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/sched/smt.h> 18 #include <linux/init.h> 19 #include <linux/kprobes.h> 20 #include <linux/kgdb.h> 21 #include <linux/mem_encrypt.h> 22 #include <linux/smp.h> 23 #include <linux/cpu.h> 24 #include <linux/io.h> 25 #include <linux/syscore_ops.h> 26 #include <linux/pgtable.h> 27 #include <linux/stackprotector.h> 28 #include <linux/utsname.h> 29 30 #include <asm/alternative.h> 31 #include <asm/cmdline.h> 32 #include <asm/perf_event.h> 33 #include <asm/mmu_context.h> 34 #include <asm/doublefault.h> 35 #include <asm/archrandom.h> 36 #include <asm/hypervisor.h> 37 #include <asm/processor.h> 38 #include <asm/tlbflush.h> 39 #include <asm/debugreg.h> 40 #include <asm/sections.h> 41 #include <asm/vsyscall.h> 42 #include <linux/topology.h> 43 #include <linux/cpumask.h> 44 #include <linux/atomic.h> 45 #include <asm/proto.h> 46 #include <asm/setup.h> 47 #include <asm/apic.h> 48 #include <asm/desc.h> 49 #include <asm/fpu/api.h> 50 #include <asm/mtrr.h> 51 #include <asm/hwcap2.h> 52 #include <linux/numa.h> 53 #include <asm/numa.h> 54 #include <asm/asm.h> 55 #include <asm/bugs.h> 56 #include <asm/cpu.h> 57 #include <asm/mce.h> 58 #include <asm/msr.h> 59 #include <asm/cacheinfo.h> 60 #include <asm/memtype.h> 61 #include <asm/microcode.h> 62 #include <asm/intel-family.h> 63 #include <asm/cpu_device_id.h> 64 #include <asm/fred.h> 65 #include <asm/uv/uv.h> 66 #include <asm/ia32.h> 67 #include <asm/set_memory.h> 68 #include <asm/traps.h> 69 #include <asm/sev.h> 70 #include <asm/tdx.h> 71 72 #include "cpu.h" 73 74 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info); 75 EXPORT_PER_CPU_SYMBOL(cpu_info); 76 77 u32 elf_hwcap2 __read_mostly; 78 79 /* Number of siblings per CPU package */ 80 unsigned int __max_threads_per_core __ro_after_init = 1; 81 EXPORT_SYMBOL(__max_threads_per_core); 82 83 unsigned int __max_dies_per_package __ro_after_init = 1; 84 EXPORT_SYMBOL(__max_dies_per_package); 85 86 unsigned int __max_logical_packages __ro_after_init = 1; 87 EXPORT_SYMBOL(__max_logical_packages); 88 89 unsigned int __num_cores_per_package __ro_after_init = 1; 90 EXPORT_SYMBOL(__num_cores_per_package); 91 92 unsigned int __num_threads_per_package __ro_after_init = 1; 93 EXPORT_SYMBOL(__num_threads_per_package); 94 95 static struct ppin_info { 96 int feature; 97 int msr_ppin_ctl; 98 int msr_ppin; 99 } ppin_info[] = { 100 [X86_VENDOR_INTEL] = { 101 .feature = X86_FEATURE_INTEL_PPIN, 102 .msr_ppin_ctl = MSR_PPIN_CTL, 103 .msr_ppin = MSR_PPIN 104 }, 105 [X86_VENDOR_AMD] = { 106 .feature = X86_FEATURE_AMD_PPIN, 107 .msr_ppin_ctl = MSR_AMD_PPIN_CTL, 108 .msr_ppin = MSR_AMD_PPIN 109 }, 110 }; 111 112 static const struct x86_cpu_id ppin_cpuids[] = { 113 X86_MATCH_FEATURE(X86_FEATURE_AMD_PPIN, &ppin_info[X86_VENDOR_AMD]), 114 X86_MATCH_FEATURE(X86_FEATURE_INTEL_PPIN, &ppin_info[X86_VENDOR_INTEL]), 115 116 /* Legacy models without CPUID enumeration */ 117 X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE_X, &ppin_info[X86_VENDOR_INTEL]), 118 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, &ppin_info[X86_VENDOR_INTEL]), 119 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_D, &ppin_info[X86_VENDOR_INTEL]), 120 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, &ppin_info[X86_VENDOR_INTEL]), 121 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_X, &ppin_info[X86_VENDOR_INTEL]), 122 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_X, &ppin_info[X86_VENDOR_INTEL]), 123 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_D, &ppin_info[X86_VENDOR_INTEL]), 124 X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, &ppin_info[X86_VENDOR_INTEL]), 125 X86_MATCH_INTEL_FAM6_MODEL(EMERALDRAPIDS_X, &ppin_info[X86_VENDOR_INTEL]), 126 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNL, &ppin_info[X86_VENDOR_INTEL]), 127 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNM, &ppin_info[X86_VENDOR_INTEL]), 128 129 {} 130 }; 131 132 static void ppin_init(struct cpuinfo_x86 *c) 133 { 134 const struct x86_cpu_id *id; 135 unsigned long long val; 136 struct ppin_info *info; 137 138 id = x86_match_cpu(ppin_cpuids); 139 if (!id) 140 return; 141 142 /* 143 * Testing the presence of the MSR is not enough. Need to check 144 * that the PPIN_CTL allows reading of the PPIN. 145 */ 146 info = (struct ppin_info *)id->driver_data; 147 148 if (rdmsrl_safe(info->msr_ppin_ctl, &val)) 149 goto clear_ppin; 150 151 if ((val & 3UL) == 1UL) { 152 /* PPIN locked in disabled mode */ 153 goto clear_ppin; 154 } 155 156 /* If PPIN is disabled, try to enable */ 157 if (!(val & 2UL)) { 158 wrmsrl_safe(info->msr_ppin_ctl, val | 2UL); 159 rdmsrl_safe(info->msr_ppin_ctl, &val); 160 } 161 162 /* Is the enable bit set? */ 163 if (val & 2UL) { 164 c->ppin = __rdmsr(info->msr_ppin); 165 set_cpu_cap(c, info->feature); 166 return; 167 } 168 169 clear_ppin: 170 clear_cpu_cap(c, info->feature); 171 } 172 173 static void default_init(struct cpuinfo_x86 *c) 174 { 175 #ifdef CONFIG_X86_64 176 cpu_detect_cache_sizes(c); 177 #else 178 /* Not much we can do here... */ 179 /* Check if at least it has cpuid */ 180 if (c->cpuid_level == -1) { 181 /* No cpuid. It must be an ancient CPU */ 182 if (c->x86 == 4) 183 strcpy(c->x86_model_id, "486"); 184 else if (c->x86 == 3) 185 strcpy(c->x86_model_id, "386"); 186 } 187 #endif 188 } 189 190 static const struct cpu_dev default_cpu = { 191 .c_init = default_init, 192 .c_vendor = "Unknown", 193 .c_x86_vendor = X86_VENDOR_UNKNOWN, 194 }; 195 196 static const struct cpu_dev *this_cpu = &default_cpu; 197 198 DEFINE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page) = { .gdt = { 199 #ifdef CONFIG_X86_64 200 /* 201 * We need valid kernel segments for data and code in long mode too 202 * IRET will check the segment types kkeil 2000/10/28 203 * Also sysret mandates a special GDT layout 204 * 205 * TLS descriptors are currently at a different place compared to i386. 206 * Hopefully nobody expects them at a fixed place (Wine?) 207 */ 208 [GDT_ENTRY_KERNEL32_CS] = GDT_ENTRY_INIT(DESC_CODE32, 0, 0xfffff), 209 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(DESC_CODE64, 0, 0xfffff), 210 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(DESC_DATA64, 0, 0xfffff), 211 [GDT_ENTRY_DEFAULT_USER32_CS] = GDT_ENTRY_INIT(DESC_CODE32 | DESC_USER, 0, 0xfffff), 212 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(DESC_DATA64 | DESC_USER, 0, 0xfffff), 213 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(DESC_CODE64 | DESC_USER, 0, 0xfffff), 214 #else 215 [GDT_ENTRY_KERNEL_CS] = GDT_ENTRY_INIT(DESC_CODE32, 0, 0xfffff), 216 [GDT_ENTRY_KERNEL_DS] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff), 217 [GDT_ENTRY_DEFAULT_USER_CS] = GDT_ENTRY_INIT(DESC_CODE32 | DESC_USER, 0, 0xfffff), 218 [GDT_ENTRY_DEFAULT_USER_DS] = GDT_ENTRY_INIT(DESC_DATA32 | DESC_USER, 0, 0xfffff), 219 /* 220 * Segments used for calling PnP BIOS have byte granularity. 221 * They code segments and data segments have fixed 64k limits, 222 * the transfer segment sizes are set at run time. 223 */ 224 [GDT_ENTRY_PNPBIOS_CS32] = GDT_ENTRY_INIT(DESC_CODE32_BIOS, 0, 0xffff), 225 [GDT_ENTRY_PNPBIOS_CS16] = GDT_ENTRY_INIT(DESC_CODE16, 0, 0xffff), 226 [GDT_ENTRY_PNPBIOS_DS] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0xffff), 227 [GDT_ENTRY_PNPBIOS_TS1] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0), 228 [GDT_ENTRY_PNPBIOS_TS2] = GDT_ENTRY_INIT(DESC_DATA16, 0, 0), 229 /* 230 * The APM segments have byte granularity and their bases 231 * are set at run time. All have 64k limits. 232 */ 233 [GDT_ENTRY_APMBIOS_BASE] = GDT_ENTRY_INIT(DESC_CODE32_BIOS, 0, 0xffff), 234 [GDT_ENTRY_APMBIOS_BASE+1] = GDT_ENTRY_INIT(DESC_CODE16, 0, 0xffff), 235 [GDT_ENTRY_APMBIOS_BASE+2] = GDT_ENTRY_INIT(DESC_DATA32_BIOS, 0, 0xffff), 236 237 [GDT_ENTRY_ESPFIX_SS] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff), 238 [GDT_ENTRY_PERCPU] = GDT_ENTRY_INIT(DESC_DATA32, 0, 0xfffff), 239 #endif 240 } }; 241 EXPORT_PER_CPU_SYMBOL_GPL(gdt_page); 242 243 #ifdef CONFIG_X86_64 244 static int __init x86_nopcid_setup(char *s) 245 { 246 /* nopcid doesn't accept parameters */ 247 if (s) 248 return -EINVAL; 249 250 /* do not emit a message if the feature is not present */ 251 if (!boot_cpu_has(X86_FEATURE_PCID)) 252 return 0; 253 254 setup_clear_cpu_cap(X86_FEATURE_PCID); 255 pr_info("nopcid: PCID feature disabled\n"); 256 return 0; 257 } 258 early_param("nopcid", x86_nopcid_setup); 259 #endif 260 261 static int __init x86_noinvpcid_setup(char *s) 262 { 263 /* noinvpcid doesn't accept parameters */ 264 if (s) 265 return -EINVAL; 266 267 /* do not emit a message if the feature is not present */ 268 if (!boot_cpu_has(X86_FEATURE_INVPCID)) 269 return 0; 270 271 setup_clear_cpu_cap(X86_FEATURE_INVPCID); 272 pr_info("noinvpcid: INVPCID feature disabled\n"); 273 return 0; 274 } 275 early_param("noinvpcid", x86_noinvpcid_setup); 276 277 #ifdef CONFIG_X86_32 278 static int cachesize_override = -1; 279 static int disable_x86_serial_nr = 1; 280 281 static int __init cachesize_setup(char *str) 282 { 283 get_option(&str, &cachesize_override); 284 return 1; 285 } 286 __setup("cachesize=", cachesize_setup); 287 288 /* Standard macro to see if a specific flag is changeable */ 289 static inline int flag_is_changeable_p(u32 flag) 290 { 291 u32 f1, f2; 292 293 /* 294 * Cyrix and IDT cpus allow disabling of CPUID 295 * so the code below may return different results 296 * when it is executed before and after enabling 297 * the CPUID. Add "volatile" to not allow gcc to 298 * optimize the subsequent calls to this function. 