xref: /linux/arch/x86/kernel/cpu/common.c (revision ef69f8d2ff09518657c3ecaf2db8408c16549829)
1 #include <linux/bootmem.h>
2 #include <linux/linkage.h>
3 #include <linux/bitops.h>
4 #include <linux/kernel.h>
5 #include <linux/export.h>
6 #include <linux/percpu.h>
7 #include <linux/string.h>
8 #include <linux/ctype.h>
9 #include <linux/delay.h>
10 #include <linux/sched/mm.h>
11 #include <linux/sched/clock.h>
12 #include <linux/sched/task.h>
13 #include <linux/init.h>
14 #include <linux/kprobes.h>
15 #include <linux/kgdb.h>
16 #include <linux/smp.h>
17 #include <linux/io.h>
18 #include <linux/syscore_ops.h>
19 
20 #include <asm/stackprotector.h>
21 #include <asm/perf_event.h>
22 #include <asm/mmu_context.h>
23 #include <asm/archrandom.h>
24 #include <asm/hypervisor.h>
25 #include <asm/processor.h>
26 #include <asm/tlbflush.h>
27 #include <asm/debugreg.h>
28 #include <asm/sections.h>
29 #include <asm/vsyscall.h>
30 #include <linux/topology.h>
31 #include <linux/cpumask.h>
32 #include <asm/pgtable.h>
33 #include <linux/atomic.h>
34 #include <asm/proto.h>
35 #include <asm/setup.h>
36 #include <asm/apic.h>
37 #include <asm/desc.h>
38 #include <asm/fpu/internal.h>
39 #include <asm/mtrr.h>
40 #include <asm/hwcap2.h>
41 #include <linux/numa.h>
42 #include <asm/asm.h>
43 #include <asm/bugs.h>
44 #include <asm/cpu.h>
45 #include <asm/mce.h>
46 #include <asm/msr.h>
47 #include <asm/pat.h>
48 #include <asm/microcode.h>
49 #include <asm/microcode_intel.h>
50 
51 #ifdef CONFIG_X86_LOCAL_APIC
52 #include <asm/uv/uv.h>
53 #endif
54 
55 #include "cpu.h"
56 
57 u32 elf_hwcap2 __read_mostly;
58 
59 /* all of these masks are initialized in setup_cpu_local_masks() */
60 cpumask_var_t cpu_initialized_mask;
61 cpumask_var_t cpu_callout_mask;
62 cpumask_var_t cpu_callin_mask;
63 
64 /* representing cpus for which sibling maps can be computed */
65 cpumask_var_t cpu_sibling_setup_mask;
66 
67 /* correctly size the local cpu masks */
68 void __init setup_cpu_local_masks(void)
69 {
70 	alloc_bootmem_cpumask_var(&cpu_initialized_mask);
71 	alloc_bootmem_cpumask_var(&cpu_callin_mask);
72 	alloc_bootmem_cpumask_var(&cpu_callout_mask);
73 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
74 }
75 
76 static void default_init(struct cpuinfo_x86 *c)
77 {
78 #ifdef CONFIG_X86_64
79 	cpu_detect_cache_sizes(c);
80 #else
81 	/* Not much we can do here... */
82 	/* Check if at least it has cpuid */
83 	if (c->cpuid_level == -1) {
84 		/* No cpuid. It must be an ancient CPU */
85 		if (c->x86 == 4)
86 			strcpy(c->x86_model_id, "486");
87 		else if (c->x86 == 3)
88 			strcpy(c->x86_model_id, "386");
89 	}
90 #endif
91 }
92 
93 static const struct cpu_dev default_cpu = {
94 	.c_init		= default_init,
95 	.c_vendor	= "Unknown",
96 	.c_x86_vendor	= X86_VENDOR_UNKNOWN,
97 };
98 
99 static const struct cpu_dev *this_cpu = &default_cpu;
100 
101 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
102 #ifdef CONFIG_X86_64
103 	/*
104 	 * We need valid kernel segments for data and code in long mode too
105 	 * IRET will check the segment types  kkeil 2000/10/28
106 	 * Also sysret mandates a special GDT layout
107 	 *
108 	 * TLS descriptors are currently at a different place compared to i386.
109 	 * Hopefully nobody expects them at a fixed place (Wine?)
110 	 */
111 	[GDT_ENTRY_KERNEL32_CS]		= GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
112 	[GDT_ENTRY_KERNEL_CS]		= GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
113 	[GDT_ENTRY_KERNEL_DS]		= GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
114 	[GDT_ENTRY_DEFAULT_USER32_CS]	= GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
115 	[GDT_ENTRY_DEFAULT_USER_DS]	= GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
116 	[GDT_ENTRY_DEFAULT_USER_CS]	= GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
117 #else
118 	[GDT_ENTRY_KERNEL_CS]		= GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
119 	[GDT_ENTRY_KERNEL_DS]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
120 	[GDT_ENTRY_DEFAULT_USER_CS]	= GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
121 	[GDT_ENTRY_DEFAULT_USER_DS]	= GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
122 	/*
123 	 * Segments used for calling PnP BIOS have byte granularity.
124 	 * They code segments and data segments have fixed 64k limits,
125 	 * the transfer segment sizes are set at run time.
126 	 */
127 	/* 32-bit code */
128 	[GDT_ENTRY_PNPBIOS_CS32]	= GDT_ENTRY_INIT(0x409a, 0, 0xffff),
129 	/* 16-bit code */
130 	[GDT_ENTRY_PNPBIOS_CS16]	= GDT_ENTRY_INIT(0x009a, 0, 0xffff),
131 	/* 16-bit data */
132 	[GDT_ENTRY_PNPBIOS_DS]		= GDT_ENTRY_INIT(0x0092, 0, 0xffff),
133 	/* 16-bit data */
134 	[GDT_ENTRY_PNPBIOS_TS1]		= GDT_ENTRY_INIT(0x0092, 0, 0),
135 	/* 16-bit data */
136 	[GDT_ENTRY_PNPBIOS_TS2]		= GDT_ENTRY_INIT(0x0092, 0, 0),
137 	/*
138 	 * The APM segments have byte granularity and their bases
139 	 * are set at run time.  All have 64k limits.
