Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0-only
   2/* cpu_feature_enabled() cannot be used this early */
   3#define USE_EARLY_PGTABLE_L5
   4
   5#include <linux/memblock.h>
   6#include <linux/linkage.h>
   7#include <linux/bitops.h>
   8#include <linux/kernel.h>
   9#include <linux/export.h>
  10#include <linux/percpu.h>
  11#include <linux/string.h>
  12#include <linux/ctype.h>
  13#include <linux/delay.h>
  14#include <linux/sched/mm.h>
  15#include <linux/sched/clock.h>
  16#include <linux/sched/task.h>
  17#include <linux/init.h>
  18#include <linux/kprobes.h>
  19#include <linux/kgdb.h>
  20#include <linux/smp.h>
  21#include <linux/io.h>
  22#include <linux/syscore_ops.h>
  23
  24#include <asm/stackprotector.h>
  25#include <asm/perf_event.h>
  26#include <asm/mmu_context.h>
  27#include <asm/archrandom.h>
  28#include <asm/hypervisor.h>
  29#include <asm/processor.h>
  30#include <asm/tlbflush.h>
  31#include <asm/debugreg.h>
  32#include <asm/sections.h>
  33#include <asm/vsyscall.h>
  34#include <linux/topology.h>
  35#include <linux/cpumask.h>
  36#include <asm/pgtable.h>
  37#include <linux/atomic.h>
  38#include <asm/proto.h>
  39#include <asm/setup.h>
  40#include <asm/apic.h>
  41#include <asm/desc.h>
  42#include <asm/fpu/internal.h>
  43#include <asm/mtrr.h>
  44#include <asm/hwcap2.h>
  45#include <linux/numa.h>
  46#include <asm/asm.h>
  47#include <asm/bugs.h>
  48#include <asm/cpu.h>
  49#include <asm/mce.h>
  50#include <asm/msr.h>
  51#include <asm/pat.h>
  52#include <asm/microcode.h>
  53#include <asm/microcode_intel.h>
  54#include <asm/intel-family.h>
  55#include <asm/cpu_device_id.h>
  56
  57#ifdef CONFIG_X86_LOCAL_APIC
  58#include <asm/uv/uv.h>
  59#endif
  60
  61#include "cpu.h"
  62
  63u32 elf_hwcap2 __read_mostly;
  64
  65/* all of these masks are initialized in setup_cpu_local_masks() */
  66cpumask_var_t cpu_initialized_mask;
  67cpumask_var_t cpu_callout_mask;
  68cpumask_var_t cpu_callin_mask;
  69
  70/* representing cpus for which sibling maps can be computed */
  71cpumask_var_t cpu_sibling_setup_mask;
  72
  73/* Number of siblings per CPU package */
  74int smp_num_siblings = 1;
  75EXPORT_SYMBOL(smp_num_siblings);
  76
  77/* Last level cache ID of each logical CPU */
  78DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
  79
  80/* correctly size the local cpu masks */
  81void __init setup_cpu_local_masks(void)
  82{
  83	alloc_bootmem_cpumask_var(&cpu_initialized_mask);
  84	alloc_bootmem_cpumask_var(&cpu_callin_mask);
  85	alloc_bootmem_cpumask_var(&cpu_callout_mask);
  86	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
  87}
  88
  89static void default_init(struct cpuinfo_x86 *c)
  90{
  91#ifdef CONFIG_X86_64
  92	cpu_detect_cache_sizes(c);
  93#else
  94	/* Not much we can do here... */
  95	/* Check if at least it has cpuid */
  96	if (c->cpuid_level == -1) {
  97		/* No cpuid. It must be an ancient CPU */
  98		if (c->x86 == 4)
  99			strcpy(c->x86_model_id, "486");
 100		else if (c->x86 == 3)
 101			strcpy(c->x86_model_id, "386");
 102	}
 103#endif
 104}
 105
 106static const struct cpu_dev default_cpu = {
 107	.c_init		= default_init,
 108	.c_vendor	= "Unknown",
 109	.c_x86_vendor	= X86_VENDOR_UNKNOWN,
 110};
 111
 112static const struct cpu_dev *this_cpu = &default_cpu;
 113
 114DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = {
 115#ifdef CONFIG_X86_64
 116	/*
 117	 * We need valid kernel segments for data and code in long mode too
 118	 * IRET will check the segment types  kkeil 2000/10/28
 119	 * Also sysret mandates a special GDT layout
 120	 *
 121	 * TLS descriptors are currently at a different place compared to i386.
 122	 * Hopefully nobody expects them at a fixed place (Wine?)
 123	 */
 124	[GDT_ENTRY_KERNEL32_CS]		= GDT_ENTRY_INIT(0xc09b, 0, 0xfffff),
 125	[GDT_ENTRY_KERNEL_CS]		= GDT_ENTRY_INIT(0xa09b, 0, 0xfffff),
 126	[GDT_ENTRY_KERNEL_DS]		= GDT_ENTRY_INIT(0xc093, 0, 0xfffff),
 127	[GDT_ENTRY_DEFAULT_USER32_CS]	= GDT_ENTRY_INIT(0xc0fb, 0, 0xfffff),
 128	[GDT_ENTRY_DEFAULT_USER_DS]	= GDT_ENTRY_INIT(0xc0f3, 0, 0xfffff),
 129	[GDT_ENTRY_DEFAULT_USER_CS]	= GDT_ENTRY_INIT(0xa0fb, 0, 0xfffff),
 130#else
 131	[GDT_ENTRY_KERNEL_CS]		= GDT_ENTRY_INIT(0xc09a, 0, 0xfffff),
 132	[GDT_ENTRY_KERNEL_DS]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
 133	[GDT_ENTRY_DEFAULT_USER_CS]	= GDT_ENTRY_INIT(0xc0fa, 0, 0xfffff),
 134	[GDT_ENTRY_DEFAULT_USER_DS]	= GDT_ENTRY_INIT(0xc0f2, 0, 0xfffff),
 135	/*
 136	 * Segments used for calling PnP BIOS have byte granularity.
 137	 * They code segments and data segments have fixed 64k limits,
 138	 * the transfer segment sizes are set at run time.
 139	 */
 140	/* 32-bit code */
 141	[GDT_ENTRY_PNPBIOS_CS32]	= GDT_ENTRY_INIT(0x409a, 0, 0xffff),
 142	/* 16-bit code */
 143	[GDT_ENTRY_PNPBIOS_CS16]	= GDT_ENTRY_INIT(0x009a, 0, 0xffff),
 144	/* 16-bit data */
 145	[GDT_ENTRY_PNPBIOS_DS]		= GDT_ENTRY_INIT(0x0092, 0, 0xffff),
 146	/* 16-bit data */
 147	[GDT_ENTRY_PNPBIOS_TS1]		= GDT_ENTRY_INIT(0x0092, 0, 0),
 148	/* 16-bit data */
 149	[GDT_ENTRY_PNPBIOS_TS2]		= GDT_ENTRY_INIT(0x0092, 0, 0),
 150	/*
 151	 * The APM segments have byte granularity and their bases
 152	 * are set at run time.  All have 64k limits.
 153	 */
 154	/* 32-bit code */
 155	[GDT_ENTRY_APMBIOS_BASE]	= GDT_ENTRY_INIT(0x409a, 0, 0xffff),
 156	/* 16-bit code */
 157	[GDT_ENTRY_APMBIOS_BASE+1]	= GDT_ENTRY_INIT(0x009a, 0, 0xffff),
 158	/* data */
 159	[GDT_ENTRY_APMBIOS_BASE+2]	= GDT_ENTRY_INIT(0x4092, 0, 0xffff),
 160
 161	[GDT_ENTRY_ESPFIX_SS]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
 162	[GDT_ENTRY_PERCPU]		= GDT_ENTRY_INIT(0xc092, 0, 0xfffff),
 163	GDT_STACK_CANARY_INIT
 164#endif
 165} };
 166EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
 167
 168static int __init x86_mpx_setup(char *s)
 169{
 170	/* require an exact match without trailing characters */
 171	if (strlen(s))
 172		return 0;
 173
 174	/* do not emit a message if the feature is not present */
 175	if (!boot_cpu_has(X86_FEATURE_MPX))
 176		return 1;
 177
 178	setup_clear_cpu_cap(X86_FEATURE_MPX);
 179	pr_info("nompx: Intel Memory Protection Extensions (MPX) disabled\n");
 180	return 1;
 181}
 182__setup("nompx", x86_mpx_setup);
 183
 184#ifdef CONFIG_X86_64
 185static int __init x86_nopcid_setup(char *s)
 186{
 187	/* nopcid doesn't accept parameters */
 188	if (s)
 189		return -EINVAL;
 190
 191	/* do not emit a message if the feature is not present */
 192	if (!boot_cpu_has(X86_FEATURE_PCID))
 193		return 0;
 194
 195	setup_clear_cpu_cap(X86_FEATURE_PCID);
 196	pr_info("nopcid: PCID feature disabled\n");
 197	return 0;
 198}
 199early_param("nopcid", x86_nopcid_setup);
 200#endif
 201
 202static int __init x86_noinvpcid_setup(char *s)
 203{
 204	/* noinvpcid doesn't accept parameters */
 205	if (s)
 206		return -EINVAL;
 207
 208	/* do not emit a message if the feature is not present */
 209	if (!boot_cpu_has(X86_FEATURE_INVPCID))
 210		return 0;
 211
 212	setup_clear_cpu_cap(X86_FEATURE_INVPCID);
 213	pr_info("noinvpcid: INVPCID feature disabled\n");
 214	return 0;
 215}
 216early_param("noinvpcid", x86_noinvpcid_setup);
 217
 218#ifdef CONFIG_X86_32
 219static int cachesize_override = -1;
 220static int disable_x86_serial_nr = 1;
 221
 222static int __init cachesize_setup(char *str)
 223{
 224	get_option(&str, &cachesize_override);
 225	return 1;
 226}
 227__setup("cachesize=", cachesize_setup);
 228
 229static int __init x86_sep_setup(char *s)
 230{
 231	setup_clear_cpu_cap(X86_FEATURE_SEP);
 232	return 1;
 233}
 234__setup("nosep", x86_sep_setup);
 235
 236/* Standard macro to see if a specific flag is changeable */
 237static inline int flag_is_changeable_p(u32 flag)
 238{
 239	u32 f1, f2;
 240
 241	/*
 242	 * Cyrix and IDT cpus allow disabling of CPUID
 243	 * so the code below may return different results
 244	 * when it is executed before and after enabling
 245	 * the CPUID. Add "volatile" to not allow gcc to
 246	 * optimize the subsequent calls to this function.
 247	 */
 248	asm volatile ("pushfl		\n\t"
 249		      "pushfl		\n\t"
 250		      "popl %0		\n\t"
 251		      "movl %0, %1	\n\t"
 252		      "xorl %2, %0	\n\t"
 253		      "pushl %0		\n\t"
 254		      "popfl		\n\t"
 255		      "pushfl		\n\t"
 256		      "popl %0		\n\t"
 257		      "popfl		\n\t"
 258
 259		      : "=&r" (f1), "=&r" (f2)
 260		      : "ir" (flag));
 261
 262	return ((f1^f2) & flag) != 0;
 263}
 264
 265/* Probe for the CPUID instruction */
 266int have_cpuid_p(void)
 267{
 268	return flag_is_changeable_p(X86_EFLAGS_ID);
 269}
 270
 271static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
 272{
 273	unsigned long lo, hi;
 274
 275	if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr)
 276		return;
 277
 278	/* Disable processor serial number: */
 279
 280	rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
 281	lo |= 0x200000;
 282	wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi);
 283
 284	pr_notice("CPU serial number disabled.\n");
 285	clear_cpu_cap(c, X86_FEATURE_PN);
 286
 287	/* Disabling the serial number may affect the cpuid level */
 288	c->cpuid_level = cpuid_eax(0);
 289}
 290
 291static int __init x86_serial_nr_setup(char *s)
 292{
 293	disable_x86_serial_nr = 0;
 294	return 1;
 295}
 296__setup("serialnumber", x86_serial_nr_setup);
 297#else
 298static inline int flag_is_changeable_p(u32 flag)
 299{
 300	return 1;
 301}
 302static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
 303{
 304}
 305#endif
 306
 307static __init int setup_disable_smep(char *arg)
 308{
 309	setup_clear_cpu_cap(X86_FEATURE_SMEP);
 310	/* Check for things that depend on SMEP being enabled: */
 311	check_mpx_erratum(&boot_cpu_data);
 312	return 1;
 313}
 314__setup("nosmep", setup_disable_smep);
 315
 316static __always_inline void setup_smep(struct cpuinfo_x86 *c)
 317{
 318	if (cpu_has(c, X86_FEATURE_SMEP))
 319		cr4_set_bits(X86_CR4_SMEP);
 320}
 321
 322static __init int setup_disable_smap(char *arg)
 323{
 324	setup_clear_cpu_cap(X86_FEATURE_SMAP);
 325	return 1;
 326}
 327__setup("nosmap", setup_disable_smap);
 328
 329static __always_inline void setup_smap(struct cpuinfo_x86 *c)
 330{
 331	unsigned long eflags = native_save_fl();
 332
 333	/* This should have been cleared long ago */
 334	BUG_ON(eflags & X86_EFLAGS_AC);
 335
 336	if (cpu_has(c, X86_FEATURE_SMAP)) {
 337#ifdef CONFIG_X86_SMAP
 338		cr4_set_bits(X86_CR4_SMAP);
 339#else
 340		cr4_clear_bits(X86_CR4_SMAP);
 341#endif
 342	}
 343}
 344
 345static __always_inline void setup_umip(struct cpuinfo_x86 *c)
 346{
 347	/* Check the boot processor, plus build option for UMIP. */
 348	if (!cpu_feature_enabled(X86_FEATURE_UMIP))
 349		goto out;
 350
 351	/* Check the current processor's cpuid bits. */
 352	if (!cpu_has(c, X86_FEATURE_UMIP))
 353		goto out;
 354
 355	cr4_set_bits(X86_CR4_UMIP);
 356
 357	pr_info_once("x86/cpu: User Mode Instruction Prevention (UMIP) activated\n");
 358
 359	return;
 360
 361out:
 362	/*
 363	 * Make sure UMIP is disabled in case it was enabled in a
 364	 * previous boot (e.g., via kexec).
 365	 */
 366	cr4_clear_bits(X86_CR4_UMIP);
 367}
 368
 369static DEFINE_STATIC_KEY_FALSE_RO(cr_pinning);
 370static unsigned long cr4_pinned_bits __ro_after_init;
 371
 372void native_write_cr0(unsigned long val)
 373{
 374	unsigned long bits_missing = 0;
 375
 376set_register:
 377	asm volatile("mov %0,%%cr0": "+r" (val), "+m" (__force_order));
 378
 379	if (static_branch_likely(&cr_pinning)) {
 380		if (unlikely((val & X86_CR0_WP) != X86_CR0_WP)) {
 381			bits_missing = X86_CR0_WP;
 382			val |= bits_missing;
 383			goto set_register;
 384		}
 385		/* Warn after we've set the missing bits. */
 386		WARN_ONCE(bits_missing, "CR0 WP bit went missing!?\n");
 387	}
 388}
 389EXPORT_SYMBOL(native_write_cr0);
 390
 391void native_write_cr4(unsigned long val)
 392{
 393	unsigned long bits_missing = 0;
 394
 395set_register:
 396	asm volatile("mov %0,%%cr4": "+r" (val), "+m" (cr4_pinned_bits));
 397
 398	if (static_branch_likely(&cr_pinning)) {
 399		if (unlikely((val & cr4_pinned_bits) != cr4_pinned_bits)) {
 400			bits_missing = ~val & cr4_pinned_bits;
 401			val |= bits_missing;
 402			goto set_register;
 403		}
 404		/* Warn after we've set the missing bits. */
 405		WARN_ONCE(bits_missing, "CR4 bits went missing: %lx!?\n",
 406			  bits_missing);
 407	}
 408}
 409EXPORT_SYMBOL(native_write_cr4);
 410
 411void cr4_init(void)
 412{
 413	unsigned long cr4 = __read_cr4();
 414
 415	if (boot_cpu_has(X86_FEATURE_PCID))
 416		cr4 |= X86_CR4_PCIDE;
 417	if (static_branch_likely(&cr_pinning))
 418		cr4 |= cr4_pinned_bits;
 419
 420	__write_cr4(cr4);
 421
 422	/* Initialize cr4 shadow for this CPU. */
 423	this_cpu_write(cpu_tlbstate.cr4, cr4);
 424}
 425
 426/*
 427 * Once CPU feature detection is finished (and boot params have been
 428 * parsed), record any of the sensitive CR bits that are set, and
 429 * enable CR pinning.
 430 */
 431static void __init setup_cr_pinning(void)
 432{
 433	unsigned long mask;
 434
 435	mask = (X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_UMIP);
 436	cr4_pinned_bits = this_cpu_read(cpu_tlbstate.cr4) & mask;
 437	static_key_enable(&cr_pinning.key);
 438}
 439
 440/*
 441 * Protection Keys are not available in 32-bit mode.
 442 */
 443static bool pku_disabled;
 444
 445static __always_inline void setup_pku(struct cpuinfo_x86 *c)
 446{
 447	struct pkru_state *pk;
 448
 449	/* check the boot processor, plus compile options for PKU: */
 450	if (!cpu_feature_enabled(X86_FEATURE_PKU))
 451		return;
 452	/* checks the actual processor's cpuid bits: */
 453	if (!cpu_has(c, X86_FEATURE_PKU))
 454		return;
 455	if (pku_disabled)
 456		return;
 457
 458	cr4_set_bits(X86_CR4_PKE);
 459	pk = get_xsave_addr(&init_fpstate.xsave, XFEATURE_PKRU);
 460	if (pk)
 461		pk->pkru = init_pkru_value;
 462	/*
 463	 * Seting X86_CR4_PKE will cause the X86_FEATURE_OSPKE
 464	 * cpuid bit to be set.  We need to ensure that we
 465	 * update that bit in this CPU's "cpu_info".
 466	 */
 467	get_cpu_cap(c);
 468}
 469
 470#ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
 471static __init int setup_disable_pku(char *arg)
 472{
 473	/*
 474	 * Do not clear the X86_FEATURE_PKU bit.  All of the
 475	 * runtime checks are against OSPKE so clearing the
 476	 * bit does nothing.
 477	 *
 478	 * This way, we will see "pku" in cpuinfo, but not
 479	 * "ospke", which is exactly what we want.  It shows
 480	 * that the CPU has PKU, but the OS has not enabled it.
 481	 * This happens to be exactly how a system would look
 482	 * if we disabled the config option.
 483	 */
 484	pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n");
 485	pku_disabled = true;
 486	return 1;
 487}
 488__setup("nopku", setup_disable_pku);
 489#endif /* CONFIG_X86_64 */
 490
 491/*
 492 * Some CPU features depend on higher CPUID levels, which may not always
 493 * be available due to CPUID level capping or broken virtualization
 494 * software.  Add those features to this table to auto-disable them.
 495 */
 496struct cpuid_dependent_feature {
 497	u32 feature;
 498	u32 level;
 499};
 500
 501static const struct cpuid_dependent_feature
 502cpuid_dependent_features[] = {
 503	{ X86_FEATURE_MWAIT,		0x00000005 },
 504	{ X86_FEATURE_DCA,		0x00000009 },
 505	{ X86_FEATURE_XSAVE,		0x0000000d },
 506	{ 0, 0 }
 507};
 508
 509static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn)
 510{
 511	const struct cpuid_dependent_feature *df;
 512
 513	for (df = cpuid_dependent_features; df->feature; df++) {
 514
 515		if (!cpu_has(c, df->feature))
 516			continue;
 517		/*
 518		 * Note: cpuid_level is set to -1 if unavailable, but
 519		 * extended_extended_level is set to 0 if unavailable
 520		 * and the legitimate extended levels are all negative
 521		 * when signed; hence the weird messing around with
 522		 * signs here...
 523		 */
 524		if (!((s32)df->level < 0 ?
 525		     (u32)df->level > (u32)c->extended_cpuid_level :
 526		     (s32)df->level > (s32)c->cpuid_level))
 527			continue;
 528
 529		clear_cpu_cap(c, df->feature);
 530		if (!warn)
 531			continue;
 532
 533		pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n",
 534			x86_cap_flag(df->feature), df->level);
 535	}
 536}
 537
 538/*
 539 * Naming convention should be: <Name> [(<Codename>)]
 540 * This table only is used unless init_<vendor>() below doesn't set it;
 541 * in particular, if CPUID levels 0x80000002..4 are supported, this
 542 * isn't used
 543 */
 544
 545/* Look up CPU names by table lookup. */
 546static const char *table_lookup_model(struct cpuinfo_x86 *c)
 547{
 548#ifdef CONFIG_X86_32
 549	const struct legacy_cpu_model_info *info;
 550
 551	if (c->x86_model >= 16)
 552		return NULL;	/* Range check */
 553
 554	if (!this_cpu)
 555		return NULL;
 556
 557	info = this_cpu->legacy_models;
 558
 559	while (info->family) {
 560		if (info->family == c->x86)
 561			return info->model_names[c->x86_model];
 562		info++;
 563	}
 564#endif
 565	return NULL;		/* Not found */
 566}
 567
 568__u32 cpu_caps_cleared[NCAPINTS + NBUGINTS];
 569__u32 cpu_caps_set[NCAPINTS + NBUGINTS];
 570
 571void load_percpu_segment(int cpu)
 572{
 573#ifdef CONFIG_X86_32
 574	loadsegment(fs, __KERNEL_PERCPU);
 575#else
 576	__loadsegment_simple(gs, 0);
 577	wrmsrl(MSR_GS_BASE, cpu_kernelmode_gs_base(cpu));
 578#endif
 579	load_stack_canary_segment();
 580}
 581
 582#ifdef CONFIG_X86_32
 583/* The 32-bit entry code needs to find cpu_entry_area. */
 584DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area);
 585#endif
 586
 587/* Load the original GDT from the per-cpu structure */
 588void load_direct_gdt(int cpu)
 589{
 590	struct desc_ptr gdt_descr;
 591
 592	gdt_descr.address = (long)get_cpu_gdt_rw(cpu);
 593	gdt_descr.size = GDT_SIZE - 1;
 594	load_gdt(&gdt_descr);
 595}
 596EXPORT_SYMBOL_GPL(load_direct_gdt);
 597
 598/* Load a fixmap remapping of the per-cpu GDT */
 599void load_fixmap_gdt(int cpu)
 600{
 601	struct desc_ptr gdt_descr;
 602
 603	gdt_descr.address = (long)get_cpu_gdt_ro(cpu);
 604	gdt_descr.size = GDT_SIZE - 1;
 605	load_gdt(&gdt_descr);
 606}
 607EXPORT_SYMBOL_GPL(load_fixmap_gdt);
 608
 609/*
 610 * Current gdt points %fs at the "master" per-cpu area: after this,
 611 * it's on the real one.
 612 */
 613void switch_to_new_gdt(int cpu)
 614{
 615	/* Load the original GDT */
 616	load_direct_gdt(cpu);
 
