xref: /openbmc/linux/arch/x86/kernel/cpu/common.c (revision 293d5b43)
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.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/bugs.h>
41 #include <asm/cpu.h>
42 #include <asm/mce.h>
43 #include <asm/msr.h>
44 #include <asm/pat.h>
45 #include <asm/microcode.h>
46 #include <asm/microcode_intel.h>
47 
48 #ifdef CONFIG_X86_LOCAL_APIC
49 #include <asm/uv/uv.h>
50 #endif
51 
52 #include "cpu.h"
53 
54 /* all of these masks are initialized in setup_cpu_local_masks() */
55 cpumask_var_t cpu_initialized_mask;
56 cpumask_var_t cpu_callout_mask;
57 cpumask_var_t cpu_callin_mask;
58 
59 /* representing cpus for which sibling maps can be computed */
60 cpumask_var_t cpu_sibling_setup_mask;
61 
62 /* correctly size the local cpu masks */
63 void __init setup_cpu_local_masks(void)
64 {
65 	alloc_bootmem_cpumask_var(&cpu_initialized_mask);
66 	alloc_bootmem_cpumask_var(&cpu_callin_mask);
67 	alloc_bootmem_cpumask_var(&cpu_callout_mask);
68 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
69 }
70 
71 static void default_init(struct cpuinfo_x86 *c)
72 {
73 #ifdef CONFIG_X86_64
74 	cpu_detect_cache_sizes(c);
75 #else
76 	/* Not much we can do here... */
77 	/* Check if at least it has cpuid */
78 	if (c->cpuid_level == -1) {
79 		/* No cpuid. It must be an ancient CPU */
80 		if (c->x86 == 4)
81 			strcpy(c->x86_model_id, "486");
82 		else if (c->x86 == 3)
83 			strcpy(c->x86_model_id, "386");
84 	}
85 #endif
86 }
87 
88 static const struct cpu_dev default_cpu = {
89 	.c_init		= default_init,
90 	.c_vendor	= "Unknown",
91 	.c_x86_vendor	= X86_VENDOR_UNKNOWN,
92 };
93 
94 static const struct cpu_dev *this_cpu = &default_cpu;
95 
96 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
97 #ifdef CONFIG_X86_64
98 	/*
99 	 * We need valid kernel segments for data and code in long mode too
100 	 * IRET will check the segment types  kkeil 2000/10/28
101 	 * Also sysret mandates a special GDT layout
102 	 *
103 	 * TLS descriptors are currently at a different place compared to i386.
104 	 * Hopefully nobody expects them at a fixed place (Wine?)
105 	 */
106 	[GDT_ENTRY_KERNEL32_CS]		= GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
107 	[GDT_ENTRY_KERNEL_CS]		= GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
108 	[GDT_ENTRY_KERNEL_DS]		= GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
109 	[GDT_ENTRY_DEFAULT_USER32_CS]	= GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
110 	[GDT_ENTRY_DEFAULT_USER_DS]	= GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
111 	[GDT_ENTRY_DEFAULT_USER_CS]	= GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
112 #else
113 	[GDT_ENTRY_KERNEL_CS]		= GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
114 	[GDT_ENTRY_KERNEL_DS]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
115 	[GDT_ENTRY_DEFAULT_USER_CS]	= GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
116 	[GDT_ENTRY_DEFAULT_USER_DS]	= GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
117 	/*
118 	 * Segments used for calling PnP BIOS have byte granularity.
119 	 * They code segments and data segments have fixed 64k limits,
120 	 * the transfer segment sizes are set at run time.
121 	 */
122 	/* 32-bit code */
123 	[GDT_ENTRY_PNPBIOS_CS32]	= GDT_ENTRY_INIT(0x409a, 0, 0xffff),
124 	/* 16-bit code */
125 	[GDT_ENTRY_PNPBIOS_CS16]	= GDT_ENTRY_INIT(0x009a, 0, 0xffff),
126 	/* 16-bit data */
127 	[GDT_ENTRY_PNPBIOS_DS]		= GDT_ENTRY_INIT(0x0092, 0, 0xffff),
128 	/* 16-bit data */
129 	[GDT_ENTRY_PNPBIOS_TS1]		= GDT_ENTRY_INIT(0x0092, 0, 0),
130 	/* 16-bit data */
131 	[GDT_ENTRY_PNPBIOS_TS2]		= GDT_ENTRY_INIT(0x0092, 0, 0),
132 	/*
133 	 * The APM segments have byte granularity and their bases
134 	 * are set at run time.  All have 64k limits.
135 	 */
136 	/* 32-bit code */
137 	[GDT_ENTRY_APMBIOS_BASE]	= GDT_ENTRY_INIT(0x409a, 0, 0xffff),
138 	/* 16-bit code */
139 	[GDT_ENTRY_APMBIOS_BASE+1]	= GDT_ENTRY_INIT(0x009a, 0, 0xffff),
140 	/* data */
141 	[GDT_ENTRY_APMBIOS_BASE+2]	= GDT_ENTRY_INIT(0x4092, 0, 0xffff),
142 
143 	[GDT_ENTRY_ESPFIX_SS]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
144 	[GDT_ENTRY_PERCPU]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
145 	GDT_STACK_CANARY_INIT
146 #endif
147 } };
148 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
149 
150 static int __init x86_mpx_setup(char *s)
151 {
152 	/* require an exact match without trailing characters */
153 	if (strlen(s))
154 		return 0;
155 
156 	/* do not emit a message if the feature is not present */
157 	if (!boot_cpu_has(X86_FEATURE_MPX))
158 		return 1;
159 
160 	setup_clear_cpu_cap(X86_FEATURE_MPX);
161 	pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
162 	return 1;
163 }
164 __setup("nompx", x86_mpx_setup);
165 
166 static int __init x86_noinvpcid_setup(char *s)
167 {
168 	/* noinvpcid doesn't accept parameters */
169 	if (s)
170 		return -EINVAL;
171 
172 	/* do not emit a message if the feature is not present */
173 	if (!boot_cpu_has(X86_FEATURE_INVPCID))
174 		return 0;
175 
176 	setup_clear_cpu_cap(X86_FEATURE_INVPCID);
177 	pr_info("noinvpcid: INVPCID feature disabled\n");
178 	return 0;
179 }
180 early_param("noinvpcid", x86_noinvpcid_setup);
181 
182 #ifdef CONFIG_X86_32
183 static int cachesize_override = -1;
184 static int disable_x86_serial_nr = 1;
185 
186 static int __init cachesize_setup(char *str)
187 {
188 	get_option(&str, &cachesize_override);
189 	return 1;
190 }
191 __setup("cachesize=", cachesize_setup);
192 
193 static int __init x86_sep_setup(char *s)
194 {
195 	setup_clear_cpu_cap(X86_FEATURE_SEP);
196 	return 1;
197 }
198 __setup("nosep", x86_sep_setup);
199 
200 /* Standard macro to see if a specific flag is changeable */
201 static inline int flag_is_changeable_p(u32 flag)
202 {
203 	u32 f1, f2;
204 
205 	/*
206 	 * Cyrix and IDT cpus allow disabling of CPUID
207 	 * so the code below may return different results
208 	 * when it is executed before and after enabling
209 	 * the CPUID. Add "volatile" to not allow gcc to
210 	 * optimize the subsequent calls to this function.
