1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * x86 SMP booting functions 4 * 5 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk> 6 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com> 7 * Copyright 2001 Andi Kleen, SuSE Labs. 8 * 9 * Much of the core SMP work is based on previous work by Thomas Radke, to 10 * whom a great many thanks are extended. 11 * 12 * Thanks to Intel for making available several different Pentium, 13 * Pentium Pro and Pentium-II/Xeon MP machines. 14 * Original development of Linux SMP code supported by Caldera. 15 * 16 * Fixes 17 * Felix Koop : NR_CPUS used properly 18 * Jose Renau : Handle single CPU case. 19 * Alan Cox : By repeated request 8) - Total BogoMIPS report. 20 * Greg Wright : Fix for kernel stacks panic. 21 * Erich Boleyn : MP v1.4 and additional changes. 22 * Matthias Sattler : Changes for 2.1 kernel map. 23 * Michel Lespinasse : Changes for 2.1 kernel map. 24 * Michael Chastain : Change trampoline.S to gnu as. 25 * Alan Cox : Dumb bug: 'B' step PPro's are fine 26 * Ingo Molnar : Added APIC timers, based on code 27 * from Jose Renau 28 * Ingo Molnar : various cleanups and rewrites 29 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug. 30 * Maciej W. Rozycki : Bits for genuine 82489DX APICs 31 * Andi Kleen : Changed for SMP boot into long mode. 32 * Martin J. Bligh : Added support for multi-quad systems 33 * Dave Jones : Report invalid combinations of Athlon CPUs. 34 * Rusty Russell : Hacked into shape for new "hotplug" boot process. 35 * Andi Kleen : Converted to new state machine. 36 * Ashok Raj : CPU hotplug support 37 * Glauber Costa : i386 and x86_64 integration 38 */ 39 40 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 41 42 #include <linux/init.h> 43 #include <linux/smp.h> 44 #include <linux/export.h> 45 #include <linux/sched.h> 46 #include <linux/sched/topology.h> 47 #include <linux/sched/hotplug.h> 48 #include <linux/sched/task_stack.h> 49 #include <linux/percpu.h> 50 #include <linux/memblock.h> 51 #include <linux/err.h> 52 #include <linux/nmi.h> 53 #include <linux/tboot.h> 54 #include <linux/gfp.h> 55 #include <linux/cpuidle.h> 56 #include <linux/kexec.h> 57 #include <linux/numa.h> 58 #include <linux/pgtable.h> 59 #include <linux/overflow.h> 60 #include <linux/stackprotector.h> 61 #include <linux/cpuhotplug.h> 62 #include <linux/mc146818rtc.h> 63 #include <linux/acpi.h> 64 65 #include <asm/acpi.h> 66 #include <asm/cacheinfo.h> 67 #include <asm/cpuid/api.h> 68 #include <asm/desc.h> 69 #include <asm/nmi.h> 70 #include <asm/irq.h> 71 #include <asm/realmode.h> 72 #include <asm/cpu.h> 73 #include <asm/numa.h> 74 #include <asm/tlbflush.h> 75 #include <asm/mtrr.h> 76 #include <asm/mwait.h> 77 #include <asm/apic.h> 78 #include <asm/io_apic.h> 79 #include <asm/fpu/api.h> 80 #include <asm/setup.h> 81 #include <asm/uv/uv.h> 82 #include <asm/microcode.h> 83 #include <asm/i8259.h> 84 #include <asm/misc.h> 85 #include <asm/qspinlock.h> 86 #include <asm/intel-family.h> 87 #include <asm/cpu_device_id.h> 88 #include <asm/spec-ctrl.h> 89 #include <asm/hw_irq.h> 90 #include <asm/stackprotector.h> 91 #include <asm/sev.h> 92 #include <asm/spec-ctrl.h> 93 94 /* representing HT siblings of each logical CPU */ 95 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map); 96 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map); 97 98 /* representing HT and core siblings of each logical CPU */ 99 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map); 100 EXPORT_PER_CPU_SYMBOL(cpu_core_map); 101 102 /* representing HT, core, and die siblings of each logical CPU */ 103 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map); 104 EXPORT_PER_CPU_SYMBOL(cpu_die_map); 105 106 /* Representing CPUs for which sibling maps can be computed */ 107 static cpumask_var_t cpu_sibling_setup_mask; 108 109 struct mwait_cpu_dead { 110 unsigned int control; 111 unsigned int status; 112 }; 113 114 #define CPUDEAD_MWAIT_WAIT 0xDEADBEEF 115 #define CPUDEAD_MWAIT_KEXEC_HLT 0x4A17DEAD 116 117 /* 118 * Cache line aligned data for mwait_play_dead(). Separate on purpose so 119 * that it's unlikely to be touched by other CPUs. 120 */ 121 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead); 122 123 /* Maximum number of SMT threads on any online core */ 124 int __read_mostly __max_smt_threads = 1; 125 126 /* Flag to indicate if a complete sched domain rebuild is required */ 127 bool x86_topology_update; 128 129 int arch_update_cpu_topology(void) 130 { 131 int retval = x86_topology_update; 132 133 x86_topology_update = false; 134 return retval; 135 } 136 137 static unsigned int smpboot_warm_reset_vector_count; 138 139 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip) 140 { 141 unsigned long flags; 142 143 spin_lock_irqsave(&rtc_lock, flags); 144 if (!smpboot_warm_reset_vector_count++) { 145 CMOS_WRITE(0xa, 0xf); 146 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4; 147 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf; 148 } 149 spin_unlock_irqrestore(&rtc_lock, flags); 150 } 151 152 static inline void smpboot_restore_warm_reset_vector(void) 153 { 154 unsigned long flags; 155 156 /* 157 * Paranoid: Set warm reset code and vector here back 158 * to default values. 159 */ 160 spin_lock_irqsave(&rtc_lock, flags); 161 if (!