140 	 */
141 	/* 32-bit code */
142 	[GDT_ENTRY_APMBIOS_BASE]	= GDT_ENTRY_INIT(0x409a, 0, 0xffff),
143 	/* 16-bit code */
144 	[GDT_ENTRY_APMBIOS_BASE+1]	= GDT_ENTRY_INIT(0x009a, 0, 0xffff),
145 	/* data */
146 	[GDT_ENTRY_APMBIOS_BASE+2]	= GDT_ENTRY_INIT(0x4092, 0, 0xffff),
147 
148 	[GDT_ENTRY_ESPFIX_SS]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
149 	[GDT_ENTRY_PERCPU]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
150 	GDT_STACK_CANARY_INIT
151 #endif
152 } };
153 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
154 
155 static int __init x86_mpx_setup(char *s)
156 {
157 	/* require an exact match without trailing characters */
158 	if (strlen(s))
159 		return 0;
160 
161 	/* do not emit a message if the feature is not present */
162 	if (!boot_cpu_has(X86_FEATURE_MPX))
163 		return 1;
164 
165 	setup_clear_cpu_cap(X86_FEATURE_MPX);
166 	pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
167 	return 1;
168 }
169 __setup("nompx", x86_mpx_setup);
170 
171 #ifdef CONFIG_X86_64
172 static int __init x86_nopcid_setup(char *s)
173 {
174 	/* nopcid doesn't accept parameters */
175 	if (s)
176 		return -EINVAL;
177 
178 	/* do not emit a message if the feature is not present */
179 	if (!boot_cpu_has(X86_FEATURE_PCID))
180 		return 0;
181 
182 	setup_clear_cpu_cap(X86_FEATURE_PCID);
183 	pr_info("nopcid: PCID feature disabled\n");
184 	return 0;
185 }
186 early_param("nopcid", x86_nopcid_setup);
187 #endif
188 
189 static int __init x86_noinvpcid_setup(char *s)
190 {
191 	/* noinvpcid doesn't accept parameters */
192 	if (s)
193 		return -EINVAL;
194 
195 	/* do not emit a message if the feature is not present */
196 	if (!boot_cpu_has(X86_FEATURE_INVPCID))
197 		return 0;
198 
199 	setup_clear_cpu_cap(X86_FEATURE_INVPCID);
200 	pr_info("noinvpcid: INVPCID feature disabled\n");
201 	return 0;
202 }
203 early_param("noinvpcid", x86_noinvpcid_setup);
204 
205 #ifdef CONFIG_X86_32
206 static int cachesize_override = -1;
207 static int disable_x86_serial_nr = 1;
208 
209 static int __init cachesize_setup(char *str)
210 {
211 	get_option(&str, &cachesize_override);
212 	return 1;
213 }
214 __setup("cachesize=", cachesize_setup);
215 
216 static int __init x86_sep_setup(char *s)
217 {
218 	setup_clear_cpu_cap(X86_FEATURE_SEP);
219 	return 1;
220 }
221 __setup("nosep", x86_sep_setup);
222 
223 /* Standard macro to see if a specific flag is changeable */
224 static inline int flag_is_changeable_p(u32 flag)
225 {
226 	u32 f1, f2;
227 
228 	/*
229 	 * Cyrix and IDT cpus allow disabling of CPUID
230 	 * so the code below may return different results
231 	 * when it is executed before and after enabling
232 	 * the CPUID. Add "volatile" to not allow gcc to
233 	 * optimize the subsequent calls to this function.
234 	 */
235 	asm volatile ("pushfl		\n\t"
236 		      "pushfl		\n\t"
237 		      "popl %0		\n\t"
238 		      "movl %0, %1	\n\t"
239 		      "xorl %2, %0	\n\t"
240 		      "pushl %0		\n\t"
241 		      "popfl		\n\t"
242 		      "pushfl		\n\t"
243 		      "popl %0		\n\t"
244 		      "popfl		\n\t"
245 
246 		      : "=&r" (f1), "=&r" (f2)
247 		      : "ir" (flag));
248 
249 	return ((f1^f2) & flag) != 0;
250 }
251 
252 /* Probe for the CPUID instruction */
253 int have_cpuid_p(void)
254 {
255 	return flag_is_changeable_p(X86_EFLAGS_ID);
256 }
257 
258 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
259 {
260 	unsigned long lo, hi;
261 
262 	if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
263 		return;
264 
265 	/* Disable processor serial number: */
266 
267 	rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
268 	lo |= 0x200000;
269 	wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
270 
271 	pr_notice("CPU serial number disabled.\n");
272 	clear_cpu_cap(c, X86_FEATURE_PN);
273 
274 	/* Disabling the serial number may affect the cpuid level */
275 	c->cpuid_level = cpuid_eax(0);
276 }
277 
278 static int __init x86_serial_nr_setup(char *s)
279 {
280 	disable_x86_serial_nr = 0;
281 	return 1;
282 }
283 __setup("serialnumber", x86_serial_nr_setup);
284 #else
285 static inline int flag_is_changeable_p(u32 flag)
286 {
287 	return 1;
288 }
289 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
290 {
291 }
292 #endif
293 
294 static __init int setup_disable_smep(char *arg)
295 {
296 	setup_clear_cpu_cap(X86_FEATURE_SMEP);
297 	/* Check for things that depend on SMEP being enabled: */
298 	check_mpx_erratum(&boot_cpu_data);
299 	return 1;
300 }
301 __setup("nosmep", setup_disable_smep);
302 
303 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
304 {
305 	if (cpu_has(c, X86_FEATURE_SMEP))
306 		cr4_set_bits(X86_CR4_SMEP);
307 }
308 
309 static __init int setup_disable_smap(char *arg)
310 {
311 	setup_clear_cpu_cap(X86_FEATURE_SMAP);
312 	return 1;
313 }
314 __setup("nosmap", setup_disable_smap);
315 
316 static __always_inline void setup_smap(struct cpuinfo_x86 *c)
317 {
318 	unsigned long eflags = native_save_fl();
319 
320 	/* This should have been cleared long ago */
321 	BUG_ON(eflags & X86_EFLAGS_AC);
322 
323 	if (cpu_has(c, X86_FEATURE_SMAP)) {
324 #ifdef CONFIG_X86_SMAP
325 		cr4_set_bits(X86_CR4_SMAP);
326 #else
327 		cr4_clear_bits(X86_CR4_SMAP);
328 #endif
329 	}
330 }
331 
332 static __always_inline void setup_umip(struct cpuinfo_x86 *c)
333 {
334 	/* Check the boot processor, plus build option for UMIP. */
335 	if (!cpu_feature_enabled(X86_FEATURE_UMIP))
336 		goto out;
337 
338 	/* Check the current processor's cpuid bits. */
339 	if (!cpu_has(c, X86_FEATURE_UMIP))
340 		goto out;
341 
342 	cr4_set_bits(X86_CR4_UMIP);
343 
344 	pr_info("x86/cpu: Activated the Intel User Mode Instruction Prevention (UMIP) CPU feature\n");
345 
346 	return;
347 
348 out:
349 	/*
350 	 * Make sure UMIP is disabled in case it was enabled in a
351 	 * previous boot (e.g., via kexec).
352 	 */
353 	cr4_clear_bits(X86_CR4_UMIP);
354 }
355 
356 /*
357  * Protection Keys are not available in 32-bit mode.
358  */
359 static bool pku_disabled;
360 
361 static __always_inline void setup_pku(struct cpuinfo_x86 *c)
362 {
363 	/* check the boot processor, plus compile options for PKU: */
364 	if (!cpu_feature_enabled(X86_FEATURE_PKU))
365 		return;
366 	/* checks the actual processor's cpuid bits: */
367 	if (!cpu_has(c, X86_FEATURE_PKU))
368 		return;
369 	if (pku_disabled)
370 		return;
371 
372 	cr4_set_bits(X86_CR4_PKE);
373 	/*
374 	 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
375 	 * cpuid bit to be set.  We need to ensure that we
376 	 * update that bit in this CPU's "cpu_info".
377 	 */
378 	get_cpu_cap(c);
379 }
380 
381 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
382 static __init int setup_disable_pku(char *arg)
383 {
384 	/*
385 	 * Do not clear the X86_FEATURE_PKU bit.  All of the
386 	 * runtime checks are against OSPKE so clearing the
387 	 * bit does nothing.
388 	 *
389 	 * This way, we will see "pku" in cpuinfo, but not
390 	 * "ospke", which is exactly what we want.  It shows
391 	 * that the CPU has PKU, but the OS has not enabled it.