211 	 */
212 	asm volatile ("pushfl		\n\t"
213 		      "pushfl		\n\t"
214 		      "popl %0		\n\t"
215 		      "movl %0, %1	\n\t"
216 		      "xorl %2, %0	\n\t"
217 		      "pushl %0		\n\t"
218 		      "popfl		\n\t"
219 		      "pushfl		\n\t"
220 		      "popl %0		\n\t"
221 		      "popfl		\n\t"
222 
223 		      : "=&r" (f1), "=&r" (f2)
224 		      : "ir" (flag));
225 
226 	return ((f1^f2) & flag) != 0;
227 }
228 
229 /* Probe for the CPUID instruction */
230 int have_cpuid_p(void)
231 {
232 	return flag_is_changeable_p(X86_EFLAGS_ID);
233 }
234 
235 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
236 {
237 	unsigned long lo, hi;
238 
239 	if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
240 		return;
241 
242 	/* Disable processor serial number: */
243 
244 	rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
245 	lo |= 0x200000;
246 	wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
247 
248 	pr_notice("CPU serial number disabled.\n");
249 	clear_cpu_cap(c, X86_FEATURE_PN);
250 
251 	/* Disabling the serial number may affect the cpuid level */
252 	c->cpuid_level = cpuid_eax(0);
253 }
254 
255 static int __init x86_serial_nr_setup(char *s)
256 {
257 	disable_x86_serial_nr = 0;
258 	return 1;
259 }
260 __setup("serialnumber", x86_serial_nr_setup);
261 #else
262 static inline int flag_is_changeable_p(u32 flag)
263 {
264 	return 1;
265 }
266 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
267 {
268 }
269 #endif
270 
271 static __init int setup_disable_smep(char *arg)
272 {
273 	setup_clear_cpu_cap(X86_FEATURE_SMEP);
274 	/* Check for things that depend on SMEP being enabled: */
275 	check_mpx_erratum(&boot_cpu_data);
276 	return 1;
277 }
278 __setup("nosmep", setup_disable_smep);
279 
280 static __always_inline void setup_smep(struct cpuinfo_x86 *c)
281 {
282 	if (cpu_has(c, X86_FEATURE_SMEP))
283 		cr4_set_bits(X86_CR4_SMEP);
284 }
285 
286 static __init int setup_disable_smap(char *arg)
287 {
288 	setup_clear_cpu_cap(X86_FEATURE_SMAP);
289 	return 1;
290 }
291 __setup("nosmap", setup_disable_smap);
292 
293 static __always_inline void setup_smap(struct cpuinfo_x86 *c)
294 {
295 	unsigned long eflags = native_save_fl();
296 
297 	/* This should have been cleared long ago */
298 	BUG_ON(eflags & X86_EFLAGS_AC);
299 
300 	if (cpu_has(c, X86_FEATURE_SMAP)) {
301 #ifdef CONFIG_X86_SMAP
302 		cr4_set_bits(X86_CR4_SMAP);
303 #else
304 		cr4_clear_bits(X86_CR4_SMAP);
305 #endif
306 	}
307 }
308 
309 /*
310  * Protection Keys are not available in 32-bit mode.
311  */
312 static bool pku_disabled;
313 
314 static __always_inline void setup_pku(struct cpuinfo_x86 *c)
315 {
316 	/* check the boot processor, plus compile options for PKU: */
317 	if (!cpu_feature_enabled(X86_FEATURE_PKU))
318 		return;
319 	/* checks the actual processor's cpuid bits: */
320 	if (!cpu_has(c, X86_FEATURE_PKU))
321 		return;
322 	if (pku_disabled)
323 		return;
324 
325 	cr4_set_bits(X86_CR4_PKE);
326 	/*
327 	 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
328 	 * cpuid bit to be set.  We need to ensure that we
329 	 * update that bit in this CPU's "cpu_info".
330 	 */
331 	get_cpu_cap(c);
332 }
333 
334 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
335 static __init int setup_disable_pku(char *arg)
336 {
337 	/*
338 	 * Do not clear the X86_FEATURE_PKU bit.  All of the
339 	 * runtime checks are against OSPKE so clearing the
340 	 * bit does nothing.
341 	 *
342 	 * This way, we will see "pku" in cpuinfo, but not
343 	 * "ospke", which is exactly what we want.  It shows
344 	 * that the CPU has PKU, but the OS has not enabled it.
345 	 * This happens to be exactly how a system would look
346 	 * if we disabled the config option.
347 	 */
348 	pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
349 	pku_disabled = true;
350 	return 1;
351 }
352 __setup("nopku", setup_disable_pku);
353 #endif /* CONFIG_X86_64 */
354 
355 /*
356  * Some CPU features depend on higher CPUID levels, which may not always
357  * be available due to CPUID level capping or broken virtualization
358  * software.  Add those features to this table to auto-disable them.
359  */
360 struct cpuid_dependent_feature {
361 	u32 feature;
362 	u32 level;
363 };
364 
365 static const struct cpuid_dependent_feature
366 cpuid_dependent_features[] = {
367 	{ X86_FEATURE_MWAIT,		0x00000005 },
368 	{ X86_FEATURE_DCA,		0x00000009 },
369 	{ X86_FEATURE_XSAVE,		0x0000000d },
370 	{ 0, 0 }
371 };
372 
373 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
374 {
375 	const struct cpuid_dependent_feature *df;
376 
377 	for (df = cpuid_dependent_features; df->feature; df++) {
378 
379 		if (!cpu_has(c, df->feature))
380 			continue;
381 		/*
382 		 * Note: cpuid_level is set to -1 if unavailable, but
383 		 * extended_extended_level is set to 0 if unavailable
384 		 * and the legitimate extended levels are all negative
385 		 * when signed; hence the weird messing around with
386 		 * signs here...
387 		 */
388 		if (!((s32)df->level < 0 ?