--smpboot_warm_reset_vector_count) { 162 CMOS_WRITE(0, 0xf); 163 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0; 164 } 165 spin_unlock_irqrestore(&rtc_lock, flags); 166 167 } 168 169 /* Run the next set of setup steps for the upcoming CPU */ 170 static void ap_starting(void) 171 { 172 int cpuid = smp_processor_id(); 173 174 /* Mop up eventual mwait_play_dead() wreckage */ 175 this_cpu_write(mwait_cpu_dead.status, 0); 176 this_cpu_write(mwait_cpu_dead.control, 0); 177 178 /* 179 * If woken up by an INIT in an 82489DX configuration the alive 180 * synchronization guarantees that the CPU does not reach this 181 * point before an INIT_deassert IPI reaches the local APIC, so it 182 * is now safe to touch the local APIC. 183 * 184 * Set up this CPU, first the APIC, which is probably redundant on 185 * most boards. 186 */ 187 apic_ap_setup(); 188 189 /* Save the processor parameters. */ 190 identify_secondary_cpu(cpuid); 191 192 /* 193 * The topology information must be up to date before 194 * notify_cpu_starting(). 195 */ 196 set_cpu_sibling_map(cpuid); 197 198 ap_init_aperfmperf(); 199 200 pr_debug("Stack at about %p\n", &cpuid); 201 202 wmb(); 203 204 /* 205 * This runs the AP through all the cpuhp states to its target 206 * state CPUHP_ONLINE. 207 */ 208 notify_cpu_starting(cpuid); 209 } 210 211 static void ap_calibrate_delay(void) 212 { 213 /* 214 * Calibrate the delay loop and update loops_per_jiffy in cpu_data. 215 * identify_secondary_cpu() stored a value that is close but not as 216 * accurate as the value just calculated. 217 * 218 * As this is invoked after the TSC synchronization check, 219 * calibrate_delay_is_known() will skip the calibration routine 220 * when TSC is synchronized across sockets. 221 */ 222 calibrate_delay(); 223 cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy; 224 } 225 226 /* 227 * Activate a secondary processor. 228 */ 229 static void notrace __noendbr start_secondary(void *unused) 230 { 231 /* 232 * Don't put *anything* except direct CPU state initialization 233 * before cpu_init(), SMP booting is too fragile that we want to 234 * limit the things done here to the most necessary things. 235 */ 236 cr4_init(); 237 238 /* 239 * 32-bit specific. 64-bit reaches this code with the correct page 240 * table established. Yet another historical divergence. 241 */ 242 if (IS_ENABLED(CONFIG_X86_32)) { 243 /* switch away from the initial page table */ 244 load_cr3(swapper_pg_dir); 245 __flush_tlb_all(); 246 } 247 248 cpu_init_exception_handling(false); 249 250 /* 251 * Load the microcode before reaching the AP alive synchronization 252 * point below so it is not part of the full per CPU serialized 253 * bringup part when "parallel" bringup is enabled. 254 * 255 * That's even safe when hyperthreading is enabled in the CPU as 256 * the core code starts the primary threads first and leaves the 257 * secondary threads waiting for SIPI. Loading microcode on 258 * physical cores concurrently is a safe operation. 259 * 260 * This covers both the Intel specific issue that concurrent 261 * microcode loading on SMT siblings must be prohibited and the 262 * vendor independent issue`that microcode loading which changes 263 * CPUID, MSRs etc. must be strictly serialized to maintain 264 * software state correctness. 265 */ 266 load_ucode_ap(); 267 268 /* 269 * Synchronization point with the hotplug core. Sets this CPUs 270 * synchronization state to ALIVE and spin-waits for the control CPU to 271 * release this CPU for further bringup. 272 */ 273 cpuhp_ap_sync_alive(); 274 275 cpu_init(); 276 fpu__init_cpu(); 277 rcutree_report_cpu_starting(raw_smp_processor_id()); 278 x86_cpuinit.early_percpu_clock_init(); 279 280 ap_starting(); 281 282 /* Check TSC synchronization with the control CPU. */ 283 check_tsc_sync_target(); 284 285 /* 286 * Calibrate the delay loop after the TSC synchronization check. 287 * This allows to skip the calibration when TSC is synchronized 288 * across sockets. 289 */ 290 ap_calibrate_delay(); 291 292 speculative_store_bypass_ht_init(); 293 294 /* 295 * Lock vector_lock, set CPU online and bring the vector 296 * allocator online. Online must be set with vector_lock held 297 * to prevent a concurrent irq setup/teardown from seeing a 298 * half valid vector space. 299 */ 300 lock_vector_lock(); 301 set_cpu_online(smp_processor_id(), true); 302 lapic_online(); 303 unlock_vector_lock(); 304 x86_platform.nmi_init(); 305 306 /* enable local interrupts */ 307 local_irq_enable(); 308 309 x86_cpuinit.setup_percpu_clockev(); 310 311 wmb(); 312 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE); 313 } 314 ANNOTATE_NOENDBR_SYM(start_secondary); 315 316 static bool 317 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o) 318 { 319 int cpu1 = c->cpu_index, cpu2 = o->cpu_index; 320 321 return (cpu_to_node(cpu1) == cpu_to_node(cpu2)); 322 } 323 324 static bool 325 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name) 326 { 327 int cpu1 = c->cpu_index, cpu2 = o->cpu_index; 328 329 return !WARN_ONCE(!topology_same_node(c, o), 330 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! " 331 "[node: %d != %d]. Ignoring dependency.