(&init_mm, curr);
1936
1937 /*
1938 * Initialize the TSS. sp0 points to the entry trampoline stack
1939 * regardless of what task is running.
1940 */
1941 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1942 load_TR_desc();
1943 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1944
1945 load_mm_ldt(&init_mm);
1946
1947 t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1948
1949#ifdef CONFIG_DOUBLEFAULT
1950 /* Set up doublefault TSS pointer in the GDT */
1951 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1952#endif
1953
1954 clear_all_debug_regs();
1955 dbg_restore_debug_regs();
1956
1957 fpu__init_cpu();
1958
1959 load_fixmap_gdt(cpu);
1960}
1961#endif
1962
1963/*
1964 * The microcode loader calls this upon late microcode load to recheck features,
1965 * only when microcode has been updated. Caller holds microcode_mutex and CPU
1966 * hotplug lock.
1967 */
1968void microcode_check(void)
1969{
1970 struct cpuinfo_x86 info;
1971
1972 perf_check_microcode();
1973
1974 /* Reload CPUID max function as it might've changed. */
1975 info.cpuid_level = cpuid_eax(0);
1976
1977 /*
1978 * Copy all capability leafs to pick up the synthetic ones so that
1979 * memcmp() below doesn't fail on that. The ones coming from CPUID will
1980 * get overwritten in get_cpu_cap().
1981 */
1982 memcpy(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability));
1983
1984 get_cpu_cap(&info);
1985
1986 if (!memcmp(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability)))
1987 return;
1988
1989 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
1990 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");
1991}
1992
1993/*
1994 * Invoked from core CPU hotplug code after hotplug operations
1995 */
1996void arch_smt_update(void)
1997{
1998 /* Handle the speculative execution misfeatures */
1999 cpu_bugs_smt_update();
2000 /* Check whether IPI broadcasting can be enabled */
2001 apic_smt_update();
2002}
1#include <linux/bootmem.h>
2#include <linux/linkage.h>
3#include <linux/bitops.h>
4#include <linux/kernel.h>
5#include <linux/module.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.h>
11#include <linux/init.h>
12#include <linux/kprobes.h>
13#include <linux/kgdb.h>
14#include <linux/smp.h>
15#include <linux/io.h>
16#include <linux/syscore_ops.h>
17
18#include <asm/stackprotector.h>
19#include <asm/perf_event.h>
20#include <asm/mmu_context.h>
21#include <asm/archrandom.h>
22#include <asm/hypervisor.h>
23#include <asm/processor.h>
24#include <asm/tlbflush.h>
25#include <asm/debugreg.h>
26#include <asm/sections.h>
27#include <asm/vsyscall.h>
28#include <linux/topology.h>
29#include <linux/cpumask.h>
30#include <asm/pgtable.h>
31#include <linux/atomic.h>
32#include <asm/proto.h>
33#include <asm/setup.h>
34#include <asm/apic.h>
35#include <asm/desc.h>
36#include <asm/fpu/internal.h>
37#include <asm/mtrr.h>
38#include <linux/numa.h>
39#include <asm/asm.h>
40#include <asm/cpu.h>
41#include <asm/mce.h>
42#include <asm/msr.h>
43#include <asm/pat.h>
44#include <asm/microcode.h>
45#include <asm/microcode_intel.h>
46
47#ifdef CONFIG_X86_LOCAL_APIC
48#include <asm/uv/uv.h>
49#endif
50
51#include "cpu.h"
52
53/* all of these masks are initialized in setup_cpu_local_masks() */
54cpumask_var_t cpu_initialized_mask;
55cpumask_var_t cpu_callout_mask;
56cpumask_var_t cpu_callin_mask;
57
58/* representing cpus for which sibling maps can be computed */
59cpumask_var_t cpu_sibling_setup_mask;
60
61/* correctly size the local cpu masks */
62void __init setup_cpu_local_masks(void)
63{
64 alloc_bootmem_cpumask_var(&cpu_initialized_mask);
65 alloc_bootmem_cpumask_var(&cpu_callin_mask);
66 alloc_bootmem_cpumask_var(&cpu_callout_mask);
67 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
68}
69
70static void default_init(struct cpuinfo_x86 *c)
71{
72#ifdef CONFIG_X86_64
73 cpu_detect_cache_sizes(c);
74#else
75 /* Not much we can do here... */
76 /* Check if at least it has cpuid */
77 if (c->cpuid_level == -1) {
78 /* No cpuid. It must be an ancient CPU */
79 if (c->x86 == 4)
80 strcpy(c->x86_model_id, "486");
81 else if (c->x86 == 3)
82 strcpy(c->x86_model_id, "386");
83 }
84#endif
85}
86
87static const struct cpu_dev default_cpu = {
88 .c_init = default_init,
89 .c_vendor = "Unknown",
90 .c_x86_vendor = X86_VENDOR_UNKNOWN,
91};
92
93static const struct cpu_dev *this_cpu = &default_cpu;
94
95DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
96#ifdef CONFIG_X86_64
97 /*
98 * We need valid kernel segments for data and code in long mode too
99 * IRET will check the segment types kkeil 2000/10/28
100 * Also sysret mandates a special GDT layout
101 *
102 * TLS descriptors are currently at a different place compared to i386.
103 * Hopefully nobody expects them at a fixed place (Wine?)
104 */
105 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
106 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
107 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
108 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
109 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
110 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
111#else
112 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
113 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
114 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
115 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
116 /*
117 * Segments used for calling PnP BIOS have byte granularity.
118 * They code segments and data segments have fixed 64k limits,
119 * the transfer segment sizes are set at run time.
120 */
121 /* 32-bit code */
122 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
123 /* 16-bit code */
124 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
125 /* 16-bit data */
126 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(0x0092, 0, 0xffff),
127 /* 16-bit data */
128 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(0x0092, 0, 0),
129 /* 16-bit data */
130 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(0x0092, 0, 0),
131 /*
132 * The APM segments have byte granularity and their bases
133 * are set at run time. All have 64k limits.