392 	 * This happens to be exactly how a system would look
393 	 * if we disabled the config option.
394 	 */
395 	pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
396 	pku_disabled = true;
397 	return 1;
398 }
399 __setup("nopku", setup_disable_pku);
400 #endif /* CONFIG_X86_64 */
401 
402 /*
403  * Some CPU features depend on higher CPUID levels, which may not always
404  * be available due to CPUID level capping or broken virtualization
405  * software.  Add those features to this table to auto-disable them.
406  */
407 struct cpuid_dependent_feature {
408 	u32 feature;
409 	u32 level;
410 };
411 
412 static const struct cpuid_dependent_feature
413 cpuid_dependent_features[] = {
414 	{ X86_FEATURE_MWAIT,		0x00000005 },
415 	{ X86_FEATURE_DCA,		0x00000009 },
416 	{ X86_FEATURE_XSAVE,		0x0000000d },
417 	{ 0, 0 }
418 };
419 
420 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
421 {
422 	const struct cpuid_dependent_feature *df;
423 
424 	for (df = cpuid_dependent_features; df->feature; df++) {
425 
426 		if (!cpu_has(c, df->feature))
427 			continue;
428 		/*
429 		 * Note: cpuid_level is set to -1 if unavailable, but
430 		 * extended_extended_level is set to 0 if unavailable
431 		 * and the legitimate extended levels are all negative
432 		 * when signed; hence the weird messing around with
433 		 * signs here...
434 		 */
435 		if (!((s32)df->level < 0 ?
436 		     (u32)df->level > (u32)c->extended_cpuid_level :
437 		     (s32)df->level > (s32)c->cpuid_level))
438 			continue;
439 
440 		clear_cpu_cap(c, df->feature);
441 		if (!warn)
442 			continue;
443 
444 		pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
445 			x86_cap_flag(df->feature), df->level);
446 	}
447 }
448 
449 /*
450  * Naming convention should be: <Name> [(<Codename>)]
451  * This table only is used unless init_<vendor>() below doesn't set it;
452  * in particular, if CPUID levels 0x80000002..4 are supported, this
453  * isn't used
454  */
455 
456 /* Look up CPU names by table lookup. */
457 static const char *table_lookup_model(struct cpuinfo_x86 *c)
458 {
459 #ifdef CONFIG_X86_32
460 	const struct legacy_cpu_model_info *info;
461 
462 	if (c->x86_model >= 16)
463 		return NULL;	/* Range check */
464 
465 	if (!this_cpu)
466 		return NULL;
467 
468 	info = this_cpu->legacy_models;
469 
470 	while (info->family) {
471 		if (info->family == c->x86)
472 			return info->model_names[c->x86_model];
473 		info++;
474 	}
475 #endif
476 	return NULL;		/* Not found */
477 }
478 
479 __u32 cpu_caps_cleared[NCAPINTS];
480 __u32 cpu_caps_set[NCAPINTS];
481 
482 void load_percpu_segment(int cpu)
483 {
484 #ifdef CONFIG_X86_32
485 	loadsegment(fs, __KERNEL_PERCPU);
486 #else
487 	__loadsegment_simple(gs, 0);
488 	wrmsrl(MSR_GS_BASE, (unsigned long)per_cpu(irq_stack_union.gs_base, cpu));
489 #endif
490 	load_stack_canary_segment();
491 }
492 
493 /* Setup the fixmap mapping only once per-processor */
494 static inline void setup_fixmap_gdt(int cpu)
495 {
496 #ifdef CONFIG_X86_64
497 	/* On 64-bit systems, we use a read-only fixmap GDT. */
498 	pgprot_t prot = PAGE_KERNEL_RO;
499 #else
500 	/*
501 	 * On native 32-bit systems, the GDT cannot be read-only because
502 	 * our double fault handler uses a task gate, and entering through
503 	 * a task gate needs to change an available TSS to busy.  If the GDT
504 	 * is read-only, that will triple fault.
505 	 *
506 	 * On Xen PV, the GDT must be read-only because the hypervisor requires
507 	 * it.
508 	 */
509 	pgprot_t prot = boot_cpu_has(X86_FEATURE_XENPV) ?
510 		PAGE_KERNEL_RO : PAGE_KERNEL;
511 #endif
512 
513 	__set_fixmap(get_cpu_gdt_ro_index(cpu), get_cpu_gdt_paddr(cpu), prot);
514 }
515 
516 /* Load the original GDT from the per-cpu structure */
517 void load_direct_gdt(int cpu)
518 {
519 	struct desc_ptr gdt_descr;
520 
521 	gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
522 	gdt_descr.size = GDT_SIZE - 1;
523 	load_gdt(&gdt_descr);
524 }
525 EXPORT_SYMBOL_GPL(load_direct_gdt);
526 
527 /* Load a fixmap remapping of the per-cpu GDT */
528 void load_fixmap_gdt(int cpu)
529 {
530 	struct desc_ptr gdt_descr;
531 
532 	gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
533 	gdt_descr.size = GDT_SIZE - 1;
534 	load_gdt(&gdt_descr);
535 }
536 EXPORT_SYMBOL_GPL(load_fixmap_gdt);
537 
538 /*
539  * Current gdt points %fs at the "master" per-cpu area: after this,
540  * it's on the real one.