\n", 332 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2)); 333 } 334 335 #define link_mask(mfunc, c1, c2) \ 336 do { \ 337 cpumask_set_cpu((c1), mfunc(c2)); \ 338 cpumask_set_cpu((c2), mfunc(c1)); \ 339 } while (0) 340 341 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o) 342 { 343 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) { 344 int cpu1 = c->cpu_index, cpu2 = o->cpu_index; 345 346 if (c->topo.pkg_id == o->topo.pkg_id && 347 c->topo.die_id == o->topo.die_id && 348 c->topo.amd_node_id == o->topo.amd_node_id && 349 per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) { 350 if (c->topo.core_id == o->topo.core_id) 351 return topology_sane(c, o, "smt"); 352 353 if ((c->topo.cu_id != 0xff) && 354 (o->topo.cu_id != 0xff) && 355 (c->topo.cu_id == o->topo.cu_id)) 356 return topology_sane(c, o, "smt"); 357 } 358 359 } else if (c->topo.pkg_id == o->topo.pkg_id && 360 c->topo.die_id == o->topo.die_id && 361 c->topo.core_id == o->topo.core_id) { 362 return topology_sane(c, o, "smt"); 363 } 364 365 return false; 366 } 367 368 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o) 369 { 370 if (c->topo.pkg_id != o->topo.pkg_id || c->topo.die_id != o->topo.die_id) 371 return false; 372 373 if (cpu_feature_enabled(X86_FEATURE_TOPOEXT) && topology_amd_nodes_per_pkg() > 1) 374 return c->topo.amd_node_id == o->topo.amd_node_id; 375 376 return true; 377 } 378 379 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o) 380 { 381 int cpu1 = c->cpu_index, cpu2 = o->cpu_index; 382 383 /* If the arch didn't set up l2c_id, fall back to SMT */ 384 if (per_cpu_l2c_id(cpu1) == BAD_APICID) 385 return match_smt(c, o); 386 387 /* Do not match if L2 cache id does not match: */ 388 if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2)) 389 return false; 390 391 return topology_sane(c, o, "l2c"); 392 } 393 394 /* 395 * Unlike the other levels, we do not enforce keeping a 396 * multicore group inside a NUMA node. If this happens, we will 397 * discard the MC level of the topology later. 398 */ 399 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o) 400 { 401 if (c->topo.pkg_id == o->topo.pkg_id) 402 return true; 403 return false; 404 } 405 406 /* 407 * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs. 408 * 409 * Any Intel CPU that has multiple nodes per package and does not 410 * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology. 411 * 412 * When in SNC mode, these CPUs enumerate an LLC that is shared 413 * by multiple NUMA nodes. The LLC is shared for off-package data 414 * access but private to the NUMA node (half of the package) for 415 * on-package access. CPUID (the source of the information about 416 * the LLC) can only enumerate the cache as shared or unshared, 417 * but not this particular configuration. 418 */ 419 420 static const struct x86_cpu_id intel_cod_cpu[] = { 421 X86_MATCH_VFM(INTEL_HASWELL_X, 0), /* COD */ 422 X86_MATCH_VFM(INTEL_BROADWELL_X, 0), /* COD */ 423 X86_MATCH_VFM(INTEL_ANY, 1), /* SNC */ 424 {} 425 }; 426 427 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o) 428 { 429 const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu); 430 int cpu1 = c->cpu_index, cpu2 = o->cpu_index; 431 bool intel_snc = id && id->driver_data; 432 433 /* Do not match if we do not have a valid APICID for cpu: */ 434 if (per_cpu_llc_id(cpu1) == BAD_APICID) 435 return false; 436 437 /* Do not match if LLC id does not match: */ 438 if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2)) 439 return false; 440 441 /* 442 * Allow the SNC topology without warning. Return of false 443 * means 'c' does not share the LLC of 'o'. This will be 444 * reflected to userspace. 445 */ 446 if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc) 447 return false; 448 449 return topology_sane(c, o, "llc"); 450 } 451 452 453 static inline int x86_sched_itmt_flags(void) 454 { 455 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0; 456 } 457 458 #ifdef CONFIG_SCHED_MC 459 static int x86_core_flags(void) 460 { 461 return cpu_core_flags() | x86_sched_itmt_flags(); 462 } 463 #endif 464 #ifdef CONFIG_SCHED_CLUSTER 465 static int x86_cluster_flags(void) 466 { 467 return cpu_cluster_flags() | x86_sched_itmt_flags(); 468 } 469 #endif 470 471 static struct sched_domain_topology_level x86_topology[] = { 472 SDTL_INIT(tl_smt_mask, cpu_smt_flags, SMT), 473 #ifdef CONFIG_SCHED_CLUSTER 474 SDTL_INIT(tl_cls_mask, x86_cluster_flags, CLS), 475 #endif 476 #ifdef CONFIG_SCHED_MC 477 SDTL_INIT(tl_mc_mask, x86_core_flags, MC), 478 #endif 479 SDTL_INIT(tl_pkg_mask, x86_sched_itmt_flags, PKG), 480 { NULL }, 481 }; 482 483 static void __init build_sched_topology(void) 484 { 485 struct sched_domain_topology_level *topology = x86_topology; 486 487 /* 488 * When there is NUMA topology inside the package invalidate the 489 * PKG domain since the NUMA domains will auto-magically create the 490 * right spanning domains based on the SLIT. 491 */ 492 if (topology_num_nodes_per_package() > 1) { 493 unsigned int pkgdom = ARRAY_SIZE(x86_topology) - 2; 494 495 memset(&x86_topology[pkgdom], 0, sizeof(x86_topology[pkgdom])); 496 } 497 498 /* 499 * Drop the SMT domains if there is only one thread per-core 500 * since it'll get degenerated by the scheduler anyways. 501 */ 502 if (cpu_smt_num_threads <= 1) 503 ++topology; 504 505 set_sched_topology(topology); 506 } 507 508 #ifdef CONFIG_NUMA 509 /* 510 * Test if the on-trace cluster at (N,N) is symmetric. 511 * Uses upper triangle iteration to avoid obvious duplicates. 512 */ 513 static bool slit_cluster_symmetric(int N) 514 { 515 int u = topology_num_nodes_per_package(); 516 517 for (int k = 0; k < u; k++) { 518 for (int l = k; l < u; l++) { 519 if (node_distance(N + k, N + l) != 520 node_distance(N + l, N + k)) 521 return false; 522 } 523 } 524 525 return true; 526 } 527 528 /* 529 * Return the package-id of the cluster, or ~0 if indeterminate. 