 
 
 617	/* Reload the per-cpu base */
 
 618	load_percpu_segment(cpu);
 619}
 620
 621static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
 622
 623static void get_model_name(struct cpuinfo_x86 *c)
 624{
 625	unsigned int *v;
 626	char *p, *q, *s;
 627
 628	if (c->extended_cpuid_level < 0x80000004)
 629		return;
 630
 631	v = (unsigned int *)c->x86_model_id;
 632	cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]);
 633	cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]);
 634	cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]);
 635	c->x86_model_id[48] = 0;
 636
 637	/* Trim whitespace */
 638	p = q = s = &c->x86_model_id[0];
 639
 640	while (*p == ' ')
 641		p++;
 642
 643	while (*p) {
 644		/* Note the last non-whitespace index */
 645		if (!isspace(*p))
 646			s = q;
 647
 648		*q++ = *p++;
 649	}
 650
 651	*(s + 1) = '\0';
 652}
 653
 654void detect_num_cpu_cores(struct cpuinfo_x86 *c)
 655{
 656	unsigned int eax, ebx, ecx, edx;
 657
 658	c->x86_max_cores = 1;
 659	if (!IS_ENABLED(CONFIG_SMP) || c->cpuid_level < 4)
 660		return;
 661
 662	cpuid_count(4, 0, &eax, &ebx, &ecx, &edx);
 663	if (eax & 0x1f)
 664		c->x86_max_cores = (eax >> 26) + 1;
 665}
 666
 667void cpu_detect_cache_sizes(struct cpuinfo_x86 *c)
 668{
 669	unsigned int n, dummy, ebx, ecx, edx, l2size;
 670
 671	n = c->extended_cpuid_level;
 672
 673	if (n >= 0x80000005) {
 674		cpuid(0x80000005, &dummy, &ebx, &ecx, &edx);
 675		c->x86_cache_size = (ecx>>24) + (edx>>24);
 676#ifdef CONFIG_X86_64
 677		/* On K8 L1 TLB is inclusive, so don't count it */
 678		c->x86_tlbsize = 0;
 679#endif
 680	}
 681
 682	if (n < 0x80000006)	/* Some chips just has a large L1. */
 683		return;
 684
 685	cpuid(0x80000006, &dummy, &ebx, &ecx, &edx);
 686	l2size = ecx >> 16;
 687
 688#ifdef CONFIG_X86_64
 689	c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff);
 690#else
 691	/* do processor-specific cache resizing */
 692	if (this_cpu->legacy_cache_size)
 693		l2size = this_cpu->legacy_cache_size(c, l2size);
 694
 695	/* Allow user to override all this if necessary. */
 696	if (cachesize_override != -1)
 697		l2size = cachesize_override;
 698
 699	if (l2size == 0)
 700		return;		/* Again, no L2 cache is possible */
 701#endif
 702
 703	c->x86_cache_size = l2size;
 704}
 705
 706u16 __read_mostly tlb_lli_4k[NR_INFO];
 707u16 __read_mostly tlb_lli_2m[NR_INFO];
 708u16 __read_mostly tlb_lli_4m[NR_INFO];
 709u16 __read_mostly tlb_lld_4k[NR_INFO];
 710u16 __read_mostly tlb_lld_2m[NR_INFO];
 711u16 __read_mostly tlb_lld_4m[NR_INFO];
 712u16 __read_mostly tlb_lld_1g[NR_INFO];
 713
 714static void cpu_detect_tlb(struct cpuinfo_x86 *c)
 715{
 716	if (this_cpu->c_detect_tlb)
 717		this_cpu->c_detect_tlb(c);
 718
 719	pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n",
 720		tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES],
 721		tlb_lli_4m[ENTRIES]);
 722
 723	pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n",
 724		tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES],
 725		tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]);
 726}
 727
 728int detect_ht_early(struct cpuinfo_x86 *c)
 729{
 730#ifdef CONFIG_SMP
 731	u32 eax, ebx, ecx, edx;
 