134 */
135 /* 32-bit code */
136 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(0x409a, 0, 0xffff),
137 /* 16-bit code */
138 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(0x009a, 0, 0xffff),
139 /* data */
140 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(0x4092, 0, 0xffff),
141
142 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
143 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
144 GDT_STACK_CANARY_INIT
145#endif
146} };
147EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
148
149static int __init x86_mpx_setup(char *s)
150{
151 /* require an exact match without trailing characters */
152 if (strlen(s))
153 return 0;
154
155 /* do not emit a message if the feature is not present */
156 if (!boot_cpu_has(X86_FEATURE_MPX))
157 return 1;
158
159 setup_clear_cpu_cap(X86_FEATURE_MPX);
160 pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
161 return 1;
162}
163__setup("nompx", x86_mpx_setup);
164
165static int __init x86_noinvpcid_setup(char *s)
166{
167 /* noinvpcid doesn't accept parameters */
168 if (s)
169 return -EINVAL;
170
171 /* do not emit a message if the feature is not present */
172 if (!boot_cpu_has(X86_FEATURE_INVPCID))
173 return 0;
174
175 setup_clear_cpu_cap(X86_FEATURE_INVPCID);
176 pr_info("noinvpcid: INVPCID feature disabled\n");
177 return 0;
178}
179early_param("noinvpcid", x86_noinvpcid_setup);
180
181#ifdef CONFIG_X86_32
182static int cachesize_override = -1;
183static int disable_x86_serial_nr = 1;
184
185static int __init cachesize_setup(char *str)
186{
187 get_option(&str, &cachesize_override);
188 return 1;
189}
190__setup("cachesize=", cachesize_setup);
191
192static int __init x86_sep_setup(char *s)
193{
194 setup_clear_cpu_cap(X86_FEATURE_SEP);
195 return 1;
196}
197__setup("nosep", x86_sep_setup);
198
199/* Standard macro to see if a specific flag is changeable */
200static inline int flag_is_changeable_p(u32 flag)
201{
202 u32 f1, f2;
203
204 /*
205 * Cyrix and IDT cpus allow disabling of CPUID
206 * so the code below may return different results
207 * when it is executed before and after enabling
208 * the CPUID. Add "volatile" to not allow gcc to
209 * optimize the subsequent calls to this function.
210 */
211 asm volatile ("pushfl \n\t"
212 "pushfl \n\t"
213 "popl %0 \n\t"
214 "movl %0, %1 \n\t"
215 "xorl %2, %0 \n\t"
216 "pushl %0 \n\t"
217 "popfl \n\t"
218 "pushfl \n\t"
219 "popl %0 \n\t"
220 "popfl \n\t"
221
222 : "=&r" (f1), "=&r" (f2)
223 : "ir" (flag));
224
225 return ((f1^f2) & flag) != 0;
226}
227
228/* Probe for the CPUID instruction */
229int have_cpuid_p(void)
230{
231 return flag_is_changeable_p(X86_EFLAGS_ID);
232}
233
234static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
235{
236 unsigned long lo, hi;
237
238 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
239 return;
240
241 /* Disable processor serial number: */
242
243 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
244 lo |= 0x200000;
245 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
246
247 pr_notice("CPU serial number disabled.\n");
248 clear_cpu_cap(c, X86_FEATURE_PN);
249
250 /* Disabling the serial number may affect the cpuid level */
251 c->cpuid_level = cpuid_eax(0);
252}
253
254static int __init x86_serial_nr_setup(char *s)
255{
256 disable_x86_serial_nr = 0;
257 return 1;
258}
259__setup("serialnumber", x86_serial_nr_setup);
260#else
261static inline int flag_is_changeable_p(u32 flag)
262{
263 return 1;
264}
265static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
266{
267}
268#endif
269
270static __init int setup_disable_smep(char *arg)
271{
272 setup_clear_cpu_cap(X86_FEATURE_SMEP);
273 return 1;
274}
275__setup("nosmep", setup_disable_smep);
276
277static __always_inline void setup_smep(struct cpuinfo_x86 *c)
278{
279 if (cpu_has(c, X86_FEATURE_SMEP))
280 cr4_set_bits(X86_CR4_SMEP);
281}
282
283static __init int setup_disable_smap(char *arg)
284{
285 setup_clear_cpu_cap(X86_FEATURE_SMAP);
286 return 1;
287}
288__setup("nosmap", setup_disable_smap);
289
290static __always_inline void setup_smap(struct cpuinfo_x86 *c)
291{
292 unsigned long eflags = native_save_fl();
293
294 /* This should have been cleared long ago */
295 BUG_ON(eflags & X86_EFLAGS_AC);
296
297 if (cpu_has(c, X86_FEATURE_SMAP)) {
298#ifdef CONFIG_X86_SMAP
299 cr4_set_bits(X86_CR4_SMAP);
300#else
301 cr4_clear_bits(X86_CR4_SMAP);
302#endif
303 }
304}
305
306/*
307 * Protection Keys are not available in 32-bit mode.
308 */
309static bool pku_disabled;
310
311static __always_inline void setup_pku(struct cpuinfo_x86 *c)
312{
313 if (!cpu_has(c, X86_FEATURE_PKU))
314 return;
315 if (pku_disabled)
316 return;
317
318 cr4_set_bits(X86_CR4_PKE);
319 /*
320 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
321 * cpuid bit to be set. We need to ensure that we
322 * update that bit in this CPU's "cpu_info".
323 */
324 get_cpu_cap(c);
325}
326
327#ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
328static __init int setup_disable_pku(char *arg)
329{
330 /*
331 * Do not clear the X86_FEATURE_PKU bit. All of the
332 * runtime checks are against OSPKE so clearing the
333 * bit does nothing.
334 *
335 * This way, we will see "pku" in cpuinfo, but not
336 * "ospke", which is exactly what we want. It shows
337 * that the CPU has PKU, but the OS has not enabled it.
338 * This happens to be exactly how a system would look
339 * if we disabled the config option.
340 */
341 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
342 pku_disabled = true;
343 return 1;
344}
345__setup("nopku", setup_disable_pku);
346#endif /* CONFIG_X86_64 */
347
348/*
349 * Some CPU features depend on higher CPUID levels, which may not always
350 * be available due to CPUID level capping or broken virtualization
351 * software. Add those features to this table to auto-disable them.