389 		     (u32)df->level > (u32)c->extended_cpuid_level :
390 		     (s32)df->level > (s32)c->cpuid_level))
391 			continue;
392 
393 		clear_cpu_cap(c, df->feature);
394 		if (!warn)
395 			continue;
396 
397 		pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
398 			x86_cap_flag(df->feature), df->level);
399 	}
400 }
401 
402 /*
403  * Naming convention should be: <Name> [(<Codename>)]
404  * This table only is used unless init_<vendor>() below doesn't set it;
405  * in particular, if CPUID levels 0x80000002..4 are supported, this
406  * isn't used
407  */
408 
409 /* Look up CPU names by table lookup. */
410 static const char *table_lookup_model(struct cpuinfo_x86 *c)
411 {
412 #ifdef CONFIG_X86_32
413 	const struct legacy_cpu_model_info *info;
414 
415 	if (c->x86_model >= 16)
416 		return NULL;	/* Range check */
417 
418 	if (!this_cpu)
419 		return NULL;
420 
421 	info = this_cpu->legacy_models;
422 
423 	while (info->family) {
424 		if (info->family == c->x86)
425 			return info->model_names[c->x86_model];
426 		info++;
427 	}
428 #endif
429 	return NULL;		/* Not found */
430 }
431 
432 __u32 cpu_caps_cleared[NCAPINTS];
433 __u32 cpu_caps_set[NCAPINTS];
434 
435 void load_percpu_segment(int cpu)
436 {
437 #ifdef CONFIG_X86_32
438 	loadsegment(fs, __KERNEL_PERCPU);
439 #else
440 	__loadsegment_simple(gs, 0);
441 	wrmsrl(MSR_GS_BASE, (unsigned long)per_cpu(irq_stack_union.gs_base, cpu));
442 #endif
443 	load_stack_canary_segment();
444 }
445 
446 /*
447  * Current gdt points %fs at the "master" per-cpu area: after this,
448  * it's on the real one.
449  */
450 void switch_to_new_gdt(int cpu)
451 {
452 	struct desc_ptr gdt_descr;
453 
454 	gdt_descr.address = (long)get_cpu_gdt_table(cpu);
455 	gdt_descr.size = GDT_SIZE - 1;
456 	load_gdt(&gdt_descr);
457 	/* Reload the per-cpu base */
458 
459 	load_percpu_segment(cpu);
460 }
461 
462 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
463 
464 static void get_model_name(struct cpuinfo_x86 *c)
465 {
466 	unsigned int *v;
467 	char *p, *q, *s;
468 
469 	if (c->extended_cpuid_level < 0x80000004)
470 		return;
471 
472 	v = (unsigned int *)c->x86_model_id;
473 	cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
474 	cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
475 	cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
476 	c->x86_model_id[48] = 0;
477 
478 	/* Trim whitespace */
479 	p = q = s = &c->x86_model_id[0];
480 
481 	while (*p == ' ')
482 		p++;
483 
484 	while (*p) {
485 		/* Note the last non-whitespace index */
486 		if (!isspace(*p))
487 			s = q;
488 
489 		*q++ = *p++;
490 	}
491 
492 	*(s + 1) = '\0';
493 }
494 
495 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
496 {
497 	unsigned int n, dummy, ebx, ecx, edx, l2size;
498 
499 	n = c->extended_cpuid_level;
500 
501 	if (n >= 0x80000005) {
502 		cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
503 		c->x86_cache_size = (ecx>>24) + (edx>>24);
504 #ifdef CONFIG_X86_64
505 		/* On K8 L1 TLB is inclusive, so don't count it */
506 		c->x86_tlbsize = 0;
507 #endif
508 	}
509 
510 	if (n < 0x80000006)	/* Some chips just has a large L1. */
511 		return;
512 
513 	cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
514 	l2size = ecx >> 16;
515 
516 #ifdef CONFIG_X86_64
517 	c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
518 #else
519 	/* do processor-specific cache resizing */
520 	if (this_cpu->legacy_cache_size)
521 		l2size = this_cpu->legacy_cache_size(c, l2size);
522 
523 	/* Allow user to override all this if necessary. */
524 	if (cachesize_override != -1)
525 		l2size = cachesize_override;
526 
527 	if (l2size == 0)
528 		return;		/* Again, no L2 cache is possible */
529 #endif
530 
531 	c->x86_cache_size = l2size;
532 }
533 
534 u16 __read_mostly tlb_lli_4k[NR_INFO];
535 u16 __read_mostly tlb_lli_2m[NR_INFO];
536 u16 __read_mostly tlb_lli_4m[NR_INFO];
537 u16 __read_mostly tlb_lld_4k[NR_INFO];
538 u16 __read_mostly tlb_lld_2m[NR_INFO];
539 u16 __read_mostly tlb_lld_4m[NR_INFO];
540 u16 __read_mostly tlb_lld_1g[NR_INFO];
541 
542 static void cpu_detect_tlb(struct cpuinfo_x86 *c)
543 {
544 	if (this_cpu->c_detect_tlb)
545 		this_cpu->c_detect_tlb(c);
546 
547 	pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
548 		tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
549 		tlb_lli_4m[ENTRIES]);
550 
551 	pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
552 		tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
553 		tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
554 }
555 
556 void detect_ht(struct cpuinfo_x86 *c)
557 {
558 #ifdef CONFIG_SMP
559 	u32 eax, ebx, ecx, edx;
560 	int index_msb, core_bits;
561 	static bool printed;
562 
563 	if (!cpu_has(c, X86_FEATURE_HT))
564 		return;
565 
566 	if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
567 		goto out;
568 
569 	if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
570 		return;
571 
572 	cpuid(1, &eax, &ebx, &ecx, &edx);
573 
574 	smp_num_siblings = (ebx & 0xff0000) >> 16;
575 
576 	if (smp_num_siblings == 1) {
577 		pr_info_once("CPU0: Hyper-Threading is disabled\n");
578 		goto out;
579 	}
580 
581 	if (smp_num_siblings <= 1)
582 		goto out;
583 
584 	index_msb = get_count_order(smp_num_siblings);
585 	c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
586 
587 	smp_num_siblings = smp_num_siblings / c->x86_max_cores;
588 
589 	index_msb = get_count_order(smp_num_siblings);
590 
591 	core_bits = get_count_order(c->x86_max_cores);
592 
593 	c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
594 				       ((1 << core_bits) - 1);
595 
596 out:
597 	if (!printed && (c->x86_max_cores * smp_num_siblings) > 1) {
598 		pr_info("CPU: Physical Processor ID: %d\n",
599 			c->phys_proc_id);
600 		pr_info("CPU: Processor Core ID: %d\n",
601 			c->cpu_core_id);
602 		printed = 1;
603 	}
604 #endif
605 }
606 
607 static void get_cpu_vendor(struct cpuinfo_x86 *c)
608 {
609 	char *v = c->x86_vendor_id;
610 	int i;
611 
612 	for (i = 0; i < X86_VENDOR_NUM; i++) {
613 		if (!cpu_devs[i])
614 			break;
615 
616 		if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
617 		    (cpu_devs[i]->c_ident[1] &&
618 		     !strcmp(v, cpu_devs[i]->c_ident[1]))) {
619 
620 			this_cpu = cpu_devs[i];
621 			c->x86_vendor = this_cpu->c_x86_vendor;
622 			return;
623 		}
624 	}
625 
626 	pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
627 		    "CPU: Your system may be unstable.\n", v);
628 
629 	c->x86_vendor = X86_VENDOR_UNKNOWN;
630 	this_cpu = &default_cpu;
631 }
632 
633 void cpu_detect(struct cpuinfo_x86 *c)
634 {
635 	/* Get vendor name */
636 	cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
637 	      (unsigned int *)&c->x86_vendor_id[0],
638 	      (unsigned int *)&c->x86_vendor_id[8],
639 	      (unsigned int *)&c->x86_vendor_id[4]);
640 
641 	c->x86 = 4;
642 	/* Intel-defined flags: level 0x00000001 */
643 	if (c->cpuid_level >= 0x00000001) {
644 		u32 junk, tfms, cap0, misc;
645 
646 		cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
647 		c->x86		= x86_family(tfms);
648 		c->x86_model	= x86_model(tfms);
649 		c->x86_mask	= x86_stepping(tfms);
650 
651 		if (cap0 & (1<<19)) {
652 			c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
653 			c->x86_cache_alignment = c->x86_clflush_size;
654 		}
655 	}
656 }
657 
658 void get_cpu_cap(struct cpuinfo_x86 *c)
659 {
660 	u32 eax, ebx, ecx, edx;
661 
662 	/* Intel-defined flags: level 0x00000001 */
663 	if (c->cpuid_level >= 0x00000001) {
664 		cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
665 
666 		c->x86_capability[CPUID_1_ECX] = ecx;
667 		c->x86_capability[CPUID_1_EDX] = edx;
668 	}
669 
670 	/* Additional Intel-defined flags: level 0x00000007 */
671 	if (c->cpuid_level >= 0x00000007) {
672 		cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
673 
674 		c->x86_capability[CPUID_7_0_EBX] = ebx;
675 
676 		c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
677 		c->x86_capability[CPUID_7_ECX] = ecx;
678 	}
679 
680 	/* Extended state features: level 0x0000000d */
681 	if (c->cpuid_level >= 0x0000000d) {
682 		cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
683 
684 		c->x86_capability[CPUID_D_1_EAX] = eax;
685 	}
686 
687 	/* Additional Intel-defined flags: level 0x0000000F */
688 	if (c->cpuid_level >= 0x0000000F) {
689 
690 		/* QoS sub-leaf, EAX=0Fh, ECX=0 */
691 		cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
692 		c->x86_capability[CPUID_F_0_EDX] = edx;
693 
694 		if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
695 			/* will be overridden if occupancy monitoring exists */
696 			c->x86_cache_max_rmid = ebx;
697 
698 			/* QoS sub-leaf, EAX=0Fh, ECX=1 */
699 			cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
700 			c->x86_capability[CPUID_F_1_EDX] = edx;
701 
702 			if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
703 			      ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
704 			       (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
705 				c->x86_cache_max_rmid = ecx;
706 				c->x86_cache_occ_scale = ebx;
707 			}
708 		} else {
709 			c->x86_cache_max_rmid = -1;
710 			c->x86_cache_occ_scale = -1;
711 		}
712 	}
713 
714 	/* AMD-defined flags: level 0x80000001 */
715 	eax = cpuid_eax(0x80000000);
716 	c->extended_cpuid_level = eax;
717 
718 	if ((eax & 0xffff0000) == 0x80000000) {
719 		if (eax >= 0x80000001) {
720 			cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
721 
722 			c->x86_capability[CPUID_8000_0001_ECX] = ecx;
723 			c->x86_capability[CPUID_8000_0001_EDX] = edx;
724 		}
725 	}
726 
727 	if (c->extended_cpuid_level >= 0x80000007) {
728 		cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
729 
730 		c->x86_capability[CPUID_8000_0007_EBX] = ebx;
731 		c->x86_power = edx;
732 	}
733 
734 	if (c->extended_cpuid_level >= 0x80000008) {
735 		cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
736 
737 		c->x86_virt_bits = (eax >> 8) & 0xff;
738 		c->x86_phys_bits = eax & 0xff;
739 		c->x86_capability[CPUID_8000_0008_EBX] = ebx;
740 	}
741 #ifdef CONFIG_X86_32
742 	else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
743 		c->x86_phys_bits = 36;
744 #endif
745 
746 	if (c->extended_cpuid_level >= 0x8000000a)
747 		c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
748 
749 	init_scattered_cpuid_features(c);
750 }
751 
752 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
753 {
754 #ifdef CONFIG_X86_32
755 	int i;
756 
757 	/*
758 	 * First of all, decide if this is a 486 or higher
759 	 * It's a 486 if we can modify the AC flag
760 	 */
761 	if (flag_is_changeable_p(X86_EFLAGS_AC))
762 		c->x86 = 4;
763 	else
764 		c->x86 = 3;
765 
766 	for (i = 0; i < X86_VENDOR_NUM; i++)
767 		if (cpu_devs[i] && cpu_devs[i]->c_identify) {
768 			c->x86_vendor_id[0] = 0;
769 			cpu_devs[i]->c_identify(c);
770 			if (c->x86_vendor_id[0]) {
771 				get_cpu_vendor(c);
772 				break;
773 			}
774 		}
775 #endif
776 }
777 
778 /*
779  * Do minimum CPU detection early.
780  * Fields really needed: vendor, cpuid_level, family, model, mask,
781  * cache alignment.
782  * The others are not touched to avoid unwanted side effects.
783  *
784  * WARNING: this function is only called on the BP.  Don't add code here
785  * that is supposed to run on all CPUs.