299 */ 300 asm volatile ("pushfl \n\t" 301 "pushfl \n\t" 302 "popl %0 \n\t" 303 "movl %0, %1 \n\t" 304 "xorl %2, %0 \n\t" 305 "pushl %0 \n\t" 306 "popfl \n\t" 307 "pushfl \n\t" 308 "popl %0 \n\t" 309 "popfl \n\t" 310 311 : "=&r" (f1), "=&r" (f2) 312 : "ir" (flag)); 313 314 return ((f1^f2) & flag) != 0; 315 } 316 317 /* Probe for the CPUID instruction */ 318 int have_cpuid_p(void) 319 { 320 return flag_is_changeable_p(X86_EFLAGS_ID); 321 } 322 323 static void squash_the_stupid_serial_number(struct cpuinfo_x86 *c) 324 { 325 unsigned long lo, hi; 326 327 if (!cpu_has(c, X86_FEATURE_PN) || !disable_x86_serial_nr) 328 return; 329 330 /* Disable processor serial number: */ 331 332 rdmsr(MSR_IA32_BBL_CR_CTL, lo, hi); 333 lo |= 0x200000; 334 wrmsr(MSR_IA32_BBL_CR_CTL, lo, hi); 335 336 pr_notice("CPU serial number disabled.\n"); 337 clear_cpu_cap(c, X86_FEATURE_PN); 338 339 /* Disabling the serial number may affect the cpuid level */ 340 c->cpuid_level = cpuid_eax(0); 341 } 342 343 static int __init x86_serial_nr_setup(char *s) 344 { 345 disable_x86_serial_nr = 0; 346 return 1; 347 } 348 __setup("serialnumber", x86_serial_nr_setup); 349 #else 350 static inline int flag_is_changeable_p(u32 flag) 351 { 352 return 1; 353 } 354 static inline void squash_the_stupid_serial_number(struct cpuinfo_x86 *c) 355 { 356 } 357 #endif 358 359 static __always_inline void setup_smep(struct cpuinfo_x86 *c) 360 { 361 if (cpu_has(c, X86_FEATURE_SMEP)) 362 cr4_set_bits(X86_CR4_SMEP); 363 } 364 365 static __always_inline void setup_smap(struct cpuinfo_x86 *c) 366 { 367 unsigned long eflags = native_save_fl(); 368 369 /* This should have been cleared long ago */ 370 BUG_ON(eflags & X86_EFLAGS_AC); 371 372 if (cpu_has(c, X86_FEATURE_SMAP)) 373 cr4_set_bits(X86_CR4_SMAP); 374 } 375 376 static __always_inline void setup_umip(struct cpuinfo_x86 *c) 377 { 378 /* Check the boot processor, plus build option for UMIP. */ 379 if (!cpu_feature_enabled(X86_FEATURE_UMIP)) 380 goto out; 381 382 /* Check the current processor's cpuid bits. */ 383 if (!cpu_has(c, X86_FEATURE_UMIP)) 384 goto out; 385 386 cr4_set_bits(X86_CR4_UMIP); 387 388 pr_info_once("x86/cpu: User Mode Instruction Prevention (UMIP) activated\n"); 389 390 return; 391 392 out: 393 /* 394 * Make sure UMIP is disabled in case it was enabled in a 395 * previous boot (e.g., via kexec). 396 */ 397 cr4_clear_bits(X86_CR4_UMIP); 398 } 399 400 /* These bits should not change their value after CPU init is finished. */ 401 static const unsigned long cr4_pinned_mask = X86_CR4_SMEP | X86_CR4_SMAP | X86_CR4_UMIP | 402 X86_CR4_FSGSBASE | X86_CR4_CET | X86_CR4_FRED; 403 static DEFINE_STATIC_KEY_FALSE_RO(cr_pinning); 404 static unsigned long cr4_pinned_bits __ro_after_init; 405 406 void native_write_cr0(unsigned long val) 407 { 408 unsigned long bits_missing = 0; 409 410 set_register: 411 asm volatile("mov %0,%%cr0": "+r" (val) : : "memory"); 412 413 if (static_branch_likely(&cr_pinning)) { 414 if (unlikely((val & X86_CR0_WP) != X86_CR0_WP)) { 415 bits_missing = X86_CR0_WP; 416 val |= bits_missing; 417 goto set_register; 418 } 419 /* Warn after we've set the missing bits. */ 420 WARN_ONCE(bits_missing, "CR0 WP bit went missing!?\n"); 421 } 422 } 423 EXPORT_SYMBOL(native_write_cr0); 424 425 void __no_profile native_write_cr4(unsigned long val) 426 { 427 unsigned long bits_changed = 0; 428 429 set_register: 430 asm volatile("mov %0,%%cr4": "+r" (val) : : "memory"); 431 432 if (static_branch_likely(&cr_pinning)) { 433 if (unlikely((val & cr4_pinned_mask) != cr4_pinned_bits)) { 434 bits_changed = (val & cr4_pinned_mask) ^ cr4_pinned_bits; 435 val = (val & ~cr4_pinned_mask) | cr4_pinned_bits; 436 goto set_register; 437 } 438 /* Warn after we've corrected the changed bits. */ 439 WARN_ONCE(bits_changed, "pinned CR4 bits changed: 0x%lx!?\n", 440 bits_changed); 441 } 442 } 443 #if IS_MODULE(CONFIG_LKDTM) 444 EXPORT_SYMBOL_GPL(native_write_cr4); 445 #endif 446 447 void cr4_update_irqsoff(unsigned long set, unsigned long clear) 448 { 449 unsigned long newval, cr4 = this_cpu_read(cpu_tlbstate.cr4); 450 451 lockdep_assert_irqs_disabled(); 452 453 newval = (cr4 & ~clear) | set; 454 if (newval != cr4) { 455 this_cpu_write(cpu_tlbstate.cr4, newval); 456 __write_cr4(newval); 457 } 458 } 459 EXPORT_SYMBOL(cr4_update_irqsoff); 460 461 /* Read the CR4 shadow. */ 462 unsigned long cr4_read_shadow(void) 463 { 464 return this_cpu_read(cpu_tlbstate.cr4); 465 } 466 EXPORT_SYMBOL_GPL(cr4_read_shadow); 467 468 void cr4_init(void) 469 { 470 unsigned long cr4 = __read_cr4(); 471 472 if (boot_cpu_has(X86_FEATURE_PCID)) 473 cr4 |= X86_CR4_PCIDE; 474 if (static_branch_likely(&cr_pinning)) 475 cr4 = (cr4 & ~cr4_pinned_mask) | cr4_pinned_bits; 476 477 __write_cr4(cr4); 478 479 /* Initialize cr4 shadow for this CPU. */ 480 this_cpu_write(cpu_tlbstate.cr4, cr4); 481 } 482 483 /* 484 * Once CPU feature detection is finished (and boot params have been 485 * parsed), record any of the sensitive CR bits that are set, and 486 * enable CR pinning. 487 */ 488 static void __init setup_cr_pinning(void) 489 { 490 cr4_pinned_bits = this_cpu_read(cpu_tlbstate.cr4) & cr4_pinned_mask; 491 static_key_enable(&cr_pinning.key); 492 } 493 494 static __init int x86_nofsgsbase_setup(char *arg) 495 { 496 /* Require an exact match without trailing characters. */ 497 if (strlen(arg)) 498 return 0; 499 500 /* Do not emit a message if the feature is not present. */ 501 if (!boot_cpu_has(X86_FEATURE_FSGSBASE)) 502 return 1; 503 504 setup_clear_cpu_cap(X86_FEATURE_FSGSBASE); 505 pr_info("FSGSBASE disabled via kernel command line\n"); 506 return 1; 507 } 508 __setup("nofsgsbase", x86_nofsgsbase_setup); 509 510 /* 511 * Protection Keys are not available in 32-bit mode. 512 */ 513 static bool pku_disabled; 514 515 static __always_inline void setup_pku(struct cpuinfo_x86 *c) 516 { 517 if (c == &boot_cpu_data) { 518 if (pku_disabled || !cpu_feature_enabled(X86_FEATURE_PKU)) 519 return; 520 /* 521 * Setting CR4.PKE will cause the X86_FEATURE_OSPKE cpuid 522 * bit to be set. Enforce it. 523 */ 524 setup_force_cpu_cap(X86_FEATURE_OSPKE); 525 526 } else if (!cpu_feature_enabled(X86_FEATURE_OSPKE)) { 527 return; 528 } 529 530 cr4_set_bits(X86_CR4_PKE); 531 /* Load the default PKRU value */ 532 pkru_write_default(); 533 } 534 535 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS 536 static __init int setup_disable_pku(char *arg) 537 { 538 /* 539 * Do not clear the X86_FEATURE_PKU bit. All of the 540 * runtime checks are against OSPKE so clearing the 541 * bit does nothing. 542 * 543 * This way, we will see "pku" in cpuinfo, but not 544 * "ospke", which is exactly what we want. It shows 545 * that the CPU has PKU, but the OS has not enabled it. 546 * This happens to be exactly how a system would look 547 * if we disabled the config option. 548 */ 549 pr_info("x86: 'nopku' specified, disabling Memory Protection Keys\n"); 550 pku_disabled = true; 551 return 1; 552 } 553 __setup("nopku", setup_disable_pku); 554 #endif 555 556 #ifdef CONFIG_X86_KERNEL_IBT 557 558 __noendbr u64 ibt_save(bool disable) 559 { 560 u64 msr = 0; 561 562 if (cpu_feature_enabled(X86_FEATURE_IBT)) { 563 rdmsrl(MSR_IA32_S_CET, msr); 564 if (disable) 565 wrmsrl(MSR_IA32_S_CET, msr & ~CET_ENDBR_EN); 566 } 567 568 return msr; 569 } 570 571 __noendbr void ibt_restore(u64 save) 572 { 573 u64 msr; 574 575 if (cpu_feature_enabled(X86_FEATURE_IBT)) { 576 rdmsrl(MSR_IA32_S_CET, msr); 577 msr &= ~CET_ENDBR_EN; 578 msr |= (save & CET_ENDBR_EN); 579 wrmsrl(MSR_IA32_S_CET, msr); 580 } 581 } 582 583 #endif 584 585 static __always_inline void setup_cet(struct cpuinfo_x86 *c) 586 { 587 bool user_shstk, kernel_ibt; 588 589 if (!IS_ENABLED(CONFIG_X86_CET)) 590 return; 591 592 kernel_ibt = HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT); 593 user_shstk = cpu_feature_enabled(X86_FEATURE_SHSTK) && 594 IS_ENABLED(CONFIG_X86_USER_SHADOW_STACK); 595 596 if (!kernel_ibt && !user_shstk) 597 return; 598 599 if (user_shstk) 600 set_cpu_cap(c, X86_FEATURE_USER_SHSTK); 601 602 if (kernel_ibt) 603 wrmsrl(MSR_IA32_S_CET, CET_ENDBR_EN); 604 else 605 wrmsrl(MSR_IA32_S_CET, 0); 606 607 cr4_set_bits(X86_CR4_CET); 608 609 if (kernel_ibt && ibt_selftest()) { 610 pr_err("IBT selftest: Failed!\n"); 611 wrmsrl(MSR_IA32_S_CET, 0); 612 setup_clear_cpu_cap(X86_FEATURE_IBT); 613 } 614 } 615 616 __noendbr void cet_disable(void) 617 { 618 if (!(cpu_feature_enabled(X86_FEATURE_IBT) || 619 cpu_feature_enabled(X86_FEATURE_SHSTK))) 620 return; 621 622 wrmsrl(MSR_IA32_S_CET, 0); 623 wrmsrl(MSR_IA32_U_CET, 0); 624 } 625 626 /* 627 * Some CPU features depend on higher CPUID levels, which may not always 628 * be available due to CPUID level capping or broken virtualization 629 * software. Add those features to this table to auto-disable them. 630 */ 631 struct cpuid_dependent_feature { 632 u32 feature; 633 u32 level; 634 }; 635 636 static const struct cpuid_dependent_feature 637 cpuid_dependent_features[] = { 638 { X86_FEATURE_MWAIT, 0x00000005 }, 639 { X86_FEATURE_DCA, 0x00000009 }, 640 { X86_FEATURE_XSAVE, 0x0000000d }, 641 { 0, 0 } 642 }; 643 644 static void filter_cpuid_features(struct cpuinfo_x86 *c, bool warn) 645 { 646 const struct cpuid_dependent_feature *df; 647 648 for (df = cpuid_dependent_features; df->feature; df++) { 649 650 if (!cpu_has(c, df->feature)) 651 continue; 652 /* 653 * Note: cpuid_level is set to -1 if unavailable, but 654 * extended_extended_level is set to 0 if unavailable 655 * and the legitimate extended levels are all negative 656 * when signed; hence the weird messing around with 657 * signs here... 658 */ 659 if (!