 
 732
 733	if (!cpu_has(c, X86_FEATURE_HT))
 734		return -1;
 735
 736	if (cpu_has(c, X86_FEATURE_CMP_LEGACY))
 737		return -1;
 738
 739	if (cpu_has(c, X86_FEATURE_XTOPOLOGY))
 740		return -1;
 741
 742	cpuid(1, &eax, &ebx, &ecx, &edx);
 743
 744	smp_num_siblings = (ebx & 0xff0000) >> 16;
 745	if (smp_num_siblings == 1)
 746		pr_info_once("CPU0: Hyper-Threading is disabled\n");
 747#endif
 748	return 0;
 749}
 750
 751void detect_ht(struct cpuinfo_x86 *c)
 752{
 753#ifdef CONFIG_SMP
 754	int index_msb, core_bits;
 755
 756	if (detect_ht_early(c) < 0)
 757		return;
 758
 759	index_msb = get_count_order(smp_num_siblings);
 760	c->phys_proc_id = apic->phys_pkg_id(c->initial_apicid, index_msb);
 761
 762	smp_num_siblings = smp_num_siblings / c->x86_max_cores;
 763
 764	index_msb = get_count_order(smp_num_siblings);
 765
 766	core_bits = get_count_order(c->x86_max_cores);
 767
 768	c->cpu_core_id = apic->phys_pkg_id(c->initial_apicid, index_msb) &
 769				       ((1 << core_bits) - 1);
 
 
 
 
 
 
 
 
 
 770#endif
 771}
 772
 773static void get_cpu_vendor(struct cpuinfo_x86 *c)
 774{
 775	char *v = c->x86_vendor_id;
 776	int i;
 777
 778	for (i = 0; i < X86_VENDOR_NUM; i++) {
 779		if (!cpu_devs[i])
 780			break;
 781
 782		if (!strcmp(v, cpu_devs[i]->c_ident[0]) ||
 783		    (cpu_devs[i]->c_ident[1] &&
 784		     !strcmp(v, cpu_devs[i]->c_ident[1]))) {
 785
 786			this_cpu = cpu_devs[i];
 787			c->x86_vendor = this_cpu->c_x86_vendor;
 788			return;
 789		}
 790	}
 791
 792	pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \
 793		    "CPU: Your system may be unstable.\n", v);
 794
 795	c->x86_vendor = X86_VENDOR_UNKNOWN;
 796	this_cpu = &default_cpu;
 797}
 798
 799void cpu_detect(struct cpuinfo_x86 *c)
 800{
 801	/* Get vendor name */
 802	cpuid(0x00000000, (unsigned int *)&c->cpuid_level,
 803	      (unsigned int *)&c->x86_vendor_id[0],
 804	      (unsigned int *)&c->x86_vendor_id[8],
 805	      (unsigned int *)&c->x86_vendor_id[4]);
 806
 807	c->x86 = 4;
 808	/* Intel-defined flags: level 0x00000001 */
 809	if (c->cpuid_level >= 0x00000001) {
 810		u32 junk, tfms, cap0, misc;
 811
 812		cpuid(0x00000001, &tfms, &misc, &junk, &cap0);
 813		c->x86		= x86_family(tfms);
 814		c->x86_model	= x86_model(tfms);
 815		c->x86_stepping	= x86_stepping(tfms);
 816
 817		if (cap0 & (1<<19)) {
 818			c->x86_clflush_size = ((misc >> 8) & 0xff) * 8;
 819			c->x86_cache_alignment = c->x86_clflush_size;
 820		}
 821	}
 822}
 823
 824static void apply_forced_caps(struct cpuinfo_x86 *c)
 825{
 826	int i;
 827
 828	for (i = 0; i < NCAPINTS + NBUGINTS; i++) {
 829		c->x86_capability[i] &= ~cpu_caps_cleared[i];
 830		c->x86_capability[i] |= cpu_caps_set[i];
 831	}
 832}
 833
 834static void init_speculation_control(struct cpuinfo_x86 *c)
 835{
 836	/*
 837	 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support,
 838	 * and they also have a different bit for STIBP support. Also,
 839	 * a hypervisor might have set the individual AMD bits even on
 840	 * Intel CPUs, for finer-grained selection of what's available.
 841	 */
 842	if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) {
 843		set_cpu_cap(c, X86_FEATURE_IBRS);
 844		set_cpu_cap(c, X86_FEATURE_IBPB);
 845		set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
 846	}
 847
 848	if (cpu_has(c, X86_FEATURE_INTEL_STIBP))
 849		set_cpu_cap(c, X86_FEATURE_STIBP);
 850
 851	if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) ||
 852	    cpu_has(c, X86_FEATURE_VIRT_SSBD))
 853		set_cpu_cap(c, X86_FEATURE_SSBD);
 854
 855	if (cpu_has(c, X86_FEATURE_AMD_IBRS)) {
 856		set_cpu_cap(c, X86_FEATURE_IBRS);
 857		set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
 858	}
 859
 860	if (cpu_has(c, X86_FEATURE_AMD_IBPB))
 861		set_cpu_cap(c, X86_FEATURE_IBPB);
 862
 863	if (cpu_has(c, X86_FEATURE_AMD_STIBP)) {
 864		set_cpu_cap(c, X86_FEATURE_STIBP);
 865		set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
 866	}
 867
 868	if (cpu_has(c, X86_FEATURE_AMD_SSBD)) {
 869		set_cpu_cap(c, X86_FEATURE_SSBD);
 870		set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL);
 871		clear_cpu_cap(c, X86_FEATURE_VIRT_SSBD);
 872	}
 873}
 874
 875static void init_cqm(struct cpuinfo_x86 *c)
 876{
 877	if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
 878		c->x86_cache_max_rmid  = -1;
 879		c->x86_cache_occ_scale = -1;
 880		return;
 881	}
 882
 883	/* will be overridden if occupancy monitoring exists */
 884	c->x86_cache_max_rmid = cpuid_ebx(0xf);
 885
 886	if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
 887	    cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
 888	    cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
 889		u32 eax, ebx, ecx, edx;
 890
 891		/* QoS sub-leaf, EAX=0Fh, ECX=1 */
 892		cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
 893
 894		c->x86_cache_max_rmid  = ecx;
 895		c->x86_cache_occ_scale = ebx;
 896	}
 897}
 898
 899void get_cpu_cap(struct cpuinfo_x86 *c)
 900{
 901	u32 eax, ebx, ecx, edx;
 902
 903	/* Intel-defined flags: level 0x00000001 */
 904	if (c->cpuid_level >= 0x00000001) {
 905		cpuid(0x00000001, &eax, &ebx, &ecx, &edx);
 906
 907		c->x86_capability[CPUID_1_ECX] = ecx;
 908		c->x86_capability[CPUID_1_EDX] = edx;
 909	}
 910
 911	/* Thermal and Power Management Leaf: level 0x00000006 (eax) */
 912	if (c->cpuid_level >= 0x00000006)
 913		c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
 914
 915	/* Additional Intel-defined flags: level 0x00000007 */
 916	if (c->cpuid_level >= 0x00000007) {
 917		cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
 918		c->x86_capability[CPUID_7_0_EBX] = ebx;
 919		c->x86_capability[CPUID_7_ECX] = ecx;
 920		c->x86_capability[CPUID_7_EDX] = edx;
 921
 922		/* Check valid sub-leaf index before accessing it */
 923		if (eax >= 1) {
 924			cpuid_count(0x00000007, 1, &eax, &ebx, &ecx, &edx);
 925			c->x86_capability[CPUID_7_1_EAX] = eax;
 926		}
 927	}
 928
 929	/* Extended state features: level 0x0000000d */
 930	if (c->cpuid_level >= 0x0000000d) {
 931		cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
 932
 933		c->x86_capability[CPUID_D_1_EAX] = eax;
 934	}
 935
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 936	/* AMD-defined flags: level 0x80000001 */
 937	eax = cpuid_eax(0x80000000);
 938	c->extended_cpuid_level = eax;
 939
 940	if ((eax & 0xffff0000) == 0x80000000) {
 941		if (eax >= 0x80000001) {
 942			cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
 943
 944			c->x86_capability[CPUID_8000_0001_ECX] = ecx;
 945			c->x86_capability[CPUID_8000_0001_EDX] = edx;
 946		}
 947	}
 948
 949	if (c->extended_cpuid_level >= 0x80000007) {
 950		cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
 951
 952		c->x86_capability[CPUID_8000_0007_EBX] = ebx;
 953		c->x86_power = edx;
 954	}
 955
 956	if (c->extended_cpuid_level >= 0x80000008) {
 957		cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
 958		c->x86_capability[CPUID_8000_0008_EBX] = ebx;
 959	}
 960
 961	if (c->extended_cpuid_level >= 0x8000000a)
 962		c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
 963
 964	init_scattered_cpuid_features(c);
 965	init_speculation_control(c);
 966	init_cqm(c);
 967
 968	/*
 969	 * Clear/Set all flags overridden by options, after probe.
 970	 * This needs to happen each time we re-probe, which may happen
 971	 * several times during CPU initialization.
 972	 */
 973	apply_forced_caps(c);
 974}
 975
 976void get_cpu_address_sizes(struct cpuinfo_x86 *c)
 977{
 978	u32 eax, ebx, ecx, edx;
 979
 980	if (c->extended_cpuid_level >= 0x80000008) {
 981		cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
 982
 983		c->x86_virt_bits = (eax >> 8) & 0xff;
 984		c->x86_phys_bits = eax & 0xff;
 
 985	}
 986#ifdef CONFIG_X86_32
 987	else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
 988		c->x86_phys_bits = 36;
 989#endif
 990	c->x86_cache_bits = c->x86_phys_bits;
 
 
 