352 */
353struct cpuid_dependent_feature {
354 u32 feature;
355 u32 level;
356};
357
358static const struct cpuid_dependent_feature
359cpuid_dependent_features[] = {
360 { X86_FEATURE_MWAIT, 0x00000005 },
361 { X86_FEATURE_DCA, 0x00000009 },
362 { X86_FEATURE_XSAVE, 0x0000000d },
363 { 0, 0 }
364};
365
366static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
367{
368 const struct cpuid_dependent_feature *df;
369
370 for (df = cpuid_dependent_features; df->feature; df++) {
371
372 if (!cpu_has(c, df->feature))
373 continue;
374 /*
375 * Note: cpuid_level is set to -1 if unavailable, but
376 * extended_extended_level is set to 0 if unavailable
377 * and the legitimate extended levels are all negative
378 * when signed; hence the weird messing around with
379 * signs here...
380 */
381 if (!((s32)df->level < 0 ?
382 (u32)df->level > (u32)c->extended_cpuid_level :
383 (s32)df->level > (s32)c->cpuid_level))
384 continue;
385
386 clear_cpu_cap(c, df->feature);
387 if (!warn)
388 continue;
389
390 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
391 x86_cap_flag(df->feature), df->level);
392 }
393}
394
395/*
396 * Naming convention should be: <Name> [(<Codename>)]
397 * This table only is used unless init_<vendor>() below doesn't set it;
398 * in particular, if CPUID levels 0x80000002..4 are supported, this
399 * isn't used
400 */
401
402/* Look up CPU names by table lookup. */
403static const char *table_lookup_model(struct cpuinfo_x86 *c)
404{
405#ifdef CONFIG_X86_32
406 const struct legacy_cpu_model_info *info;
407
408 if (c->x86_model >= 16)
409 return NULL; /* Range check */
410
411 if (!this_cpu)
412 return NULL;
413
414 info = this_cpu->legacy_models;
415
416 while (info->family) {
417 if (info->family == c->x86)
418 return info->model_names[c->x86_model];
419 info++;
420 }
421#endif
422 return NULL; /* Not found */
423}
424
425__u32 cpu_caps_cleared[NCAPINTS];
426__u32 cpu_caps_set[NCAPINTS];
427
428void load_percpu_segment(int cpu)
429{
430#ifdef CONFIG_X86_32
431 loadsegment(fs, __KERNEL_PERCPU);
432#else
433 loadsegment(gs, 0);
434 wrmsrl(MSR_GS_BASE, (unsigned long)per_cpu(irq_stack_union.gs_base, cpu));
435#endif
436 load_stack_canary_segment();
437}
438
439/*
440 * Current gdt points %fs at the "master" per-cpu area: after this,
441 * it's on the real one.
442 */
443void switch_to_new_gdt(int cpu)
444{
445 struct desc_ptr gdt_descr;
446
447 gdt_descr.address = (long)get_cpu_gdt_table(cpu);
448 gdt_descr.size = GDT_SIZE - 1;
449 load_gdt(&gdt_descr);
450 /* Reload the per-cpu base */
451
452 load_percpu_segment(cpu);
453}
454
455static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
456
457static void get_model_name(struct cpuinfo_x86 *c)
458{
459 unsigned int *v;
460 char *p, *q, *s;
461
462 if (c->extended_cpuid_level < 0x80000004)
463 return;
464
465 v = (unsigned int *)c->x86_model_id;
466 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
467 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
468 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
469 c->x86_model_id[48] = 0;
470
471 /* Trim whitespace */
472 p = q = s = &c->x86_model_id[0];
473
474 while (*p == ' ')
475 p++;
476
477 while (*p) {
478 /* Note the last non-whitespace index */
479 if (!isspace(*p))
480 s = q;
481
482 *q++ = *p++;
483 }
484
485 *(s + 1) = '\0';
486}
487
488void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
489{
490 unsigned int n, dummy, ebx, ecx, edx, l2size;
491
492 n = c->extended_cpuid_level;
493
494 if (n >= 0x80000005) {
495 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
496 c->x86_cache_size = (ecx>>24) + (edx>>24);
497#ifdef CONFIG_X86_64
498 /* On K8 L1 TLB is inclusive, so don't count it */
499 c->x86_tlbsize = 0;
500#endif
501 }
502
503 if (n < 0x80000006) /* Some chips just has a large L1. */
504 return;
505
506 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
507 l2size = ecx >> 16;
508
509#ifdef CONFIG_X86_64
510 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
511#else
512 /* do processor-specific cache resizing */
513 if (this_cpu->legacy_cache_size)
514 l2size = this_cpu->legacy_cache_size(c, l2size);
515
516 /* Allow user to override all this if necessary. */
517 if (cachesize_override != -1)
518 l2size = cachesize_override;
519
520 if (l2size == 0)
521 return; /* Again, no L2 cache is possible */
522#endif
523
524 c->x86_cache_size = l2size;
525}
526
527u16 __read_mostly tlb_lli_4k[NR_INFO];
528u16 __read_mostly tlb_lli_2m[NR_INFO];
529u16 __read_mostly tlb_lli_4m[NR_INFO];
530u16 __read_mostly tlb_lld_4k[NR_INFO];
531u16 __read_mostly tlb_lld_2m[NR_INFO];
532u16 __read_mostly tlb_lld_4m[NR_INFO];
533u16 __read_mostly tlb_lld_1g[NR_INFO];
534
535static void cpu_detect_tlb(struct cpuinfo_x86 *c)
536{
537 if (this_cpu->c_detect_tlb)
538 this_cpu->c_detect_tlb(c);
539
540 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
541 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
542 tlb_lli_4m[ENTRIES]);
543
544 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
545 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
546 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
547}
548
549void detect_ht(struct cpuinfo_x86 *c)
550{
551#ifdef CONFIG_SMP
552 u32 eax, ebx, ecx, edx;
553 int index_msb, core_bits;
554 static bool printed;
555
556 if (!cpu_has(c, X86_FEATURE_HT))
557 return;
558
559 if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
560 goto out;
561
562 if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
563 return;
564
565 cpuid(1, &eax, &ebx, &ecx, &edx);
566
567 smp_num_siblings = (ebx & 0xff0000) >> 16;
568
569 if (smp_num_siblings == 1) {
570 pr_info_once("CPU0: Hyper-Threading is disabled\n");