530 * Each node in the on-trace cluster should have the same package-id. 531 */ 532 static u32 slit_cluster_package(int N) 533 { 534 int u = topology_num_nodes_per_package(); 535 u32 pkg_id = ~0; 536 537 for (int n = 0; n < u; n++) { 538 const struct cpumask *cpus = cpumask_of_node(N + n); 539 int cpu; 540 541 for_each_cpu(cpu, cpus) { 542 u32 id = topology_logical_package_id(cpu); 543 544 if (pkg_id == ~0) 545 pkg_id = id; 546 if (pkg_id != id) 547 return ~0; 548 } 549 } 550 551 return pkg_id; 552 } 553 554 /* 555 * Validate the SLIT table is of the form expected for SNC, specifically: 556 * 557 * - each on-trace cluster should be symmetric, 558 * - each on-trace cluster should have a unique package-id. 559 * 560 * If you NUMA_EMU on top of SNC, you get to keep the pieces. 561 */ 562 static bool slit_validate(void) 563 { 564 int u = topology_num_nodes_per_package(); 565 u32 pkg_id, prev_pkg_id = ~0; 566 567 for (int pkg = 0; pkg < topology_max_packages(); pkg++) { 568 int n = pkg * u; 569 570 /* 571 * Ensure the on-trace cluster is symmetric and each cluster 572 * has a different package id. 573 */ 574 if (!slit_cluster_symmetric(n)) 575 return false; 576 pkg_id = slit_cluster_package(n); 577 if (pkg_id == ~0) 578 return false; 579 if (pkg && pkg_id == prev_pkg_id) 580 return false; 581 582 prev_pkg_id = pkg_id; 583 } 584 585 return true; 586 } 587 588 /* 589 * Compute a sanitized SLIT table for SNC; notably SNC-3 can end up with 590 * asymmetric off-trace clusters, reflecting physical assymmetries. However 591 * this leads to 'unfortunate' sched_domain configurations. 592 * 593 * For example dual socket GNR with SNC-3: 594 * 595 * node distances: 596 * node 0 1 2 3 4 5 597 * 0: 10 15 17 21 28 26 598 * 1: 15 10 15 23 26 23 599 * 2: 17 15 10 26 23 21 600 * 3: 21 28 26 10 15 17 601 * 4: 23 26 23 15 10 15 602 * 5: 26 23 21 17 15 10 603 * 604 * Fix things up by averaging out the off-trace clusters; resulting in: 605 * 606 * node 0 1 2 3 4 5 607 * 0: 10 15 17 24 24 24 608 * 1: 15 10 15 24 24 24 609 * 2: 17 15 10 24 24 24 610 * 3: 24 24 24 10 15 17 611 * 4: 24 24 24 15 10 15 612 * 5: 24 24 24 17 15 10 613 */ 614 static int slit_cluster_distance(int i, int j) 615 { 616 static int slit_valid = -1; 617 int u = topology_num_nodes_per_package(); 618 long d = 0; 619 int x, y; 620 621 if (slit_valid < 0) { 622 slit_valid = slit_validate(); 623 if (!slit_valid) 624 pr_err(FW_BUG "SLIT table doesn't have the expected form for SNC -- fixup disabled!\n"); 625 else 626 pr_info("Fixing up SNC SLIT table.\n"); 627 } 628 629 /* 630 * Is this a unit cluster on the trace? 631 */ 632 if ((i / u) == (j / u) || !slit_valid) 633 return node_distance(i, j); 634 635 /* 636 * Off-trace cluster. 637 * 638 * Notably average out the symmetric pair of off-trace clusters to 639 * ensure the resulting SLIT table is symmetric. 640 */ 641 x = i - (i % u); 642 y = j - (j % u); 643 644 for (i = x; i < x + u; i++) { 645 for (j = y; j < y + u; j++) { 646 d += node_distance(i, j); 647 d += node_distance(j, i); 648 } 649 } 650 651 return d / (2*u*u); 652 } 653 654 int arch_sched_node_distance(int from, int to) 655 { 656 int d = node_distance(from, to); 657 658 switch (boot_cpu_data.x86_vfm) { 659 case INTEL_GRANITERAPIDS_X: 660 case INTEL_ATOM_DARKMONT_X: 661 if (topology_max_packages() == 1 || 662 topology_num_nodes_per_package() < 3) 663 return d; 664 665 /* 666 * Handle SNC-3 asymmetries. 667 */ 668 return slit_cluster_distance(from, to); 669 } 670 return d; 671 } 672 #endif /* CONFIG_NUMA */ 673 674 void set_cpu_sibling_map(int cpu) 675 { 676 bool has_smt = __max_threads_per_core > 1; 677 bool has_mp = has_smt || topology_num_cores_per_package() > 1; 678 struct cpuinfo_x86 *c = &cpu_data(cpu); 679 struct cpuinfo_x86 *o; 680 int i, threads; 681 682 cpumask_set_cpu(cpu, cpu_sibling_setup_mask); 683 684 if (!has_mp) { 685 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu)); 686 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu)); 687 cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu)); 688 cpumask_set_cpu(cpu, topology_core_cpumask(cpu)); 689 cpumask_set_cpu(cpu, topology_die_cpumask(cpu)); 690 c->booted_cores = 1; 691 return; 692 } 693 694 for_each_cpu(i, cpu_sibling_setup_mask) { 695 o = &cpu_data(i); 696 697 if (match_pkg(c, o) && !topology_same_node(c, o)) 698 WARN_ON_ONCE(topology_num_nodes_per_package() == 1); 699 700 if ((i == cpu) || (has_smt && match_smt(c, o))) 701 link_mask(topology_sibling_cpumask, cpu, i); 702 703 if ((i == cpu) || (has_mp && match_llc(c, o))) 704 link_mask(cpu_llc_shared_mask, cpu, i); 705 706 if ((i == cpu) || (has_mp && match_l2c(c, o))) 707 link_mask(cpu_l2c_shared_mask, cpu, i); 708 709 if ((i == cpu) || (has_mp && match_die(c, o))) 710 link_mask(topology_die_cpumask, cpu, i); 711 } 712 713 threads = cpumask_weight(topology_sibling_cpumask(cpu)); 714 if (threads > __max_smt_threads) 715 __max_smt_threads = threads; 716 717 for_each_cpu(i, topology_sibling_cpumask(cpu)) 718 cpu_data(i).smt_active = threads > 1; 719 720 /* 721 * This needs a separate iteration over the cpus because we rely on all 722 * topology_sibling_cpumask links to be set-up. 723 */ 724 for_each_cpu(i, cpu_sibling_setup_mask) { 725 o = &cpu_data(i); 726 727 if ((i == cpu) || (has_mp && match_pkg(c, o))) { 728 