((s32)df->level < 0 ? 660 (u32)df->level > (u32)c->extended_cpuid_level : 661 (s32)df->level > (s32)c->cpuid_level)) 662 continue; 663 664 clear_cpu_cap(c, df->feature); 665 if (!warn) 666 continue; 667 668 pr_warn("CPU: CPU feature " X86_CAP_FMT " disabled, no CPUID level 0x%x\n", 669 x86_cap_flag(df->feature), df->level); 670 } 671 } 672 673 /* 674 * Naming convention should be: <Name> [(<Codename>)] 675 * This table only is used unless init_<vendor>() below doesn't set it; 676 * in particular, if CPUID levels 0x80000002..4 are supported, this 677 * isn't used 678 */ 679 680 /* Look up CPU names by table lookup. */ 681 static const char *table_lookup_model(struct cpuinfo_x86 *c) 682 { 683 #ifdef CONFIG_X86_32 684 const struct legacy_cpu_model_info *info; 685 686 if (c->x86_model >= 16) 687 return NULL; /* Range check */ 688 689 if (!this_cpu) 690 return NULL; 691 692 info = this_cpu->legacy_models; 693 694 while (info->family) { 695 if (info->family == c->x86) 696 return info->model_names[c->x86_model]; 697 info++; 698 } 699 #endif 700 return NULL; /* Not found */ 701 } 702 703 /* Aligned to unsigned long to avoid split lock in atomic bitmap ops */ 704 __u32 cpu_caps_cleared[NCAPINTS + NBUGINTS] __aligned(sizeof(unsigned long)); 705 __u32 cpu_caps_set[NCAPINTS + NBUGINTS] __aligned(sizeof(unsigned long)); 706 707 #ifdef CONFIG_X86_32 708 /* The 32-bit entry code needs to find cpu_entry_area. */ 709 DEFINE_PER_CPU(struct cpu_entry_area *, cpu_entry_area); 710 #endif 711 712 /* Load the original GDT from the per-cpu structure */ 713 void load_direct_gdt(int cpu) 714 { 715 struct desc_ptr gdt_descr; 716 717 gdt_descr.address = (long)get_cpu_gdt_rw(cpu); 718 gdt_descr.size = GDT_SIZE - 1; 719 load_gdt(&gdt_descr); 720 } 721 EXPORT_SYMBOL_GPL(load_direct_gdt); 722 723 /* Load a fixmap remapping of the per-cpu GDT */ 724 void load_fixmap_gdt(int cpu) 725 { 726 struct desc_ptr gdt_descr; 727 728 gdt_descr.address = (long)get_cpu_gdt_ro(cpu); 729 gdt_descr.size = GDT_SIZE - 1; 730 load_gdt(&gdt_descr); 731 } 732 EXPORT_SYMBOL_GPL(load_fixmap_gdt); 733 734 /** 735 * switch_gdt_and_percpu_base - Switch to direct GDT and runtime per CPU base 736 * @cpu: The CPU number for which this is invoked 737 * 738 * Invoked during early boot to switch from early GDT and early per CPU to 739 * the direct GDT and the runtime per CPU area. On 32-bit the percpu base 740 * switch is implicit by loading the direct GDT. On 64bit this requires 741 * to update GSBASE. 742 */ 743 void __init switch_gdt_and_percpu_base(int cpu) 744 { 745 load_direct_gdt(cpu); 746 747 #ifdef CONFIG_X86_64 748 /* 749 * No need to load %gs. It is already correct. 750 * 751 * Writing %gs on 64bit would zero GSBASE which would make any per 752 * CPU operation up to the point of the wrmsrl() fault. 753 * 754 * Set GSBASE to the new offset. Until the wrmsrl() happens the 755 * early mapping is still valid. That means the GSBASE update will 756 * lose any prior per CPU data which was not copied over in 757 * setup_per_cpu_areas(). 758 * 759 * This works even with stackprotector enabled because the 760 * per CPU stack canary is 0 in both per CPU areas. 761 */ 762 wrmsrl(MSR_GS_BASE, cpu_kernelmode_gs_base(cpu)); 763 #else 764 /* 765 * %fs is already set to __KERNEL_PERCPU, but after switching GDT 766 * it is required to load FS again so that the 'hidden' part is 767 * updated from the new GDT. Up to this point the early per CPU 768 * translation is active. Any content of the early per CPU data 769 * which was not copied over in setup_per_cpu_areas() is lost. 770 */ 771 loadsegment(fs, __KERNEL_PERCPU); 772 #endif 773 } 774 775 static const struct cpu_dev *cpu_devs[X86_VENDOR_NUM] = {}; 776 777 static void get_model_name(struct cpuinfo_x86 *c) 778 { 779 unsigned int *v; 780 char *p, *q, *s; 781 782 if (c->extended_cpuid_level < 0x80000004) 783 return; 784 785 v = (unsigned int *)c->x86_model_id; 786 cpuid(0x80000002, &v[0], &v[1], &v[2], &v[3]); 787 cpuid(0x80000003, &v[4], &v[5], &v[6], &v[7]); 788 cpuid(0x80000004, &v[8], &v[9], &v[10], &v[11]); 789 c->x86_model_id[48] = 0; 790 791 /* Trim whitespace */ 792 p = q = s = &c->x86_model_id[0]; 793 794 while (*p == ' ') 795 p++; 796 797 while (*p) { 798 /* Note the last non-whitespace index */ 799 if (!isspace(*p)) 800 s = q; 801 802 *q++ = *p++; 803 } 804 805 *(s + 1) = '\0'; 806 } 807 808 void cpu_detect_cache_sizes(struct cpuinfo_x86 *c) 809 { 810 unsigned int n, dummy, ebx, ecx, edx, l2size; 811 812 n = c->extended_cpuid_level; 813 814 if (n >= 0x80000005) { 815 cpuid(0x80000005, &dummy, &ebx, &ecx, &edx); 816 c->x86_cache_size = (ecx>>24) + (edx>>24); 817 #ifdef CONFIG_X86_64 818 /* On K8 L1 TLB is inclusive, so don't count it */ 819 c->x86_tlbsize = 0; 820 #endif 821 } 822 823 if (n < 0x80000006) /* Some chips just has a large L1. */ 824 return; 825 826 cpuid(0x80000006, &dummy, &ebx, &ecx, &edx); 827 l2size = ecx >> 16; 828 829 #ifdef CONFIG_X86_64 830 c->x86_tlbsize += ((ebx >> 16) & 0xfff) + (ebx & 0xfff); 831 #else 832 /* do processor-specific cache resizing */ 833 if (this_cpu->legacy_cache_size) 834 l2size = this_cpu->legacy_cache_size(c, l2size); 835 836 /* Allow user to override all this if necessary. */ 837 if (cachesize_override != -1) 838 l2size = cachesize_override; 839 840 if (l2size == 0) 841 return; /* Again, no L2 cache is possible */ 842 #endif 843 844 c->x86_cache_size = l2size; 845 } 846 847 u16 __read_mostly tlb_lli_4k[NR_INFO]; 848 u16 __read_mostly tlb_lli_2m[NR_INFO]; 849 u16 __read_mostly tlb_lli_4m[NR_INFO]; 850 u16 __read_mostly tlb_lld_4k[NR_INFO]; 851 u16 __read_mostly tlb_lld_2m[NR_INFO]; 852 u16 __read_mostly tlb_lld_4m[NR_INFO]; 853 u16 __read_mostly tlb_lld_1g[NR_INFO]; 854 855 static void cpu_detect_tlb(struct cpuinfo_x86 *c) 856 { 857 if (this_cpu->c_detect_tlb) 858 this_cpu->c_detect_tlb(c); 859 860 pr_info("Last level iTLB entries: 4KB %d, 2MB %d, 4MB %d\n", 861 tlb_lli_4k[ENTRIES], tlb_lli_2m[ENTRIES], 862 tlb_lli_4m[ENTRIES]); 863 864 pr_info("Last level dTLB entries: 4KB %d, 2MB %d, 4MB %d, 1GB %d\n", 865 tlb_lld_4k[ENTRIES], tlb_lld_2m[ENTRIES], 866 tlb_lld_4m[ENTRIES], tlb_lld_1g[ENTRIES]); 867 } 868 869 static void get_cpu_vendor(struct cpuinfo_x86 *c) 870 { 871 char *v = c->x86_vendor_id; 872 int i; 873 874 for (i = 0; i < X86_VENDOR_NUM; i++) { 875 if (!cpu_devs[i]) 876 break; 877 878 if (!strcmp(v, cpu_devs[i]->c_ident[0]) || 879 (cpu_devs[i]->c_ident[1] && 880 !strcmp(v, cpu_devs[i]->c_ident[1]))) { 881 882 this_cpu = cpu_devs[i]; 883 c->x86_vendor = this_cpu->c_x86_vendor; 884 return; 885 } 886 } 887 888 pr_err_once("CPU: vendor_id '%s' unknown, using generic init.\n" \ 889 "CPU: Your system may be unstable.\n", v); 890 891 c->x86_vendor = X86_VENDOR_UNKNOWN; 892 this_cpu = &default_cpu; 893 } 894 895 void cpu_detect(struct cpuinfo_x86 *c) 896 { 897 /* Get vendor name */ 898 cpuid(0x00000000, (unsigned int *)&c->cpuid_level, 899 (unsigned int *)&c->x86_vendor_id[0], 900 (unsigned int *)&c->x86_vendor_id[8], 901 (unsigned int *)&c->x86_vendor_id[4]); 902 903 c->x86 = 4; 904 /* Intel-defined flags: level 0x00000001 */ 905 if (c->cpuid_level >= 0x00000001) { 906 u32 junk, tfms, cap0, misc; 907 908 cpuid(0x00000001, &tfms, &misc, &junk, &cap0); 909 c->x86 = x86_family(tfms); 910 c->x86_model = x86_model(tfms); 911 c->x86_stepping = x86_stepping(tfms); 912 913 if (cap0 & (1<<19)) { 914 c->x86_clflush_size = ((misc >> 8) & 0xff) * 8; 915 c->x86_cache_alignment = c->x86_clflush_size; 916 } 917 } 918 } 919 920 static void apply_forced_caps(struct cpuinfo_x86 *c) 921 { 922 int i; 923 924 for (i = 0; i < NCAPINTS + NBUGINTS; i++) { 925 c->x86_capability[i] &= ~cpu_caps_cleared[i]; 926 c->x86_capability[i] |= cpu_caps_set[i]; 927 } 928 } 929 930 static void init_speculation_control(struct cpuinfo_x86 *c) 931 { 932 /* 933 * The Intel SPEC_CTRL CPUID bit implies IBRS and IBPB support, 934 * and they also have a different bit for STIBP support. Also, 935 * a hypervisor might have set the individual AMD bits even on 936 * Intel CPUs, for finer-grained selection of what's available. 