571 goto out;
572 }
573
574 if (smp_num_siblings <= 1)
575 goto out;
576
577 index_msb = get_count_order(smp_num_siblings);
578 c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
579
580 smp_num_siblings = smp_num_siblings / c->x86_max_cores;
581
582 index_msb = get_count_order(smp_num_siblings);
583
584 core_bits = get_count_order(c->x86_max_cores);
585
586 c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
587 ((1 << core_bits) - 1);
588
589out:
590 if (!printed && (c->x86_max_cores * smp_num_siblings) > 1) {
591 pr_info("CPU: Physical Processor ID: %d\n",
592 c->phys_proc_id);
593 pr_info("CPU: Processor Core ID: %d\n",
594 c->cpu_core_id);
595 printed = 1;
596 }
597#endif
598}
599
600static void get_cpu_vendor(struct cpuinfo_x86 *c)
601{
602 char *v = c->x86_vendor_id;
603 int i;
604
605 for (i = 0; i < X86_VENDOR_NUM; i++) {
606 if (!cpu_devs[i])
607 break;
608
609 if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
610 (cpu_devs[i]->c_ident[1] &&
611 !strcmp(v, cpu_devs[i]->c_ident[1]))) {
612
613 this_cpu = cpu_devs[i];
614 c->x86_vendor = this_cpu->c_x86_vendor;
615 return;
616 }
617 }
618
619 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
620 "CPU: Your system may be unstable.\n", v);
621
622 c->x86_vendor = X86_VENDOR_UNKNOWN;
623 this_cpu = &default_cpu;
624}
625
626void cpu_detect(struct cpuinfo_x86 *c)
627{
628 /* Get vendor name */
629 cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
630 (unsigned int *)&c->x86_vendor_id[0],
631 (unsigned int *)&c->x86_vendor_id[8],
632 (unsigned int *)&c->x86_vendor_id[4]);
633
634 c->x86 = 4;
635 /* Intel-defined flags: level 0x00000001 */
636 if (c->cpuid_level >= 0x00000001) {
637 u32 junk, tfms, cap0, misc;
638
639 cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
640 c->x86 = x86_family(tfms);
641 c->x86_model = x86_model(tfms);
642 c->x86_mask = x86_stepping(tfms);
643
644 if (cap0 & (1<<19)) {
645 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
646 c->x86_cache_alignment = c->x86_clflush_size;
647 }
648 }
649}
650
651void get_cpu_cap(struct cpuinfo_x86 *c)
652{
653 u32 eax, ebx, ecx, edx;
654
655 /* Intel-defined flags: level 0x00000001 */
656 if (c->cpuid_level >= 0x00000001) {
657 cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
658
659 c->x86_capability[CPUID_1_ECX] = ecx;
660 c->x86_capability[CPUID_1_EDX] = edx;
661 }
662
663 /* Additional Intel-defined flags: level 0x00000007 */
664 if (c->cpuid_level >= 0x00000007) {
665 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
666
667 c->x86_capability[CPUID_7_0_EBX] = ebx;
668
669 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
670 c->x86_capability[CPUID_7_ECX] = ecx;
671 }
672
673 /* Extended state features: level 0x0000000d */
674 if (c->cpuid_level >= 0x0000000d) {
675 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
676
677 c->x86_capability[CPUID_D_1_EAX] = eax;
678 }
679
680 /* Additional Intel-defined flags: level 0x0000000F */
681 if (c->cpuid_level >= 0x0000000F) {
682
683 /* QoS sub-leaf, EAX=0Fh, ECX=0 */
684 cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
685 c->x86_capability[CPUID_F_0_EDX] = edx;
686
687 if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
688 /* will be overridden if occupancy monitoring exists */
689 c->x86_cache_max_rmid = ebx;
690
691 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
692 cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
693 c->x86_capability[CPUID_F_1_EDX] = edx;
694
695 if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
696 ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
697 (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
698 c->x86_cache_max_rmid = ecx;
699 c->x86_cache_occ_scale = ebx;
700 }
701 } else {
702 c->x86_cache_max_rmid = -1;
703 c->x86_cache_occ_scale = -1;
704 }
705 }
706
707 /* AMD-defined flags: level 0x80000001 */
708 eax = cpuid_eax(0x80000000);
709 c->extended_cpuid_level = eax;
710
711 if ((eax & 0xffff0000) == 0x80000000) {
712 if (eax >= 0x80000001) {
713 cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
714
715 c->x86_capability[CPUID_8000_0001_ECX] = ecx;
716 c->x86_capability[CPUID_8000_0001_EDX] = edx;
717 }
718 }
719
720 if (c->extended_cpuid_level >= 0x80000008) {
721 cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
722
723 c->x86_virt_bits = (eax >> 8) & 0xff;
724 c->x86_phys_bits = eax & 0xff;
725 c->x86_capability[CPUID_8000_0008_EBX] = ebx;
726 }
727#ifdef CONFIG_X86_32
728 else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
729 c->x86_phys_bits = 36;
730#endif
731
732 if (c->extended_cpuid_level >= 0x80000007)
733 c->x86_power = cpuid_edx(0x80000007);
734
735 if (c->extended_cpuid_level >= 0x8000000a)
736 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
737
738 init_scattered_cpuid_features(c);
739}
740
741static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
742{
743#ifdef CONFIG_X86_32
744 int i;
745
746 /*
747 * First of all, decide if this is a 486 or higher
748 * It's a 486 if we can modify the AC flag
749 */
750 if (flag_is_changeable_p(X86_EFLAGS_AC))
751 c->x86 = 4;
752 else
753 c->x86 = 3;
754
755 for (i = 0; i < X86_VENDOR_NUM; i++)
756 if (cpu_devs[i] && cpu_devs[i]->c_identify) {
757 c->x86_vendor_id[0] = 0;
758 cpu_devs[i]->c_identify(c);
759 if (c->x86_vendor_id[0]) {
760 get_cpu_vendor(c);
761 break;
762 }
763 }
764#endif
765}
766
767/*
768 * Do minimum CPU detection early.
769 * Fields really needed: vendor, cpuid_level, family, model, mask,
770 * cache alignment.
771 * The others are not touched to avoid unwanted side effects.
772 *
773 * WARNING: this function is only called on the BP. Don't add code here
774 * that is supposed to run on all CPUs.