786  */
787 static void __init early_identify_cpu(struct cpuinfo_x86 *c)
788 {
789 #ifdef CONFIG_X86_64
790 	c->x86_clflush_size = 64;
791 	c->x86_phys_bits = 36;
792 	c->x86_virt_bits = 48;
793 #else
794 	c->x86_clflush_size = 32;
795 	c->x86_phys_bits = 32;
796 	c->x86_virt_bits = 32;
797 #endif
798 	c->x86_cache_alignment = c->x86_clflush_size;
799 
800 	memset(&c->x86_capability, 0, sizeof c->x86_capability);
801 	c->extended_cpuid_level = 0;
802 
803 	if (!have_cpuid_p())
804 		identify_cpu_without_cpuid(c);
805 
806 	/* cyrix could have cpuid enabled via c_identify()*/
807 	if (!have_cpuid_p())
808 		return;
809 
810 	cpu_detect(c);
811 	get_cpu_vendor(c);
812 	get_cpu_cap(c);
813 
814 	if (this_cpu->c_early_init)
815 		this_cpu->c_early_init(c);
816 
817 	c->cpu_index = 0;
818 	filter_cpuid_features(c, false);
819 
820 	if (this_cpu->c_bsp_init)
821 		this_cpu->c_bsp_init(c);
822 
823 	setup_force_cpu_cap(X86_FEATURE_ALWAYS);
824 	fpu__init_system(c);
825 }
826 
827 void __init early_cpu_init(void)
828 {
829 	const struct cpu_dev *const *cdev;
830 	int count = 0;
831 
832 #ifdef CONFIG_PROCESSOR_SELECT
833 	pr_info("KERNEL supported cpus:\n");
834 #endif
835 
836 	for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
837 		const struct cpu_dev *cpudev = *cdev;
838 
839 		if (count >= X86_VENDOR_NUM)
840 			break;
841 		cpu_devs[count] = cpudev;
842 		count++;
843 
844 #ifdef CONFIG_PROCESSOR_SELECT
845 		{
846 			unsigned int j;
847 
848 			for (j = 0; j < 2; j++) {
849 				if (!cpudev->c_ident[j])
850 					continue;
851 				pr_info("  %s %s\n", cpudev->c_vendor,
852 					cpudev->c_ident[j]);
853 			}
854 		}
855 #endif
856 	}
857 	early_identify_cpu(&boot_cpu_data);
858 }
859 
860 /*
861  * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
862  * unfortunately, that's not true in practice because of early VIA
863  * chips and (more importantly) broken virtualizers that are not easy
864  * to detect. In the latter case it doesn't even *fail* reliably, so
865  * probing for it doesn't even work. Disable it completely on 32-bit
866  * unless we can find a reliable way to detect all the broken cases.
867  * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
868  */
869 static void detect_nopl(struct cpuinfo_x86 *c)
870 {
871 #ifdef CONFIG_X86_32
872 	clear_cpu_cap(c, X86_FEATURE_NOPL);
873 #else
874 	set_cpu_cap(c, X86_FEATURE_NOPL);
875 #endif
876 }
877 
878 static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
879 {
880 #ifdef CONFIG_X86_64
881 	/*
882 	 * Empirically, writing zero to a segment selector on AMD does
883 	 * not clear the base, whereas writing zero to a segment
884 	 * selector on Intel does clear the base.  Intel's behavior
885 	 * allows slightly faster context switches in the common case
886 	 * where GS is unused by the prev and next threads.
887 	 *
888 	 * Since neither vendor documents this anywhere that I can see,
889 	 * detect it directly instead of hardcoding the choice by
890 	 * vendor.
891 	 *
892 	 * I've designated AMD's behavior as the "bug" because it's
893 	 * counterintuitive and less friendly.
894 	 */
895 
896 	unsigned long old_base, tmp;
897 	rdmsrl(MSR_FS_BASE, old_base);
898 	wrmsrl(MSR_FS_BASE, 1);
899 	loadsegment(fs, 0);
900 	rdmsrl(MSR_FS_BASE, tmp);
901 	if (tmp != 0)
902 		set_cpu_bug(c, X86_BUG_NULL_SEG);
903 	wrmsrl(MSR_FS_BASE, old_base);
904 #endif
905 }
906 
907 static void generic_identify(struct cpuinfo_x86 *c)
908 {
909 	c->extended_cpuid_level = 0;
910 
911 	if (!have_cpuid_p())
912 		identify_cpu_without_cpuid(c);
913 
914 	/* cyrix could have cpuid enabled via c_identify()*/
915 	if (!have_cpuid_p())
916 		return;
917 
918 	cpu_detect(c);
919 
920 	get_cpu_vendor(c);
921 
922 	get_cpu_cap(c);
923 
924 	if (c->cpuid_level >= 0x00000001) {
925 		c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
926 #ifdef CONFIG_X86_32
927 # ifdef CONFIG_SMP
928 		c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
929 # else
930 		c->apicid = c->initial_apicid;
931 # endif
932 #endif
933 		c->phys_proc_id = c->initial_apicid;
934 	}
935 
936 	get_model_name(c); /* Default name */
937 
938 	detect_nopl(c);
939 
940 	detect_null_seg_behavior(c);
941 
942 	/*
943 	 * ESPFIX is a strange bug.  All real CPUs have it.  Paravirt
944 	 * systems that run Linux at CPL > 0 may or may not have the
945 	 * issue, but, even if they have the issue, there's absolutely
946 	 * nothing we can do about it because we can't use the real IRET
947 	 * instruction.
948 	 *
949 	 * NB: For the time being, only 32-bit kernels support
950 	 * X86_BUG_ESPFIX as such.  64-bit kernels directly choose
951 	 * whether to apply espfix using paravirt hooks.  If any
952 	 * non-paravirt system ever shows up that does *not* have the
953 	 * ESPFIX issue, we can change this.
954 	 */
955 #ifdef CONFIG_X86_32
956 # ifdef CONFIG_PARAVIRT
957 	do {
958 		extern void native_iret(void);
959 		if (pv_cpu_ops.iret == native_iret)
960 			set_cpu_bug(c, X86_BUG_ESPFIX);
961 	} while (0);
962 # else
963 	set_cpu_bug(c, X86_BUG_ESPFIX);
964 # endif
965 #endif
966 }
967 
968 static void x86_init_cache_qos(struct cpuinfo_x86 *c)
969 {
970 	/*
971 	 * The heavy lifting of max_rmid and cache_occ_scale are handled
972 	 * in get_cpu_cap().  Here we just set the max_rmid for the boot_cpu
973 	 * in case CQM bits really aren't there in this CPU.
974 	 */
975 	if (c != &boot_cpu_data) {
976 		boot_cpu_data.x86_cache_max_rmid =
977 			min(boot_cpu_data.x86_cache_max_rmid,
978 			    c->x86_cache_max_rmid);
979 	}
980 }
981 
982 /*
983  * This does the hard work of actually picking apart the CPU stuff...
984  */
985 static void identify_cpu(struct cpuinfo_x86 *c)
986 {
987 	int i;
988 
989 	c->loops_per_jiffy = loops_per_jiffy;
990 	c->x86_cache_size = -1;
991 	c->x86_vendor = X86_VENDOR_UNKNOWN;
992 	c->x86_model = c->x86_mask = 0;	/* So far unknown... */
993 	c->x86_vendor_id[0] = '\0'; /* Unset */
994 	c->x86_model_id[0] = '\0';  /* Unset */
995 	c->x86_max_cores = 1;
996 	c->x86_coreid_bits = 0;
997 #ifdef CONFIG_X86_64
998 	c->x86_clflush_size = 64;
999 	c->x86_phys_bits = 36;
1000 	c->x86_virt_bits = 48;
1001 #else
1002 	c->cpuid_level = -1;	/* CPUID not detected */
1003 	c->x86_clflush_size = 32;
1004 	c->x86_phys_bits = 32;
1005 	c->x86_virt_bits = 32;
1006 #endif
1007 	c->x86_cache_alignment = c->x86_clflush_size;
1008 	memset(&c->x86_capability, 0, sizeof c->x86_capability);
1009 
1010 	generic_identify(c);
1011 
1012 	if (this_cpu->c_identify)
1013 		this_cpu->c_identify(c);
1014 
1015 	/* Clear/Set all flags overridden by options, after probe */
1016 	for (i = 0; i < NCAPINTS; i++) {
1017 		c->x86_capability[i] &= ~cpu_caps_cleared[i];
1018 		c->x86_capability[i] |= cpu_caps_set[i];
1019 	}
1020 
1021 #ifdef CONFIG_X86_64
1022 	c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1023 #endif
1024 
1025 	/*
1026 	 * Vendor-specific initialization.  In this section we
1027 	 * canonicalize the feature flags, meaning if there are
1028 	 * features a certain CPU supports which CPUID doesn't
1029 	 * tell us, CPUID claiming incorrect flags, or other bugs,
1030 	 * we handle them here.