541  */
542 void switch_to_new_gdt(int cpu)
543 {
544 	/* Load the original GDT */
545 	load_direct_gdt(cpu);
546 	/* Reload the per-cpu base */
547 	load_percpu_segment(cpu);
548 }
549 
550 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
551 
552 static void get_model_name(struct cpuinfo_x86 *c)
553 {
554 	unsigned int *v;
555 	char *p, *q, *s;
556 
557 	if (c->extended_cpuid_level < 0x80000004)
558 		return;
559 
560 	v = (unsigned int *)c->x86_model_id;
561 	cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
562 	cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
563 	cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
564 	c->x86_model_id[48] = 0;
565 
566 	/* Trim whitespace */
567 	p = q = s = &c->x86_model_id[0];
568 
569 	while (*p == ' ')
570 		p++;
571 
572 	while (*p) {
573 		/* Note the last non-whitespace index */
574 		if (!isspace(*p))
575 			s = q;
576 
577 		*q++ = *p++;
578 	}
579 
580 	*(s + 1) = '\0';
581 }
582 
583 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
584 {
585 	unsigned int n, dummy, ebx, ecx, edx, l2size;
586 
587 	n = c->extended_cpuid_level;
588 
589 	if (n >= 0x80000005) {
590 		cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
591 		c->x86_cache_size = (ecx>>24) + (edx>>24);
592 #ifdef CONFIG_X86_64
593 		/* On K8 L1 TLB is inclusive, so don't count it */
594 		c->x86_tlbsize = 0;
595 #endif
596 	}
597 
598 	if (n < 0x80000006)	/* Some chips just has a large L1. */
599 		return;
600 
601 	cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
602 	l2size = ecx >> 16;
603 
604 #ifdef CONFIG_X86_64
605 	c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
606 #else
607 	/* do processor-specific cache resizing */
608 	if (this_cpu->legacy_cache_size)
609 		l2size = this_cpu->legacy_cache_size(c, l2size);
610 
611 	/* Allow user to override all this if necessary. */
612 	if (cachesize_override != -1)
613 		l2size = cachesize_override;
614 
615 	if (l2size == 0)
616 		return;		/* Again, no L2 cache is possible */
617 #endif
618 
619 	c->x86_cache_size = l2size;
620 }
621 
622 u16 __read_mostly tlb_lli_4k[NR_INFO];
623 u16 __read_mostly tlb_lli_2m[NR_INFO];
624 u16 __read_mostly tlb_lli_4m[NR_INFO];
625 u16 __read_mostly tlb_lld_4k[NR_INFO];
626 u16 __read_mostly tlb_lld_2m[NR_INFO];
627 u16 __read_mostly tlb_lld_4m[NR_INFO];
628 u16 __read_mostly tlb_lld_1g[NR_INFO];
629 
630 static void cpu_detect_tlb(struct cpuinfo_x86 *c)
631 {
632 	if (this_cpu->c_detect_tlb)
633 		this_cpu->c_detect_tlb(c);
634 
635 	pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
636 		tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
637 		tlb_lli_4m[ENTRIES]);
638 
639 	pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
640 		tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
641 		tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
642 }
643 
644 void detect_ht(struct cpuinfo_x86 *c)
645 {
646 #ifdef CONFIG_SMP
647 	u32 eax, ebx, ecx, edx;
648 	int index_msb, core_bits;
649 	static bool printed;
650 
651 	if (!cpu_has(c, X86_FEATURE_HT))
652 		return;
653 
654 	if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
655 		goto out;
656 
657 	if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
658 		return;
659 
660 	cpuid(1, &eax, &ebx, &ecx, &edx);
661 
662 	smp_num_siblings = (ebx & 0xff0000) >> 16;
663 
664 	if (smp_num_siblings == 1) {
665 		pr_info_once("CPU0: Hyper-Threading is disabled\n");
666 		goto out;
667 	}
668 
669 	if (smp_num_siblings <= 1)
670 		goto out;
671 
672 	index_msb = get_count_order(smp_num_siblings);
673 	c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
674 
675 	smp_num_siblings = smp_num_siblings / c->x86_max_cores;
676 
677 	index_msb = get_count_order(smp_num_siblings);
678 
679 	core_bits = get_count_order(c->x86_max_cores);
680 
681 	c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
682 				       ((1 << core_bits) - 1);
683 
684 out:
685 	if (!printed && (c->x86_max_cores * smp_num_siblings) > 1) {
686 		pr_info("CPU: Physical Processor ID: %d\n",
687 			c->phys_proc_id);
688 		pr_info("CPU: Processor Core ID: %d\n",
689 			c->cpu_core_id);
690 		printed = 1;
691 	}
692 #endif
693 }
694 
695 static void get_cpu_vendor(struct cpuinfo_x86 *c)
696 {
697 	char *v = c->x86_vendor_id;
698 	int i;
699 
700 	for (i = 0; i < X86_VENDOR_NUM; i++) {
701 		if (!cpu_devs[i])
702 			break;
703 
704 		if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
705 		    (cpu_devs[i]->c_ident[1] &&
706 		     !strcmp(v, cpu_devs[i]->c_ident[1]))) {
707 
708 			this_cpu = cpu_devs[i];
709 			c->x86_vendor = this_cpu->c_x86_vendor;
710 			return;
711 		}
712 	}
713 
714 	pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
715 		    "CPU: Your system may be unstable.\n", v);
716 
717 	c->x86_vendor = X86_VENDOR_UNKNOWN;
718 	this_cpu = &default_cpu;
719 }
720 
721 void cpu_detect(struct cpuinfo_x86 *c)
722 {
723 	/* Get vendor name */
724 	cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
725 	      (unsigned int *)&c->x86_vendor_id[0],
726 	      (unsigned int *)&c->x86_vendor_id[8],
727 	      (unsigned int *)&c->x86_vendor_id[4]);
728 
729 	c->x86 = 4;
730 	/* Intel-defined flags: level 0x00000001 */
731 	if (c->cpuid_level >= 0x00000001) {
732 		u32 junk, tfms, cap0, misc;
733 
734 		cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
735 		c->x86		= x86_family(tfms);
736 		c->x86_model	= x86_model(tfms);
737 		c->x86_mask	= x86_stepping(tfms);
738 
739 		if (cap0 & (1<<19)) {
740 			c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
741 			c->x86_cache_alignment = c->x86_clflush_size;
742 		}
743 	}
744 }
745 
746 static void apply_forced_caps(struct cpuinfo_x86 *c)
747 {
748 	int i;
749 
750 	for (i = 0; i < NCAPINTS; i++) {
751 		c->x86_capability[i] &= ~cpu_caps_cleared[i];
752 		c->x86_capability[i] |= cpu_caps_set[i];
753 	}
754 }
755 
756 void get_cpu_cap(struct cpuinfo_x86 *c)
757 {
758 	u32 eax, ebx, ecx, edx;
759 
760 	/* Intel-defined flags: level 0x00000001 */
761 	if (c->cpuid_level >= 0x00000001) {
762 		cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
763 
764 		c->x86_capability[CPUID_1_ECX] = ecx;
765 		c->x86_capability[CPUID_1_EDX] = edx;
766 	}
767 
768 	/* Thermal and Power Management Leaf: level 0x00000006 (eax) */
769 	if (c->cpuid_level >= 0x00000006)
770 		c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
771 
772 	/* Additional Intel-defined flags: level 0x00000007 */
773 	if (c->cpuid_level >= 0x00000007) {
774 		cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
775 		c->x86_capability[CPUID_7_0_EBX] = ebx;
776 		c->x86_capability[CPUID_7_ECX] = ecx;
777 	}
778 
779 	/* Extended state features: level 0x0000000d */
780 	if (c->cpuid_level >= 0x0000000d) {
781 		cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
782 
783 		c->x86_capability[CPUID_D_1_EAX] = eax;
784 	}
785 
786 	/* Additional Intel-defined flags: level 0x0000000F */
787 	if (c->cpuid_level >= 0x0000000F) {
788 
789 		/* QoS sub-leaf, EAX=0Fh, ECX=0 */
790 		cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
791 		c->x86_capability[CPUID_F_0_EDX] = edx;
792 
793 		if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
794 			/* will be overridden if occupancy monitoring exists */
795 			c->x86_cache_max_rmid = ebx;
796 
797 			/* QoS sub-leaf, EAX=0Fh, ECX=1 */
798 			cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
799 			c->x86_capability[CPUID_F_1_EDX] = edx;
800 
801 			if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
802 			      ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
803 			       (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
804 				c->x86_cache_max_rmid = ecx;
805 				c->x86_cache_occ_scale = ebx;
806 			}
807 		} else {
808 			c->x86_cache_max_rmid = -1;
809 			c->x86_cache_occ_scale = -1;
810 		}
811 	}
812 
813 	/* AMD-defined flags: level 0x80000001 */
814 	eax = cpuid_eax(0x80000000);
815 	c->extended_cpuid_level = eax;
816 
817 	if ((eax & 0xffff0000) == 0x80000000) {
818 		if (eax >= 0x80000001) {
819 			cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
820 
821 			c->x86_capability[CPUID_8000_0001_ECX] = ecx;
822 			c->x86_capability[CPUID_8000_0001_EDX] = edx;
823 		}
824 	}
825 
826 	if (c->extended_cpuid_level >= 0x80000007) {
827 		cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
828 
829 		c->x86_capability[CPUID_8000_0007_EBX] = ebx;
830 		c->x86_power = edx;
831 	}
832 
833 	if (c->extended_cpuid_level >= 0x80000008) {
834 		cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
835 
836 		c->x86_virt_bits = (eax >> 8) & 0xff;
837 		c->x86_phys_bits = eax & 0xff;
838 		c->x86_capability[CPUID_8000_0008_EBX] = ebx;
839 	}
840 #ifdef CONFIG_X86_32
841 	else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
842 		c->x86_phys_bits = 36;
843 #endif
844 
845 	if (c->extended_cpuid_level >= 0x8000000a)
846 		c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
847 
848 	init_scattered_cpuid_features(c);
849 
850 	/*
851 	 * Clear/Set all flags overridden by options, after probe.