937 */ 938 if (cpu_has(c, X86_FEATURE_SPEC_CTRL)) { 939 set_cpu_cap(c, X86_FEATURE_IBRS); 940 set_cpu_cap(c, X86_FEATURE_IBPB); 941 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL); 942 } 943 944 if (cpu_has(c, X86_FEATURE_INTEL_STIBP)) 945 set_cpu_cap(c, X86_FEATURE_STIBP); 946 947 if (cpu_has(c, X86_FEATURE_SPEC_CTRL_SSBD) || 948 cpu_has(c, X86_FEATURE_VIRT_SSBD)) 949 set_cpu_cap(c, X86_FEATURE_SSBD); 950 951 if (cpu_has(c, X86_FEATURE_AMD_IBRS)) { 952 set_cpu_cap(c, X86_FEATURE_IBRS); 953 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL); 954 } 955 956 if (cpu_has(c, X86_FEATURE_AMD_IBPB)) 957 set_cpu_cap(c, X86_FEATURE_IBPB); 958 959 if (cpu_has(c, X86_FEATURE_AMD_STIBP)) { 960 set_cpu_cap(c, X86_FEATURE_STIBP); 961 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL); 962 } 963 964 if (cpu_has(c, X86_FEATURE_AMD_SSBD)) { 965 set_cpu_cap(c, X86_FEATURE_SSBD); 966 set_cpu_cap(c, X86_FEATURE_MSR_SPEC_CTRL); 967 clear_cpu_cap(c, X86_FEATURE_VIRT_SSBD); 968 } 969 } 970 971 void get_cpu_cap(struct cpuinfo_x86 *c) 972 { 973 u32 eax, ebx, ecx, edx; 974 975 /* Intel-defined flags: level 0x00000001 */ 976 if (c->cpuid_level >= 0x00000001) { 977 cpuid(0x00000001, &eax, &ebx, &ecx, &edx); 978 979 c->x86_capability[CPUID_1_ECX] = ecx; 980 c->x86_capability[CPUID_1_EDX] = edx; 981 } 982 983 /* Thermal and Power Management Leaf: level 0x00000006 (eax) */ 984 if (c->cpuid_level >= 0x00000006) 985 c->x86_capability[CPUID_6_EAX] = cpuid_eax(0x00000006); 986 987 /* Additional Intel-defined flags: level 0x00000007 */ 988 if (c->cpuid_level >= 0x00000007) { 989 cpuid_count(0x00000007, 0, &eax, &ebx, &ecx, &edx); 990 c->x86_capability[CPUID_7_0_EBX] = ebx; 991 c->x86_capability[CPUID_7_ECX] = ecx; 992 c->x86_capability[CPUID_7_EDX] = edx; 993 994 /* Check valid sub-leaf index before accessing it */ 995 if (eax >= 1) { 996 cpuid_count(0x00000007, 1, &eax, &ebx, &ecx, &edx); 997 c->x86_capability[CPUID_7_1_EAX] = eax; 998 } 999 } 1000 1001 /* Extended state features: level 0x0000000d */ 1002 if (c->cpuid_level >= 0x0000000d) { 1003 cpuid_count(0x0000000d, 1, &eax, &ebx, &ecx, &edx); 1004 1005 c->x86_capability[CPUID_D_1_EAX] = eax; 1006 } 1007 1008 /* AMD-defined flags: level 0x80000001 */ 1009 eax = cpuid_eax(0x80000000); 1010 c->extended_cpuid_level = eax; 1011 1012 if ((eax & 0xffff0000) == 0x80000000) { 1013 if (eax >= 0x80000001) { 1014 cpuid(0x80000001, &eax, &ebx, &ecx, &edx); 1015 1016 c->x86_capability[CPUID_8000_0001_ECX] = ecx; 1017 c->x86_capability[CPUID_8000_0001_EDX] = edx; 1018 } 1019 } 1020 1021 if (c->extended_cpuid_level >= 0x80000007) { 1022 cpuid(0x80000007, &eax, &ebx, &ecx, &edx); 1023 1024 c->x86_capability[CPUID_8000_0007_EBX] = ebx; 1025 c->x86_power = edx; 1026 } 1027 1028 if (c->extended_cpuid_level >= 0x80000008) { 1029 cpuid(0x80000008, &eax, &ebx, &ecx, &edx); 1030 c->x86_capability[CPUID_8000_0008_EBX] = ebx; 1031 } 1032 1033 if (c->extended_cpuid_level >= 0x8000000a) 1034 c->x86_capability[CPUID_8000_000A_EDX] = cpuid_edx(0x8000000a); 1035 1036 if (c->extended_cpuid_level >= 0x8000001f) 1037 c->x86_capability[CPUID_8000_001F_EAX] = cpuid_eax(0x8000001f); 1038 1039 if (c->extended_cpuid_level >= 0x80000021) 1040 c->x86_capability[CPUID_8000_0021_EAX] = cpuid_eax(0x80000021); 1041 1042 init_scattered_cpuid_features(c); 1043 init_speculation_control(c); 1044 1045 /* 1046 * Clear/Set all flags overridden by options, after probe. 1047 * This needs to happen each time we re-probe, which may happen 1048 * several times during CPU initialization. 1049 */ 1050 apply_forced_caps(c); 1051 } 1052 1053 void get_cpu_address_sizes(struct cpuinfo_x86 *c) 1054 { 1055 u32 eax, ebx, ecx, edx; 1056 bool vp_bits_from_cpuid = true; 1057 1058 if (!cpu_has(c, X86_FEATURE_CPUID) || 1059 (c->extended_cpuid_level < 0x80000008)) 1060 vp_bits_from_cpuid = false; 1061 1062 if (vp_bits_from_cpuid) { 1063 cpuid(0x80000008, &eax, &ebx, &ecx, &edx); 1064 1065 c->x86_virt_bits = (eax >> 8) & 0xff; 1066 c->x86_phys_bits = eax & 0xff; 1067 } else { 1068 if (IS_ENABLED(CONFIG_X86_64)) { 1069 c->x86_clflush_size = 64; 1070 c->x86_phys_bits = 36; 1071 c->x86_virt_bits = 48; 1072 } else { 1073 c->x86_clflush_size = 32; 1074 c->x86_virt_bits = 32; 1075 c->x86_phys_bits = 32; 1076 1077 if (cpu_has(c, X86_FEATURE_PAE) || 1078 cpu_has(c, X86_FEATURE_PSE36)) 1079 c->x86_phys_bits = 36; 1080 } 1081 } 1082 c->x86_cache_bits = c->x86_phys_bits; 1083 c->x86_cache_alignment = c->x86_clflush_size; 1084 } 1085 1086 static void identify_cpu_without_cpuid(struct cpuinfo_x86 *c) 1087 { 1088 #ifdef CONFIG_X86_32 1089 int i; 1090 1091 /* 1092 * First of all, decide if this is a 486 or higher 1093 * It's a 486 if we can modify the AC flag 1094 */ 1095 if (flag_is_changeable_p(X86_EFLAGS_AC)) 1096 c->x86 = 4; 1097 else 1098 c->x86 = 3; 1099 1100 for (i = 0; i < X86_VENDOR_NUM; i++) 1101 if (cpu_devs[i] && cpu_devs[i]->c_identify) { 1102 c->x86_vendor_id[0] = 0; 1103 cpu_devs[i]->c_identify(c); 1104 if (c->x86_vendor_id[0]) { 1105 get_cpu_vendor(c); 1106 break; 1107 } 1108 } 1109 #endif 1110 } 1111 1112 #define NO_SPECULATION BIT(0) 1113 #define NO_MELTDOWN BIT(1) 1114 #define NO_SSB BIT(2) 1115 #define NO_L1TF BIT(3) 1116 #define NO_MDS BIT(4) 1117 #define MSBDS_ONLY BIT(5) 1118 #define NO_SWAPGS BIT(6) 1119 #define NO_ITLB_MULTIHIT BIT(7) 1120 #define NO_SPECTRE_V2 BIT(8) 1121 #define NO_MMIO BIT(9) 1122 #define NO_EIBRS_PBRSB BIT(10) 1123 #define NO_BHI BIT(11) 1124 1125 #define VULNWL(vendor, family, model, whitelist) \ 1126 X86_MATCH_VENDOR_FAM_MODEL(vendor, family, model, whitelist) 1127 1128 #define VULNWL_INTEL(model, whitelist) \ 1129 VULNWL(INTEL, 6, INTEL_FAM6_##model, whitelist) 1130 1131 #define VULNWL_AMD(family, whitelist) \ 1132 VULNWL(AMD, family, X86_MODEL_ANY, whitelist) 1133 1134 #define VULNWL_HYGON(family, whitelist) \ 1135 VULNWL(HYGON, family, X86_MODEL_ANY, whitelist) 1136 1137 static const __initconst struct x86_cpu_id cpu_vuln_whitelist[] = { 1138 VULNWL(ANY, 4, X86_MODEL_ANY, NO_SPECULATION), 1139 VULNWL(CENTAUR, 5, X86_MODEL_ANY, NO_SPECULATION), 1140 VULNWL(INTEL, 5, X86_MODEL_ANY, NO_SPECULATION), 1141 VULNWL(NSC, 5, X86_MODEL_ANY, NO_SPECULATION), 1142 VULNWL(VORTEX, 5, X86_MODEL_ANY, NO_SPECULATION), 1143 VULNWL(VORTEX, 6, X86_MODEL_ANY, NO_SPECULATION), 1144 1145 /* Intel Family 6 */ 1146 VULNWL_INTEL(TIGERLAKE, NO_MMIO), 1147 VULNWL_INTEL(TIGERLAKE_L, NO_MMIO), 1148 VULNWL_INTEL(ALDERLAKE, NO_MMIO), 1149 VULNWL_INTEL(ALDERLAKE_L, NO_MMIO), 1150 1151 VULNWL_INTEL(ATOM_SALTWELL, NO_SPECULATION | NO_ITLB_MULTIHIT), 1152 VULNWL_INTEL(ATOM_SALTWELL_TABLET, NO_SPECULATION | NO_ITLB_MULTIHIT), 1153 VULNWL_INTEL(ATOM_SALTWELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT), 1154 VULNWL_INTEL(ATOM_BONNELL, NO_SPECULATION | NO_ITLB_MULTIHIT), 1155 VULNWL_INTEL(ATOM_BONNELL_MID, NO_SPECULATION | NO_ITLB_MULTIHIT), 1156 1157 VULNWL_INTEL(ATOM_SILVERMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT), 1158 VULNWL_INTEL(ATOM_SILVERMONT_D, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT), 1159 VULNWL_INTEL(ATOM_SILVERMONT_MID, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT), 1160 VULNWL_INTEL(ATOM_AIRMONT, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT), 1161 VULNWL_INTEL(XEON_PHI_KNL, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT), 1162 VULNWL_INTEL(XEON_PHI_KNM, NO_SSB | NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT), 1163 1164 VULNWL_INTEL(CORE_YONAH, NO_SSB), 1165 1166 VULNWL_INTEL(ATOM_AIRMONT_MID, NO_L1TF | MSBDS_ONLY | NO_SWAPGS | NO_ITLB_MULTIHIT), 1167 VULNWL_INTEL(ATOM_AIRMONT_NP, NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT), 1168 1169 VULNWL_INTEL(ATOM_GOLDMONT, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO), 1170 VULNWL_INTEL(ATOM_GOLDMONT_D, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO), 1171 VULNWL_INTEL(ATOM_GOLDMONT_PLUS, NO_MDS | NO_L1TF | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB), 1172 1173 /* 1174 * Technically, swapgs isn't serializing on AMD (despite it previously 1175 * being documented as such in the APM). But according to AMD, %gs is 1176 * updated non-speculatively, and the issuing of %gs-relative memory 1177 * operands will be blocked until the %gs update completes, which is 1178 * good enough for our purposes. 1179 */ 1180 1181 VULNWL_INTEL(ATOM_TREMONT, NO_EIBRS_PBRSB), 1182 VULNWL_INTEL(ATOM_TREMONT_L, NO_EIBRS_PBRSB), 1183 VULNWL_INTEL(ATOM_TREMONT_D, NO_ITLB_MULTIHIT | NO_EIBRS_PBRSB), 1184 1185 /* AMD Family 0xf - 0x12 */ 1186 VULNWL_AMD(0x0f, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_BHI), 1187 VULNWL_AMD(0x10, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_BHI), 1188 VULNWL_AMD(0x11, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_BHI), 1189 VULNWL_AMD(0x12, NO_MELTDOWN | NO_SSB | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_BHI), 1190 1191 /* FAMILY_ANY must be last, otherwise 0x0f - 0x12 matches won't work */ 1192 VULNWL_AMD(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB | NO_BHI), 1193 VULNWL_HYGON(X86_FAMILY_ANY, NO_MELTDOWN | NO_L1TF | NO_MDS | NO_SWAPGS | NO_ITLB_MULTIHIT | NO_MMIO | NO_EIBRS_PBRSB | NO_BHI), 1194 1195 /* Zhaoxin Family 7 */ 1196 VULNWL(CENTAUR, 7, X86_MODEL_ANY, NO_SPECTRE_V2 | NO_SWAPGS | NO_MMIO | NO_BHI), 1197 VULNWL(ZHAOXIN, 7, X86_MODEL_ANY, NO_SPECTRE_V2 | NO_SWAPGS | NO_MMIO | NO_BHI), 1198 {} 1199 }; 1200 1201 #define VULNBL(vendor, family, model, blacklist) \ 1202 X86_MATCH_VENDOR_FAM_MODEL(vendor, family, model, blacklist) 1203 1204 #define VULNBL_INTEL_STEPPINGS(model, steppings, issues) \ 1205 X86_MATCH_VENDOR_FAM_MODEL_STEPPINGS_FEATURE(INTEL, 6, \ 1206 INTEL_FAM6_##model, steppings, \ 1207 X86_FEATURE_ANY, issues) 1208 1209 #define VULNBL_AMD(family, blacklist) \ 1210 VULNBL(AMD, family, X86_MODEL_ANY, blacklist) 1211 1212 #define VULNBL_HYGON(family, blacklist) \ 1213 VULNBL(HYGON, family, X86_MODEL_ANY, blacklist) 1214 1215 #define SRBDS BIT(0) 1216 /* CPU is affected by X86_BUG_MMIO_STALE_DATA */ 1217 #define MMIO BIT(1) 1218 /* CPU is affected by Shared Buffers Data Sampling (SBDS), a variant of X86_BUG_MMIO_STALE_DATA */ 1219 #define MMIO_SBDS BIT(2) 1220 /* CPU is affected by RETbleed, speculating where you would not expect it */ 1221 #define RETBLEED BIT(3) 1222 /* CPU is affected by SMT (cross-thread) return predictions */ 