1031 	 *
1032 	 * At the end of this section, c->x86_capability better
1033 	 * indicate the features this CPU genuinely supports!
1034 	 */
1035 	if (this_cpu->c_init)
1036 		this_cpu->c_init(c);
1037 
1038 	/* Disable the PN if appropriate */
1039 	squash_the_stupid_serial_number(c);
1040 
1041 	/* Set up SMEP/SMAP */
1042 	setup_smep(c);
1043 	setup_smap(c);
1044 
1045 	/*
1046 	 * The vendor-specific functions might have changed features.
1047 	 * Now we do "generic changes."
1048 	 */
1049 
1050 	/* Filter out anything that depends on CPUID levels we don't have */
1051 	filter_cpuid_features(c, true);
1052 
1053 	/* If the model name is still unset, do table lookup. */
1054 	if (!c->x86_model_id[0]) {
1055 		const char *p;
1056 		p = table_lookup_model(c);
1057 		if (p)
1058 			strcpy(c->x86_model_id, p);
1059 		else
1060 			/* Last resort... */
1061 			sprintf(c->x86_model_id, "%02x/%02x",
1062 				c->x86, c->x86_model);
1063 	}
1064 
1065 #ifdef CONFIG_X86_64
1066 	detect_ht(c);
1067 #endif
1068 
1069 	init_hypervisor(c);
1070 	x86_init_rdrand(c);
1071 	x86_init_cache_qos(c);
1072 	setup_pku(c);
1073 
1074 	/*
1075 	 * Clear/Set all flags overridden by options, need do it
1076 	 * before following smp all cpus cap AND.
1077 	 */
1078 	for (i = 0; i < NCAPINTS; i++) {
1079 		c->x86_capability[i] &= ~cpu_caps_cleared[i];
1080 		c->x86_capability[i] |= cpu_caps_set[i];
1081 	}
1082 
1083 	/*
1084 	 * On SMP, boot_cpu_data holds the common feature set between
1085 	 * all CPUs; so make sure that we indicate which features are
1086 	 * common between the CPUs.  The first time this routine gets
1087 	 * executed, c == &boot_cpu_data.
1088 	 */
1089 	if (c != &boot_cpu_data) {
1090 		/* AND the already accumulated flags with these */
1091 		for (i = 0; i < NCAPINTS; i++)
1092 			boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1093 
1094 		/* OR, i.e. replicate the bug flags */
1095 		for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1096 			c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1097 	}
1098 
1099 	/* Init Machine Check Exception if available. */
1100 	mcheck_cpu_init(c);
1101 
1102 	select_idle_routine(c);
1103 
1104 #ifdef CONFIG_NUMA
1105 	numa_add_cpu(smp_processor_id());
1106 #endif
1107 	/* The boot/hotplug time assigment got cleared, restore it */
1108 	c->logical_proc_id = topology_phys_to_logical_pkg(c->phys_proc_id);
1109 }
1110 
1111 /*
1112  * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1113  * on 32-bit kernels:
1114  */
1115 #ifdef CONFIG_X86_32
1116 void enable_sep_cpu(void)
1117 {
1118 	struct tss_struct *tss;
1119 	int cpu;
1120 
1121 	if (!boot_cpu_has(X86_FEATURE_SEP))
1122 		return;
1123 
1124 	cpu = get_cpu();
1125 	tss = &per_cpu(cpu_tss, cpu);
1126 
1127 	/*
1128 	 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1129 	 * see the big comment in struct x86_hw_tss's definition.
1130 	 */
1131 
1132 	tss->x86_tss.ss1 = __KERNEL_CS;
1133 	wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1134 
1135 	wrmsr(MSR_IA32_SYSENTER_ESP,
1136 	      (unsigned long)tss + offsetofend(struct tss_struct, SYSENTER_stack),
1137 	      0);
1138 
1139 	wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1140 
1141 	put_cpu();
1142 }
1143 #endif
1144 
1145 void __init identify_boot_cpu(void)
1146 {
1147 	identify_cpu(&boot_cpu_data);
1148 	init_amd_e400_c1e_mask();
1149 #ifdef CONFIG_X86_32
1150 	sysenter_setup();
1151 	enable_sep_cpu();
1152 #endif
1153 	cpu_detect_tlb(&boot_cpu_data);
1154 }
1155 
1156 void identify_secondary_cpu(struct cpuinfo_x86 *c)
1157 {
1158 	BUG_ON(c == &boot_cpu_data);
1159 	identify_cpu(c);
1160 #ifdef CONFIG_X86_32
1161 	enable_sep_cpu();
1162 #endif
1163 	mtrr_ap_init();
1164 }
1165 
1166 struct msr_range {
1167 	unsigned	min;
1168 	unsigned	max;
1169 };
1170 
1171 static const struct msr_range msr_range_array[] = {
1172 	{ 0x00000000, 0x00000418},
1173 	{ 0xc0000000, 0xc000040b},
1174 	{ 0xc0010000, 0xc0010142},
1175 	{ 0xc0011000, 0xc001103b},
1176 };
1177 
1178 static void __print_cpu_msr(void)
1179 {
1180 	unsigned index_min, index_max;
1181 	unsigned index;
1182 	u64 val;
1183 	int i;
1184 
1185 	for (i = 0; i < ARRAY_SIZE(msr_range_array); i++) {
1186 		index_min = msr_range_array[i].min;
1187 		index_max = msr_range_array[i].max;
1188 
1189 		for (index = index_min; index < index_max; index++) {
1190 			if (rdmsrl_safe(index, &val))
1191 				continue;
1192 			pr_info(" MSR%08x: %016llx\n", index, val);
1193 		}
1194 	}
1195 }
1196 
1197 static int show_msr;
1198 
1199 static __init int setup_show_msr(char *arg)
1200 {
1201 	int num;
1202 
1203 	get_option(&arg, &num);
1204 
1205 	if (num > 0)
1206 		show_msr = num;
1207 	return 1;
1208 }
1209 __setup("show_msr=", setup_show_msr);
1210 
1211 static __init int setup_noclflush(char *arg)
1212 {
1213 	setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1214 	setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1215 	return 1;
1216 }
1217 __setup("noclflush", setup_noclflush);
1218 
1219 void print_cpu_info(struct cpuinfo_x86 *c)
1220 {
1221 	const char *vendor = NULL;
1222 
1223 	if (c->x86_vendor < X86_VENDOR_NUM) {
1224 		vendor = this_cpu->c_vendor;
1225 	} else {
1226 		if (c->cpuid_level >= 0)
1227 			vendor = c->x86_vendor_id;
1228 	}
1229 
1230 	if (vendor && !strstr(c->x86_model_id, vendor))
1231 		pr_cont("%s ", vendor);
1232 
1233 	if (c->x86_model_id[0])
1234 		pr_cont("%s", c->x86_model_id);
1235 	else
1236 		pr_cont("%d86", c->x86);
1237 
1238 	pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1239 
1240 	if (c->x86_mask || c->cpuid_level >= 0)
1241 		pr_cont(", stepping: 0x%x)\n", c->x86_mask);
1242 	else