775 */
776static void __init early_identify_cpu(struct cpuinfo_x86 *c)
777{
778#ifdef CONFIG_X86_64
779 c->x86_clflush_size = 64;
780 c->x86_phys_bits = 36;
781 c->x86_virt_bits = 48;
782#else
783 c->x86_clflush_size = 32;
784 c->x86_phys_bits = 32;
785 c->x86_virt_bits = 32;
786#endif
787 c->x86_cache_alignment = c->x86_clflush_size;
788
789 memset(&c->x86_capability, 0, sizeof c->x86_capability);
790 c->extended_cpuid_level = 0;
791
792 if (!have_cpuid_p())
793 identify_cpu_without_cpuid(c);
794
795 /* cyrix could have cpuid enabled via c_identify()*/
796 if (!have_cpuid_p())
797 return;
798
799 cpu_detect(c);
800 get_cpu_vendor(c);
801 get_cpu_cap(c);
802
803 if (this_cpu->c_early_init)
804 this_cpu->c_early_init(c);
805
806 c->cpu_index = 0;
807 filter_cpuid_features(c, false);
808
809 if (this_cpu->c_bsp_init)
810 this_cpu->c_bsp_init(c);
811
812 setup_force_cpu_cap(X86_FEATURE_ALWAYS);
813 fpu__init_system(c);
814}
815
816void __init early_cpu_init(void)
817{
818 const struct cpu_dev *const *cdev;
819 int count = 0;
820
821#ifdef CONFIG_PROCESSOR_SELECT
822 pr_info("KERNEL supported cpus:\n");
823#endif
824
825 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
826 const struct cpu_dev *cpudev = *cdev;
827
828 if (count >= X86_VENDOR_NUM)
829 break;
830 cpu_devs[count] = cpudev;
831 count++;
832
833#ifdef CONFIG_PROCESSOR_SELECT
834 {
835 unsigned int j;
836
837 for (j = 0; j < 2; j++) {
838 if (!cpudev->c_ident[j])
839 continue;
840 pr_info(" %s %s\n", cpudev->c_vendor,
841 cpudev->c_ident[j]);
842 }
843 }
844#endif
845 }
846 early_identify_cpu(&boot_cpu_data);
847}
848
849/*
850 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
851 * unfortunately, that's not true in practice because of early VIA
852 * chips and (more importantly) broken virtualizers that are not easy
853 * to detect. In the latter case it doesn't even *fail* reliably, so
854 * probing for it doesn't even work. Disable it completely on 32-bit
855 * unless we can find a reliable way to detect all the broken cases.
856 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
857 */
858static void detect_nopl(struct cpuinfo_x86 *c)
859{
860#ifdef CONFIG_X86_32
861 clear_cpu_cap(c, X86_FEATURE_NOPL);
862#else
863 set_cpu_cap(c, X86_FEATURE_NOPL);
864#endif
865
866 /*
867 * ESPFIX is a strange bug. All real CPUs have it. Paravirt
868 * systems that run Linux at CPL > 0 may or may not have the
869 * issue, but, even if they have the issue, there's absolutely
870 * nothing we can do about it because we can't use the real IRET
871 * instruction.
872 *
873 * NB: For the time being, only 32-bit kernels support
874 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose
875 * whether to apply espfix using paravirt hooks. If any
876 * non-paravirt system ever shows up that does *not* have the
877 * ESPFIX issue, we can change this.
878 */
879#ifdef CONFIG_X86_32
880#ifdef CONFIG_PARAVIRT
881 do {
882 extern void native_iret(void);
883 if (pv_cpu_ops.iret == native_iret)
884 set_cpu_bug(c, X86_BUG_ESPFIX);
885 } while (0);
886#else
887 set_cpu_bug(c, X86_BUG_ESPFIX);
888#endif
889#endif
890}
891
892static void generic_identify(struct cpuinfo_x86 *c)
893{
894 c->extended_cpuid_level = 0;
895
896 if (!have_cpuid_p())
897 identify_cpu_without_cpuid(c);
898
899 /* cyrix could have cpuid enabled via c_identify()*/
900 if (!have_cpuid_p())
901 return;
902
903 cpu_detect(c);
904
905 get_cpu_vendor(c);
906
907 get_cpu_cap(c);
908
909 if (c->cpuid_level >= 0x00000001) {
910 c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
911#ifdef CONFIG_X86_32
912# ifdef CONFIG_SMP
913 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
914# else
915 c->apicid = c->initial_apicid;
916# endif
917#endif
918 c->phys_proc_id = c->initial_apicid;
919 }
920
921 get_model_name(c); /* Default name */
922
923 detect_nopl(c);
924}
925
926static void x86_init_cache_qos(struct cpuinfo_x86 *c)
927{
928 /*
929 * The heavy lifting of max_rmid and cache_occ_scale are handled
930 * in get_cpu_cap(). Here we just set the max_rmid for the boot_cpu
931 * in case CQM bits really aren't there in this CPU.
932 */
933 if (c != &boot_cpu_data) {
934 boot_cpu_data.x86_cache_max_rmid =
935 min(boot_cpu_data.x86_cache_max_rmid,
936 c->x86_cache_max_rmid);
937 }
938}
939
940/*
941 * This does the hard work of actually picking apart the CPU stuff...
942 */
943static void identify_cpu(struct cpuinfo_x86 *c)
944{
945 int i;
946
947 c->loops_per_jiffy = loops_per_jiffy;
948 c->x86_cache_size = -1;
949 c->x86_vendor = X86_VENDOR_UNKNOWN;
950 c->x86_model = c->x86_mask = 0; /* So far unknown... */
951 c->x86_vendor_id[0] = '\0'; /* Unset */
952 c->x86_model_id[0] = '\0'; /* Unset */
953 c->x86_max_cores = 1;
954 c->x86_coreid_bits = 0;
955#ifdef CONFIG_X86_64
956 c->x86_clflush_size = 64;
957 c->x86_phys_bits = 36;
958 c->x86_virt_bits = 48;
959#else
960 c->cpuid_level = -1; /* CPUID not detected */
961 c->x86_clflush_size = 32;
962 c->x86_phys_bits = 32;
963 c->x86_virt_bits = 32;
964#endif
965 c->x86_cache_alignment = c->x86_clflush_size;
966 memset(&c->x86_capability, 0, sizeof c->x86_capability);
967
968 generic_identify(c);
969
970 if (this_cpu->c_identify)
971 this_cpu->c_identify(c);
972
973 /* Clear/Set all flags overridden by options, after probe */
974 for (i = 0; i < NCAPINTS; i++) {
975 c->x86_capability[i] &= ~cpu_caps_cleared[i];
976 c->x86_capability[i] |= cpu_caps_set[i];
977 }
978
979#ifdef CONFIG_X86_64
980 c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
981#endif
982
983 /*
984 * Vendor-specific initialization. In this section we
985 * canonicalize the feature flags, meaning if there are
986 * features a certain CPU supports which CPUID doesn't
987 * tell us, CPUID claiming incorrect flags, or other bugs,
988 * we handle them here.