1223 #define SMT_RSB BIT(4) 1224 /* CPU is affected by SRSO */ 1225 #define SRSO BIT(5) 1226 /* CPU is affected by GDS */ 1227 #define GDS BIT(6) 1228 /* CPU is affected by Register File Data Sampling */ 1229 #define RFDS BIT(7) 1230 1231 static const struct x86_cpu_id cpu_vuln_blacklist[] __initconst = { 1232 VULNBL_INTEL_STEPPINGS(IVYBRIDGE, X86_STEPPING_ANY, SRBDS), 1233 VULNBL_INTEL_STEPPINGS(HASWELL, X86_STEPPING_ANY, SRBDS), 1234 VULNBL_INTEL_STEPPINGS(HASWELL_L, X86_STEPPING_ANY, SRBDS), 1235 VULNBL_INTEL_STEPPINGS(HASWELL_G, X86_STEPPING_ANY, SRBDS), 1236 VULNBL_INTEL_STEPPINGS(HASWELL_X, X86_STEPPING_ANY, MMIO), 1237 VULNBL_INTEL_STEPPINGS(BROADWELL_D, X86_STEPPING_ANY, MMIO), 1238 VULNBL_INTEL_STEPPINGS(BROADWELL_G, X86_STEPPING_ANY, SRBDS), 1239 VULNBL_INTEL_STEPPINGS(BROADWELL_X, X86_STEPPING_ANY, MMIO), 1240 VULNBL_INTEL_STEPPINGS(BROADWELL, X86_STEPPING_ANY, SRBDS), 1241 VULNBL_INTEL_STEPPINGS(SKYLAKE_X, X86_STEPPING_ANY, MMIO | RETBLEED | GDS), 1242 VULNBL_INTEL_STEPPINGS(SKYLAKE_L, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS), 1243 VULNBL_INTEL_STEPPINGS(SKYLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS), 1244 VULNBL_INTEL_STEPPINGS(KABYLAKE_L, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS), 1245 VULNBL_INTEL_STEPPINGS(KABYLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS | SRBDS), 1246 VULNBL_INTEL_STEPPINGS(CANNONLAKE_L, X86_STEPPING_ANY, RETBLEED), 1247 VULNBL_INTEL_STEPPINGS(ICELAKE_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS), 1248 VULNBL_INTEL_STEPPINGS(ICELAKE_D, X86_STEPPING_ANY, MMIO | GDS), 1249 VULNBL_INTEL_STEPPINGS(ICELAKE_X, X86_STEPPING_ANY, MMIO | GDS), 1250 VULNBL_INTEL_STEPPINGS(COMETLAKE, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS), 1251 VULNBL_INTEL_STEPPINGS(COMETLAKE_L, X86_STEPPINGS(0x0, 0x0), MMIO | RETBLEED), 1252 VULNBL_INTEL_STEPPINGS(COMETLAKE_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED | GDS), 1253 VULNBL_INTEL_STEPPINGS(TIGERLAKE_L, X86_STEPPING_ANY, GDS), 1254 VULNBL_INTEL_STEPPINGS(TIGERLAKE, X86_STEPPING_ANY, GDS), 1255 VULNBL_INTEL_STEPPINGS(LAKEFIELD, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RETBLEED), 1256 VULNBL_INTEL_STEPPINGS(ROCKETLAKE, X86_STEPPING_ANY, MMIO | RETBLEED | GDS), 1257 VULNBL_INTEL_STEPPINGS(ALDERLAKE, X86_STEPPING_ANY, RFDS), 1258 VULNBL_INTEL_STEPPINGS(ALDERLAKE_L, X86_STEPPING_ANY, RFDS), 1259 VULNBL_INTEL_STEPPINGS(RAPTORLAKE, X86_STEPPING_ANY, RFDS), 1260 VULNBL_INTEL_STEPPINGS(RAPTORLAKE_P, X86_STEPPING_ANY, RFDS), 1261 VULNBL_INTEL_STEPPINGS(RAPTORLAKE_S, X86_STEPPING_ANY, RFDS), 1262 VULNBL_INTEL_STEPPINGS(ATOM_GRACEMONT, X86_STEPPING_ANY, RFDS), 1263 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RFDS), 1264 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT_D, X86_STEPPING_ANY, MMIO | RFDS), 1265 VULNBL_INTEL_STEPPINGS(ATOM_TREMONT_L, X86_STEPPING_ANY, MMIO | MMIO_SBDS | RFDS), 1266 VULNBL_INTEL_STEPPINGS(ATOM_GOLDMONT, X86_STEPPING_ANY, RFDS), 1267 VULNBL_INTEL_STEPPINGS(ATOM_GOLDMONT_D, X86_STEPPING_ANY, RFDS), 1268 VULNBL_INTEL_STEPPINGS(ATOM_GOLDMONT_PLUS, X86_STEPPING_ANY, RFDS), 1269 1270 VULNBL_AMD(0x15, RETBLEED), 1271 VULNBL_AMD(0x16, RETBLEED), 1272 VULNBL_AMD(0x17, RETBLEED | SMT_RSB | SRSO), 1273 VULNBL_HYGON(0x18, RETBLEED | SMT_RSB | SRSO), 1274 VULNBL_AMD(0x19, SRSO), 1275 {} 1276 }; 1277 1278 static bool __init cpu_matches(const struct x86_cpu_id *table, unsigned long which) 1279 { 1280 const struct x86_cpu_id *m = x86_match_cpu(table); 1281 1282 return m && !!(m->driver_data & which); 1283 } 1284 1285 u64 x86_read_arch_cap_msr(void) 1286 { 1287 u64 x86_arch_cap_msr = 0; 1288 1289 if (boot_cpu_has(X86_FEATURE_ARCH_CAPABILITIES)) 1290 rdmsrl(MSR_IA32_ARCH_CAPABILITIES, x86_arch_cap_msr); 1291 1292 return x86_arch_cap_msr; 1293 } 1294 1295 static bool arch_cap_mmio_immune(u64 x86_arch_cap_msr) 1296 { 1297 return (x86_arch_cap_msr & ARCH_CAP_FBSDP_NO && 1298 x86_arch_cap_msr & ARCH_CAP_PSDP_NO && 1299 x86_arch_cap_msr & ARCH_CAP_SBDR_SSDP_NO); 1300 } 1301 1302 static bool __init vulnerable_to_rfds(u64 x86_arch_cap_msr) 1303 { 1304 /* The "immunity" bit trumps everything else: */ 1305 if (x86_arch_cap_msr & ARCH_CAP_RFDS_NO) 1306 return false; 1307 1308 /* 1309 * VMMs set ARCH_CAP_RFDS_CLEAR for processors not in the blacklist to 1310 * indicate that mitigation is needed because guest is running on a 1311 * vulnerable hardware or may migrate to such hardware: 1312 */ 1313 if (x86_arch_cap_msr & ARCH_CAP_RFDS_CLEAR) 1314 return true; 1315 1316 /* Only consult the blacklist when there is no enumeration: */ 1317 return cpu_matches(cpu_vuln_blacklist, RFDS); 1318 } 1319 1320 static void __init cpu_set_bug_bits(struct cpuinfo_x86 *c) 1321 { 1322 u64 x86_arch_cap_msr = x86_read_arch_cap_msr(); 1323 1324 /* Set ITLB_MULTIHIT bug if cpu is not in the whitelist and not mitigated */ 1325 if (!cpu_matches(cpu_vuln_whitelist, NO_ITLB_MULTIHIT) && 1326 !(x86_arch_cap_msr & ARCH_CAP_PSCHANGE_MC_NO)) 1327 setup_force_cpu_bug(X86_BUG_ITLB_MULTIHIT); 1328 1329 if (cpu_matches(cpu_vuln_whitelist, NO_SPECULATION)) 1330 return; 1331 1332 setup_force_cpu_bug(X86_BUG_SPECTRE_V1); 1333 1334 if (!cpu_matches(cpu_vuln_whitelist, NO_SPECTRE_V2)) 1335 setup_force_cpu_bug(X86_BUG_SPECTRE_V2); 1336 1337 if (!cpu_matches(cpu_vuln_whitelist, NO_SSB) && 1338 !(x86_arch_cap_msr & ARCH_CAP_SSB_NO) && 1339 !cpu_has(c, X86_FEATURE_AMD_SSB_NO)) 1340 setup_force_cpu_bug(X86_BUG_SPEC_STORE_BYPASS); 1341 1342 /* 1343 * AMD's AutoIBRS is equivalent to Intel's eIBRS - use the Intel feature 1344 * flag and protect from vendor-specific bugs via the whitelist. 1345 * 1346 * Don't use AutoIBRS when SNP is enabled because it degrades host 1347 * userspace indirect branch performance. 1348 */ 1349 if ((x86_arch_cap_msr & ARCH_CAP_IBRS_ALL) || 1350 (cpu_has(c, X86_FEATURE_AUTOIBRS) && 1351 !cpu_feature_enabled(X86_FEATURE_SEV_SNP))) { 1352 setup_force_cpu_cap(X86_FEATURE_IBRS_ENHANCED); 1353 if (!cpu_matches(cpu_vuln_whitelist, NO_EIBRS_PBRSB) && 1354 !(x86_arch_cap_msr & ARCH_CAP_PBRSB_NO)) 1355 setup_force_cpu_bug(X86_BUG_EIBRS_PBRSB); 1356 } 1357 1358 if (!cpu_matches(cpu_vuln_whitelist, NO_MDS) && 1359 !(x86_arch_cap_msr & ARCH_CAP_MDS_NO)) { 1360 setup_force_cpu_bug(X86_BUG_MDS); 1361 if (cpu_matches(cpu_vuln_whitelist, MSBDS_ONLY)) 1362 setup_force_cpu_bug(X86_BUG_MSBDS_ONLY); 1363 } 1364 1365 if (!cpu_matches(cpu_vuln_whitelist, NO_SWAPGS)) 1366 setup_force_cpu_bug(X86_BUG_SWAPGS); 1367 1368 /* 1369 * When the CPU is not mitigated for TAA (TAA_NO=0) set TAA bug when: 1370 * - TSX is supported or 1371 * - TSX_CTRL is present 1372 * 1373 * TSX_CTRL check is needed for cases when TSX could be disabled before 1374 * the kernel boot e.g. kexec. 1375 * TSX_CTRL check alone is not sufficient for cases when the microcode 1376 * update is not present or running as guest that don't get TSX_CTRL. 1377 */ 1378 if (!(x86_arch_cap_msr & ARCH_CAP_TAA_NO) && 1379 (cpu_has(c, X86_FEATURE_RTM) || 1380 (x86_arch_cap_msr & ARCH_CAP_TSX_CTRL_MSR))) 1381 setup_force_cpu_bug(X86_BUG_TAA); 1382 1383 /* 1384 * SRBDS affects CPUs which support RDRAND or RDSEED and are listed 1385 * in the vulnerability blacklist. 1386 * 1387 * Some of the implications and mitigation of Shared Buffers Data 1388 * Sampling (SBDS) are similar to SRBDS. Give SBDS same treatment as 1389 * SRBDS. 1390 */ 1391 if ((cpu_has(c, X86_FEATURE_RDRAND) || 1392 cpu_has(c, X86_FEATURE_RDSEED)) && 1393 cpu_matches(cpu_vuln_blacklist, SRBDS | MMIO_SBDS)) 1394 setup_force_cpu_bug(X86_BUG_SRBDS); 1395 1396 /* 1397 * Processor MMIO Stale Data bug enumeration 1398 * 1399 * Affected CPU list is generally enough to enumerate the vulnerability, 1400 * but for virtualization case check for ARCH_CAP MSR bits also, VMM may 1401 * not want the guest to enumerate the bug. 1402 * 1403 * Set X86_BUG_MMIO_UNKNOWN for CPUs that are neither in the blacklist, 1404 * nor in the whitelist and also don't enumerate MSR ARCH_CAP MMIO bits. 1405 */ 1406 if (!arch_cap_mmio_immune(x86_arch_cap_msr)) { 1407 if (cpu_matches(cpu_vuln_blacklist, MMIO)) 1408 setup_force_cpu_bug(X86_BUG_MMIO_STALE_DATA); 1409 else if (!cpu_matches(cpu_vuln_whitelist, NO_MMIO)) 1410 setup_force_cpu_bug(X86_BUG_MMIO_UNKNOWN); 1411 } 1412 1413 if (!cpu_has(c, X86_FEATURE_BTC_NO)) { 1414 if (cpu_matches(cpu_vuln_blacklist, RETBLEED) || (x86_arch_cap_msr & ARCH_CAP_RSBA)) 1415 setup_force_cpu_bug(X86_BUG_RETBLEED); 1416 } 1417 1418 if (cpu_matches(cpu_vuln_blacklist, SMT_RSB)) 1419 setup_force_cpu_bug(X86_BUG_SMT_RSB); 1420 1421 if (!cpu_has(c, X86_FEATURE_SRSO_NO)) { 1422 if (cpu_matches(cpu_vuln_blacklist, SRSO)) 1423 setup_force_cpu_bug(X86_BUG_SRSO); 1424 } 1425 1426 /* 1427 * Check if CPU is vulnerable to GDS. If running in a virtual machine on 1428 * an affected processor, the VMM may have disabled the use of GATHER by 1429 * disabling AVX2. The only way to do this in HW is to clear XCR0[2], 1430 * which means that AVX will be disabled. 1431 */ 1432 if (cpu_matches(cpu_vuln_blacklist, GDS) && !(x86_arch_cap_msr & ARCH_CAP_GDS_NO) && 1433 boot_cpu_has(X86_FEATURE_AVX)) 1434 setup_force_cpu_bug(X86_BUG_GDS); 1435 1436 if (vulnerable_to_rfds(x86_arch_cap_msr)) 1437 setup_force_cpu_bug(X86_BUG_RFDS); 1438 1439 /* When virtualized, eIBRS could be hidden, assume vulnerable */ 1440 if (!