852 	 * This needs to happen each time we re-probe, which may happen
853 	 * several times during CPU initialization.
854 	 */
855 	apply_forced_caps(c);
856 }
857 
858 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
859 {
860 #ifdef CONFIG_X86_32
861 	int i;
862 
863 	/*
864 	 * First of all, decide if this is a 486 or higher
865 	 * It's a 486 if we can modify the AC flag
866 	 */
867 	if (flag_is_changeable_p(X86_EFLAGS_AC))
868 		c->x86 = 4;
869 	else
870 		c->x86 = 3;
871 
872 	for (i = 0; i < X86_VENDOR_NUM; i++)
873 		if (cpu_devs[i] && cpu_devs[i]->c_identify) {
874 			c->x86_vendor_id[0] = 0;
875 			cpu_devs[i]->c_identify(c);
876 			if (c->x86_vendor_id[0]) {
877 				get_cpu_vendor(c);
878 				break;
879 			}
880 		}
881 #endif
882 }
883 
884 /*
885  * Do minimum CPU detection early.
886  * Fields really needed: vendor, cpuid_level, family, model, mask,
887  * cache alignment.
888  * The others are not touched to avoid unwanted side effects.
889  *
890  * WARNING: this function is only called on the boot CPU.  Don't add code
891  * here that is supposed to run on all CPUs.
892  */
893 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
894 {
895 #ifdef CONFIG_X86_64
896 	c->x86_clflush_size = 64;
897 	c->x86_phys_bits = 36;
898 	c->x86_virt_bits = 48;
899 #else
900 	c->x86_clflush_size = 32;
901 	c->x86_phys_bits = 32;
902 	c->x86_virt_bits = 32;
903 #endif
904 	c->x86_cache_alignment = c->x86_clflush_size;
905 
906 	memset(&c->x86_capability, 0, sizeof c->x86_capability);
907 	c->extended_cpuid_level = 0;
908 
909 	/* cyrix could have cpuid enabled via c_identify()*/
910 	if (have_cpuid_p()) {
911 		cpu_detect(c);
912 		get_cpu_vendor(c);
913 		get_cpu_cap(c);
914 		setup_force_cpu_cap(X86_FEATURE_CPUID);
915 
916 		if (this_cpu->c_early_init)
917 			this_cpu->c_early_init(c);
918 
919 		c->cpu_index = 0;
920 		filter_cpuid_features(c, false);
921 
922 		if (this_cpu->c_bsp_init)
923 			this_cpu->c_bsp_init(c);
924 	} else {
925 		identify_cpu_without_cpuid(c);
926 		setup_clear_cpu_cap(X86_FEATURE_CPUID);
927 	}
928 
929 	setup_force_cpu_cap(X86_FEATURE_ALWAYS);
930 	fpu__init_system(c);
931 
932 #ifdef CONFIG_X86_32
933 	/*
934 	 * Regardless of whether PCID is enumerated, the SDM says
935 	 * that it can't be enabled in 32-bit mode.
936 	 */
937 	setup_clear_cpu_cap(X86_FEATURE_PCID);
938 #endif
939 }
940 
941 void __init early_cpu_init(void)
942 {
943 	const struct cpu_dev *const *cdev;
944 	int count = 0;
945 
946 #ifdef CONFIG_PROCESSOR_SELECT
947 	pr_info("KERNEL supported cpus:\n");
948 #endif
949 
950 	for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
951 		const struct cpu_dev *cpudev = *cdev;
952 
953 		if (count >= X86_VENDOR_NUM)
954 			break;
955 		cpu_devs[count] = cpudev;
956 		count++;
957 
958 #ifdef CONFIG_PROCESSOR_SELECT
959 		{
960 			unsigned int j;
961 
962 			for (j = 0; j < 2; j++) {
963 				if (!cpudev->c_ident[j])
964 					continue;
965 				pr_info("  %s %s\n", cpudev->c_vendor,
966 					cpudev->c_ident[j]);
967 			}
968 		}
969 #endif
970 	}
971 	early_identify_cpu(&boot_cpu_data);
972 }
973 
974 /*
975  * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
976  * unfortunately, that's not true in practice because of early VIA
977  * chips and (more importantly) broken virtualizers that are not easy
978  * to detect. In the latter case it doesn't even *fail* reliably, so
979  * probing for it doesn't even work. Disable it completely on 32-bit
980  * unless we can find a reliable way to detect all the broken cases.
981  * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
982  */
983 static void detect_nopl(struct cpuinfo_x86 *c)
984 {
985 #ifdef CONFIG_X86_32
986 	clear_cpu_cap(c, X86_FEATURE_NOPL);
987 #else
988 	set_cpu_cap(c, X86_FEATURE_NOPL);
989 #endif
990 }
991 
992 static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
993 {
994 #ifdef CONFIG_X86_64
995 	/*
996 	 * Empirically, writing zero to a segment selector on AMD does
997 	 * not clear the base, whereas writing zero to a segment
998 	 * selector on Intel does clear the base.  Intel's behavior
999 	 * allows slightly faster context switches in the common case
1000 	 * where GS is unused by the prev and next threads.
1001 	 *
1002 	 * Since neither vendor documents this anywhere that I can see,
1003 	 * detect it directly instead of hardcoding the choice by
1004 	 * vendor.
1005 	 *
1006 	 * I've designated AMD's behavior as the "bug" because it's
1007 	 * counterintuitive and less friendly.
1008 	 */
1009 
1010 	unsigned long old_base, tmp;
1011 	rdmsrl(MSR_FS_BASE, old_base);
1012 	wrmsrl(MSR_FS_BASE, 1);
1013 	loadsegment(fs, 0);
1014 	rdmsrl(MSR_FS_BASE, tmp);
1015 	if (tmp != 0)
1016 		set_cpu_bug(c, X86_BUG_NULL_SEG);
1017 	wrmsrl(MSR_FS_BASE, old_base);
1018 #endif
1019 }
1020 
1021 static void generic_identify(struct cpuinfo_x86 *c)
1022 {
1023 	c->extended_cpuid_level = 0;
1024 
1025 	if (!have_cpuid_p())
1026 		identify_cpu_without_cpuid(c);
1027 
1028 	/* cyrix could have cpuid enabled via c_identify()*/
1029 	if (!have_cpuid_p())
1030 		return;
1031 
1032 	cpu_detect(c);
1033 
1034 	get_cpu_vendor(c);
1035 
1036 	get_cpu_cap(c);
1037 
1038 	if (c->cpuid_level >= 0x00000001) {
1039 		c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1040 #ifdef CONFIG_X86_32
1041 # ifdef CONFIG_SMP
1042 		c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1043 # else
1044 		c->apicid = c->initial_apicid;
1045 # endif
1046 #endif
1047 		c->phys_proc_id = c->initial_apicid;
1048 	}
1049 
1050 	get_model_name(c); /* Default name */
1051 
1052 	detect_nopl(c);
1053 
1054 	detect_null_seg_behavior(c);
1055 
1056 	/*
1057 	 * ESPFIX is a strange bug.  All real CPUs have it.  Paravirt
1058 	 * systems that run Linux at CPL > 0 may or may not have the
1059 	 * issue, but, even if they have the issue, there's absolutely
1060 	 * nothing we can do about it because we can't use the real IRET
1061 	 * instruction.