 
 991}
 992
 993static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
 994{
 995#ifdef CONFIG_X86_32
 996	int i;
 997
 998	/*
 999	 * First of all, decide if this is a 486 or higher
1000	 * It's a 486 if we can modify the AC flag
1001	 */
1002	if (flag_is_changeable_p(X86_EFLAGS_AC))
1003		c->x86 = 4;
1004	else
1005		c->x86 = 3;
1006
1007	for (i = 0; i < X86_VENDOR_NUM; i++)
1008		if (cpu_devs[i] && cpu_devs[i]->c_identify) {
1009			c->x86_vendor_id[0] = 0;
1010			cpu_devs[i]->c_identify(c);
1011			if (c->x86_vendor_id[0]) {
1012				get_cpu_vendor(c);
1013				break;
1014			}
1015		}
1016#endif
1017}
1018
1019#define NO_SPECULATION		BIT(0)
1020#define NO_MELTDOWN		BIT(1)
1021#define NO_SSB			BIT(2)
1022#define NO_L1TF			BIT(3)
1023#define NO_MDS			BIT(4)
1024#define MSBDS_ONLY		BIT(5)
1025#define NO_SWAPGS		BIT(6)
1026#define NO_ITLB_MULTIHIT	BIT(7)
1027
1028#define VULNWL(_vendor, _family, _model, _whitelist)	\
1029	{ X86_VENDOR_##_vendor, _family, _model, X86_FEATURE_ANY, _whitelist }
1030
1031#define VULNWL_INTEL(model, whitelist)		\
1032	VULNWL(INTEL, 6, INTEL_FAM6_##model, whitelist)
1033
1034#define VULNWL_AMD(family, whitelist)		\
1035	VULNWL(AMD, family, X86_MODEL_ANY, whitelist)
1036
1037#define VULNWL_HYGON(family, whitelist)		\
1038	VULNWL(HYGON, family, X86_MODEL_ANY, whitelist)
1039
1040static const __initconst struct x86_cpu_id cpu_vuln_whitelist[] = {
1041	VULNWL(ANY,	4, X86_MODEL_ANY,	NO_SPECULATION),
1042	VULNWL(CENTAUR,	5, X86_MODEL_ANY,	NO_SPECULATION),
1043	VULNWL(INTEL,	5, X86_MODEL_ANY,	NO_SPECULATION),
1044	VULNWL(NSC,	5, X86_MODEL_ANY,	NO_SPECULATION),
1045
1046	/* Intel Family 6 */
1047	VULNWL_INTEL(ATOM_SALTWELL,		NO_SPECULATION | NO_ITLB_MULTIHIT),
1048	VULNWL_INTEL(ATOM_SALTWELL_TABLET,	NO_SPECULATION | NO_ITLB_MULTIHIT),
1049	VULNWL_INTEL(ATOM_SALTWELL_MID,		NO_SPECULATION | NO_ITLB_MULTIHIT),
1050	VULNWL_INTEL(ATOM_BONNELL,		NO_SPECULATION | NO_ITLB_MULTIHIT),
1051	VULNWL_INTEL(ATOM_BONNELL_MID,		NO_SPECULATION | NO_ITLB_MULTIHIT),
1052
1053	VULNWL_INTEL(ATOM_SILVERMONT,		NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1054	VULNWL_INTEL(ATOM_SILVERMONT_D,		NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1055	VULNWL_INTEL(ATOM_SILVERMONT_MID,	NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1056	VULNWL_INTEL(ATOM_AIRMONT,		NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1057	VULNWL_INTEL(XEON_PHI_KNL,		NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1058	VULNWL_INTEL(XEON_PHI_KNM,		NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1059
1060	VULNWL_INTEL(CORE_YONAH,		NO_SSB),
1061
1062	VULNWL_INTEL(ATOM_AIRMONT_MID,		NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT),
1063	VULNWL_INTEL(ATOM_AIRMONT_NP,		NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1064
1065	VULNWL_INTEL(ATOM_GOLDMONT,		NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1066	VULNWL_INTEL(ATOM_GOLDMONT_D,		NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1067	VULNWL_INTEL(ATOM_GOLDMONT_PLUS,	NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT),
1068
1069	/*
1070	 * Technically, swapgs isn't serializing on AMD (despite it previously
1071	 * being documented as such in the APM).  But according to AMD, %gs is
1072	 * updated non-speculatively, and the issuing of %gs-relative memory
1073	 * operands will be blocked until the %gs update completes, which is
1074	 * good enough for our purposes.
1075	 */
1076
1077	VULNWL_INTEL(ATOM_TREMONT_D,		NO_ITLB_MULTIHIT),
1078
1079	/* AMD Family 0xf - 0x12 */
1080	VULNWL_AMD(0x0f,	NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1081	VULNWL_AMD(0x10,	NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1082	VULNWL_AMD(0x11,	NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1083	VULNWL_AMD(0x12,	NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1084
1085	/* FAMILY_ANY must be last, otherwise 0x0f - 0x12 matches won't work */
1086	VULNWL_AMD(X86_FAMILY_ANY,	NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1087	VULNWL_HYGON(X86_FAMILY_ANY,	NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT),
1088	{}
1089};
1090
1091static bool __init cpu_matches(unsigned long which)
1092{
1093	const struct x86_cpu_id *m = x86_match_cpu(cpu_vuln_whitelist);
1094
1095	return m && !!(m->driver_data & which);
1096}
1097
1098u64 x86_read_arch_cap_msr(void)
1099{
1100	u64 ia32_cap = 0;
1101
1102	if (boot_cpu_has(X86_FEATURE_ARCH_CAPABILITIES))
1103		rdmsrl(MSR_IA32_ARCH_CAPABILITIES, ia32_cap);
1104
1105	return ia32_cap;
1106}
1107
1108static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c)
1109{
1110	u64 ia32_cap = x86_read_arch_cap_msr();
1111
1112	/* Set ITLB_MULTIHIT bug if cpu is not in the whitelist and not mitigated */
1113	if (!cpu_matches(NO_ITLB_MULTIHIT) && !(ia32_cap & ARCH_CAP_PSCHANGE_MC_NO))
1114		setup_force_cpu_bug(X86_BUG_ITLB_MULTIHIT);
1115
1116	if (cpu_matches(NO_SPECULATION))
1117		return;
1118
1119	setup_force_cpu_bug(X86_BUG_SPECTRE_V1);
1120	setup_force_cpu_bug(X86_BUG_SPECTRE_V2);
1121
1122	if (!cpu_matches(NO_SSB) && !(ia32_cap & ARCH_CAP_SSB_NO) &&
1123	   !cpu_has(c, X86_FEATURE_AMD_SSB_NO))
1124		setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS);
1125
1126	if (ia32_cap & ARCH_CAP_IBRS_ALL)
1127		setup_force_cpu_cap(X86_FEATURE_IBRS_ENHANCED);
1128
1129	if (!cpu_matches(NO_MDS) && !(ia32_cap & ARCH_CAP_MDS_NO)) {
1130		setup_force_cpu_bug(X86_BUG_MDS);
1131		if (cpu_matches(MSBDS_ONLY))
1132			setup_force_cpu_bug(X86_BUG_MSBDS_ONLY);
1133	}
1134
1135	if (!cpu_matches(NO_SWAPGS))
1136		setup_force_cpu_bug(X86_BUG_SWAPGS);
1137
1138	/*
1139	 * When the CPU is not mitigated for TAA (TAA_NO=0) set TAA bug when:
1140	 *	- TSX is supported or
1141	 *	- TSX_CTRL is present
1142	 *
1143	 * TSX_CTRL check is needed for cases when TSX could be disabled before
1144	 * the kernel boot e.g. kexec.
1145	 * TSX_CTRL check alone is not sufficient for cases when the microcode
1146	 * update is not present or running as guest that don't get TSX_CTRL.
1147	 */
1148	if (!(ia32_cap & ARCH_CAP_TAA_NO) &&
1149	    (cpu_has(c, X86_FEATURE_RTM) ||
1150	     (ia32_cap & ARCH_CAP_TSX_CTRL_MSR)))
1151		setup_force_cpu_bug(X86_BUG_TAA);
1152
1153	if (cpu_matches(NO_MELTDOWN))
1154		return;
1155
1156	/* Rogue Data Cache Load? No! */
1157	if (ia32_cap & ARCH_CAP_RDCL_NO)
1158		return;
1159
1160	setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN);
1161
1162	if (cpu_matches(NO_L1TF))
1163		return;
1164
1165	setup_force_cpu_bug(X86_BUG_L1TF);
1166}
1167
1168/*
1169 * The NOPL instruction is supposed to exist on all CPUs of family >= 6;
1170 * unfortunately, that's not true in practice because of early VIA
1171 * chips and (more importantly) broken virtualizers that are not easy
1172 * to detect. In the latter case it doesn't even *fail* reliably, so
1173 * probing for it doesn't even work. Disable it completely on 32-bit
1174 * unless we can find a reliable way to detect all the broken cases.
1175 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has().
1176 */
1177static void detect_nopl(void)
1178{
1179#ifdef CONFIG_X86_32
1180	setup_clear_cpu_cap(X86_FEATURE_NOPL);
1181#else
1182	setup_force_cpu_cap(X86_FEATURE_NOPL);
1183#endif
1184}
1185
1186/*
1187 * Do minimum CPU detection early.
1188 * Fields really needed: vendor, cpuid_level, family, model, mask,
1189 * cache alignment.
1190 * The others are not touched to avoid unwanted side effects.
1191 *
1192 * WARNING: this function is only called on the boot CPU.  Don't add code
1193 * here that is supposed to run on all CPUs.
1194 */
1195static void __init early_identify_cpu(struct cpuinfo_x86 *c)
1196{
1197#ifdef CONFIG_X86_64
1198	c->x86_clflush_size = 64;
1199	c->x86_phys_bits = 36;
1200	c->x86_virt_bits = 48;
1201#else
1202	c->x86_clflush_size = 32;
1203	c->x86_phys_bits = 32;
1204	c->x86_virt_bits = 32;
1205#endif
1206	c->x86_cache_alignment = c->x86_clflush_size;
1207
1208	memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1209	c->extended_cpuid_level = 0;
1210
1211	if (!have_cpuid_p())
1212		identify_cpu_without_cpuid(c);
1213
1214	/* cyrix could have cpuid enabled via c_identify()*/
1215	if (have_cpuid_p()) {
1216		cpu_detect(c);
1217		get_cpu_vendor(c);
1218		get_cpu_cap(c);
1219		get_cpu_address_sizes(c);
1220		setup_force_cpu_cap(X86_FEATURE_CPUID);
1221
1222		if (this_cpu->c_early_init)
1223			this_cpu->c_early_init(c);
1224
1225		c->cpu_index = 0;
1226		filter_cpuid_features(c, false);
1227
1228		if (this_cpu->c_bsp_init)
1229			this_cpu->c_bsp_init(c);
1230	} else {
1231		setup_clear_cpu_cap(X86_FEATURE_CPUID);
1232	}
1233
1234	setup_force_cpu_cap(X86_FEATURE_ALWAYS);
1235
1236	cpu_set_bug_bits(c);
1237
1238	fpu__init_system(c);
1239
1240#ifdef CONFIG_X86_32
1241	/*
1242	 * Regardless of whether PCID is enumerated, the SDM says
1243	 * that it can't be enabled in 32-bit mode.
1244	 */
1245	setup_clear_cpu_cap(X86_FEATURE_PCID);
1246#endif
1247
1248	/*
1249	 * Later in the boot process pgtable_l5_enabled() relies on
1250	 * cpu_feature_enabled(X86_FEATURE_LA57). If 5-level paging is not
1251	 * enabled by this point we need to clear the feature bit to avoid
1252	 * false-positives at the later stage.
1253	 *
1254	 * pgtable_l5_enabled() can be false here for several reasons:
1255	 *  - 5-level paging is disabled compile-time;
1256	 *  - it's 32-bit kernel;
1257	 *  - machine doesn't support 5-level paging;
1258	 *  - user specified 'no5lvl' in kernel command line.
1259	 */
1260	if (!pgtable_l5_enabled())
1261		setup_clear_cpu_cap(X86_FEATURE_LA57);
1262
1263	detect_nopl();
1264}
1265
1266void __init early_cpu_init(void)
1267{
1268	const struct cpu_dev *const *cdev;
1269	int count = 0;
1270
1271#ifdef CONFIG_PROCESSOR_SELECT
1272	pr_info("KERNEL supported cpus:\n");
1273#endif
1274
1275	for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) {
1276		const struct cpu_dev *cpudev = *cdev;
1277
1278		if (count >= X86_VENDOR_NUM)
1279			break;
1280		cpu_devs[count] = cpudev;
1281		count++;
1282
1283#ifdef CONFIG_PROCESSOR_SELECT
1284		{
1285			unsigned int j;
1286
1287			for (j = 0; j < 2; j++) {
1288				if (!cpudev->c_ident[j])
1289					continue;
1290				pr_info("  %s %s\n", cpudev->c_vendor,
1291					cpudev->c_ident[j]);
1292			}
1293		}
1294#endif
1295	}
1296	early_identify_cpu(&boot_cpu_data);
1297}
1298
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1299static void detect_null_seg_behavior(struct cpuinfo_x86 *c)
1300{
1301#ifdef CONFIG_X86_64
1302	/*
1303	 * Empirically, writing zero to a segment selector on AMD does
1304	 * not clear the base, whereas writing zero to a segment
1305	 * selector on Intel does clear the base.  Intel's behavior
1306	 * allows slightly faster context switches in the common case
1307	 * where GS is unused by the prev and next threads.
1308	 *
1309	 * Since neither vendor documents this anywhere that I can see,
1310	 * detect it directly instead of hardcoding the choice by
1311	 * vendor.
1312	 *
1313	 * I've designated AMD's behavior as the "bug" because it's
1314	 * counterintuitive and less friendly.
1315	 */
1316
1317	unsigned long old_base, tmp;
1318	rdmsrl(MSR_FS_BASE, old_base);
1319	wrmsrl(MSR_FS_BASE, 1);
1320	loadsegment(fs, 0);
1321	rdmsrl(MSR_FS_BASE, tmp);
1322	if (tmp != 0)
1323		set_cpu_bug(c, X86_BUG_NULL_SEG);
1324	wrmsrl(MSR_FS_BASE, old_base);
1325#endif
1326}
1327
1328static void generic_identify(struct cpuinfo_x86 *c)
1329{
1330	c->extended_cpuid_level = 0;
1331
1332	if (!have_cpuid_p())
1333		identify_cpu_without_cpuid(c);
1334
1335	/* cyrix could have cpuid enabled via c_identify()*/
1336	if (!have_cpuid_p())
1337		return;
1338
1339	cpu_detect(c);
1340
1341	get_cpu_vendor(c);
1342
1343	get_cpu_cap(c);
1344
1345	get_cpu_address_sizes(c);
1346
1347	if (c->cpuid_level >= 0x00000001) {
1348		c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF;
1349#ifdef CONFIG_X86_32
1350# ifdef CONFIG_SMP
1351		c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1352# else
1353		c->apicid = c->initial_apicid;
1354# endif
1355#endif
1356		c->phys_proc_id = c->initial_apicid;
1357	}
1358
1359	get_model_name(c); /* Default name */
1360
 