989 *
990 * At the end of this section, c->x86_capability better
991 * indicate the features this CPU genuinely supports!
992 */
993 if (this_cpu->c_init)
994 this_cpu->c_init(c);
995
996 /* Disable the PN if appropriate */
997 squash_the_stupid_serial_number(c);
998
999 /* Set up SMEP/SMAP */
1000 setup_smep(c);
1001 setup_smap(c);
1002
1003 /*
1004 * The vendor-specific functions might have changed features.
1005 * Now we do "generic changes."
1006 */
1007
1008 /* Filter out anything that depends on CPUID levels we don't have */
1009 filter_cpuid_features(c, true);
1010
1011 /* If the model name is still unset, do table lookup. */
1012 if (!c->x86_model_id[0]) {
1013 const char *p;
1014 p = table_lookup_model(c);
1015 if (p)
1016 strcpy(c->x86_model_id, p);
1017 else
1018 /* Last resort... */
1019 sprintf(c->x86_model_id, "%02x/%02x",
1020 c->x86, c->x86_model);
1021 }
1022
1023#ifdef CONFIG_X86_64
1024 detect_ht(c);
1025#endif
1026
1027 init_hypervisor(c);
1028 x86_init_rdrand(c);
1029 x86_init_cache_qos(c);
1030 setup_pku(c);
1031
1032 /*
1033 * Clear/Set all flags overridden by options, need do it
1034 * before following smp all cpus cap AND.
1035 */
1036 for (i = 0; i < NCAPINTS; i++) {
1037 c->x86_capability[i] &= ~cpu_caps_cleared[i];
1038 c->x86_capability[i] |= cpu_caps_set[i];
1039 }
1040
1041 /*
1042 * On SMP, boot_cpu_data holds the common feature set between
1043 * all CPUs; so make sure that we indicate which features are
1044 * common between the CPUs. The first time this routine gets
1045 * executed, c == &boot_cpu_data.
1046 */
1047 if (c != &boot_cpu_data) {
1048 /* AND the already accumulated flags with these */
1049 for (i = 0; i < NCAPINTS; i++)
1050 boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1051
1052 /* OR, i.e. replicate the bug flags */
1053 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1054 c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1055 }
1056
1057 /* Init Machine Check Exception if available. */
1058 mcheck_cpu_init(c);
1059
1060 select_idle_routine(c);
1061
1062#ifdef CONFIG_NUMA
1063 numa_add_cpu(smp_processor_id());
1064#endif
1065 /* The boot/hotplug time assigment got cleared, restore it */
1066 c->logical_proc_id = topology_phys_to_logical_pkg(c->phys_proc_id);
1067}
1068
1069/*
1070 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1071 * on 32-bit kernels:
1072 */
1073#ifdef CONFIG_X86_32
1074void enable_sep_cpu(void)
1075{
1076 struct tss_struct *tss;
1077 int cpu;
1078
1079 cpu = get_cpu();
1080 tss = &per_cpu(cpu_tss, cpu);
1081
1082 if (!boot_cpu_has(X86_FEATURE_SEP))
1083 goto out;
1084
1085 /*
1086 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1087 * see the big comment in struct x86_hw_tss's definition.
1088 */
1089
1090 tss->x86_tss.ss1 = __KERNEL_CS;
1091 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1092
1093 wrmsr(MSR_IA32_SYSENTER_ESP,
1094 (unsigned long)tss + offsetofend(struct tss_struct, SYSENTER_stack),
1095 0);
1096
1097 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1098
1099out:
1100 put_cpu();
1101}
1102#endif
1103
1104void __init identify_boot_cpu(void)
1105{
1106 identify_cpu(&boot_cpu_data);
1107 init_amd_e400_c1e_mask();
1108#ifdef CONFIG_X86_32
1109 sysenter_setup();
1110 enable_sep_cpu();
1111#endif
1112 cpu_detect_tlb(&boot_cpu_data);
1113}
1114
1115void identify_secondary_cpu(struct cpuinfo_x86 *c)
1116{
1117 BUG_ON(c == &boot_cpu_data);
1118 identify_cpu(c);
1119#ifdef CONFIG_X86_32
1120 enable_sep_cpu();
1121#endif
1122 mtrr_ap_init();
1123}
1124
1125struct msr_range {
1126 unsigned min;
1127 unsigned max;
1128};
1129
1130static const struct msr_range msr_range_array[] = {
1131 { 0x00000000, 0x00000418},
1132 { 0xc0000000, 0xc000040b},
1133 { 0xc0010000, 0xc0010142},
1134 { 0xc0011000, 0xc001103b},
1135};
1136
1137static void __print_cpu_msr(void)
1138{
1139 unsigned index_min, index_max;
1140 unsigned index;
1141 u64 val;
1142 int i;
1143
1144 for (i = 0; i < ARRAY_SIZE(msr_range_array); i++) {
1145 index_min = msr_range_array[i].min;
1146 index_max = msr_range_array[i].max;
1147
1148 for (index = index_min; index < index_max; index++) {
1149 if (rdmsrl_safe(index, &val))
1150 continue;
1151 pr_info(" MSR%08x: %016llx\n", index, val);
1152 }
1153 }
1154}
1155
1156static int show_msr;
1157
1158static __init int setup_show_msr(char *arg)
1159{
1160 int num;
1161
1162 get_option(&arg, &num);
1163
1164 if (num > 0)
1165 show_msr = num;
1166 return 1;
1167}
1168__setup("show_msr=", setup_show_msr);
1169
1170static __init int setup_noclflush(char *arg)
1171{
1172 setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1173 setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1174 return 1;
1175}
1176__setup("noclflush", setup_noclflush);
1177
1178void print_cpu_info(struct cpuinfo_x86 *c)
1179{
1180 const char *vendor = NULL;
1181
1182 if (c->x86_vendor < X86_VENDOR_NUM) {
1183 vendor = this_cpu->c_vendor;
1184 } else {
1185 if (c->cpuid_level >= 0)
1186 vendor = c->x86_vendor_id;
1187 }
1188
1189 if (vendor && !strstr(c->x86_model_id, vendor))
1190 pr_cont("%s ", vendor);
1191
1192 if (c->x86_model_id[0])
1193 pr_cont("%s", c->x86_model_id);
1194 else
1195 pr_cont("%d86", c->x86);
1196
1197 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1198
1199 if (c->x86_mask || c->cpuid_level >= 0)
1200 pr_cont(", stepping: 0x%x)\n", c->x86_mask);
1201 else
1202 pr_cont(")\n");