(x86_arch_cap_msr & ARCH_CAP_BHI_NO) && 1441 !cpu_matches(cpu_vuln_whitelist, NO_BHI) && 1442 (boot_cpu_has(X86_FEATURE_IBRS_ENHANCED) || 1443 boot_cpu_has(X86_FEATURE_HYPERVISOR))) 1444 setup_force_cpu_bug(X86_BUG_BHI); 1445 1446 if (cpu_matches(cpu_vuln_whitelist, NO_MELTDOWN)) 1447 return; 1448 1449 /* Rogue Data Cache Load? No! */ 1450 if (x86_arch_cap_msr & ARCH_CAP_RDCL_NO) 1451 return; 1452 1453 setup_force_cpu_bug(X86_BUG_CPU_MELTDOWN); 1454 1455 if (cpu_matches(cpu_vuln_whitelist, NO_L1TF)) 1456 return; 1457 1458 setup_force_cpu_bug(X86_BUG_L1TF); 1459 } 1460 1461 /* 1462 * The NOPL instruction is supposed to exist on all CPUs of family >= 6; 1463 * unfortunately, that's not true in practice because of early VIA 1464 * chips and (more importantly) broken virtualizers that are not easy 1465 * to detect. In the latter case it doesn't even *fail* reliably, so 1466 * probing for it doesn't even work. Disable it completely on 32-bit 1467 * unless we can find a reliable way to detect all the broken cases. 1468 * Enable it explicitly on 64-bit for non-constant inputs of cpu_has(). 1469 */ 1470 static void detect_nopl(void) 1471 { 1472 #ifdef CONFIG_X86_32 1473 setup_clear_cpu_cap(X86_FEATURE_NOPL); 1474 #else 1475 setup_force_cpu_cap(X86_FEATURE_NOPL); 1476 #endif 1477 } 1478 1479 /* 1480 * We parse cpu parameters early because fpu__init_system() is executed 1481 * before parse_early_param(). 1482 */ 1483 static void __init cpu_parse_early_param(void) 1484 { 1485 char arg[128]; 1486 char *argptr = arg, *opt; 1487 int arglen, taint = 0; 1488 1489 #ifdef CONFIG_X86_32 1490 if (cmdline_find_option_bool(boot_command_line, "no387")) 1491 #ifdef CONFIG_MATH_EMULATION 1492 setup_clear_cpu_cap(X86_FEATURE_FPU); 1493 #else 1494 pr_err("Option 'no387' required CONFIG_MATH_EMULATION enabled.\n"); 1495 #endif 1496 1497 if (cmdline_find_option_bool(boot_command_line, "nofxsr")) 1498 setup_clear_cpu_cap(X86_FEATURE_FXSR); 1499 #endif 1500 1501 if (cmdline_find_option_bool(boot_command_line, "noxsave")) 1502 setup_clear_cpu_cap(X86_FEATURE_XSAVE); 1503 1504 if (cmdline_find_option_bool(boot_command_line, "noxsaveopt")) 1505 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT); 1506 1507 if (cmdline_find_option_bool(boot_command_line, "noxsaves")) 1508 setup_clear_cpu_cap(X86_FEATURE_XSAVES); 1509 1510 if (cmdline_find_option_bool(boot_command_line, "nousershstk")) 1511 setup_clear_cpu_cap(X86_FEATURE_USER_SHSTK); 1512 1513 arglen = cmdline_find_option(boot_command_line, "clearcpuid", arg, sizeof(arg)); 1514 if (arglen <= 0) 1515 return; 1516 1517 pr_info("Clearing CPUID bits:"); 1518 1519 while (argptr) { 1520 bool found __maybe_unused = false; 1521 unsigned int bit; 1522 1523 opt = strsep(&argptr, ","); 1524 1525 /* 1526 * Handle naked numbers first for feature flags which don't 1527 * have names. 1528 */ 1529 if (!kstrtouint(opt, 10, &bit)) { 1530 if (bit < NCAPINTS * 32) { 1531 1532 /* empty-string, i.e., ""-defined feature flags */ 1533 if (!x86_cap_flags[bit]) 1534 pr_cont(" " X86_CAP_FMT_NUM, x86_cap_flag_num(bit)); 1535 else 1536 pr_cont(" " X86_CAP_FMT, x86_cap_flag(bit)); 1537 1538 setup_clear_cpu_cap(bit); 1539 taint++; 1540 } 1541 /* 1542 * The assumption is that there are no feature names with only 1543 * numbers in the name thus go to the next argument. 1544 */ 1545 continue; 1546 } 1547 1548 for (bit = 0; bit < 32 * NCAPINTS; bit++) { 1549 if (!x86_cap_flag(bit)) 1550 continue; 1551 1552 if (strcmp(x86_cap_flag(bit), opt)) 1553 continue; 1554 1555 pr_cont(" %s", opt); 1556 setup_clear_cpu_cap(bit); 1557 taint++; 1558 found = true; 1559 break; 1560 } 1561 1562 if (!found) 1563 pr_cont(" (unknown: %s)", opt); 1564 } 1565 pr_cont("\n"); 1566 1567 if (taint) 1568 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK); 1569 } 1570 1571 /* 1572 * Do minimum CPU detection early. 1573 * Fields really needed: vendor, cpuid_level, family, model, mask, 1574 * cache alignment. 1575 * The others are not touched to avoid unwanted side effects. 1576 * 1577 * WARNING: this function is only called on the boot CPU. Don't add code 1578 * here that is supposed to run on all CPUs. 1579 */ 1580 static void __init early_identify_cpu(struct cpuinfo_x86 *c) 1581 { 1582 memset(&c->x86_capability, 0, sizeof(c->x86_capability)); 1583 c->extended_cpuid_level = 0; 1584 1585 if (!have_cpuid_p()) 1586 identify_cpu_without_cpuid(c); 1587 1588 /* cyrix could have cpuid enabled via c_identify()*/ 1589 if (have_cpuid_p()) { 1590 cpu_detect(c); 1591 get_cpu_vendor(c); 1592 get_cpu_cap(c); 1593 setup_force_cpu_cap(X86_FEATURE_CPUID); 1594 get_cpu_address_sizes(c); 1595 cpu_parse_early_param(); 1596 1597 cpu_init_topology(c); 1598 1599 if (this_cpu->c_early_init) 1600 this_cpu->c_early_init(c); 1601 1602 c->cpu_index = 0; 1603 filter_cpuid_features(c, false); 1604 1605 if (this_cpu->c_bsp_init) 1606 this_cpu->c_bsp_init(c); 1607 } else { 1608 setup_clear_cpu_cap(X86_FEATURE_CPUID); 1609 get_cpu_address_sizes(c); 1610 cpu_init_topology(c); 1611 } 1612 1613 setup_force_cpu_cap(X86_FEATURE_ALWAYS); 1614 1615 cpu_set_bug_bits(c); 1616 1617 sld_setup(c); 1618 1619 #ifdef CONFIG_X86_32 1620 /* 1621 * Regardless of whether PCID is enumerated, the SDM says 1622 * that it can't be enabled in 32-bit mode. 1623 */ 1624 setup_clear_cpu_cap(X86_FEATURE_PCID); 1625 #endif 1626 1627 /* 1628 * Later in the boot process pgtable_l5_enabled() relies on 1629 * cpu_feature_enabled(X86_FEATURE_LA57). If 5-level paging is not 1630 * enabled by this point we need to clear the feature bit to avoid 1631 * false-positives at the later stage. 1632 * 1633 * pgtable_l5_enabled() can be false here for several reasons: 1634 * - 5-level paging is disabled compile-time; 1635 * - it's 32-bit kernel; 1636 * - machine doesn't support 5-level paging; 1637 * - user specified 'no5lvl' in kernel command line. 1638 */ 1639 if (!pgtable_l5_enabled()) 1640 setup_clear_cpu_cap(X86_FEATURE_LA57); 1641 1642 detect_nopl(); 1643 } 1644 1645 void __init early_cpu_init(void) 1646 { 1647 const struct cpu_dev *const *cdev; 1648 int count = 0; 1649 1650 #ifdef CONFIG_PROCESSOR_SELECT 1651 pr_info("KERNEL supported cpus:\n"); 1652 #endif 1653 1654 for (cdev = __x86_cpu_dev_start; cdev < __x86_cpu_dev_end; cdev++) { 1655 const struct cpu_dev *cpudev = *cdev; 1656 1657 if (count >= X86_VENDOR_NUM) 1658 break; 1659 cpu_devs[count] = cpudev; 1660 count++; 1661 1662 #ifdef CONFIG_PROCESSOR_SELECT 1663 { 1664 unsigned int j; 1665 1666 for (j = 0; j < 2; j++) { 1667 if (!cpudev->c_ident[j]) 1668 continue; 1669 pr_info(" %s %s\n", cpudev->c_vendor, 1670 cpudev->c_ident[j]); 1671 } 1672 } 1673 #endif 1674 } 1675 early_identify_cpu(&boot_cpu_data); 1676 } 1677 1678 static bool detect_null_seg_behavior(void) 1679 { 1680 /* 1681 * Empirically, writing zero to a segment selector on AMD does 1682 * not clear the base, whereas writing zero to a segment 1683 * selector on Intel does clear the base. Intel's behavior 1684 * allows slightly faster context switches in the common case 1685 * where GS is unused by the prev and next threads. 1686 * 1687 * Since neither vendor documents this anywhere that I can see, 1688 * detect it directly instead of hard-coding the choice by 1689 * vendor. 1690 * 1691 * I've designated AMD's behavior as the "bug" because it's 1692 * counterintuitive and less friendly. 1693 */ 1694 1695 unsigned long old_base, tmp; 1696 rdmsrl(MSR_FS_BASE, old_base); 1697 wrmsrl(MSR_FS_BASE, 1); 1698 loadsegment(fs, 0); 1699 rdmsrl(MSR_FS_BASE, tmp); 1700 wrmsrl(MSR_FS_BASE, old_base); 1701 return tmp == 0; 1702 } 1703 1704 void check_null_seg_clears_base(struct cpuinfo_x86 *c) 1705 { 1706 /* BUG_NULL_SEG is only relevant with 64bit userspace */ 1707 if (!IS_ENABLED(CONFIG_X86_64)) 1708 return; 1709 1710 if (cpu_has(c, X86_FEATURE_NULL_SEL_CLR_BASE)) 1711 return; 1712 1713 /* 1714 * CPUID bit above wasn't set. If this kernel is still running 1715 * as a HV guest, then the HV has decided not to advertize 1716 * that CPUID bit for whatever reason. For example, one 1717 * member of the migration pool might be vulnerable. Which 1718 * means, the bug is present: set the BUG flag and return. 1719 */ 1720 if (cpu_has(c, X86_FEATURE_HYPERVISOR)) { 1721 set_cpu_bug(c, X86_BUG_NULL_SEG); 1722 return; 1723 } 1724 1725 /* 1726 * Zen2 CPUs also have this behaviour, but no CPUID bit. 1727 * 0x18 is the respective family for Hygon. 1728 */ 1729 if ((c->x86 == 0x17 || c->x86 == 0x18) && 1730 detect_null_seg_behavior()) 1731 return; 1732 1733 /* All the remaining ones are affected */ 1734 set_cpu_bug(c, X86_BUG_NULL_SEG); 1735 } 1736 1737 static void generic_identify(struct cpuinfo_x86 *c) 1738 { 1739 c->extended_cpuid_level = 0; 1740 1741 if (!have_cpuid_p()) 1742 identify_cpu_without_cpuid(c); 1743 1744 /* cyrix could have cpuid enabled via c_identify()*/ 1745 if (!have_cpuid_p()) 1746 return; 1747 1748 cpu_detect(c); 1749 1750 get_cpu_vendor(c); 1751 1752 get_cpu_cap(c); 1753 1754 get_cpu_address_sizes(c); 1755 1756 get_model_name(c); /* Default name */ 1757 1758 /* 1759 * ESPFIX is a strange bug. All real CPUs have it. Paravirt 1760 * systems that run Linux at CPL > 0 may or may not have the 1761 * issue, but, even if they have the issue, there's absolutely 1762 * nothing we can do about it because we can't use the real IRET 1763 * instruction. 