link_mask(topology_core_cpumask, cpu, i); 729 730 /* 731 * Does this new cpu bringup a new core? 732 */ 733 if (threads == 1) { 734 /* 735 * for each core in package, increment 736 * the booted_cores for this new cpu 737 */ 738 if (cpumask_first( 739 topology_sibling_cpumask(i)) == i) 740 c->booted_cores++; 741 /* 742 * increment the core count for all 743 * the other cpus in this package 744 */ 745 if (i != cpu) 746 cpu_data(i).booted_cores++; 747 } else if (i != cpu && !c->booted_cores) 748 c->booted_cores = cpu_data(i).booted_cores; 749 } 750 } 751 } 752 753 /* maps the cpu to the sched domain representing multi-core */ 754 const struct cpumask *cpu_coregroup_mask(int cpu) 755 { 756 return cpu_llc_shared_mask(cpu); 757 } 758 759 const struct cpumask *cpu_clustergroup_mask(int cpu) 760 { 761 return cpu_l2c_shared_mask(cpu); 762 } 763 EXPORT_SYMBOL_GPL(cpu_clustergroup_mask); 764 765 static void impress_friends(void) 766 { 767 int cpu; 768 unsigned long bogosum = 0; 769 /* 770 * Allow the user to impress friends. 771 */ 772 pr_debug("Before bogomips\n"); 773 for_each_online_cpu(cpu) 774 bogosum += cpu_data(cpu).loops_per_jiffy; 775 776 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n", 777 num_online_cpus(), 778 bogosum/(500000/HZ), 779 (bogosum/(5000/HZ))%100); 780 781 pr_debug("Before bogocount - setting activated=1\n"); 782 } 783 784 /* 785 * The Multiprocessor Specification 1.4 (1997) example code suggests 786 * that there should be a 10ms delay between the BSP asserting INIT 787 * and de-asserting INIT, when starting a remote processor. 788 * But that slows boot and resume on modern processors, which include 789 * many cores and don't require that delay. 790 * 791 * Cmdline "cpu_init_udelay=" is available to override this delay. 792 */ 793 #define UDELAY_10MS_LEGACY 10000 794 795 static unsigned int init_udelay = UINT_MAX; 796 797 static int __init cpu_init_udelay(char *str) 798 { 799 get_option(&str, &init_udelay); 800 801 return 0; 802 } 803 early_param("cpu_init_udelay", cpu_init_udelay); 804 805 static void __init smp_set_init_udelay(void) 806 { 807 /* if cmdline changed it from default, leave it alone */ 808 if (init_udelay != UINT_MAX) 809 return; 810 811 /* if modern processor, use no delay */ 812 if ((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL && boot_cpu_data.x86_vfm >= INTEL_PENTIUM_PRO) || 813 (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON && boot_cpu_data.x86 >= 0x18) || 814 (boot_cpu_data.x86_vendor == X86_VENDOR_AMD && boot_cpu_data.x86 >= 0xF)) { 815 init_udelay = 0; 816 return; 817 } 818 /* else, use legacy delay */ 819 init_udelay = UDELAY_10MS_LEGACY; 820 } 821 822 /* 823 * Wake up AP by INIT, INIT, STARTUP sequence. 824 */ 825 static void send_init_sequence(u32 phys_apicid) 826 { 827 int maxlvt = lapic_get_maxlvt(); 828 829 /* Be paranoid about clearing APIC errors. */ 830 if (APIC_INTEGRATED(boot_cpu_apic_version)) { 831 /* Due to the Pentium erratum 3AP. */ 832 if (maxlvt > 3) 833 apic_write(APIC_ESR, 0); 834 apic_read(APIC_ESR); 835 } 836 837 /* Assert INIT on the target CPU */ 838 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid); 839 safe_apic_wait_icr_idle(); 840 841 udelay(init_udelay); 842 843 /* Deassert INIT on the target CPU */ 844 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid); 845 safe_apic_wait_icr_idle(); 846 } 847 848 /* 849 * Wake up AP by INIT, INIT, STARTUP sequence. 850 */ 851 static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip, unsigned int cpu) 852 { 853 unsigned long send_status = 0, accept_status = 0; 854 int num_starts, j, maxlvt; 855 856 preempt_disable(); 857 maxlvt = lapic_get_maxlvt(); 858 send_init_sequence(phys_apicid); 859 860 mb(); 861 862 /* 863 * Should we send STARTUP IPIs ? 864 * 865 * Determine this based on the APIC version. 866 * If we don't have an integrated APIC, don't send the STARTUP IPIs. 867 */ 868 if (APIC_INTEGRATED(boot_cpu_apic_version)) 869 num_starts = 2; 870 else 871 num_starts = 0; 872 873 /* 874 * Run STARTUP IPI loop. 875 */ 876 pr_debug("#startup loops: %d\n", num_starts); 877 878 for (j = 1; j <= num_starts; j++) { 879 pr_debug("Sending STARTUP #%d\n", j); 880 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */ 881 apic_write(APIC_ESR, 0); 882 apic_read(APIC_ESR); 883 pr_debug("After apic_write\n"); 884 885 /* 886 * STARTUP IPI 887 */ 888 889 /* Target chip */ 890 /* Boot on the stack */ 891 /* Kick the second */ 892 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12), 893 phys_apicid); 894 895 /* 896 * Give the other CPU some time to accept the IPI. 897 */ 898 if (init_udelay == 0) 899 udelay(10); 900 else 901 udelay(300); 902 903 pr_debug("Startup point 1\n"); 904 905 pr_debug("Waiting for send to finish...\n"); 906 send_status = safe_apic_wait_icr_idle(); 907 908 /* 909 * Give the other CPU some time to accept the IPI. 910 */ 911 if (init_udelay == 0) 912 udelay(10); 913 else 914 udelay(200); 915 916 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */ 917 apic_write(APIC_ESR, 0); 918 accept_status = (apic_read(APIC_ESR) & 0xEF); 919 if (send_status || accept_status) 920 break; 921 } 922 pr_debug("After Startup\n"); 923 924 if (send_status) 925 pr_err("APIC never delivered???