1203
1204 print_cpu_msr(c);
1205}
1206
1207void print_cpu_msr(struct cpuinfo_x86 *c)
1208{
1209 if (c->cpu_index < show_msr)
1210 __print_cpu_msr();
1211}
1212
1213static __init int setup_disablecpuid(char *arg)
1214{
1215 int bit;
1216
1217 if (get_option(&arg, &bit) && bit < NCAPINTS*32)
1218 setup_clear_cpu_cap(bit);
1219 else
1220 return 0;
1221
1222 return 1;
1223}
1224__setup("clearcpuid=", setup_disablecpuid);
1225
1226#ifdef CONFIG_X86_64
1227struct desc_ptr idt_descr = { NR_VECTORS * 16 - 1, (unsigned long) idt_table };
1228struct desc_ptr debug_idt_descr = { NR_VECTORS * 16 - 1,
1229 (unsigned long) debug_idt_table };
1230
1231DEFINE_PER_CPU_FIRST(union irq_stack_union,
1232 irq_stack_union) __aligned(PAGE_SIZE) __visible;
1233
1234/*
1235 * The following percpu variables are hot. Align current_task to
1236 * cacheline size such that they fall in the same cacheline.
1237 */
1238DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1239 &init_task;
1240EXPORT_PER_CPU_SYMBOL(current_task);
1241
1242DEFINE_PER_CPU(char *, irq_stack_ptr) =
1243 init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE - 64;
1244
1245DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1246
1247DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1248EXPORT_PER_CPU_SYMBOL(__preempt_count);
1249
1250/*
1251 * Special IST stacks which the CPU switches to when it calls
1252 * an IST-marked descriptor entry. Up to 7 stacks (hardware
1253 * limit), all of them are 4K, except the debug stack which
1254 * is 8K.
1255 */
1256static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
1257 [0 ... N_EXCEPTION_STACKS - 1] = EXCEPTION_STKSZ,
1258 [DEBUG_STACK - 1] = DEBUG_STKSZ
1259};
1260
1261static DEFINE_PER_CPU_PAGE_ALIGNED(char, exception_stacks
1262 [(N_EXCEPTION_STACKS - 1) * EXCEPTION_STKSZ + DEBUG_STKSZ]);
1263
1264/* May not be marked __init: used by software suspend */
1265void syscall_init(void)
1266{
1267 /*
1268 * LSTAR and STAR live in a bit strange symbiosis.
1269 * They both write to the same internal register. STAR allows to
1270 * set CS/DS but only a 32bit target. LSTAR sets the 64bit rip.
1271 */
1272 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1273 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1274
1275#ifdef CONFIG_IA32_EMULATION
1276 wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1277 /*
1278 * This only works on Intel CPUs.
1279 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1280 * This does not cause SYSENTER to jump to the wrong location, because
1281 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1282 */
1283 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1284 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1285 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1286#else
1287 wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1288 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1289 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1290 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1291#endif
1292
1293 /* Flags to clear on syscall */
1294 wrmsrl(MSR_SYSCALL_MASK,
1295 X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1296 X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1297}
1298
1299/*
1300 * Copies of the original ist values from the tss are only accessed during
1301 * debugging, no special alignment required.
1302 */
1303DEFINE_PER_CPU(struct orig_ist, orig_ist);
1304
1305static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1306DEFINE_PER_CPU(int, debug_stack_usage);
1307
1308int is_debug_stack(unsigned long addr)
1309{
1310 return __this_cpu_read(debug_stack_usage) ||
1311 (addr <= __this_cpu_read(debug_stack_addr) &&
1312 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1313}
1314NOKPROBE_SYMBOL(is_debug_stack);
1315
1316DEFINE_PER_CPU(u32, debug_idt_ctr);
1317
1318void debug_stack_set_zero(void)
1319{
1320 this_cpu_inc(debug_idt_ctr);
1321 load_current_idt();
1322}
1323NOKPROBE_SYMBOL(debug_stack_set_zero);
1324
1325void debug_stack_reset(void)
1326{
1327 if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1328 return;
1329 if (this_cpu_dec_return(debug_idt_ctr) == 0)
1330 load_current_idt();
1331}
1332NOKPROBE_SYMBOL(debug_stack_reset);
1333
1334#else /* CONFIG_X86_64 */
1335
1336DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1337EXPORT_PER_CPU_SYMBOL(current_task);
1338DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1339EXPORT_PER_CPU_SYMBOL(__preempt_count);
1340
1341/*
1342 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1343 * the top of the kernel stack. Use an extra percpu variable to track the
1344 * top of the kernel stack directly.
1345 */
1346DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1347 (unsigned long)&init_thread_union + THREAD_SIZE;
1348EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1349
1350#ifdef CONFIG_CC_STACKPROTECTOR
1351DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1352#endif
1353
1354#endif /* CONFIG_X86_64 */
1355
1356/*
1357 * Clear all 6 debug registers:
1358 */
1359static void clear_all_debug_regs(void)
1360{
1361 int i;
1362
1363 for (i = 0; i < 8; i++) {
1364 /* Ignore db4, db5 */
1365 if ((i == 4) || (i == 5))
1366 continue;
1367
1368 set_debugreg(0, i);
1369 }
1370}
1371
1372#ifdef CONFIG_KGDB
1373/*
1374 * Restore debug regs if using kgdbwait and you have a kernel debugger
1375 * connection established.