 
1361	detect_null_seg_behavior(c);
1362
1363	/*
1364	 * ESPFIX is a strange bug.  All real CPUs have it.  Paravirt
1365	 * systems that run Linux at CPL > 0 may or may not have the
1366	 * issue, but, even if they have the issue, there's absolutely
1367	 * nothing we can do about it because we can't use the real IRET
1368	 * instruction.
1369	 *
1370	 * NB: For the time being, only 32-bit kernels support
1371	 * X86_BUG_ESPFIX as such.  64-bit kernels directly choose
1372	 * whether to apply espfix using paravirt hooks.  If any
1373	 * non-paravirt system ever shows up that does *not* have the
1374	 * ESPFIX issue, we can change this.
1375	 */
1376#ifdef CONFIG_X86_32
1377# ifdef CONFIG_PARAVIRT_XXL
1378	do {
1379		extern void native_iret(void);
1380		if (pv_ops.cpu.iret == native_iret)
1381			set_cpu_bug(c, X86_BUG_ESPFIX);
1382	} while (0);
1383# else
1384	set_cpu_bug(c, X86_BUG_ESPFIX);
1385# endif
1386#endif
1387}
1388
1389static void x86_init_cache_qos(struct cpuinfo_x86 *c)
1390{
1391	/*
1392	 * The heavy lifting of max_rmid and cache_occ_scale are handled
1393	 * in get_cpu_cap().  Here we just set the max_rmid for the boot_cpu
1394	 * in case CQM bits really aren't there in this CPU.
1395	 */
1396	if (c != &boot_cpu_data) {
1397		boot_cpu_data.x86_cache_max_rmid =
1398			min(boot_cpu_data.x86_cache_max_rmid,
1399			    c->x86_cache_max_rmid);
1400	}
1401}
1402
1403/*
1404 * Validate that ACPI/mptables have the same information about the
1405 * effective APIC id and update the package map.
1406 */
1407static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
1408{
1409#ifdef CONFIG_SMP
1410	unsigned int apicid, cpu = smp_processor_id();
1411
1412	apicid = apic->cpu_present_to_apicid(cpu);
1413
1414	if (apicid != c->apicid) {
1415		pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
1416		       cpu, apicid, c->initial_apicid);
1417	}
1418	BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
1419	BUG_ON(topology_update_die_map(c->cpu_die_id, cpu));
1420#else
1421	c->logical_proc_id = 0;
1422#endif
1423}
1424
1425/*
1426 * This does the hard work of actually picking apart the CPU stuff...
1427 */
1428static void identify_cpu(struct cpuinfo_x86 *c)
1429{
1430	int i;
1431
1432	c->loops_per_jiffy = loops_per_jiffy;
1433	c->x86_cache_size = 0;
1434	c->x86_vendor = X86_VENDOR_UNKNOWN;
1435	c->x86_model = c->x86_stepping = 0;	/* So far unknown... */
1436	c->x86_vendor_id[0] = '\0'; /* Unset */
1437	c->x86_model_id[0] = '\0';  /* Unset */
1438	c->x86_max_cores = 1;
1439	c->x86_coreid_bits = 0;
1440	c->cu_id = 0xff;
1441#ifdef CONFIG_X86_64
1442	c->x86_clflush_size = 64;
1443	c->x86_phys_bits = 36;
1444	c->x86_virt_bits = 48;
1445#else
1446	c->cpuid_level = -1;	/* CPUID not detected */
1447	c->x86_clflush_size = 32;
1448	c->x86_phys_bits = 32;
1449	c->x86_virt_bits = 32;
1450#endif
1451	c->x86_cache_alignment = c->x86_clflush_size;
1452	memset(&c->x86_capability, 0, sizeof(c->x86_capability));
1453
1454	generic_identify(c);
1455
1456	if (this_cpu->c_identify)
1457		this_cpu->c_identify(c);
1458
1459	/* Clear/Set all flags overridden by options, after probe */
1460	apply_forced_caps(c);
 
 
 
1461
1462#ifdef CONFIG_X86_64
1463	c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1464#endif
1465
1466	/*
1467	 * Vendor-specific initialization.  In this section we
1468	 * canonicalize the feature flags, meaning if there are
1469	 * features a certain CPU supports which CPUID doesn't
1470	 * tell us, CPUID claiming incorrect flags, or other bugs,
1471	 * we handle them here.
1472	 *
1473	 * At the end of this section, c->x86_capability better
1474	 * indicate the features this CPU genuinely supports!
1475	 */
1476	if (this_cpu->c_init)
1477		this_cpu->c_init(c);
1478
1479	/* Disable the PN if appropriate */
1480	squash_the_stupid_serial_number(c);
1481
1482	/* Set up SMEP/SMAP/UMIP */
1483	setup_smep(c);
1484	setup_smap(c);
1485	setup_umip(c);
1486
1487	/*
1488	 * The vendor-specific functions might have changed features.
1489	 * Now we do "generic changes."
1490	 */
1491
1492	/* Filter out anything that depends on CPUID levels we don't have */
1493	filter_cpuid_features(c, true);
1494
1495	/* If the model name is still unset, do table lookup. */
1496	if (!c->x86_model_id[0]) {
1497		const char *p;
1498		p = table_lookup_model(c);
1499		if (p)
1500			strcpy(c->x86_model_id, p);
1501		else
1502			/* Last resort... */
1503			sprintf(c->x86_model_id, "%02x/%02x",
1504				c->x86, c->x86_model);
1505	}
1506
1507#ifdef CONFIG_X86_64
1508	detect_ht(c);
1509#endif
1510
 
1511	x86_init_rdrand(c);
1512	x86_init_cache_qos(c);
1513	setup_pku(c);
1514
1515	/*
1516	 * Clear/Set all flags overridden by options, need do it
1517	 * before following smp all cpus cap AND.
1518	 */
1519	apply_forced_caps(c);
 
 
 
1520
1521	/*
1522	 * On SMP, boot_cpu_data holds the common feature set between
1523	 * all CPUs; so make sure that we indicate which features are
1524	 * common between the CPUs.  The first time this routine gets
1525	 * executed, c == &boot_cpu_data.
1526	 */
1527	if (c != &boot_cpu_data) {
1528		/* AND the already accumulated flags with these */
1529		for (i = 0; i < NCAPINTS; i++)
1530			boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1531
1532		/* OR, i.e. replicate the bug flags */
1533		for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1534			c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1535	}
1536
1537	/* Init Machine Check Exception if available. */
1538	mcheck_cpu_init(c);
1539
1540	select_idle_routine(c);
1541
1542#ifdef CONFIG_NUMA
1543	numa_add_cpu(smp_processor_id());
1544#endif
1545}
1546
1547/*
1548 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1549 * on 32-bit kernels:
1550 */
1551#ifdef CONFIG_X86_32
1552void enable_sep_cpu(void)
1553{
1554	struct tss_struct *tss;
1555	int cpu;
1556
1557	if (!boot_cpu_has(X86_FEATURE_SEP))
1558		return;
1559
1560	cpu = get_cpu();
1561	tss = &per_cpu(cpu_tss_rw, cpu);
1562
1563	/*
1564	 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1565	 * see the big comment in struct x86_hw_tss's definition.
1566	 */
1567
1568	tss->x86_tss.ss1 = __KERNEL_CS;
1569	wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1570	wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0);
 
 
 
 
1571	wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1572
1573	put_cpu();
1574}
1575#endif
1576
1577void __init identify_boot_cpu(void)
1578{
1579	identify_cpu(&boot_cpu_data);
1580#ifdef CONFIG_X86_32
1581	sysenter_setup();
1582	enable_sep_cpu();
1583#endif
1584	cpu_detect_tlb(&boot_cpu_data);
1585	setup_cr_pinning();
1586
1587	tsx_init();
1588}
1589
1590void identify_secondary_cpu(struct cpuinfo_x86 *c)
1591{
1592	BUG_ON(c == &boot_cpu_data);
1593	identify_cpu(c);
1594#ifdef CONFIG_X86_32
1595	enable_sep_cpu();
1596#endif
1597	mtrr_ap_init();
1598	validate_apic_and_package_id(c);
1599	x86_spec_ctrl_setup_ap();
1600}
1601
1602static __init int setup_noclflush(char *arg)
1603{
1604	setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1605	setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1606	return 1;
1607}
1608__setup("noclflush", setup_noclflush);
1609
1610void print_cpu_info(struct cpuinfo_x86 *c)
1611{
1612	const char *vendor = NULL;
1613
1614	if (c->x86_vendor < X86_VENDOR_NUM) {
1615		vendor = this_cpu->c_vendor;
1616	} else {
1617		if (c->cpuid_level >= 0)
1618			vendor = c->x86_vendor_id;
1619	}
1620
1621	if (vendor && !strstr(c->x86_model_id, vendor))
1622		pr_cont("%s ", vendor);
1623
1624	if (c->x86_model_id[0])
1625		pr_cont("%s", c->x86_model_id);
1626	else
1627		pr_cont("%d86", c->x86);
1628
1629	pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1630
1631	if (c->x86_stepping || c->cpuid_level >= 0)
1632		pr_cont(", stepping: 0x%x)\n", c->x86_stepping);
1633	else
1634		pr_cont(")\n");
1635}
1636
1637/*
1638 * clearcpuid= was already parsed in fpu__init_parse_early_param.
1639 * But we need to keep a dummy __setup around otherwise it would
1640 * show up as an environment variable for init.
1641 */
1642static __init int setup_clearcpuid(char *arg)
1643{
 
 
 
 
 
 
 
1644	return 1;
1645}
1646__setup("clearcpuid=", setup_clearcpuid);
1647
1648#ifdef CONFIG_X86_64
1649DEFINE_PER_CPU_FIRST(struct fixed_percpu_data,
1650		     fixed_percpu_data) __aligned(PAGE_SIZE) __visible;
1651EXPORT_PER_CPU_SYMBOL_GPL(fixed_percpu_data);
 
 
 
 
 
 
 
 
1652
1653/*
1654 * The following percpu variables are hot.  Align current_task to
1655 * cacheline size such that they fall in the same cacheline.
1656 */
1657DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1658	&init_task;
1659EXPORT_PER_CPU_SYMBOL(current_task);
1660
1661DEFINE_PER_CPU(struct irq_stack *, hardirq_stack_ptr);
 
 
1662DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1663
1664DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1665EXPORT_PER_CPU_SYMBOL(__preempt_count);
1666
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1667/* May not be marked __init: used by software suspend */
1668void syscall_init(void)
1669{
1670	wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1671	wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1672
1673#ifdef CONFIG_IA32_EMULATION
1674	wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1675	/*
1676	 * This only works on Intel CPUs.
1677	 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1678	 * This does not cause SYSENTER to jump to the wrong location, because
1679	 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1680	 */
1681	wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1682	wrmsrl_safe(MSR_IA32_SYSENTER_ESP,
1683		    (unsigned long)(cpu_entry_stack(smp_processor_id()) + 1));
1684	wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1685#else
1686	wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1687	wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1688	wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1689	wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1690#endif
1691
1692	/* Flags to clear on syscall */
1693	wrmsrl(MSR_SYSCALL_MASK,
1694	       X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1695	       X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1696}
1697
 
 
 
 
 
 
 
1698DEFINE_PER_CPU(int, debug_stack_usage);
 
 
 
 
 
 
 
 
 