1243 		pr_cont(")\n");
1244 
1245 	print_cpu_msr(c);
1246 }
1247 
1248 void print_cpu_msr(struct cpuinfo_x86 *c)
1249 {
1250 	if (c->cpu_index < show_msr)
1251 		__print_cpu_msr();
1252 }
1253 
1254 static __init int setup_disablecpuid(char *arg)
1255 {
1256 	int bit;
1257 
1258 	if (get_option(&arg, &bit) && bit < NCAPINTS*32)
1259 		setup_clear_cpu_cap(bit);
1260 	else
1261 		return 0;
1262 
1263 	return 1;
1264 }
1265 __setup("clearcpuid=", setup_disablecpuid);
1266 
1267 #ifdef CONFIG_X86_64
1268 struct desc_ptr idt_descr = { NR_VECTORS * 16 - 1, (unsigned long) idt_table };
1269 struct desc_ptr debug_idt_descr = { NR_VECTORS * 16 - 1,
1270 				    (unsigned long) debug_idt_table };
1271 
1272 DEFINE_PER_CPU_FIRST(union irq_stack_union,
1273 		     irq_stack_union) __aligned(PAGE_SIZE) __visible;
1274 
1275 /*
1276  * The following percpu variables are hot.  Align current_task to
1277  * cacheline size such that they fall in the same cacheline.
1278  */
1279 DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1280 	&init_task;
1281 EXPORT_PER_CPU_SYMBOL(current_task);
1282 
1283 DEFINE_PER_CPU(char *, irq_stack_ptr) =
1284 	init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE - 64;
1285 
1286 DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1287 
1288 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1289 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1290 
1291 /*
1292  * Special IST stacks which the CPU switches to when it calls
1293  * an IST-marked descriptor entry. Up to 7 stacks (hardware
1294  * limit), all of them are 4K, except the debug stack which
1295  * is 8K.
1296  */
1297 static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
1298 	  [0 ... N_EXCEPTION_STACKS - 1]	= EXCEPTION_STKSZ,
1299 	  [DEBUG_STACK - 1]			= DEBUG_STKSZ
1300 };
1301 
1302 static DEFINE_PER_CPU_PAGE_ALIGNED(char, exception_stacks
1303 	[(N_EXCEPTION_STACKS - 1) * EXCEPTION_STKSZ + DEBUG_STKSZ]);
1304 
1305 /* May not be marked __init: used by software suspend */
1306 void syscall_init(void)
1307 {
1308 	/*
1309 	 * LSTAR and STAR live in a bit strange symbiosis.
1310 	 * They both write to the same internal register. STAR allows to
1311 	 * set CS/DS but only a 32bit target. LSTAR sets the 64bit rip.
1312 	 */
1313 	wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1314 	wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1315 
1316 #ifdef CONFIG_IA32_EMULATION
1317 	wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1318 	/*
1319 	 * This only works on Intel CPUs.
1320 	 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1321 	 * This does not cause SYSENTER to jump to the wrong location, because
1322 	 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1323 	 */
1324 	wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1325 	wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1326 	wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1327 #else
1328 	wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1329 	wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1330 	wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1331 	wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1332 #endif
1333 
1334 	/* Flags to clear on syscall */
1335 	wrmsrl(MSR_SYSCALL_MASK,
1336 	       X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1337 	       X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1338 }
1339 
1340 /*
1341  * Copies of the original ist values from the tss are only accessed during
1342  * debugging, no special alignment required.
1343  */
1344 DEFINE_PER_CPU(struct orig_ist, orig_ist);
1345 
1346 static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1347 DEFINE_PER_CPU(int, debug_stack_usage);
1348 
1349 int is_debug_stack(unsigned long addr)
1350 {
1351 	return __this_cpu_read(debug_stack_usage) ||
1352 		(addr <= __this_cpu_read(debug_stack_addr) &&
1353 		 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1354 }
1355 NOKPROBE_SYMBOL(is_debug_stack);
1356 
1357 DEFINE_PER_CPU(u32, debug_idt_ctr);
1358 
1359 void debug_stack_set_zero(void)
1360 {
1361 	this_cpu_inc(debug_idt_ctr);
1362 	load_current_idt();
1363 }
1364 NOKPROBE_SYMBOL(debug_stack_set_zero);
1365 
1366 void debug_stack_reset(void)
1367 {
1368 	if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1369 		return;
1370 	if (this_cpu_dec_return(debug_idt_ctr) == 0)
1371 		load_current_idt();
1372 }
1373 NOKPROBE_SYMBOL(debug_stack_reset);
1374 
1375 #else	/* CONFIG_X86_64 */
1376 
1377 DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1378 EXPORT_PER_CPU_SYMBOL(current_task);
1379 DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1380 EXPORT_PER_CPU_SYMBOL(__preempt_count);
1381 
1382 /*
1383  * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1384  * the top of the kernel stack.  Use an extra percpu variable to track the
1385  * top of the kernel stack directly.
1386  */
1387 DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1388 	(unsigned long)&init_thread_union + THREAD_SIZE;
1389 EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1390 
1391 #ifdef CONFIG_CC_STACKPROTECTOR
1392 DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1393 #endif
1394 
1395 #endif	/* CONFIG_X86_64 */
1396 
1397 /*
1398  * Clear all 6 debug registers:
1399  */
1400 static void clear_all_debug_regs(void)
1401 {
1402 	int i;
1403 
1404 	for (i = 0; i < 8; i++) {
1405 		/* Ignore db4, db5 */
1406 		if ((i == 4) || (i == 5))
1407 			continue;
1408 
1409 		set_debugreg(0, i);
1410 	}
1411 }
1412 
1413 #ifdef CONFIG_KGDB
1414 /*
1415  * Restore debug regs if using kgdbwait and you have a kernel debugger
1416  * connection established.