\n"); 926 if (accept_status) 927 pr_err("APIC delivery error (%lx)\n", accept_status); 928 929 preempt_enable(); 930 return (send_status | accept_status); 931 } 932 933 /* reduce the number of lines printed when booting a large cpu count system */ 934 static void announce_cpu(int cpu, int apicid) 935 { 936 static int width, node_width, first = 1; 937 static int current_node = NUMA_NO_NODE; 938 int node = early_cpu_to_node(cpu); 939 940 if (!width) 941 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */ 942 943 if (!node_width) 944 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */ 945 946 if (system_state < SYSTEM_RUNNING) { 947 if (first) 948 pr_info("x86: Booting SMP configuration:\n"); 949 950 if (node != current_node) { 951 if (current_node > (-1)) 952 pr_cont("\n"); 953 current_node = node; 954 955 printk(KERN_INFO ".... node %*s#%d, CPUs: ", 956 node_width - num_digits(node), " ", node); 957 } 958 959 /* Add padding for the BSP */ 960 if (first) 961 pr_cont("%*s", width + 1, " "); 962 first = 0; 963 964 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu); 965 } else 966 pr_info("Booting Node %d Processor %d APIC 0x%x\n", 967 node, cpu, apicid); 968 } 969 970 int common_cpu_up(unsigned int cpu, struct task_struct *idle) 971 { 972 int ret; 973 974 /* Just in case we booted with a single CPU. */ 975 alternatives_enable_smp(); 976 977 per_cpu(current_task, cpu) = idle; 978 cpu_init_stack_canary(cpu, idle); 979 980 /* Initialize the interrupt stack(s) */ 981 ret = irq_init_percpu_irqstack(cpu); 982 if (ret) 983 return ret; 984 985 #ifdef CONFIG_X86_32 986 /* Stack for startup_32 can be just as for start_secondary onwards */ 987 per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle); 988 #endif 989 return 0; 990 } 991 992 /* 993 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad 994 * (ie clustered apic addressing mode), this is a LOGICAL apic ID. 995 * Returns zero if startup was successfully sent, else error code from 996 * ->wakeup_secondary_cpu. 997 */ 998 static int do_boot_cpu(u32 apicid, unsigned int cpu, struct task_struct *idle) 999 { 1000 unsigned long start_ip = real_mode_header->trampoline_start; 1001 int ret; 1002 1003 #ifdef CONFIG_X86_64 1004 /* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */ 1005 if (apic->wakeup_secondary_cpu_64) 1006 start_ip = real_mode_header->trampoline_start64; 1007 #endif 1008 idle->thread.sp = (unsigned long)task_pt_regs(idle); 1009 initial_code = (unsigned long)start_secondary; 1010 1011 if (IS_ENABLED(CONFIG_X86_32)) { 1012 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu); 1013 initial_stack = idle->thread.sp; 1014 } else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) { 1015 smpboot_control = cpu; 1016 } 1017 1018 /* Enable the espfix hack for this CPU */ 1019 init_espfix_ap(cpu); 1020 1021 /* So we see what's up */ 1022 announce_cpu(cpu, apicid); 1023 1024 /* 1025 * This grunge runs the startup process for 1026 * the targeted processor. 1027 */ 1028 if (x86_platform.legacy.warm_reset) { 1029 1030 pr_debug("Setting warm reset code and vector.\n"); 1031 1032 smpboot_setup_warm_reset_vector(start_ip); 1033 /* 1034 * Be paranoid about clearing APIC errors. 1035 */ 1036 if (APIC_INTEGRATED(boot_cpu_apic_version)) { 1037 apic_write(APIC_ESR, 0); 1038 apic_read(APIC_ESR); 1039 } 1040 } 1041 1042 smp_mb(); 1043 1044 /* 1045 * Wake up a CPU in difference cases: 1046 * - Use a method from the APIC driver if one defined, with wakeup 1047 * straight to 64-bit mode preferred over wakeup to RM. 1048 * Otherwise, 1049 * - Use an INIT boot APIC message 1050 */ 1051 if (apic->wakeup_secondary_cpu_64) 1052 ret = apic->wakeup_secondary_cpu_64(apicid, start_ip, cpu); 1053 else if (apic->wakeup_secondary_cpu) 1054 ret = apic->wakeup_secondary_cpu(apicid, start_ip, cpu); 1055 else 1056 ret = wakeup_secondary_cpu_via_init(apicid, start_ip, cpu); 1057 1058 /* If the wakeup mechanism failed, cleanup the warm reset vector */ 1059 if (ret) 1060 arch_cpuhp_cleanup_kick_cpu(cpu); 1061 return ret; 1062 } 1063 1064 int native_kick_ap(unsigned int cpu, struct task_struct *tidle) 1065 { 1066 u32 apicid = apic->cpu_present_to_apicid(cpu); 1067 int err; 1068 1069 lockdep_assert_irqs_enabled(); 1070 1071 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu); 1072 1073 if (apicid == BAD_APICID || !apic_id_valid(apicid)) { 1074 pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid); 1075 return -EINVAL; 1076 } 1077 1078 if (!test_bit(apicid, phys_cpu_present_map)) { 1079 pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid); 1080 return -EINVAL; 1081 } 1082 1083 /* 1084 * Save current MTRR state in case it was changed since early boot 1085 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync: 1086 */ 1087 mtrr_save_state(); 1088 1089 /* the FPU context is blank, nobody can own it */ 1090 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL; 1091 1092 err = common_cpu_up(cpu, tidle); 1093 if (err) 1094 return err; 1095 1096 err = do_boot_cpu(apicid, cpu, tidle); 1097 if (err) 1098 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu); 1099 1100 return err; 1101 } 1102 1103 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle) 1104 { 1105 return smp_ops.kick_ap_alive(cpu, tidle); 1106 } 1107 1108 void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu) 1109 { 1110 /* Cleanup possible dangling ends... */ 1111 if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset) 1112 smpboot_restore_warm_reset_vector(); 1113 } 1114 1115 void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu) 1116 { 1117 if (smp_ops.cleanup_dead_cpu) 1118 smp_ops.cleanup_dead_cpu(cpu); 1119 1120 if (system_state == SYSTEM_RUNNING) 1121 pr_info("CPU %u is now offline\n", cpu); 1122 } 1123 1124 void arch_cpuhp_sync_state_poll(void) 1125 { 1126 if (smp_ops.poll_sync_state) 1127 smp_ops.poll_sync_state(); 1128 } 1129 1130 /** 1131 * arch_disable_smp_support() - Disables SMP support for x86 at boottime 1132 */ 1133 void __init arch_disable_smp_support(void) 1134 { 1135 disable_ioapic_support(); 1136 } 1137 1138 /* 1139 * Fall back to non SMP mode after errors. 