1376 */
1377static void dbg_restore_debug_regs(void)
1378{
1379 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1380 arch_kgdb_ops.correct_hw_break();
1381}
1382#else /* ! CONFIG_KGDB */
1383#define dbg_restore_debug_regs()
1384#endif /* ! CONFIG_KGDB */
1385
1386static void wait_for_master_cpu(int cpu)
1387{
1388#ifdef CONFIG_SMP
1389 /*
1390 * wait for ACK from master CPU before continuing
1391 * with AP initialization
1392 */
1393 WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1394 while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1395 cpu_relax();
1396#endif
1397}
1398
1399/*
1400 * cpu_init() initializes state that is per-CPU. Some data is already
1401 * initialized (naturally) in the bootstrap process, such as the GDT
1402 * and IDT. We reload them nevertheless, this function acts as a
1403 * 'CPU state barrier', nothing should get across.
1404 * A lot of state is already set up in PDA init for 64 bit
1405 */
1406#ifdef CONFIG_X86_64
1407
1408void cpu_init(void)
1409{
1410 struct orig_ist *oist;
1411 struct task_struct *me;
1412 struct tss_struct *t;
1413 unsigned long v;
1414 int cpu = stack_smp_processor_id();
1415 int i;
1416
1417 wait_for_master_cpu(cpu);
1418
1419 /*
1420 * Initialize the CR4 shadow before doing anything that could
1421 * try to read it.
1422 */
1423 cr4_init_shadow();
1424
1425 /*
1426 * Load microcode on this cpu if a valid microcode is available.
1427 * This is early microcode loading procedure.
1428 */
1429 load_ucode_ap();
1430
1431 t = &per_cpu(cpu_tss, cpu);
1432 oist = &per_cpu(orig_ist, cpu);
1433
1434#ifdef CONFIG_NUMA
1435 if (this_cpu_read(numa_node) == 0 &&
1436 early_cpu_to_node(cpu) != NUMA_NO_NODE)
1437 set_numa_node(early_cpu_to_node(cpu));
1438#endif
1439
1440 me = current;
1441
1442 pr_debug("Initializing CPU#%d\n", cpu);
1443
1444 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1445
1446 /*
1447 * Initialize the per-CPU GDT with the boot GDT,
1448 * and set up the GDT descriptor:
1449 */
1450
1451 switch_to_new_gdt(cpu);
1452 loadsegment(fs, 0);
1453
1454 load_current_idt();
1455
1456 memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1457 syscall_init();
1458
1459 wrmsrl(MSR_FS_BASE, 0);
1460 wrmsrl(MSR_KERNEL_GS_BASE, 0);
1461 barrier();
1462
1463 x86_configure_nx();
1464 x2apic_setup();
1465
1466 /*
1467 * set up and load the per-CPU TSS
1468 */
1469 if (!oist->ist[0]) {
1470 char *estacks = per_cpu(exception_stacks, cpu);
1471
1472 for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1473 estacks += exception_stack_sizes[v];
1474 oist->ist[v] = t->x86_tss.ist[v] =
1475 (unsigned long)estacks;
1476 if (v == DEBUG_STACK-1)
1477 per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1478 }
1479 }
1480
1481 t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1482
1483 /*
1484 * <= is required because the CPU will access up to
1485 * 8 bits beyond the end of the IO permission bitmap.
1486 */
1487 for (i = 0; i <= IO_BITMAP_LONGS; i++)
1488 t->io_bitmap[i] = ~0UL;
1489
1490 atomic_inc(&init_mm.mm_count);
1491 me->active_mm = &init_mm;
1492 BUG_ON(me->mm);
1493 enter_lazy_tlb(&init_mm, me);
1494
1495 load_sp0(t, ¤t->thread);
1496 set_tss_desc(cpu, t);
1497 load_TR_desc();
1498 load_mm_ldt(&init_mm);
1499
1500 clear_all_debug_regs();
1501 dbg_restore_debug_regs();
1502
1503 fpu__init_cpu();
1504
1505 if (is_uv_system())
1506 uv_cpu_init();
1507}
1508
1509#else
1510
1511void cpu_init(void)
1512{
1513 int cpu = smp_processor_id();
1514 struct task_struct *curr = current;
1515 struct tss_struct *t = &per_cpu(cpu_tss, cpu);
1516 struct thread_struct *thread = &curr->thread;
1517
1518 wait_for_master_cpu(cpu);
1519
1520 /*
1521 * Initialize the CR4 shadow before doing anything that could
1522 * try to read it.
1523 */
1524 cr4_init_shadow();
1525
1526 show_ucode_info_early();
1527
1528 pr_info("Initializing CPU#%d\n", cpu);
1529
1530 if (cpu_feature_enabled(X86_FEATURE_VME) ||
1531 cpu_has_tsc ||
1532 boot_cpu_has(X86_FEATURE_DE))
1533 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1534
1535 load_current_idt();
1536 switch_to_new_gdt(cpu);
1537
1538 /*
1539 * Set up and load the per-CPU TSS and LDT
1540 */
1541 atomic_inc(&init_mm.mm_count);
1542 curr->active_mm = &init_mm;
1543 BUG_ON(curr->mm);
1544 enter_lazy_tlb(&init_mm, curr);
1545
1546 load_sp0(t, thread);
1547 set_tss_desc(cpu, t);
1548 load_TR_desc();
1549 load_mm_ldt(&init_mm);
1550
1551 t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1552
1553#ifdef CONFIG_DOUBLEFAULT
1554 /* Set up doublefault TSS pointer in the GDT */
1555 __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1556#endif
1557
1558 clear_all_debug_regs();
1559 dbg_restore_debug_regs();
1560
1561 fpu__init_cpu();
1562}
1563#endif
1564
1565static void bsp_resume(void)
1566{
1567 if (this_cpu->c_bsp_resume)
1568 this_cpu->c_bsp_resume(&boot_cpu_data);
1569}
1570
1571static struct syscore_ops cpu_syscore_ops = {
1572 .resume = bsp_resume,
1573};
1574
1575static int __init init_cpu_syscore(void)
1576{
1577 register_syscore_ops(&cpu_syscore_ops);
1578 return 0;
1579}
1580core_initcall(init_cpu_syscore);