1699DEFINE_PER_CPU(u32, debug_idt_ctr);
1700
1701void debug_stack_set_zero(void)
1702{
1703	this_cpu_inc(debug_idt_ctr);
1704	load_current_idt();
1705}
1706NOKPROBE_SYMBOL(debug_stack_set_zero);
1707
1708void debug_stack_reset(void)
1709{
1710	if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1711		return;
1712	if (this_cpu_dec_return(debug_idt_ctr) == 0)
1713		load_current_idt();
1714}
1715NOKPROBE_SYMBOL(debug_stack_reset);
1716
1717#else	/* CONFIG_X86_64 */
1718
1719DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1720EXPORT_PER_CPU_SYMBOL(current_task);
1721DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1722EXPORT_PER_CPU_SYMBOL(__preempt_count);
1723
1724/*
1725 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1726 * the top of the kernel stack.  Use an extra percpu variable to track the
1727 * top of the kernel stack directly.
1728 */
1729DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1730	(unsigned long)&init_thread_union + THREAD_SIZE;
1731EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1732
1733#ifdef CONFIG_STACKPROTECTOR
1734DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1735#endif
1736
1737#endif	/* CONFIG_X86_64 */
1738
1739/*
1740 * Clear all 6 debug registers:
1741 */
1742static void clear_all_debug_regs(void)
1743{
1744	int i;
1745
1746	for (i = 0; i < 8; i++) {
1747		/* Ignore db4, db5 */
1748		if ((i == 4) || (i == 5))
1749			continue;
1750
1751		set_debugreg(0, i);
1752	}
1753}
1754
1755#ifdef CONFIG_KGDB
1756/*
1757 * Restore debug regs if using kgdbwait and you have a kernel debugger
1758 * connection established.
1759 */
1760static void dbg_restore_debug_regs(void)
1761{
1762	if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1763		arch_kgdb_ops.correct_hw_break();
1764}
1765#else /* ! CONFIG_KGDB */
1766#define dbg_restore_debug_regs()
1767#endif /* ! CONFIG_KGDB */
1768
1769static void wait_for_master_cpu(int cpu)
1770{
1771#ifdef CONFIG_SMP
1772	/*
1773	 * wait for ACK from master CPU before continuing
1774	 * with AP initialization
1775	 */
1776	WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1777	while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1778		cpu_relax();
1779#endif
1780}
1781
1782#ifdef CONFIG_X86_64
1783static void setup_getcpu(int cpu)
1784{
1785	unsigned long cpudata = vdso_encode_cpunode(cpu, early_cpu_to_node(cpu));
1786	struct desc_struct d = { };
1787
1788	if (boot_cpu_has(X86_FEATURE_RDTSCP))
1789		write_rdtscp_aux(cpudata);
1790
1791	/* Store CPU and node number in limit. */
1792	d.limit0 = cpudata;
1793	d.limit1 = cpudata >> 16;
1794
1795	d.type = 5;		/* RO data, expand down, accessed */
1796	d.dpl = 3;		/* Visible to user code */
1797	d.s = 1;		/* Not a system segment */
1798	d.p = 1;		/* Present */
1799	d.d = 1;		/* 32-bit */
1800
1801	write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_CPUNODE, &d, DESCTYPE_S);
1802}
1803#endif
1804
1805/*
1806 * cpu_init() initializes state that is per-CPU. Some data is already
1807 * initialized (naturally) in the bootstrap process, such as the GDT
1808 * and IDT. We reload them nevertheless, this function acts as a
1809 * 'CPU state barrier', nothing should get across.
 
1810 */
1811#ifdef CONFIG_X86_64
1812
1813void cpu_init(void)
1814{
1815	int cpu = raw_smp_processor_id();
1816	struct task_struct *me;
1817	struct tss_struct *t;
 
 
1818	int i;
1819
1820	wait_for_master_cpu(cpu);
1821
 
 
 
 
 
 
1822	if (cpu)
1823		load_ucode_ap();
1824
1825	t = &per_cpu(cpu_tss_rw, cpu);
 
1826
1827#ifdef CONFIG_NUMA
1828	if (this_cpu_read(numa_node) == 0 &&
1829	    early_cpu_to_node(cpu) != NUMA_NO_NODE)
1830		set_numa_node(early_cpu_to_node(cpu));
1831#endif
1832	setup_getcpu(cpu);
1833
1834	me = current;
1835
1836	pr_debug("Initializing CPU#%d\n", cpu);
1837
1838	cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1839
1840	/*
1841	 * Initialize the per-CPU GDT with the boot GDT,
1842	 * and set up the GDT descriptor:
1843	 */
1844
1845	switch_to_new_gdt(cpu);
1846	loadsegment(fs, 0);
1847
1848	load_current_idt();
1849
1850	memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1851	syscall_init();
1852
1853	wrmsrl(MSR_FS_BASE, 0);
1854	wrmsrl(MSR_KERNEL_GS_BASE, 0);
1855	barrier();
1856
1857	x86_configure_nx();
1858	x2apic_setup();
1859
1860	/*
1861	 * set up and load the per-CPU TSS
1862	 */
1863	if (!t->x86_tss.ist[0]) {
1864		t->x86_tss.ist[IST_INDEX_DF] = __this_cpu_ist_top_va(DF);
1865		t->x86_tss.ist[IST_INDEX_NMI] = __this_cpu_ist_top_va(NMI);
1866		t->x86_tss.ist[IST_INDEX_DB] = __this_cpu_ist_top_va(DB);
1867		t->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
 
 
 
 
 
1868	}
1869
1870	t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1871
1872	/*
1873	 * <= is required because the CPU will access up to
1874	 * 8 bits beyond the end of the IO permission bitmap.
1875	 */
1876	for (i = 0; i <= IO_BITMAP_LONGS; i++)
1877		t->io_bitmap[i] = ~0UL;
1878
1879	mmgrab(&init_mm);
1880	me->active_mm = &init_mm;
1881	BUG_ON(me->mm);
1882	initialize_tlbstate_and_flush();
1883	enter_lazy_tlb(&init_mm, me);
1884
1885	/*
1886	 * Initialize the TSS.  sp0 points to the entry trampoline stack
1887	 * regardless of what task is running.
1888	 */
1889	set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1890	load_TR_desc();
1891	load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1892
1893	load_mm_ldt(&init_mm);
1894
1895	clear_all_debug_regs();
1896	dbg_restore_debug_regs();
1897
1898	fpu__init_cpu();
1899
1900	if (is_uv_system())
1901		uv_cpu_init();
1902
1903	load_fixmap_gdt(cpu);
1904}
1905
1906#else
1907
1908void cpu_init(void)
1909{
1910	int cpu = smp_processor_id();
1911	struct task_struct *curr = current;
1912	struct tss_struct *t = &per_cpu(cpu_tss_rw, cpu);
 
1913
1914	wait_for_master_cpu(cpu);
1915
 
 
 
 
 
 
1916	show_ucode_info_early();
1917
1918	pr_info("Initializing CPU#%d\n", cpu);
1919
1920	if (cpu_feature_enabled(X86_FEATURE_VME) ||
1921	    boot_cpu_has(X86_FEATURE_TSC) ||
1922	    boot_cpu_has(X86_FEATURE_DE))
1923		cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1924
1925	load_current_idt();
1926	switch_to_new_gdt(cpu);
1927
1928	/*
1929	 * Set up and load the per-CPU TSS and LDT
1930	 */
1931	mmgrab(&init_mm);
1932	curr->active_mm = &init_mm;
1933	BUG_ON(curr->mm);
1934	initialize_tlbstate_and_flush();
1935	enter_lazy_tlb(&init_mm, curr);
1936
1937	/*
1938	 * Initialize the TSS.  sp0 points to the entry trampoline stack
1939	 * regardless of what task is running.
1940	 */
1941	set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
1942	load_TR_desc();
1943	load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1));
1944
1945	load_mm_ldt(&init_mm);
1946
1947	t->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET;
1948
1949#ifdef CONFIG_DOUBLEFAULT
1950	/* Set up doublefault TSS pointer in the GDT */
1951	__set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1952#endif
1953
1954	clear_all_debug_regs();
1955	dbg_restore_debug_regs();
1956
1957	fpu__init_cpu();
1958
1959	load_fixmap_gdt(cpu);
1960}
1961#endif
1962
1963/*
1964 * The microcode loader calls this upon late microcode load to recheck features,
1965 * only when microcode has been updated. Caller holds microcode_mutex and CPU
1966 * hotplug lock.
1967 */
1968void microcode_check(void)
1969{
1970	struct cpuinfo_x86 info;
1971
1972	perf_check_microcode();
1973
1974	/* Reload CPUID max function as it might've changed. */
1975	info.cpuid_level = cpuid_eax(0);
1976
1977	/*
1978	 * Copy all capability leafs to pick up the synthetic ones so that
1979	 * memcmp() below doesn't fail on that. The ones coming from CPUID will
1980	 * get overwritten in get_cpu_cap().
1981	 */
1982	memcpy(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability));
1983
1984	get_cpu_cap(&info);
1985
1986	if (!memcmp(&info.x86_capability, &boot_cpu_data.x86_capability, sizeof(info.x86_capability)))
1987		return;
1988
1989	pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n");
1990	pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n");
1991}
1992
1993/*
1994 * Invoked from core CPU hotplug code after hotplug operations
1995 */
1996void arch_smt_update(void)
 
1997{
1998	/* Handle the speculative execution misfeatures */
1999	cpu_bugs_smt_update();
2000	/* Check whether IPI broadcasting can be enabled */
2001	apic_smt_update();
2002}
v4.10.11
   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() */
  55cpumask_var_t cpu_initialized_mask;
  56cpumask_var_t cpu_callout_mask;
  57cpumask_var_t cpu_callin_mask;
  58
  59/* representing cpus for which sibling maps can be computed */
  60cpumask_var_t cpu_sibling_setup_mask;
  61
 
 
 
 
 
 
 
  62/* correctly size the local cpu masks */
  63void __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
  71static 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
  88static const struct cpu_dev default_cpu = {
  89	.c_init		= default_init,
  90	.c_vendor	= "Unknown",
  91	.c_x86_vendor	= X86_VENDOR_UNKNOWN,
  92};
  93
  94static const struct cpu_dev *this_cpu = &default_cpu;
  95
  96DEFINE_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} };
 148EXPORT_PER_CPU_SYMBOL_GPL(gdt_page);
 149
 150static 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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 166static 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}
 180early_param("noinvpcid", x86_noinvpcid_setup);
 181
 182#ifdef CONFIG_X86_32
 183static int cachesize_override = -1;
 184static int disable_x86_serial_nr = 1;
 185
 186static int __init cachesize_setup(char *str)
 187{
 188	get_option(&str, &cachesize_override);
 189	return 1;
 190}
 191__setup("cachesize=", cachesize_setup);
 192
 193static 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 */
 201static 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 */
 230int have_cpuid_p(void)
 231{
 232	return flag_is_changeable_p(X86_EFLAGS_ID);
 233}
 234
 235static 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
 255static 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
 262static inline int flag_is_changeable_p(u32 flag)
 263{
 264	return 1;
 265}
 266static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c)
 267{
 268}
 269#endif
 270
 271static __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
 280static __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
 286static __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
 293static __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 */
 312static bool pku_disabled;
 313
 314static __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
 335static __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 */
 360struct cpuid_dependent_feature {
 361	u32 feature;
 362	u32 level;
 363};
 364
 365static const struct cpuid_dependent_feature
 366cpuid_dependent_features[] = {
 367	{ X86_FEATURE_MWAIT,		0x00000005 },
 368	{ X86_FEATURE_DCA,		0x00000009 },
 369	{ X86_FEATURE_XSAVE,		0x0000000d },
 370	{ 0, 0 }
 371};
 372
 373static 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. */
 410static 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
 435void 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 */
 450void 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
 462static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {};
 463
 464static 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
 
 
 
 
 
 
 
 
 
 
 
 
 
 495void 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
 534u16 __read_mostly tlb_lli_4k[NR_INFO];
 535u16 __read_mostly tlb_lli_2m[NR_INFO];
 536u16 __read_mostly tlb_lli_4m[NR_INFO];
 537u16 __read_mostly tlb_lld_4k[NR_INFO];
 538u16 __read_mostly tlb_lld_2m[NR_INFO];
 539u16 __read_mostly tlb_lld_4m[NR_INFO];
 540u16 __read_mostly tlb_lld_1g[NR_INFO];
 541
 542static 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
 556void 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
 596out:
 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
 607static 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
 633void 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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 658void 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	/* Thermal and Power Management Leaf: level 0x00000006 (eax) */
 671	if (c->cpuid_level >= 0x00000006)
 672		c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006);
 673
 674	/* Additional Intel-defined flags: level 0x00000007 */
 675	if (c->cpuid_level >= 0x00000007) {
 676		cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx);
 677		c->x86_capability[CPUID_7_0_EBX] = ebx;
 678		c->x86_capability[CPUID_7_ECX] = ecx;
 
 
 
 
 
 
 
