kernel/cpu.c

   1 /* CPU control.
   2  * (C) 2001, 2002, 2003, 2004 Rusty Russell
   3  *
   4  * This code is licenced under the GPL.
   5  */
   6 #include <linux/sched/mm.h>
   7 #include <linux/proc_fs.h>
   8 #include <linux/smp.h>
   9 #include <linux/init.h>
  10 #include <linux/notifier.h>
  11 #include <linux/sched/signal.h>
  12 #include <linux/sched/hotplug.h>
  13 #include <linux/sched/isolation.h>
  14 #include <linux/sched/task.h>
  15 #include <linux/sched/smt.h>
  16 #include <linux/unistd.h>
  17 #include <linux/cpu.h>
  18 #include <linux/oom.h>
  19 #include <linux/rcupdate.h>
  20 #include <linux/export.h>
  21 #include <linux/bug.h>
  22 #include <linux/kthread.h>
  23 #include <linux/stop_machine.h>
  24 #include <linux/mutex.h>
  25 #include <linux/gfp.h>
  26 #include <linux/suspend.h>
  27 #include <linux/lockdep.h>
  28 #include <linux/tick.h>
  29 #include <linux/irq.h>
  30 #include <linux/nmi.h>
  31 #include <linux/smpboot.h>
  32 #include <linux/relay.h>
  33 #include <linux/slab.h>
  34 #include <linux/scs.h>
  35 #include <linux/percpu-rwsem.h>
  36 #include <linux/cpuset.h>
  37 #include <linux/random.h>
  38
  39 #include <trace/events/power.h>
  40 #define CREATE_TRACE_POINTS
  41 #include <trace/events/cpuhp.h>
  42
  43 #include "smpboot.h"
  44
  45 /**
  46  * cpuhp_cpu_state - Per cpu hotplug state storage
  47  * @state:      The current cpu state
  48  * @target:     The target state
  49  * @thread:     Pointer to the hotplug thread
  50  * @should_run: Thread should execute
  51  * @rollback:   Perform a rollback
  52  * @single:     Single callback invocation
  53  * @bringup:    Single callback bringup or teardown selector
  54  * @cb_state:   The state for a single callback (install/uninstall)
  55  * @result:     Result of the operation
  56  * @done_up:    Signal completion to the issuer of the task for cpu-up
  57  * @done_down:  Signal completion to the issuer of the task for cpu-down
  58  */
  59 struct cpuhp_cpu_state {
  60         enum cpuhp_state        state;
  61         enum cpuhp_state        target;
  62         enum cpuhp_state        fail;
  63 #ifdef CONFIG_SMP
  64         struct task_struct      *thread;
  65         bool                    should_run;
  66         bool                    rollback;
  67         bool                    single;
  68         bool                    bringup;
  69         struct hlist_node       *node;
  70         struct hlist_node       *last;
  71         enum cpuhp_state        cb_state;
  72         int                     result;
  73         struct completion       done_up;
  74         struct completion       done_down;
  75 #endif
  76 };
  77
  78 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  79         .fail = CPUHP_INVALID,
  80 };
  81
  82 #ifdef CONFIG_SMP
  83 cpumask_t cpus_booted_once_mask;
  84 #endif
  85
  86 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  87 static struct lockdep_map cpuhp_state_up_map =
  88         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  89 static struct lockdep_map cpuhp_state_down_map =
  90         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  91
  92
  93 static inline void cpuhp_lock_acquire(bool bringup)
  94 {
  95         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  96 }
  97
  98 static inline void cpuhp_lock_release(bool bringup)
  99 {
 100         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
 101 }
 102 #else
 103
 104 static inline void cpuhp_lock_acquire(bool bringup) { }
 105 static inline void cpuhp_lock_release(bool bringup) { }
 106
 107 #endif
 108
 109 /**
 110  * cpuhp_step - Hotplug state machine step
 111  * @name:       Name of the step
 112  * @startup:    Startup function of the step
 113  * @teardown:   Teardown function of the step
 114  * @cant_stop:  Bringup/teardown can't be stopped at this step
 115  */
 116 struct cpuhp_step {
 117         const char              *name;
 118         union {
 119                 int             (*single)(unsigned int cpu);
 120                 int             (*multi)(unsigned int cpu,
 121                                          struct hlist_node *node);
 122         } startup;
 123         union {
 124                 int             (*single)(unsigned int cpu);
 125                 int             (*multi)(unsigned int cpu,
 126                                          struct hlist_node *node);
 127         } teardown;
 128         struct hlist_head       list;
 129         bool                    cant_stop;
 130         bool                    multi_instance;
 131 };
 132
 133 static DEFINE_MUTEX(cpuhp_state_mutex);
 134 static struct cpuhp_step cpuhp_hp_states[];
 135
 136 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 137 {
 138         return cpuhp_hp_states + state;
 139 }
 140
 141 /**
 142  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 143  * @cpu:        The cpu for which the callback should be invoked
 144  * @state:      The state to do callbacks for
 145  * @bringup:    True if the bringup callback should be invoked
 146  * @node:       For multi-instance, do a single entry callback for install/remove
 147  * @lastp:      For multi-instance rollback, remember how far we got
 148  *
 149  * Called from cpu hotplug and from the state register machinery.