1417  */
1418 static void dbg_restore_debug_regs(void)
1419 {
1420 	if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1421 		arch_kgdb_ops.correct_hw_break();
1422 }
1423 #else /* ! CONFIG_KGDB */
1424 #define dbg_restore_debug_regs()
1425 #endif /* ! CONFIG_KGDB */
1426 
1427 static void wait_for_master_cpu(int cpu)
1428 {
1429 #ifdef CONFIG_SMP
1430 	/*
1431 	 * wait for ACK from master CPU before continuing
1432 	 * with AP initialization
1433 	 */
1434 	WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1435 	while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1436 		cpu_relax();
1437 #endif
1438 }
1439 
1440 /*
1441  * cpu_init() initializes state that is per-CPU. Some data is already
1442  * initialized (naturally) in the bootstrap process, such as the GDT
1443  * and IDT. We reload them nevertheless, this function acts as a
1444  * 'CPU state barrier', nothing should get across.
1445  * A lot of state is already set up in PDA init for 64 bit
1446  */
1447 #ifdef CONFIG_X86_64
1448 
1449 void cpu_init(void)
1450 {
1451 	struct orig_ist *oist;
1452 	struct task_struct *me;
1453 	struct tss_struct *t;
1454 	unsigned long v;
1455 	int cpu = raw_smp_processor_id();
1456 	int i;
1457 
1458 	wait_for_master_cpu(cpu);
1459 
1460 	/*
1461 	 * Initialize the CR4 shadow before doing anything that could
1462 	 * try to read it.
1463 	 */
1464 	cr4_init_shadow();
1465 
1466 	/*
1467 	 * Load microcode on this cpu if a valid microcode is available.
1468 	 * This is early microcode loading procedure.
1469 	 */
1470 	load_ucode_ap();
1471 
1472 	t = &per_cpu(cpu_tss, cpu);
1473 	oist = &per_cpu(orig_ist, cpu);
1474 
1475 #ifdef CONFIG_NUMA
1476 	if (this_cpu_read(numa_node) == 0 &&
1477 	    early_cpu_to_node(cpu) != NUMA_NO_NODE)
1478 		set_numa_node(early_cpu_to_node(cpu));
1479 #endif
1480 
1481 	me = current;
1482 
1483 	pr_debug("Initializing CPU#%d\n", cpu);
1484 
1485 	cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1486 
1487 	/*
1488 	 * Initialize the per-CPU GDT with the boot GDT,
1489 	 * and set up the GDT descriptor:
1490 	 */
1491 
1492 	switch_to_new_gdt(cpu);
1493 	loadsegment(fs, 0);
1494 
1495 	load_current_idt();
1496 
1497 	memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1498 	syscall_init();
1499 
1500 	wrmsrl(MSR_FS_BASE, 0);
1501 	wrmsrl(MSR_KERNEL_GS_BASE, 0);
1502 	barrier();
1503 
1504 	x86_configure_nx();
1505 	x2apic_setup();
1506 
1507 	/*
1508 	 * set up and load the per-CPU TSS
1509 	 */
1510 	if (!oist->ist[0]) {
1511 		char *estacks = per_cpu(exception_stacks, cpu);
1512 
1513 		for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1514 			estacks += exception_stack_sizes[v];
1515 			oist->ist[v] = t->x86_tss.ist[v] =
1516 					(unsigned long)estacks;
1517 			if (v == DEBUG_STACK-1)
1518 				per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1519 		}
1520 	}
1521 
1522 	t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1523 
1524 	/*
1525 	 * <= is required because the CPU will access up to
1526 	 * 8 bits beyond the end of the IO permission bitmap.
1527 	 */
1528 	for (i = 0; i <= IO_BITMAP_LONGS; i++)
1529 		t->io_bitmap[i] = ~0UL;
1530 
1531 	atomic_inc(&init_mm.mm_count);
1532 	me->active_mm = &init_mm;
1533 	BUG_ON(me->mm);
1534 	enter_lazy_tlb(&init_mm, me);
1535 
1536 	load_sp0(t, &current->thread);
1537 	set_tss_desc(cpu, t);
1538 	load_TR_desc();
1539 	load_mm_ldt(&init_mm);
1540 
1541 	clear_all_debug_regs();
1542 	dbg_restore_debug_regs();
1543 
1544 	fpu__init_cpu();
1545 
1546 	if (is_uv_system())
1547 		uv_cpu_init();
1548 }
1549 
1550 #else
1551 
1552 void cpu_init(void)
1553 {
1554 	int cpu = smp_processor_id();
1555 	struct task_struct *curr = current;
1556 	struct tss_struct *t = &per_cpu(cpu_tss, cpu);
1557 	struct thread_struct *thread = &curr->thread;
1558 
1559 	wait_for_master_cpu(cpu);
1560 
1561 	/*
1562 	 * Initialize the CR4 shadow before doing anything that could
1563 	 * try to read it.
1564 	 */
1565 	cr4_init_shadow();
1566 
1567 	show_ucode_info_early();
1568 
1569 	pr_info("Initializing CPU#%d\n", cpu);
1570 
1571 	if (cpu_feature_enabled(X86_FEATURE_VME) ||
1572 	    boot_cpu_has(X86_FEATURE_TSC) ||
1573 	    boot_cpu_has(X86_FEATURE_DE))
1574 		cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1575 
1576 	load_current_idt();
1577 	switch_to_new_gdt(cpu);
1578 
1579 	/*
1580 	 * Set up and load the per-CPU TSS and LDT
1581 	 */
1582 	atomic_inc(&init_mm.mm_count);
1583 	curr->active_mm = &init_mm;
1584 	BUG_ON(curr->mm);
1585 	enter_lazy_tlb(&init_mm, curr);
1586 
1587 	load_sp0(t, thread);
1588 	set_tss_desc(cpu, t);
1589 	load_TR_desc();
1590 	load_mm_ldt(&init_mm);
1591 
1592 	t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1593 
1594 #ifdef CONFIG_DOUBLEFAULT
1595 	/* Set up doublefault TSS pointer in the GDT */
1596 	__set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1597 #endif
1598 
1599 	clear_all_debug_regs();
1600 	dbg_restore_debug_regs();
1601 
1602 	fpu__init_cpu();
1603 }
1604 #endif
1605 
1606 static void bsp_resume(void)
1607 {
1608 	if (this_cpu->c_bsp_resume)
1609 		this_cpu->c_bsp_resume(&boot_cpu_data);
1610 }
1611 
1612 static struct syscore_ops cpu_syscore_ops = {
1613 	.resume		= bsp_resume,
1614 };
1615 
1616 static int __init init_cpu_syscore(void)
1617 {
1618 	register_syscore_ops(&cpu_syscore_ops);
1619 	return 0;
1620 }
1621 core_initcall(init_cpu_syscore);
1622