1062 	 *
1063 	 * NB: For the time being, only 32-bit kernels support
1064 	 * X86_BUG_ESPFIX as such.  64-bit kernels directly choose
1065 	 * whether to apply espfix using paravirt hooks.  If any
1066 	 * non-paravirt system ever shows up that does *not* have the
1067 	 * ESPFIX issue, we can change this.
1068 	 */
1069 #ifdef CONFIG_X86_32
1070 # ifdef CONFIG_PARAVIRT
1071 	do {
1072 		extern void native_iret(void);
1073 		if (pv_cpu_ops.iret == native_iret)
1074 			set_cpu_bug(c, X86_BUG_ESPFIX);
1075 	} while (0);
1076 # else
1077 	set_cpu_bug(c, X86_BUG_ESPFIX);
1078 # endif
1079 #endif
1080 }
1081 
1082 static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1083 {
1084 	/*
1085 	 * The heavy lifting of max_rmid and cache_occ_scale are handled
1086 	 * in get_cpu_cap().  Here we just set the max_rmid for the boot_cpu
1087 	 * in case CQM bits really aren't there in this CPU.
1088 	 */
1089 	if (c != &boot_cpu_data) {
1090 		boot_cpu_data.x86_cache_max_rmid =
1091 			min(boot_cpu_data.x86_cache_max_rmid,
1092 			    c->x86_cache_max_rmid);
1093 	}
1094 }
1095 
1096 /*
1097  * Validate that ACPI/mptables have the same information about the
1098  * effective APIC id and update the package map.
1099  */
1100 static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1101 {
1102 #ifdef CONFIG_SMP
1103 	unsigned int apicid, cpu = smp_processor_id();
1104 
1105 	apicid = apic->cpu_present_to_apicid(cpu);
1106 
1107 	if (apicid != c->apicid) {
1108 		pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1109 		       cpu, apicid, c->initial_apicid);
1110 	}
1111 	BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
1112 #else
1113 	c->logical_proc_id = 0;
1114 #endif
1115 }
1116 
1117 /*
1118  * This does the hard work of actually picking apart the CPU stuff...
1119  */
1120 static void identify_cpu(struct cpuinfo_x86 *c)
1121 {
1122 	int i;
1123 
1124 	c->loops_per_jiffy = loops_per_jiffy;
1125 	c->x86_cache_size = -1;
1126 	c->x86_vendor = X86_VENDOR_UNKNOWN;
1127 	c->x86_model = c->x86_mask = 0;	/* So far unknown... */
1128 	c->x86_vendor_id[0] = '\0'; /* Unset */
1129 	c->x86_model_id[0] = '\0';  /* Unset */
1130 	c->x86_max_cores = 1;
1131 	c->x86_coreid_bits = 0;
1132 	c->cu_id = 0xff;
1133 #ifdef CONFIG_X86_64
1134 	c->x86_clflush_size = 64;
1135 	c->x86_phys_bits = 36;
1136 	c->x86_virt_bits = 48;
1137 #else
1138 	c->cpuid_level = -1;	/* CPUID not detected */
1139 	c->x86_clflush_size = 32;
1140 	c->x86_phys_bits = 32;
1141 	c->x86_virt_bits = 32;
1142 #endif
1143 	c->x86_cache_alignment = c->x86_clflush_size;
1144 	memset(&c->x86_capability, 0, sizeof c->x86_capability);
1145 
1146 	generic_identify(c);
1147 
1148 	if (this_cpu->c_identify)
1149 		this_cpu->c_identify(c);
1150 
1151 	/* Clear/Set all flags overridden by options, after probe */
1152 	apply_forced_caps(c);
1153 
1154 #ifdef CONFIG_X86_64
1155 	c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1156 #endif
1157 
1158 	/*
1159 	 * Vendor-specific initialization.  In this section we
1160 	 * canonicalize the feature flags, meaning if there are
1161 	 * features a certain CPU supports which CPUID doesn't
1162 	 * tell us, CPUID claiming incorrect flags, or other bugs,
1163 	 * we handle them here.
1164 	 *
1165 	 * At the end of this section, c->x86_capability better
1166 	 * indicate the features this CPU genuinely supports!
1167 	 */
1168 	if (this_cpu->c_init)
1169 		this_cpu->c_init(c);
1170 
1171 	/* Disable the PN if appropriate */
1172 	squash_the_stupid_serial_number(c);
1173 
1174 	/* Set up SMEP/SMAP/UMIP */
1175 	setup_smep(c);
1176 	setup_smap(c);
1177 	setup_umip(c);
1178 
1179 	/*
1180 	 * The vendor-specific functions might have changed features.
1181 	 * Now we do "generic changes."
1182 	 */
1183 
1184 	/* Filter out anything that depends on CPUID levels we don't have */
1185 	filter_cpuid_features(c, true);
1186 
1187 	/* If the model name is still unset, do table lookup. */
1188 	if (!c->x86_model_id[0]) {
1189 		const char *p;
1190 		p = table_lookup_model(c);
1191 		if (p)
1192 			strcpy(c->x86_model_id, p);
1193 		else
1194 			/* Last resort... */
1195 			sprintf(c->x86_model_id, "%02x/%02x",
1196 				c->x86, c->x86_model);
1197 	}
1198 
1199 #ifdef CONFIG_X86_64
1200 	detect_ht(c);
1201 #endif
1202 
1203 	x86_init_rdrand(c);
1204 	x86_init_cache_qos(c);
1205 	setup_pku(c);
1206 
1207 	/*
1208 	 * Clear/Set all flags overridden by options, need do it
1209 	 * before following smp all cpus cap AND.
1210 	 */
1211 	apply_forced_caps(c);
1212 
1213 	/*
1214 	 * On SMP, boot_cpu_data holds the common feature set between
1215 	 * all CPUs; so make sure that we indicate which features are
1216 	 * common between the CPUs.  The first time this routine gets
1217 	 * executed, c == &boot_cpu_data.
1218 	 */
1219 	if (c != &boot_cpu_data) {
1220 		/* AND the already accumulated flags with these */
1221 		for (i = 0; i < NCAPINTS; i++)
1222 			boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1223 
1224 		/* OR, i.e. replicate the bug flags */
1225 		for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1226 			c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1227 	}
1228 
1229 	/* Init Machine Check Exception if available. */
1230 	mcheck_cpu_init(c);
1231 
1232 	select_idle_routine(c);
1233 
1234 #ifdef CONFIG_NUMA
1235 	numa_add_cpu(smp_processor_id());
1236 #endif
1237 }
1238 
1239 /*
1240  * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1241  * on 32-bit kernels:
1242  */
1243 #ifdef CONFIG_X86_32
1244 void enable_sep_cpu(void)
1245 {
1246 	struct tss_struct *tss;
1247 	int cpu;
1248 
1249 	if (!boot_cpu_has(X86_FEATURE_SEP))
1250 		return;
1251 
1252 	cpu = get_cpu();
1253 	tss = &per_cpu(cpu_tss, cpu);
1254 
1255 	/*
1256 	 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1257 	 * see the big comment in struct x86_hw_tss's definition.