 679	}
 680
 681	/* Extended state features: level 0x0000000d */
 682	if (c->cpuid_level >= 0x0000000d) {
 683		cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
 684
 685		c->x86_capability[CPUID_D_1_EAX] = eax;
 686	}
 687
 688	/* Additional Intel-defined flags: level 0x0000000F */
 689	if (c->cpuid_level >= 0x0000000F) {
 690
 691		/* QoS sub-leaf, EAX=0Fh, ECX=0 */
 692		cpuid_count(0x0000000F, 0, &eax, &ebx, &ecx, &edx);
 693		c->x86_capability[CPUID_F_0_EDX] = edx;
 694
 695		if (cpu_has(c, X86_FEATURE_CQM_LLC)) {
 696			/* will be overridden if occupancy monitoring exists */
 697			c->x86_cache_max_rmid = ebx;
 698
 699			/* QoS sub-leaf, EAX=0Fh, ECX=1 */
 700			cpuid_count(0x0000000F, 1, &eax, &ebx, &ecx, &edx);
 701			c->x86_capability[CPUID_F_1_EDX] = edx;
 702
 703			if ((cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC)) ||
 704			      ((cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL)) ||
 705			       (cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)))) {
 706				c->x86_cache_max_rmid = ecx;
 707				c->x86_cache_occ_scale = ebx;
 708			}
 709		} else {
 710			c->x86_cache_max_rmid = -1;
 711			c->x86_cache_occ_scale = -1;
 712		}
 713	}
 714
 715	/* AMD-defined flags: level 0x80000001 */
 716	eax = cpuid_eax(0x80000000);
 717	c->extended_cpuid_level = eax;
 718
 719	if ((eax & 0xffff0000) == 0x80000000) {
 720		if (eax >= 0x80000001) {
 721			cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
 722
 723			c->x86_capability[CPUID_8000_0001_ECX] = ecx;
 724			c->x86_capability[CPUID_8000_0001_EDX] = edx;
 725		}
 726	}
 727
 728	if (c->extended_cpuid_level >= 0x80000007) {
 729		cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
 730
 731		c->x86_capability[CPUID_8000_0007_EBX] = ebx;
 732		c->x86_power = edx;
 733	}
 734
 735	if (c->extended_cpuid_level >= 0x80000008) {
 736		cpuid(0x80000008, &eax, &ebx, &ecx, &edx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 737
 738		c->x86_virt_bits = (eax >> 8) & 0xff;
 739		c->x86_phys_bits = eax & 0xff;
 740		c->x86_capability[CPUID_8000_0008_EBX] = ebx;
 741	}
 742#ifdef CONFIG_X86_32
 743	else if (cpu_has(c, X86_FEATURE_PAE) || cpu_has(c, X86_FEATURE_PSE36))
 744		c->x86_phys_bits = 36;
 745#endif
 746
 747	if (c->extended_cpuid_level >= 0x8000000a)
 748		c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a);
 749
 750	init_scattered_cpuid_features(c);
 751}
 752
 753static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c)
 754{
 755#ifdef CONFIG_X86_32
 756	int i;
 757
 758	/*
 759	 * First of all, decide if this is a 486 or higher
 760	 * It's a 486 if we can modify the AC flag
 761	 */
 762	if (flag_is_changeable_p(X86_EFLAGS_AC))
 763		c->x86 = 4;
 764	else
 765		c->x86 = 3;
 766
 767	for (i = 0; i < X86_VENDOR_NUM; i++)
 768		if (cpu_devs[i] && cpu_devs[i]->c_identify) {
 769			c->x86_vendor_id[0] = 0;
 770			cpu_devs[i]->c_identify(c);
 771			if (c->x86_vendor_id[0]) {
 772				get_cpu_vendor(c);
 773				break;
 774			}
 775		}
 776#endif
 777}
 778
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 779/*
 780 * Do minimum CPU detection early.
 781 * Fields really needed: vendor, cpuid_level, family, model, mask,
 782 * cache alignment.
 783 * The others are not touched to avoid unwanted side effects.
 784 *
 785 * WARNING: this function is only called on the BP.  Don't add code here
 786 * that is supposed to run on all CPUs.
 787 */
 788static void __init early_identify_cpu(struct cpuinfo_x86 *c)
 789{
 790#ifdef CONFIG_X86_64
 791	c->x86_clflush_size = 64;
 792	c->x86_phys_bits = 36;
 793	c->x86_virt_bits = 48;
 794#else
 795	c->x86_clflush_size = 32;
 796	c->x86_phys_bits = 32;
 797	c->x86_virt_bits = 32;
 798#endif
 799	c->x86_cache_alignment = c->x86_clflush_size;
 800
 801	memset(&c->x86_capability, 0, sizeof c->x86_capability);
 802	c->extended_cpuid_level = 0;
 803
 804	if (!have_cpuid_p())
 805		identify_cpu_without_cpuid(c);
 806
 807	/* cyrix could have cpuid enabled via c_identify()*/
 808	if (have_cpuid_p()) {
 809		cpu_detect(c);
 810		get_cpu_vendor(c);
 811		get_cpu_cap(c);
 
 
 812
 813		if (this_cpu->c_early_init)
 814			this_cpu->c_early_init(c);
 815
 816		c->cpu_index = 0;
 817		filter_cpuid_features(c, false);
 818
 819		if (this_cpu->c_bsp_init)
 820			this_cpu->c_bsp_init(c);
 
 
 821	}
 822
 823	setup_force_cpu_cap(X86_FEATURE_ALWAYS);
 
 
 
 824	fpu__init_system(c);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 825}
 826
 827void __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 */
 869static 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
 878static 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
 907static 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
 968static 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 * Validate that ACPI/mptables have the same information about the
 984 * effective APIC id and update the package map.
 985 */
 986static void validate_apic_and_package_id(struct cpuinfo_x86 *c)
 987{
 988#ifdef CONFIG_SMP
 989	unsigned int apicid, cpu = smp_processor_id();
 990
 991	apicid = apic->cpu_present_to_apicid(cpu);
 992
 993	if (apicid != c->apicid) {
 994		pr_err(FW_BUG "CPU%u: APIC id mismatch. Firmware: %x APIC: %x\n",
 995		       cpu, apicid, c->initial_apicid);
 996	}
 997	BUG_ON(topology_update_package_map(c->phys_proc_id, cpu));
 
 998#else
 999	c->logical_proc_id = 0;
1000#endif
1001}
1002
1003/*
1004 * This does the hard work of actually picking apart the CPU stuff...
1005 */
1006static void identify_cpu(struct cpuinfo_x86 *c)
1007{
1008	int i;
1009
1010	c->loops_per_jiffy = loops_per_jiffy;
1011	c->x86_cache_size = -1;
1012	c->x86_vendor = X86_VENDOR_UNKNOWN;
1013	c->x86_model = c->x86_mask = 0;	/* So far unknown... */
1014	c->x86_vendor_id[0] = '\0'; /* Unset */
1015	c->x86_model_id[0] = '\0';  /* Unset */
1016	c->x86_max_cores = 1;
1017	c->x86_coreid_bits = 0;
1018	c->cu_id = 0xff;
1019#ifdef CONFIG_X86_64
1020	c->x86_clflush_size = 64;
1021	c->x86_phys_bits = 36;
1022	c->x86_virt_bits = 48;
1023#else
1024	c->cpuid_level = -1;	/* CPUID not detected */
1025	c->x86_clflush_size = 32;
1026	c->x86_phys_bits = 32;
1027	c->x86_virt_bits = 32;
1028#endif
1029	c->x86_cache_alignment = c->x86_clflush_size;
1030	memset(&c->x86_capability, 0, sizeof c->x86_capability);
1031
1032	generic_identify(c);
1033
1034	if (this_cpu->c_identify)
1035		this_cpu->c_identify(c);
1036
1037	/* Clear/Set all flags overridden by options, after probe */
1038	for (i = 0; i < NCAPINTS; i++) {
1039		c->x86_capability[i] &= ~cpu_caps_cleared[i];
1040		c->x86_capability[i] |= cpu_caps_set[i];
1041	}
1042
1043#ifdef CONFIG_X86_64
1044	c->apicid = apic->phys_pkg_id(c->initial_apicid, 0);
1045#endif
1046
1047	/*
1048	 * Vendor-specific initialization.  In this section we
1049	 * canonicalize the feature flags, meaning if there are
1050	 * features a certain CPU supports which CPUID doesn't
1051	 * tell us, CPUID claiming incorrect flags, or other bugs,
1052	 * we handle them here.
1053	 *
1054	 * At the end of this section, c->x86_capability better
1055	 * indicate the features this CPU genuinely supports!
1056	 */
1057	if (this_cpu->c_init)
1058		this_cpu->c_init(c);
1059
1060	/* Disable the PN if appropriate */
1061	squash_the_stupid_serial_number(c);
1062
1063	/* Set up SMEP/SMAP */
1064	setup_smep(c);
1065	setup_smap(c);
 
1066
1067	/*
1068	 * The vendor-specific functions might have changed features.
1069	 * Now we do "generic changes."
1070	 */
1071
1072	/* Filter out anything that depends on CPUID levels we don't have */
1073	filter_cpuid_features(c, true);
1074
1075	/* If the model name is still unset, do table lookup. */
1076	if (!c->x86_model_id[0]) {
1077		const char *p;
1078		p = table_lookup_model(c);
1079		if (p)
1080			strcpy(c->x86_model_id, p);
1081		else
1082			/* Last resort... */
1083			sprintf(c->x86_model_id, "%02x/%02x",
1084				c->x86, c->x86_model);
1085	}
1086
1087#ifdef CONFIG_X86_64
1088	detect_ht(c);
1089#endif
1090
1091	init_hypervisor(c);
1092	x86_init_rdrand(c);
1093	x86_init_cache_qos(c);
1094	setup_pku(c);
1095
1096	/*
1097	 * Clear/Set all flags overridden by options, need do it
1098	 * before following smp all cpus cap AND.
1099	 */
1100	for (i = 0; i < NCAPINTS; i++) {
1101		c->x86_capability[i] &= ~cpu_caps_cleared[i];
1102		c->x86_capability[i] |= cpu_caps_set[i];
1103	}
1104
1105	/*
1106	 * On SMP, boot_cpu_data holds the common feature set between
1107	 * all CPUs; so make sure that we indicate which features are
1108	 * common between the CPUs.  The first time this routine gets
1109	 * executed, c == &boot_cpu_data.
1110	 */
1111	if (c != &boot_cpu_data) {
1112		/* AND the already accumulated flags with these */
1113		for (i = 0; i < NCAPINTS; i++)
1114			boot_cpu_data.x86_capability[i] &= c->x86_capability[i];
1115
1116		/* OR, i.e. replicate the bug flags */
1117		for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++)
1118			c->x86_capability[i] |= boot_cpu_data.x86_capability[i];
1119	}
1120
1121	/* Init Machine Check Exception if available. */
1122	mcheck_cpu_init(c);
1123
1124	select_idle_routine(c);
1125
1126#ifdef CONFIG_NUMA
1127	numa_add_cpu(smp_processor_id());
1128#endif
1129}
1130
1131/*
1132 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions
1133 * on 32-bit kernels:
1134 */
1135#ifdef CONFIG_X86_32
1136void enable_sep_cpu(void)
1137{
1138	struct tss_struct *tss;
1139	int cpu;
1140
1141	if (!boot_cpu_has(X86_FEATURE_SEP))
1142		return;
1143
1144	cpu = get_cpu();
1145	tss = &per_cpu(cpu_tss, cpu);
1146
1147	/*
1148	 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
1149	 * see the big comment in struct x86_hw_tss's definition.
1150	 */
1151
1152	tss->x86_tss.ss1 = __KERNEL_CS;
1153	wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
1154
1155	wrmsr(MSR_IA32_SYSENTER_ESP,
1156	      (unsigned long)tss + offsetofend(struct tss_struct, SYSENTER_stack),
1157	      0);
1158
1159	wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
1160
1161	put_cpu();
1162}
1163#endif
1164
1165void __init identify_boot_cpu(void)
1166{
1167	identify_cpu(&boot_cpu_data);
1168#ifdef CONFIG_X86_32
1169	sysenter_setup();
1170	enable_sep_cpu();
1171#endif
1172	cpu_detect_tlb(&boot_cpu_data);
 
 
 
1173}
1174
1175void identify_secondary_cpu(struct cpuinfo_x86 *c)
1176{
1177	BUG_ON(c == &boot_cpu_data);
1178	identify_cpu(c);
1179#ifdef CONFIG_X86_32
1180	enable_sep_cpu();
1181#endif
1182	mtrr_ap_init();
1183	validate_apic_and_package_id(c);
 