1140 * 1141 * RED-PEN audit/test this more. I bet there is more state messed up here. 1142 */ 1143 static __init void disable_smp(void) 1144 { 1145 pr_info("SMP disabled\n"); 1146 1147 disable_ioapic_support(); 1148 topology_reset_possible_cpus_up(); 1149 1150 cpumask_set_cpu(0, topology_sibling_cpumask(0)); 1151 cpumask_set_cpu(0, topology_core_cpumask(0)); 1152 cpumask_set_cpu(0, topology_die_cpumask(0)); 1153 } 1154 1155 void __init smp_prepare_cpus_common(void) 1156 { 1157 unsigned int cpu, node; 1158 1159 /* Mark all except the boot CPU as hotpluggable */ 1160 for_each_possible_cpu(cpu) { 1161 if (cpu) 1162 per_cpu(cpu_info.cpu_index, cpu) = nr_cpu_ids; 1163 } 1164 1165 for_each_possible_cpu(cpu) { 1166 node = cpu_to_node(cpu); 1167 1168 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu), GFP_KERNEL, node); 1169 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu), GFP_KERNEL, node); 1170 zalloc_cpumask_var_node(&per_cpu(cpu_die_map, cpu), GFP_KERNEL, node); 1171 zalloc_cpumask_var_node(&per_cpu(cpu_llc_shared_map, cpu), GFP_KERNEL, node); 1172 zalloc_cpumask_var_node(&per_cpu(cpu_l2c_shared_map, cpu), GFP_KERNEL, node); 1173 } 1174 1175 set_cpu_sibling_map(0); 1176 } 1177 1178 void __init smp_prepare_boot_cpu(void) 1179 { 1180 smp_ops.smp_prepare_boot_cpu(); 1181 } 1182 1183 #ifdef CONFIG_X86_64 1184 /* Establish whether parallel bringup can be supported. */ 1185 bool __init arch_cpuhp_init_parallel_bringup(void) 1186 { 1187 if (!x86_cpuinit.parallel_bringup) { 1188 pr_info("Parallel CPU startup disabled by the platform\n"); 1189 return false; 1190 } 1191 1192 smpboot_control = STARTUP_READ_APICID; 1193 pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control); 1194 return true; 1195 } 1196 #endif 1197 1198 /* 1199 * Prepare for SMP bootup. 1200 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter 1201 * for common interface support. 1202 */ 1203 void __init native_smp_prepare_cpus(unsigned int max_cpus) 1204 { 1205 smp_prepare_cpus_common(); 1206 1207 switch (apic_intr_mode) { 1208 case APIC_PIC: 1209 case APIC_VIRTUAL_WIRE_NO_CONFIG: 1210 disable_smp(); 1211 return; 1212 case APIC_SYMMETRIC_IO_NO_ROUTING: 1213 disable_smp(); 1214 /* Setup local timer */ 1215 x86_init.timers.setup_percpu_clockev(); 1216 return; 1217 case APIC_VIRTUAL_WIRE: 1218 case APIC_SYMMETRIC_IO: 1219 break; 1220 } 1221 1222 /* Setup local timer */ 1223 x86_init.timers.setup_percpu_clockev(); 1224 1225 pr_info("CPU0: "); 1226 print_cpu_info(&cpu_data(0)); 1227 1228 uv_system_init(); 1229 1230 smp_set_init_udelay(); 1231 1232 speculative_store_bypass_ht_init(); 1233 1234 snp_set_wakeup_secondary_cpu(); 1235 } 1236 1237 void arch_thaw_secondary_cpus_begin(void) 1238 { 1239 set_cache_aps_delayed_init(true); 1240 } 1241 1242 void arch_thaw_secondary_cpus_end(void) 1243 { 1244 cache_aps_init(); 1245 } 1246 1247 /* 1248 * Early setup to make printk work. 1249 */ 1250 void __init native_smp_prepare_boot_cpu(void) 1251 { 1252 int me = smp_processor_id(); 1253 1254 /* SMP handles this from setup_per_cpu_areas() */ 1255 if (!IS_ENABLED(CONFIG_SMP)) 1256 switch_gdt_and_percpu_base(me); 1257 1258 native_pv_lock_init(); 1259 } 1260 1261 void __init native_smp_cpus_done(unsigned int max_cpus) 1262 { 1263 pr_debug("Boot done\n"); 1264 1265 build_sched_topology(); 1266 nmi_selftest(); 1267 impress_friends(); 1268 cache_aps_init(); 1269 } 1270 1271 /* correctly size the local cpu masks */ 1272 void __init setup_cpu_local_masks(void) 1273 { 1274 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask); 1275 } 1276 1277 #ifdef CONFIG_HOTPLUG_CPU 1278 1279 /* Recompute SMT state for all CPUs on offline */ 1280 static void recompute_smt_state(void) 1281 { 1282 int max_threads, cpu; 1283 1284 max_threads = 0; 1285 for_each_online_cpu (cpu) { 1286 int threads = cpumask_weight(topology_sibling_cpumask(cpu)); 1287 1288 if (threads > max_threads) 1289 max_threads = threads; 1290 } 1291 __max_smt_threads = max_threads; 1292 } 1293 1294 static void remove_siblinginfo(int cpu) 1295 { 1296 int sibling; 1297 struct cpuinfo_x86 *c = &cpu_data(cpu); 1298 1299 for_each_cpu(sibling, topology_core_cpumask(cpu)) { 1300 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling)); 1301 /*/ 1302 * last thread sibling in this cpu core going down 1303 */ 1304 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1) 1305 cpu_data(sibling).booted_cores--; 1306 } 1307 1308 for_each_cpu(sibling, topology_die_cpumask(cpu)) 1309 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling)); 1310 1311 for_each_cpu(sibling, topology_sibling_cpumask(cpu)) { 1312 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling)); 1313 if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1) 1314 cpu_data(sibling).smt_active = false; 1315 } 1316 1317 for_each_cpu(sibling, cpu_llc_shared_mask(cpu)) 1318 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling)); 