1258 	 */
1259 
1260 	tss->x86_tss.ss1 = __KERNEL_CS;
1261 	wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1262 
1263 	wrmsr(MSR_IA32_SYSENTER_ESP,
1264 	      (unsigned long)tss + offsetofend(struct tss_struct, SYSENTER_stack),
1265 	      0);
1266 
1267 	wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1268 
1269 	put_cpu();
1270 }
1271 #endif
1272 
1273 void __init identify_boot_cpu(void)
1274 {
1275 	identify_cpu(&boot_cpu_data);
1276 #ifdef CONFIG_X86_32
1277 	sysenter_setup();
1278 	enable_sep_cpu();
1279 #endif
1280 	cpu_detect_tlb(&boot_cpu_data);
1281 }
1282 
1283 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1284 {
1285 	BUG_ON(c == &boot_cpu_data);
1286 	identify_cpu(c);
1287 #ifdef CONFIG_X86_32
1288 	enable_sep_cpu();
1289 #endif
1290 	mtrr_ap_init();
1291 	validate_apic_and_package_id(c);
1292 }
1293 
1294 static __init int setup_noclflush(char *arg)
1295 {
1296 	setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1297 	setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1298 	return 1;
1299 }
1300 __setup("noclflush", setup_noclflush);
1301 
1302 void print_cpu_info(struct cpuinfo_x86 *c)
1303 {
1304 	const char *vendor = NULL;
1305 
1306 	if (c->x86_vendor < X86_VENDOR_NUM) {
1307 		vendor = this_cpu->c_vendor;
1308 	} else {
1309 		if (c->cpuid_level >= 0)
1310 			vendor = c->x86_vendor_id;
1311 	}
1312 
1313 	if (vendor && !strstr(c->x86_model_id, vendor))
1314 		pr_cont("%s ", vendor);
1315 
1316 	if (c->x86_model_id[0])
1317 		pr_cont("%s", c->x86_model_id);
1318 	else
1319 		pr_cont("%d86", c->x86);
1320 
1321 	pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1322 
1323 	if (c->x86_mask || c->cpuid_level >= 0)
1324 		pr_cont(", stepping: 0x%x)\n", c->x86_mask);
1325 	else
1326 		pr_cont(")\n");
1327 }
1328 
1329 /*
1330  * clearcpuid= was already parsed in fpu__init_parse_early_param.
1331  * But we need to keep a dummy __setup around otherwise it would
1332  * show up as an environment variable for init.
1333  */
1334 static __init int setup_clearcpuid(char *arg)
1335 {
1336 	return 1;
1337 }
1338 __setup("clearcpuid=", setup_clearcpuid);
1339 
1340 #ifdef CONFIG_X86_64
1341 DEFINE_PER_CPU_FIRST(union irq_stack_union,
1342 		     irq_stack_union) __aligned(PAGE_SIZE) __visible;
1343 
1344 /*
1345  * The following percpu variables are hot.  Align current_task to
1346  * cacheline size such that they fall in the same cacheline.
1347  */
1348 DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1349 	&init_task;
1350 EXPORT_PER_CPU_SYMBOL(current_task);
1351 
1352 DEFINE_PER_CPU(char *, irq_stack_ptr) =
1353 	init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE;
1354 
1355 DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1356 
1357 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1358 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1359 
1360 /*
1361  * Special IST stacks which the CPU switches to when it calls
1362  * an IST-marked descriptor entry. Up to 7 stacks (hardware
1363  * limit), all of them are 4K, except the debug stack which
1364  * is 8K.
1365  */
1366 static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
1367 	  [0 ... N_EXCEPTION_STACKS - 1]	= EXCEPTION_STKSZ,
1368 	  [DEBUG_STACK - 1]			= DEBUG_STKSZ
1369 };
1370 
1371 static DEFINE_PER_CPU_PAGE_ALIGNED(char, exception_stacks
1372 	[(N_EXCEPTION_STACKS - 1) * EXCEPTION_STKSZ + DEBUG_STKSZ]);
1373 
1374 /* May not be marked __init: used by software suspend */
1375 void syscall_init(void)
1376 {
1377 	wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1378 	wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1379 
1380 #ifdef CONFIG_IA32_EMULATION
1381 	wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1382 	/*
1383 	 * This only works on Intel CPUs.
1384 	 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1385 	 * This does not cause SYSENTER to jump to the wrong location, because
1386 	 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1387 	 */
1388 	wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1389 	wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1390 	wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1391 #else
1392 	wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1393 	wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1394 	wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1395 	wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1396 #endif
1397 
1398 	/* Flags to clear on syscall */
1399 	wrmsrl(MSR_SYSCALL_MASK,
1400 	       X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1401 	       X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1402 }
1403 
1404 /*
1405  * Copies of the original ist values from the tss are only accessed during
1406  * debugging, no special alignment required.
1407  */
1408 DEFINE_PER_CPU(struct orig_ist, orig_ist);
1409 
1410 static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1411 DEFINE_PER_CPU(int, debug_stack_usage);
1412 
1413 int is_debug_stack(unsigned long addr)
1414 {
1415 	return __this_cpu_read(debug_stack_usage) ||
1416 		(addr <= __this_cpu_read(debug_stack_addr) &&
1417 		 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1418 }
1419 NOKPROBE_SYMBOL(is_debug_stack);
1420 
1421 DEFINE_PER_CPU(u32, debug_idt_ctr);
1422 
1423 void debug_stack_set_zero(void)
1424 {
1425 	this_cpu_inc(debug_idt_ctr);
1426 	load_current_idt();
1427 }
1428 NOKPROBE_SYMBOL(debug_stack_set_zero);
1429 
1430 void debug_stack_reset(void)
1431 {
1432 	if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1433 		return;
1434 	if (this_cpu_dec_return(debug_idt_ctr) == 0)
1435 		load_current_idt();
1436 }
1437 NOKPROBE_SYMBOL(debug_stack_reset);
1438 
1439 #else	/* CONFIG_X86_64 */
1440 
1441 DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1442 EXPORT_PER_CPU_SYMBOL(current_task);
1443 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1444 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1445 
1446 /*
1447  * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1448  * the top of the kernel stack.  Use an extra percpu variable to track the
1449  * top of the kernel stack directly.
1450  */
1451 DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1452 	(unsigned long)&init_thread_union + THREAD_SIZE;
1453 EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1454 
1455 #ifdef CONFIG_CC_STACKPROTECTOR
1456 DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1457 #endif
1458 
1459 #endif	/* CONFIG_X86_64 */
1460 
1461 /*
1462  * Clear all 6 debug registers:
1463  */
1464 static void clear_all_debug_regs(void)
1465 {
1466 	int i;
1467 
1468 	for (i = 0; i < 8; i++) {
1469 		/* Ignore db4, db5 */
1470 		if ((i == 4) || (i == 5))
1471 			continue;
1472 
1473 		set_debugreg(0, i);
1474 	}
1475 }
1476 
1477 #ifdef CONFIG_KGDB
1478 /*
1479  * Restore debug regs if using kgdbwait and you have a kernel debugger
1480  * connection established.