1184}
1185
1186static __init int setup_noclflush(char *arg)
1187{
1188	setup_clear_cpu_cap(X86_FEATURE_CLFLUSH);
1189	setup_clear_cpu_cap(X86_FEATURE_CLFLUSHOPT);
1190	return 1;
1191}
1192__setup("noclflush", setup_noclflush);
1193
1194void print_cpu_info(struct cpuinfo_x86 *c)
1195{
1196	const char *vendor = NULL;
1197
1198	if (c->x86_vendor < X86_VENDOR_NUM) {
1199		vendor = this_cpu->c_vendor;
1200	} else {
1201		if (c->cpuid_level >= 0)
1202			vendor = c->x86_vendor_id;
1203	}
1204
1205	if (vendor && !strstr(c->x86_model_id, vendor))
1206		pr_cont("%s ", vendor);
1207
1208	if (c->x86_model_id[0])
1209		pr_cont("%s", c->x86_model_id);
1210	else
1211		pr_cont("%d86", c->x86);
1212
1213	pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model);
1214
1215	if (c->x86_mask || c->cpuid_level >= 0)
1216		pr_cont(", stepping: 0x%x)\n", c->x86_mask);
1217	else
1218		pr_cont(")\n");
1219}
1220
1221static __init int setup_disablecpuid(char *arg)
 
 
 
 
 
1222{
1223	int bit;
1224
1225	if (get_option(&arg, &bit) && bit >= 0 && bit < NCAPINTS * 32)
1226		setup_clear_cpu_cap(bit);
1227	else
1228		return 0;
1229
1230	return 1;
1231}
1232__setup("clearcpuid=", setup_disablecpuid);
1233
1234#ifdef CONFIG_X86_64
1235struct desc_ptr idt_descr __ro_after_init = {
1236	.size = NR_VECTORS * 16 - 1,
1237	.address = (unsigned long) idt_table,
1238};
1239const struct desc_ptr debug_idt_descr = {
1240	.size = NR_VECTORS * 16 - 1,
1241	.address = (unsigned long) debug_idt_table,
1242};
1243
1244DEFINE_PER_CPU_FIRST(union irq_stack_union,
1245		     irq_stack_union) __aligned(PAGE_SIZE) __visible;
1246
1247/*
1248 * The following percpu variables are hot.  Align current_task to
1249 * cacheline size such that they fall in the same cacheline.
1250 */
1251DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned =
1252	&init_task;
1253EXPORT_PER_CPU_SYMBOL(current_task);
1254
1255DEFINE_PER_CPU(char *, irq_stack_ptr) =
1256	init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE;
1257
1258DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
1259
1260DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1261EXPORT_PER_CPU_SYMBOL(__preempt_count);
1262
1263/*
1264 * Special IST stacks which the CPU switches to when it calls
1265 * an IST-marked descriptor entry. Up to 7 stacks (hardware
1266 * limit), all of them are 4K, except the debug stack which
1267 * is 8K.
1268 */
1269static const unsigned int exception_stack_sizes[N_EXCEPTION_STACKS] = {
1270	  [0 ... N_EXCEPTION_STACKS - 1]	= EXCEPTION_STKSZ,
1271	  [DEBUG_STACK - 1]			= DEBUG_STKSZ
1272};
1273
1274static DEFINE_PER_CPU_PAGE_ALIGNED(char, exception_stacks
1275	[(N_EXCEPTION_STACKS - 1) * EXCEPTION_STKSZ + DEBUG_STKSZ]);
1276
1277/* May not be marked __init: used by software suspend */
1278void syscall_init(void)
1279{
1280	wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
1281	wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
1282
1283#ifdef CONFIG_IA32_EMULATION
1284	wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
1285	/*
1286	 * This only works on Intel CPUs.
1287	 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
1288	 * This does not cause SYSENTER to jump to the wrong location, because
1289	 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
1290	 */
1291	wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
1292	wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
 
1293	wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
1294#else
1295	wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
1296	wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
1297	wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
1298	wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
1299#endif
1300
1301	/* Flags to clear on syscall */
1302	wrmsrl(MSR_SYSCALL_MASK,
1303	       X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
1304	       X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
1305}
1306
1307/*
1308 * Copies of the original ist values from the tss are only accessed during
1309 * debugging, no special alignment required.
1310 */
1311DEFINE_PER_CPU(struct orig_ist, orig_ist);
1312
1313static DEFINE_PER_CPU(unsigned long, debug_stack_addr);
1314DEFINE_PER_CPU(int, debug_stack_usage);
1315
1316int is_debug_stack(unsigned long addr)
1317{
1318	return __this_cpu_read(debug_stack_usage) ||
1319		(addr <= __this_cpu_read(debug_stack_addr) &&
1320		 addr > (__this_cpu_read(debug_stack_addr) - DEBUG_STKSZ));
1321}
1322NOKPROBE_SYMBOL(is_debug_stack);
1323
1324DEFINE_PER_CPU(u32, debug_idt_ctr);
1325
1326void debug_stack_set_zero(void)
1327{
1328	this_cpu_inc(debug_idt_ctr);
1329	load_current_idt();
1330}
1331NOKPROBE_SYMBOL(debug_stack_set_zero);
1332
1333void debug_stack_reset(void)
1334{
1335	if (WARN_ON(!this_cpu_read(debug_idt_ctr)))
1336		return;
1337	if (this_cpu_dec_return(debug_idt_ctr) == 0)
1338		load_current_idt();
1339}
1340NOKPROBE_SYMBOL(debug_stack_reset);
1341
1342#else	/* CONFIG_X86_64 */
1343
1344DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task;
1345EXPORT_PER_CPU_SYMBOL(current_task);
1346DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
1347EXPORT_PER_CPU_SYMBOL(__preempt_count);
1348
1349/*
1350 * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
1351 * the top of the kernel stack.  Use an extra percpu variable to track the
1352 * top of the kernel stack directly.
1353 */
1354DEFINE_PER_CPU(unsigned long, cpu_current_top_of_stack) =
1355	(unsigned long)&init_thread_union + THREAD_SIZE;
1356EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
1357
1358#ifdef CONFIG_CC_STACKPROTECTOR
1359DEFINE_PER_CPU_ALIGNED(struct stack_canary, stack_canary);
1360#endif
1361
1362#endif	/* CONFIG_X86_64 */
1363
1364/*
1365 * Clear all 6 debug registers:
1366 */
1367static void clear_all_debug_regs(void)
1368{
1369	int i;
1370
1371	for (i = 0; i < 8; i++) {
1372		/* Ignore db4, db5 */
1373		if ((i == 4) || (i == 5))
1374			continue;
1375
1376		set_debugreg(0, i);
1377	}
1378}
1379
1380#ifdef CONFIG_KGDB
1381/*
1382 * Restore debug regs if using kgdbwait and you have a kernel debugger
1383 * connection established.
1384 */
1385static void dbg_restore_debug_regs(void)
1386{
1387	if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break))
1388		arch_kgdb_ops.correct_hw_break();
1389}
1390#else /* ! CONFIG_KGDB */
1391#define dbg_restore_debug_regs()
1392#endif /* ! CONFIG_KGDB */
1393
1394static void wait_for_master_cpu(int cpu)
1395{
1396#ifdef CONFIG_SMP
1397	/*
1398	 * wait for ACK from master CPU before continuing
1399	 * with AP initialization
1400	 */
1401	WARN_ON(cpumask_test_and_set_cpu(cpu, cpu_initialized_mask));
1402	while (!cpumask_test_cpu(cpu, cpu_callout_mask))
1403		cpu_relax();
1404#endif
1405}
1406
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1407/*
1408 * cpu_init() initializes state that is per-CPU. Some data is already
1409 * initialized (naturally) in the bootstrap process, such as the GDT
1410 * and IDT. We reload them nevertheless, this function acts as a
1411 * 'CPU state barrier', nothing should get across.
1412 * A lot of state is already set up in PDA init for 64 bit
1413 */
1414#ifdef CONFIG_X86_64
1415
1416void cpu_init(void)
1417{
1418	struct orig_ist *oist;
1419	struct task_struct *me;
1420	struct tss_struct *t;
1421	unsigned long v;
1422	int cpu = raw_smp_processor_id();
1423	int i;
1424
1425	wait_for_master_cpu(cpu);
1426
1427	/*
1428	 * Initialize the CR4 shadow before doing anything that could
1429	 * try to read it.
1430	 */
1431	cr4_init_shadow();
1432
1433	if (cpu)
1434		load_ucode_ap();
1435
1436	t = &per_cpu(cpu_tss, cpu);
1437	oist = &per_cpu(orig_ist, cpu);
1438
1439#ifdef CONFIG_NUMA
1440	if (this_cpu_read(numa_node) == 0 &&
1441	    early_cpu_to_node(cpu) != NUMA_NO_NODE)
1442		set_numa_node(early_cpu_to_node(cpu));
1443#endif
 
1444
1445	me = current;
1446
1447	pr_debug("Initializing CPU#%d\n", cpu);
1448
1449	cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1450
1451	/*
1452	 * Initialize the per-CPU GDT with the boot GDT,
1453	 * and set up the GDT descriptor:
1454	 */
1455
1456	switch_to_new_gdt(cpu);
1457	loadsegment(fs, 0);
1458
1459	load_current_idt();
1460
1461	memset(me->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8);
1462	syscall_init();
1463
1464	wrmsrl(MSR_FS_BASE, 0);
1465	wrmsrl(MSR_KERNEL_GS_BASE, 0);
1466	barrier();
1467
1468	x86_configure_nx();
1469	x2apic_setup();
1470
1471	/*
1472	 * set up and load the per-CPU TSS
1473	 */
1474	if (!oist->ist[0]) {
1475		char *estacks = per_cpu(exception_stacks, cpu);
1476
1477		for (v = 0; v < N_EXCEPTION_STACKS; v++) {
1478			estacks += exception_stack_sizes[v];
1479			oist->ist[v] = t->x86_tss.ist[v] =
1480					(unsigned long)estacks;
1481			if (v == DEBUG_STACK-1)
1482				per_cpu(debug_stack_addr, cpu) = (unsigned long)estacks;
1483		}
1484	}
1485
1486	t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1487
1488	/*
1489	 * <= is required because the CPU will access up to
1490	 * 8 bits beyond the end of the IO permission bitmap.
1491	 */
1492	for (i = 0; i <= IO_BITMAP_LONGS; i++)
1493		t->io_bitmap[i] = ~0UL;
1494
1495	atomic_inc(&init_mm.mm_count);
1496	me->active_mm = &init_mm;
1497	BUG_ON(me->mm);
 
1498	enter_lazy_tlb(&init_mm, me);
1499
1500	load_sp0(t, &current->thread);
1501	set_tss_desc(cpu, t);
 
 
 
1502	load_TR_desc();
 
 
1503	load_mm_ldt(&init_mm);
1504
1505	clear_all_debug_regs();
1506	dbg_restore_debug_regs();
1507
1508	fpu__init_cpu();
1509
1510	if (is_uv_system())
1511		uv_cpu_init();
 
 
1512}
1513
1514#else
1515
1516void cpu_init(void)
1517{
1518	int cpu = smp_processor_id();
1519	struct task_struct *curr = current;
1520	struct tss_struct *t = &per_cpu(cpu_tss, cpu);
1521	struct thread_struct *thread = &curr->thread;
1522
1523	wait_for_master_cpu(cpu);
1524
1525	/*
1526	 * Initialize the CR4 shadow before doing anything that could
1527	 * try to read it.
1528	 */
1529	cr4_init_shadow();
1530
1531	show_ucode_info_early();
1532
1533	pr_info("Initializing CPU#%d\n", cpu);
1534
1535	if (cpu_feature_enabled(X86_FEATURE_VME) ||
1536	    boot_cpu_has(X86_FEATURE_TSC) ||
1537	    boot_cpu_has(X86_FEATURE_DE))
1538		cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE);
1539
1540	load_current_idt();
1541	switch_to_new_gdt(cpu);
1542
1543	/*
1544	 * Set up and load the per-CPU TSS and LDT
1545	 */
1546	atomic_inc(&init_mm.mm_count);
1547	curr->active_mm = &init_mm;
1548	BUG_ON(curr->mm);
 
1549	enter_lazy_tlb(&init_mm, curr);
1550
1551	load_sp0(t, thread);
1552	set_tss_desc(cpu, t);
 
 
 
1553	load_TR_desc();
 
 
1554	load_mm_ldt(&init_mm);
1555
1556	t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
1557
1558#ifdef CONFIG_DOUBLEFAULT
1559	/* Set up doublefault TSS pointer in the GDT */
1560	__set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
1561#endif
1562
1563	clear_all_debug_regs();
1564	dbg_restore_debug_regs();
1565
1566	fpu__init_cpu();
 
 
1567}
1568#endif
1569
1570static void bsp_resume(void)
 
 
 
 
 
1571{
1572	if (this_cpu->c_bsp_resume)
1573		this_cpu->c_bsp_resume(&boot_cpu_data);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1574}
1575
1576static struct syscore_ops cpu_syscore_ops = {
1577	.resume		= bsp_resume,
1578};
1579
1580static int __init init_cpu_syscore(void)
1581{
1582	register_syscore_ops(&cpu_syscore_ops);
1583	return 0;
 
 
1584}
1585core_initcall(init_cpu_syscore);