 150  */
 151 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 152                                  bool bringup, struct hlist_node *node,
 153                                  struct hlist_node **lastp)
 154 {
 155         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 156         struct cpuhp_step *step = cpuhp_get_step(state);
 157         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 158         int (*cb)(unsigned int cpu);
 159         int ret, cnt;
 160
 161         if (st->fail == state) {
 162                 st->fail = CPUHP_INVALID;
 163
 164                 if (!(bringup ? step->startup.single : step->teardown.single))
 165                         return 0;
 166
 167                 return -EAGAIN;
 168         }
 169
 170         if (!step->multi_instance) {
 171                 WARN_ON_ONCE(lastp && *lastp);
 172                 cb = bringup ? step->startup.single : step->teardown.single;
 173                 if (!cb)
 174                         return 0;
 175                 trace_cpuhp_enter(cpu, st->target, state, cb);
 176                 ret = cb(cpu);
 177                 trace_cpuhp_exit(cpu, st->state, state, ret);
 178                 return ret;
 179         }
 180         cbm = bringup ? step->startup.multi : step->teardown.multi;
 181         if (!cbm)
 182                 return 0;
 183
 184         /* Single invocation for instance add/remove */
 185         if (node) {
 186                 WARN_ON_ONCE(lastp && *lastp);
 187                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 188                 ret = cbm(cpu, node);
 189                 trace_cpuhp_exit(cpu, st->state, state, ret);
 190                 return ret;
 191         }
 192
 193         /* State transition. Invoke on all instances */
 194         cnt = 0;
 195         hlist_for_each(node, &step->list) {
 196                 if (lastp && node == *lastp)
 197                         break;
 198
 199                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 200                 ret = cbm(cpu, node);
 201                 trace_cpuhp_exit(cpu, st->state, state, ret);
 202                 if (ret) {
 203                         if (!lastp)
 204                                 goto err;
 205
 206                         *lastp = node;
 207                         return ret;
 208                 }
 209                 cnt++;
 210         }
 211         if (lastp)
 212                 *lastp = NULL;
 213         return 0;
 214 err:
 215         /* Rollback the instances if one failed */
 216         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 217         if (!cbm)
 218                 return ret;
 219
 220         hlist_for_each(node, &step->list) {
 221                 if (!cnt--)
 222                         break;
 223
 224                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 225                 ret = cbm(cpu, node);
 226                 trace_cpuhp_exit(cpu, st->state, state, ret);
 227                 /*
 228                  * Rollback must not fail,
 229                  */
 230                 WARN_ON_ONCE(ret);
 231         }
 232         return ret;
 233 }
 234
 235 #ifdef CONFIG_SMP
 236 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 237 {
 238         /*
 239          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 240          * purposes as that state is handled explicitly in cpu_down.