1481  */
1482 static void dbg_restore_debug_regs(void)
1483 {
1484 	if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1485 		arch_kgdb_ops.correct_hw_break();
1486 }
1487 #else /* ! CONFIG_KGDB */
1488 #define dbg_restore_debug_regs()
1489 #endif /* ! CONFIG_KGDB */
1490 
1491 static void wait_for_master_cpu(int cpu)
1492 {
1493 #ifdef CONFIG_SMP
1494 	/*
1495 	 * wait for ACK from master CPU before continuing
1496 	 * with AP initialization
1497 	 */
1498 	WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1499 	while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1500 		cpu_relax();
1501 #endif
1502 }
1503 
1504 /*
1505  * cpu_init() initializes state that is per-CPU. Some data is already
1506  * initialized (naturally) in the bootstrap process, such as the GDT
1507  * and IDT. We reload them nevertheless, this function acts as a
1508  * 'CPU state barrier', nothing should get across.
1509  * A lot of state is already set up in PDA init for 64 bit
1510  */
1511 #ifdef CONFIG_X86_64
1512 
1513 void cpu_init(void)
1514 {
1515 	struct orig_ist *oist;
1516 	struct task_struct *me;
1517 	struct tss_struct *t;
1518 	unsigned long v;
1519 	int cpu = raw_smp_processor_id();
1520 	int i;
1521 
1522 	wait_for_master_cpu(cpu);
1523 
1524 	/*
1525 	 * Initialize the CR4 shadow before doing anything that could
1526 	 * try to read it.
1527 	 */
1528 	cr4_init_shadow();
1529 
1530 	if (cpu)
1531 		load_ucode_ap();
1532 
1533 	t = &per_cpu(cpu_tss, cpu);
1534 	oist = &per_cpu(orig_ist, cpu);
1535 
1536 #ifdef CONFIG_NUMA
1537 	if (this_cpu_read(numa_node) == 0 &&
1538 	    early_cpu_to_node(cpu) != NUMA_NO_NODE)
1539 		set_numa_node(early_cpu_to_node(cpu));
1540 #endif
1541 
1542 	me = current;
1543 
1544 	pr_debug("Initializing CPU#%d\n", cpu);
1545 
1546 	cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1547 
1548 	/*
1549 	 * Initialize the per-CPU GDT with the boot GDT,
1550 	 * and set up the GDT descriptor:
1551 	 */
1552 
1553 	switch_to_new_gdt(cpu);
1554 	loadsegment(fs, 0);
1555 
1556 	load_current_idt();
1557 
1558 	memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1559 	syscall_init();
1560 
1561 	wrmsrl(MSR_FS_BASE, 0);
1562 	wrmsrl(MSR_KERNEL_GS_BASE, 0);
1563 	barrier();
1564 
1565 	x86_configure_nx();
1566 	x2apic_setup();
1567 
1568 	/*
1569 	 * set up and load the per-CPU TSS
1570 	 */
1571 	if (!oist->ist[0]) {
1572 		char *estacks = per_cpu(exception_stacks, cpu);
1573 
1574 		for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1575 			estacks += exception_stack_sizes[v];
1576 			oist->ist[v] = t->x86_tss.ist[v] =
1577 					(unsigned long)estacks;
1578 			if (v == DEBUG_STACK-1)
1579 				per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1580 		}
1581 	}
1582 
1583 	t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1584 
1585 	/*
1586 	 * <= is required because the CPU will access up to
1587 	 * 8 bits beyond the end of the IO permission bitmap.
1588 	 */
1589 	for (i = 0; i <= IO_BITMAP_LONGS; i++)
1590 		t->io_bitmap[i] = ~0UL;
1591 
1592 	mmgrab(&init_mm);
1593 	me->active_mm = &init_mm;
1594 	BUG_ON(me->mm);
1595 	initialize_tlbstate_and_flush();
1596 	enter_lazy_tlb(&init_mm, me);
1597 
1598 	/*
1599 	 * Initialize the TSS.  Don't bother initializing sp0, as the initial
1600 	 * task never enters user mode.
1601 	 */
1602 	set_tss_desc(cpu, t);
1603 	load_TR_desc();
1604 
1605 	load_mm_ldt(&init_mm);
1606 
1607 	clear_all_debug_regs();
1608 	dbg_restore_debug_regs();
1609 
1610 	fpu__init_cpu();
1611 
1612 	if (is_uv_system())
1613 		uv_cpu_init();
1614 
1615 	setup_fixmap_gdt(cpu);
1616 	load_fixmap_gdt(cpu);
1617 }
1618 
1619 #else
1620 
1621 void cpu_init(void)
1622 {
1623 	int cpu = smp_processor_id();
1624 	struct task_struct *curr = current;
1625 	struct tss_struct *t = &per_cpu(cpu_tss, cpu);
1626 
1627 	wait_for_master_cpu(cpu);
1628 
1629 	/*
1630 	 * Initialize the CR4 shadow before doing anything that could
1631 	 * try to read it.
1632 	 */
1633 	cr4_init_shadow();
1634 
1635 	show_ucode_info_early();
1636 
1637 	pr_info("Initializing CPU#%d\n", cpu);
1638 
1639 	if (cpu_feature_enabled(X86_FEATURE_VME) ||
1640 	    boot_cpu_has(X86_FEATURE_TSC) ||
1641 	    boot_cpu_has(X86_FEATURE_DE))
1642 		cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1643 
1644 	load_current_idt();
1645 	switch_to_new_gdt(cpu);
1646 
1647 	/*
1648 	 * Set up and load the per-CPU TSS and LDT
1649 	 */
1650 	mmgrab(&init_mm);
1651 	curr->active_mm = &init_mm;
1652 	BUG_ON(curr->mm);
1653 	initialize_tlbstate_and_flush();
1654 	enter_lazy_tlb(&init_mm, curr);
1655 
1656 	/*
1657 	 * Initialize the TSS.  Don't bother initializing sp0, as the initial
1658 	 * task never enters user mode.
1659 	 */
1660 	set_tss_desc(cpu, t);
1661 	load_TR_desc();
1662 
1663 	load_mm_ldt(&init_mm);
1664 
1665 	t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1666 
1667 #ifdef CONFIG_DOUBLEFAULT
1668 	/* Set up doublefault TSS pointer in the GDT */
1669 	__set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1670 #endif
1671 
1672 	clear_all_debug_regs();
1673 	dbg_restore_debug_regs();
1674 
1675 	fpu__init_cpu();
1676 
1677 	setup_fixmap_gdt(cpu);
1678 	load_fixmap_gdt(cpu);
1679 }
1680 #endif
1681 
1682 static void bsp_resume(void)
1683 {
1684 	if (this_cpu->c_bsp_resume)
1685 		this_cpu->c_bsp_resume(&boot_cpu_data);
1686 }
1687 
1688 static struct syscore_ops cpu_syscore_ops = {
1689 	.resume		= bsp_resume,
1690 };
1691 
1692 static int __init init_cpu_syscore(void)
1693 {
1694 	register_syscore_ops(&cpu_syscore_ops);
1695 	return 0;
1696 }
1697 core_initcall(init_cpu_syscore);
1698