1319 for_each_cpu(sibling, cpu_l2c_shared_mask(cpu)) 1320 cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling)); 1321 cpumask_clear(cpu_llc_shared_mask(cpu)); 1322 cpumask_clear(cpu_l2c_shared_mask(cpu)); 1323 cpumask_clear(topology_sibling_cpumask(cpu)); 1324 cpumask_clear(topology_core_cpumask(cpu)); 1325 cpumask_clear(topology_die_cpumask(cpu)); 1326 c->topo.core_id = 0; 1327 c->booted_cores = 0; 1328 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask); 1329 recompute_smt_state(); 1330 } 1331 1332 static void remove_cpu_from_maps(int cpu) 1333 { 1334 set_cpu_online(cpu, false); 1335 numa_remove_cpu(cpu); 1336 } 1337 1338 void cpu_disable_common(void) 1339 { 1340 int cpu = smp_processor_id(); 1341 1342 remove_siblinginfo(cpu); 1343 1344 /* 1345 * Stop allowing kernel-mode FPU. This is needed so that if the CPU is 1346 * brought online again, the initial state is not allowed: 1347 */ 1348 this_cpu_write(kernel_fpu_allowed, false); 1349 1350 /* It's now safe to remove this processor from the online map */ 1351 lock_vector_lock(); 1352 remove_cpu_from_maps(cpu); 1353 unlock_vector_lock(); 1354 fixup_irqs(); 1355 lapic_offline(); 1356 } 1357 1358 int native_cpu_disable(void) 1359 { 1360 int ret; 1361 1362 ret = lapic_can_unplug_cpu(); 1363 if (ret) 1364 return ret; 1365 1366 cpu_disable_common(); 1367 1368 /* 1369 * Disable the local APIC. Otherwise IPI broadcasts will reach 1370 * it. It still responds normally to INIT, NMI, SMI, and SIPI 1371 * messages. 1372 * 1373 * Disabling the APIC must happen after cpu_disable_common() 1374 * which invokes fixup_irqs(). 1375 * 1376 * Disabling the APIC preserves already set bits in IRR, but 1377 * an interrupt arriving after disabling the local APIC does not 1378 * set the corresponding IRR bit. 1379 * 1380 * fixup_irqs() scans IRR for set bits so it can raise a not 1381 * yet handled interrupt on the new destination CPU via an IPI 1382 * but obviously it can't do so for IRR bits which are not set. 1383 * IOW, interrupts arriving after disabling the local APIC will 1384 * be lost. 1385 */ 1386 apic_soft_disable(); 1387 1388 return 0; 1389 } 1390 1391 void play_dead_common(void) 1392 { 1393 idle_task_exit(); 1394 1395 cpuhp_ap_report_dead(); 1396 1397 local_irq_disable(); 1398 } 1399 1400 /* 1401 * We need to flush the caches before going to sleep, lest we have 1402 * dirty data in our caches when we come back up. 1403 */ 1404 void __noreturn mwait_play_dead(unsigned int eax_hint) 1405 { 1406 struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead); 1407 1408 /* Set up state for the kexec() hack below */ 1409 md->status = CPUDEAD_MWAIT_WAIT; 1410 md->control = CPUDEAD_MWAIT_WAIT; 1411 1412 wbinvd(); 1413 1414 while (1) { 1415 /* 1416 * The CLFLUSH is a workaround for erratum AAI65 for 1417 * the Xeon 7400 series. It's not clear it is actually 1418 * needed, but it should be harmless in either case. 1419 * The WBINVD is insufficient due to the spurious-wakeup 1420 * case where we return around the loop. 1421 */ 1422 mb(); 1423 clflush(md); 1424 mb(); 1425 __monitor(md, 0, 0); 1426 mb(); 1427 __mwait(eax_hint, 0); 1428 1429 if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) { 1430 /* 1431 * Kexec is about to happen. Don't go back into mwait() as 1432 * the kexec kernel might overwrite text and data including 1433 * page tables and stack. So mwait() would resume when the 1434 * monitor cache line is written to and then the CPU goes 1435 * south due to overwritten text, page tables and stack. 1436 * 1437 * Note: This does _NOT_ protect against a stray MCE, NMI, 1438 * SMI. They will resume execution at the instruction 1439 * following the HLT instruction and run into the problem 1440 * which this is trying to prevent. 1441 */ 1442 WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT); 1443 while(1) 1444 native_halt(); 1445 } 1446 } 1447 } 1448 1449 /* 1450 * Kick all "offline" CPUs out of mwait on kexec(). See comment in 1451 * mwait_play_dead(). 1452 */ 1453 void smp_kick_mwait_play_dead(void) 1454 { 1455 u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT; 1456 struct mwait_cpu_dead *md; 1457 unsigned int cpu, i; 1458 1459 for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) { 1460 md = per_cpu_ptr(&mwait_cpu_dead, cpu); 1461 1462 /* Does it sit in mwait_play_dead() ? */ 1463 if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT) 1464 continue; 1465 1466 /* Wait up to 5ms */ 1467 for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) { 1468 /* Bring it out of mwait */ 1469 WRITE_ONCE(md->control, newstate); 1470 udelay(5); 1471 } 1472 1473 if (READ_ONCE(md->status) != newstate) 1474 pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu); 1475 } 1476 } 1477 1478 void __noreturn hlt_play_dead(void) 1479 { 1480 if (__this_cpu_read(cpu_info.x86) >= 4) 1481 wbinvd(); 1482 1483 while (1) 1484 native_halt(); 1485 } 1486 1487 void __noreturn native_play_dead(void) 1488 { 1489 if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS)) 1490 __update_spec_ctrl(0); 1491 1492 play_dead_common(); 1493 tboot_shutdown(TB_SHUTDOWN_WFS); 1494 1495 /* Below returns only on error. */ 1496 cpuidle_play_dead(); 1497 hlt_play_dead(); 1498 } 1499 1500 #else /* ... !CONFIG_HOTPLUG_CPU */ 1501 int native_cpu_disable(void) 1502 { 1503 return -ENOSYS; 1504 } 1505 1506 void __noreturn native_play_dead(void) 1507 { 1508 BUG(); 1509 } 1510 1511 #endif 1512