1764 * 1765 * NB: For the time being, only 32-bit kernels support 1766 * X86_BUG_ESPFIX as such. 64-bit kernels directly choose 1767 * whether to apply espfix using paravirt hooks. If any 1768 * non-paravirt system ever shows up that does *not* have the 1769 * ESPFIX issue, we can change this. 1770 */ 1771 #ifdef CONFIG_X86_32 1772 set_cpu_bug(c, X86_BUG_ESPFIX); 1773 #endif 1774 } 1775 1776 /* 1777 * This does the hard work of actually picking apart the CPU stuff... 1778 */ 1779 static void identify_cpu(struct cpuinfo_x86 *c) 1780 { 1781 int i; 1782 1783 c->loops_per_jiffy = loops_per_jiffy; 1784 c->x86_cache_size = 0; 1785 c->x86_vendor = X86_VENDOR_UNKNOWN; 1786 c->x86_model = c->x86_stepping = 0; /* So far unknown... */ 1787 c->x86_vendor_id[0] = '\0'; /* Unset */ 1788 c->x86_model_id[0] = '\0'; /* Unset */ 1789 #ifdef CONFIG_X86_64 1790 c->x86_clflush_size = 64; 1791 c->x86_phys_bits = 36; 1792 c->x86_virt_bits = 48; 1793 #else 1794 c->cpuid_level = -1; /* CPUID not detected */ 1795 c->x86_clflush_size = 32; 1796 c->x86_phys_bits = 32; 1797 c->x86_virt_bits = 32; 1798 #endif 1799 c->x86_cache_alignment = c->x86_clflush_size; 1800 memset(&c->x86_capability, 0, sizeof(c->x86_capability)); 1801 #ifdef CONFIG_X86_VMX_FEATURE_NAMES 1802 memset(&c->vmx_capability, 0, sizeof(c->vmx_capability)); 1803 #endif 1804 1805 generic_identify(c); 1806 1807 cpu_parse_topology(c); 1808 1809 if (this_cpu->c_identify) 1810 this_cpu->c_identify(c); 1811 1812 /* Clear/Set all flags overridden by options, after probe */ 1813 apply_forced_caps(c); 1814 1815 /* 1816 * Set default APIC and TSC_DEADLINE MSR fencing flag. AMD and 1817 * Hygon will clear it in ->c_init() below. 1818 */ 1819 set_cpu_cap(c, X86_FEATURE_APIC_MSRS_FENCE); 1820 1821 /* 1822 * Vendor-specific initialization. In this section we 1823 * canonicalize the feature flags, meaning if there are 1824 * features a certain CPU supports which CPUID doesn't 1825 * tell us, CPUID claiming incorrect flags, or other bugs, 1826 * we handle them here. 1827 * 1828 * At the end of this section, c->x86_capability better 1829 * indicate the features this CPU genuinely supports! 1830 */ 1831 if (this_cpu->c_init) 1832 this_cpu->c_init(c); 1833 1834 /* Disable the PN if appropriate */ 1835 squash_the_stupid_serial_number(c); 1836 1837 /* Set up SMEP/SMAP/UMIP */ 1838 setup_smep(c); 1839 setup_smap(c); 1840 setup_umip(c); 1841 1842 /* Enable FSGSBASE instructions if available. */ 1843 if (cpu_has(c, X86_FEATURE_FSGSBASE)) { 1844 cr4_set_bits(X86_CR4_FSGSBASE); 1845 elf_hwcap2 |= HWCAP2_FSGSBASE; 1846 } 1847 1848 /* 1849 * The vendor-specific functions might have changed features. 1850 * Now we do "generic changes." 1851 */ 1852 1853 /* Filter out anything that depends on CPUID levels we don't have */ 1854 filter_cpuid_features(c, true); 1855 1856 /* If the model name is still unset, do table lookup. */ 1857 if (!c->x86_model_id[0]) { 1858 const char *p; 1859 p = table_lookup_model(c); 1860 if (p) 1861 strcpy(c->x86_model_id, p); 1862 else 1863 /* Last resort... */ 1864 sprintf(c->x86_model_id, "%02x/%02x", 1865 c->x86, c->x86_model); 1866 } 1867 1868 x86_init_rdrand(c); 1869 setup_pku(c); 1870 setup_cet(c); 1871 1872 /* 1873 * Clear/Set all flags overridden by options, need do it 1874 * before following smp all cpus cap AND. 1875 */ 1876 apply_forced_caps(c); 1877 1878 /* 1879 * On SMP, boot_cpu_data holds the common feature set between 1880 * all CPUs; so make sure that we indicate which features are 1881 * common between the CPUs. The first time this routine gets 1882 * executed, c == &boot_cpu_data. 1883 */ 1884 if (c != &boot_cpu_data) { 1885 /* AND the already accumulated flags with these */ 1886 for (i = 0; i < NCAPINTS; i++) 1887 boot_cpu_data.x86_capability[i] &= c->x86_capability[i]; 1888 1889 /* OR, i.e. replicate the bug flags */ 1890 for (i = NCAPINTS; i < NCAPINTS + NBUGINTS; i++) 1891 c->x86_capability[i] |= boot_cpu_data.x86_capability[i]; 1892 } 1893 1894 ppin_init(c); 1895 1896 /* Init Machine Check Exception if available. */ 1897 mcheck_cpu_init(c); 1898 1899 #ifdef CONFIG_NUMA 1900 numa_add_cpu(smp_processor_id()); 1901 #endif 1902 } 1903 1904 /* 1905 * Set up the CPU state needed to execute SYSENTER/SYSEXIT instructions 1906 * on 32-bit kernels: 1907 */ 1908 #ifdef CONFIG_X86_32 1909 void enable_sep_cpu(void) 1910 { 1911 struct tss_struct *tss; 1912 int cpu; 1913 1914 if (!boot_cpu_has(X86_FEATURE_SEP)) 1915 return; 1916 1917 cpu = get_cpu(); 1918 tss = &per_cpu(cpu_tss_rw, cpu); 1919 1920 /* 1921 * We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field -- 1922 * see the big comment in struct x86_hw_tss's definition. 1923 */ 1924 1925 tss->x86_tss.ss1 = __KERNEL_CS; 1926 wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0); 1927 wrmsr(MSR_IA32_SYSENTER_ESP, (unsigned long)(cpu_entry_stack(cpu) + 1), 0); 1928 wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0); 1929 1930 put_cpu(); 1931 } 1932 #endif 1933 1934 static __init void identify_boot_cpu(void) 1935 { 1936 identify_cpu(&boot_cpu_data); 1937 if (HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT)) 1938 pr_info("CET detected: Indirect Branch Tracking enabled\n"); 1939 #ifdef CONFIG_X86_32 1940 enable_sep_cpu(); 1941 #endif 1942 cpu_detect_tlb(&boot_cpu_data); 1943 setup_cr_pinning(); 1944 1945 tsx_init(); 1946 tdx_init(); 1947 lkgs_init(); 1948 } 1949 1950 void identify_secondary_cpu(struct cpuinfo_x86 *c) 1951 { 1952 BUG_ON(c == &boot_cpu_data); 1953 identify_cpu(c); 1954 #ifdef CONFIG_X86_32 1955 enable_sep_cpu(); 1956 #endif 1957 x86_spec_ctrl_setup_ap(); 1958 update_srbds_msr(); 1959 if (boot_cpu_has_bug(X86_BUG_GDS)) 1960 update_gds_msr(); 1961 1962 tsx_ap_init(); 1963 } 1964 1965 void print_cpu_info(struct cpuinfo_x86 *c) 1966 { 1967 const char *vendor = NULL; 1968 1969 if (c->x86_vendor < X86_VENDOR_NUM) { 1970 vendor = this_cpu->c_vendor; 1971 } else { 1972 if (c->cpuid_level >= 0) 1973 vendor = c->x86_vendor_id; 1974 } 1975 1976 if (vendor && !strstr(c->x86_model_id, vendor)) 1977 pr_cont("%s ", vendor); 1978 1979 if (c->x86_model_id[0]) 1980 pr_cont("%s", c->x86_model_id); 1981 else 1982 pr_cont("%d86", c->x86); 1983 1984 pr_cont(" (family: 0x%x, model: 0x%x", c->x86, c->x86_model); 1985 1986 if (c->x86_stepping || c->cpuid_level >= 0) 1987 pr_cont(", stepping: 0x%x)\n", c->x86_stepping); 1988 else 1989 pr_cont(")\n"); 1990 } 1991 1992 /* 1993 * clearcpuid= was already parsed in cpu_parse_early_param(). This dummy 1994 * function prevents it from becoming an environment variable for init. 1995 */ 1996 static __init int setup_clearcpuid(char *arg) 1997 { 1998 return 1; 1999 } 2000 __setup("clearcpuid=", setup_clearcpuid); 2001 2002 DEFINE_PER_CPU_ALIGNED(struct pcpu_hot, pcpu_hot) = { 2003 .current_task = &init_task, 2004 .preempt_count = INIT_PREEMPT_COUNT, 2005 .top_of_stack = TOP_OF_INIT_STACK, 2006 }; 2007 EXPORT_PER_CPU_SYMBOL(pcpu_hot); 2008 EXPORT_PER_CPU_SYMBOL(const_pcpu_hot); 2009 2010 #ifdef CONFIG_X86_64 2011 DEFINE_PER_CPU_FIRST(struct fixed_percpu_data, 2012 fixed_percpu_data) __aligned(PAGE_SIZE) __visible; 2013 EXPORT_PER_CPU_SYMBOL_GPL(fixed_percpu_data); 2014 2015 static void wrmsrl_cstar(unsigned long val) 2016 { 2017 /* 2018 * Intel CPUs do not support 32-bit SYSCALL. Writing to MSR_CSTAR 2019 * is so far ignored by the CPU, but raises a #VE trap in a TDX 2020 * guest. Avoid the pointless write on all Intel CPUs. 2021 */ 2022 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) 2023 wrmsrl(MSR_CSTAR, val); 2024 } 2025 2026 static inline void idt_syscall_init(void) 2027 { 2028 wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64); 2029 2030 if (ia32_enabled()) { 2031 wrmsrl_cstar((unsigned long)entry_SYSCALL_compat); 2032 /* 2033 * This only works on Intel CPUs. 2034 * On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP. 2035 * This does not cause SYSENTER to jump to the wrong location, because 2036 * AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit). 2037 */ 2038 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS); 2039 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 2040 (unsigned long)(cpu_entry_stack(smp_processor_id()) + 1)); 2041 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat); 2042 } else { 2043 wrmsrl_cstar((unsigned long)entry_SYSCALL32_ignore); 2044 wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG); 2045 wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL); 2046 wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL); 2047 } 2048 2049 /* 2050 * Flags to clear on syscall; clear as much as possible 2051 * to minimize user space-kernel interference. 2052 */ 2053 wrmsrl(MSR_SYSCALL_MASK, 2054 X86_EFLAGS_CF|X86_EFLAGS_PF|X86_EFLAGS_AF| 2055 X86_EFLAGS_ZF|X86_EFLAGS_SF|X86_EFLAGS_TF| 2056 X86_EFLAGS_IF|X86_EFLAGS_DF|X86_EFLAGS_OF| 2057 X86_EFLAGS_IOPL|X86_EFLAGS_NT|X86_EFLAGS_RF| 2058 X86_EFLAGS_AC|X86_EFLAGS_ID); 2059 } 2060 2061 /* May not be marked __init: used by software suspend */ 2062 void syscall_init(void) 2063 { 2064 /* The default user and kernel segments */ 2065 wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS); 2066 2067 /* 2068 * Except the IA32_STAR MSR, there is NO need to setup SYSCALL and 2069 * SYSENTER MSRs for FRED, because FRED uses the ring 3 FRED 2070 * entrypoint for SYSCALL and SYSENTER, and ERETU is the only legit 2071 * instruction to return to ring 3 (both sysexit and sysret cause 2072 * #UD when FRED is enabled). 2073 */ 2074 if (!cpu_feature_enabled(X86_FEATURE_FRED)) 2075 idt_syscall_init(); 2076 } 2077 2078 #else /* CONFIG_X86_64 */ 2079 2080 #ifdef CONFIG_STACKPROTECTOR 2081 DEFINE_PER_CPU(unsigned long, __stack_chk_guard); 2082 EXPORT_PER_CPU_SYMBOL(__stack_chk_guard); 2083 #endif 2084 2085 #endif /* CONFIG_X86_64 */ 2086 2087 /* 2088 * Clear all 6 debug registers: 2089 */ 2090 static void clear_all_debug_regs(void) 2091 { 2092 int i; 2093 2094 for (i = 0; i < 8; i++) { 2095 /* Ignore db4, db5 */ 2096 if ((i == 4) || (i == 5)) 2097 continue; 2098 2099 set_debugreg(0, i); 2100 } 2101 } 2102 2103 #ifdef CONFIG_KGDB 2104 /* 2105 * Restore debug regs if using kgdbwait and you have a kernel debugger 2106 * connection established. 2107 */ 2108 static void dbg_restore_debug_regs(void) 2109 { 2110 if (unlikely(kgdb_connected && arch_kgdb_ops.correct_hw_break)) 2111 arch_kgdb_ops.correct_hw_break(); 2112 } 2113 #else /* ! CONFIG_KGDB */ 2114 #define dbg_restore_debug_regs() 2115 #endif /* ! CONFIG_KGDB */ 2116 2117 static inline void setup_getcpu(int cpu) 2118 { 2119 unsigned long cpudata = vdso_encode_cpunode(cpu, early_cpu_to_node(cpu)); 2120 struct desc_struct d = { }; 2121 2122 if (boot_cpu_has(X86_FEATURE_RDTSCP) || boot_cpu_has(X86_FEATURE_RDPID)) 2123 wrmsr(MSR_TSC_AUX, cpudata, 0); 2124 2125 /* Store CPU and node number in limit. */ 2126 d.limit0 = cpudata; 2127 d.limit1 = cpudata >> 16; 2128 2129 d.type = 5; /* RO data, expand down, accessed */ 2130 d.dpl = 3; /* Visible to user code */ 2131 d.s = 1; /* Not a system segment */ 2132 d.p = 1; /* Present */ 2133 d.d = 1; /* 32-bit */ 2134 2135 write_gdt_entry(get_cpu_gdt_rw(cpu), GDT_ENTRY_CPUNODE, &d, DESCTYPE_S); 2136 } 2137 2138 #ifdef CONFIG_X86_64 2139 static inline void tss_setup_ist(struct tss_struct *tss) 2140 { 2141 /* Set up the per-CPU TSS IST stacks */ 2142 tss->x86_tss.ist[IST_INDEX_DF] = __this_cpu_ist_top_va(DF); 2143 tss->x86_tss.ist[IST_INDEX_NMI] = __this_cpu_ist_top_va(NMI); 2144 tss->x86_tss.ist[IST_INDEX_DB] = __this_cpu_ist_top_va(DB); 2145 tss->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE); 2146 /* Only mapped when SEV-ES is active */ 2147 tss->x86_tss.ist[IST_INDEX_VC] = __this_cpu_ist_top_va(VC); 2148 } 2149 #else /* CONFIG_X86_64 */ 2150 static inline void tss_setup_ist(struct tss_struct *tss) { } 2151 #endif /* !CONFIG_X86_64 */ 2152 2153 static inline void tss_setup_io_bitmap(struct tss_struct *tss) 2154 { 2155 tss->x86_tss.io_bitmap_base = IO_BITMAP_OFFSET_INVALID; 2156 2157 #ifdef CONFIG_X86_IOPL_IOPERM 2158 tss->io_bitmap.prev_max = 0; 2159 tss->io_bitmap.prev_sequence = 0; 2160 memset(tss->io_bitmap.bitmap, 0xff, sizeof(tss->io_bitmap.bitmap)); 2161 /* 2162 * Invalidate the extra array entry past the end of the all 2163 * permission bitmap as required by the hardware. 2164 */ 2165 tss->io_bitmap.mapall[IO_BITMAP_LONGS] = ~0UL; 2166 #endif 2167 } 2168 2169 /* 2170 * Setup everything needed to handle exceptions from the IDT, including the IST 2171 * exceptions which use paranoid_entry(). 2172 */ 2173 void cpu_init_exception_handling(void) 2174 { 2175 struct tss_struct *tss = this_cpu_ptr(&cpu_tss_rw); 2176 int cpu = raw_smp_processor_id(); 2177 2178 /* paranoid_entry() gets the CPU number from the GDT */ 2179 setup_getcpu(cpu); 2180 2181 /* For IDT mode, IST vectors need to be set in TSS. */ 2182 if (!cpu_feature_enabled(X86_FEATURE_FRED)) 2183 tss_setup_ist(tss); 2184 tss_setup_io_bitmap(tss); 2185 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss); 2186 2187 load_TR_desc(); 2188 2189 /* GHCB needs to be setup to handle #VC. */ 2190 setup_ghcb(); 2191 2192 if (cpu_feature_enabled(X86_FEATURE_FRED)) 2193 cpu_init_fred_exceptions(); 2194 else 2195 load_current_idt(); 2196 } 2197 2198 /* 2199 * cpu_init() initializes state that is per-CPU. Some data is already 2200 * initialized (naturally) in the bootstrap process, such as the GDT. We 2201 * reload it nevertheless, this function acts as a 'CPU state barrier', 2202 * nothing should get across. 2203 */ 2204 void cpu_init(void) 2205 { 2206 struct task_struct *cur = current; 2207 int cpu = raw_smp_processor_id(); 2208 2209 #ifdef CONFIG_NUMA 2210 if (this_cpu_read(numa_node) == 0 && 2211 early_cpu_to_node(cpu) != NUMA_NO_NODE) 2212 set_numa_node(early_cpu_to_node(cpu)); 2213 #endif 2214 pr_debug("Initializing CPU#%d\n", cpu); 2215 2216 if (IS_ENABLED(CONFIG_X86_64) || cpu_feature_enabled(X86_FEATURE_VME) || 2217 boot_cpu_has(X86_FEATURE_TSC) || boot_cpu_has(X86_FEATURE_DE)) 2218 cr4_clear_bits(X86_CR4_VME|X86_CR4_PVI|X86_CR4_TSD|X86_CR4_DE); 2219 2220 if (IS_ENABLED(CONFIG_X86_64)) { 2221 loadsegment(fs, 0); 2222 memset(cur->thread.tls_array, 0, GDT_ENTRY_TLS_ENTRIES * 8); 2223 syscall_init(); 2224 2225 wrmsrl(MSR_FS_BASE, 0); 2226 wrmsrl(MSR_KERNEL_GS_BASE, 0); 2227 barrier(); 2228 2229 x2apic_setup(); 2230 } 2231 2232 mmgrab(&init_mm); 2233 cur->active_mm = &init_mm; 2234 BUG_ON(cur->mm); 2235 initialize_tlbstate_and_flush(); 2236 enter_lazy_tlb(&init_mm, cur); 2237 2238 /* 2239 * sp0 points to the entry trampoline stack regardless of what task 2240 * is running. 2241 */ 2242 load_sp0((unsigned long)(cpu_entry_stack(cpu) + 1)); 2243 2244 load_mm_ldt(&init_mm); 2245 2246 clear_all_debug_regs(); 2247 dbg_restore_debug_regs(); 2248 2249 doublefault_init_cpu_tss(); 2250 2251 if (is_uv_system()) 2252 uv_cpu_init(); 2253 2254 load_fixmap_gdt(cpu); 2255 } 2256 2257 #ifdef CONFIG_MICROCODE_LATE_LOADING 2258 /** 2259 * store_cpu_caps() - Store a snapshot of CPU capabilities 2260 * @curr_info: Pointer where to store it 2261 * 2262 * Returns: None 2263 */ 2264 void store_cpu_caps(struct cpuinfo_x86 *curr_info) 2265 { 2266 /* Reload CPUID max function as it might've changed. */ 2267 curr_info->cpuid_level = cpuid_eax(0); 2268 2269 /* Copy all capability leafs and pick up the synthetic ones. */ 2270 memcpy(&curr_info->x86_capability, &boot_cpu_data.x86_capability, 2271 sizeof(curr_info->x86_capability)); 2272 2273 /* Get the hardware CPUID leafs */ 2274 get_cpu_cap(curr_info); 2275 } 2276 2277 /** 2278 * microcode_check() - Check if any CPU capabilities changed after an update. 2279 * @prev_info: CPU capabilities stored before an update. 2280 * 2281 * The microcode loader calls this upon late microcode load to recheck features, 2282 * only when microcode has been updated. Caller holds and CPU hotplug lock. 2283 * 2284 * Return: None 2285 */ 2286 void microcode_check(struct cpuinfo_x86 *prev_info) 2287 { 2288 struct cpuinfo_x86 curr_info; 2289 2290 perf_check_microcode(); 2291 2292 amd_check_microcode(); 2293 2294 store_cpu_caps(&curr_info); 2295 2296 if (!memcmp(&prev_info->x86_capability, &curr_info.x86_capability, 2297 sizeof(prev_info->x86_capability))) 2298 return; 2299 2300 pr_warn("x86/CPU: CPU features have changed after loading microcode, but might not take effect.\n"); 2301 pr_warn("x86/CPU: Please consider either early loading through initrd/built-in or a potential BIOS update.\n"); 2302 } 2303 #endif 2304 2305 /* 2306 * Invoked from core CPU hotplug code after hotplug operations 2307 */ 2308 void arch_smt_update(void) 2309 { 2310 /* Handle the speculative execution misfeatures */ 2311 cpu_bugs_smt_update(); 2312 /* Check whether IPI broadcasting can be enabled */ 2313 apic_smt_update(); 2314 } 2315 2316 void __init arch_cpu_finalize_init(void) 2317 { 2318 struct cpuinfo_x86 *c = this_cpu_ptr(&cpu_info); 2319 2320 identify_boot_cpu(); 2321 2322 select_idle_routine(); 2323 2324 /* 2325 * identify_boot_cpu() initialized SMT support information, let the 2326 * core code know. 2327 */ 2328 cpu_smt_set_num_threads(__max_threads_per_core, __max_threads_per_core); 2329 2330 if (!IS_ENABLED(CONFIG_SMP)) { 2331 pr_info("CPU: "); 2332 print_cpu_info(&boot_cpu_data); 2333 } 2334 2335 cpu_select_mitigations(); 2336 2337 arch_smt_update(); 2338 2339 if (IS_ENABLED(CONFIG_X86_32)) { 2340 /* 2341 * Check whether this is a real i386 which is not longer 2342 * supported and fixup the utsname. 2343 */ 2344 if (boot_cpu_data.x86 < 4) 2345 panic("Kernel requires i486+ for 'invlpg' and other features"); 2346 2347 init_utsname()->machine[1] = 2348 '0' + (boot_cpu_data.x86 > 6 ? 6 : boot_cpu_data.x86); 2349 } 2350 2351 /* 2352 * Must be before alternatives because it might set or clear 2353 * feature bits. 2354 */ 2355 fpu__init_system(); 2356 fpu__init_cpu(); 2357 2358 /* 2359 * Ensure that access to the per CPU representation has the initial 2360 * boot CPU configuration. 2361 */ 2362 *c = boot_cpu_data; 2363 c->initialized = true; 2364 2365 alternative_instructions(); 2366 2367 if (IS_ENABLED(CONFIG_X86_64)) { 2368 /* 2369 * Make sure the first 2MB area is not mapped by huge pages 2370 * There are typically fixed size MTRRs in there and overlapping 2371 * MTRRs into large pages causes slow downs. 2372 * 2373 * Right now we don't do that with gbpages because there seems 2374 * very little benefit for that case. 2375 */ 2376 if (!direct_gbpages) 2377 set_memory_4k((unsigned long)__va(0), 1); 2378 } else { 2379 fpu__init_check_bugs(); 2380 } 2381 2382 /* 2383 * This needs to be called before any devices perform DMA 2384 * operations that might use the SWIOTLB bounce buffers. It will 2385 * mark the bounce buffers as decrypted so that their usage will 2386 * not cause "plain-text" data to be decrypted when accessed. It 2387 * must be called after late_time_init() so that Hyper-V x86/x64 2388 * hypercalls work when the SWIOTLB bounce buffers are decrypted. 2389 */ 2390 mem_encrypt_init(); 2391 } 2392