 241          */
 242         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 243 }
 244
 245 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 246 {
 247         struct completion *done = bringup ? &st->done_up : &st->done_down;
 248         wait_for_completion(done);
 249 }
 250
 251 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 252 {
 253         struct completion *done = bringup ? &st->done_up : &st->done_down;
 254         complete(done);
 255 }
 256
 257 /*
 258  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 259  */
 260 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 261 {
 262         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 263 }
 264
 265 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 266 static DEFINE_MUTEX(cpu_add_remove_lock);
 267 bool cpuhp_tasks_frozen;
 268 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 269
 270 /*
 271  * The following two APIs (cpu_maps_update_begin/done) must be used when
 272  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 273  */
 274 void cpu_maps_update_begin(void)
 275 {
 276         mutex_lock(&cpu_add_remove_lock);
 277 }
 278
 279 void cpu_maps_update_done(void)
 280 {
 281         mutex_unlock(&cpu_add_remove_lock);
 282 }
 283
 284 /*
 285  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 286  * Should always be manipulated under cpu_add_remove_lock
 287  */
 288 static int cpu_hotplug_disabled;
 289
 290 #ifdef CONFIG_HOTPLUG_CPU
 291
 292 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 293
 294 void cpus_read_lock(void)
 295 {
 296         percpu_down_read(&cpu_hotplug_lock);
 297 }
 298 EXPORT_SYMBOL_GPL(cpus_read_lock);
 299
 300 int cpus_read_trylock(void)
 301 {
 302         return percpu_down_read_trylock(&cpu_hotplug_lock);
 303 }
 304 EXPORT_SYMBOL_GPL(cpus_read_trylock);
 305
 306 void cpus_read_unlock(void)
 307 {
 308         percpu_up_read(&cpu_hotplug_lock);
 309 }
 310 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 311
 312 void cpus_write_lock(void)
 313 {
 314         percpu_down_write(&cpu_hotplug_lock);
 315 }
 316
 317 void cpus_write_unlock(void)
 318 {
 319         percpu_up_write(&cpu_hotplug_lock);
 320 }
 321
 322 void lockdep_assert_cpus_held(void)
 323 {
 324         /*
 325          * We can't have hotplug operations before userspace starts running,
 326          * and some init codepaths will knowingly not take the hotplug lock.
 327          * This is all valid, so mute lockdep until it makes sense to report
 328          * unheld locks.
 329          */
 330         if (system_state < SYSTEM_RUNNING)
 331                 return;
 332
 333         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 334 }
 335
 336 static void lockdep_acquire_cpus_lock(void)
 337 {
 338         rwsem_acquire(&cpu_hotplug_lock.dep_map, 0, 0, _THIS_IP_);
 339 }
 340
 341 static void lockdep_release_cpus_lock(void)
 342 {
 343         rwsem_release(&cpu_hotplug_lock.dep_map, _THIS_IP_);
 344 }
 345
 346 /*
 347  * Wait for currently running CPU hotplug operations to complete (if any) and
 348  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 349  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 350  * hotplug path before performing hotplug operations. So acquiring that lock
 351  * guarantees mutual exclusion from any currently running hotplug operations.
 352  */
 353 void cpu_hotplug_disable(void)
 354 {
 355         cpu_maps_update_begin();
 356         cpu_hotplug_disabled++;
 357         cpu_maps_update_done();
 358 }
 359 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 360
 361 static void __cpu_hotplug_enable(void)
 362 {
 363         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 364                 return;
 365         cpu_hotplug_disabled--;
 366 }
 367
 368 void cpu_hotplug_enable(void)
 369 {
 370         cpu_maps_update_begin();
 371         __cpu_hotplug_enable();
 372         cpu_maps_update_done();
 373 }
 374 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 375
 376 #else
 377
 378 static void lockdep_acquire_cpus_lock(void)
 379 {
 380 }
 381
 382 static void lockdep_release_cpus_lock(void)
 383 {
 384 }
 385
 386 #endif  /* CONFIG_HOTPLUG_CPU */
 387
 388 /*
 389  * Architectures that need SMT-specific errata handling during SMT hotplug
 390  * should override this.
 391  */
 392 void __weak arch_smt_update(void) { }
 393
 394 #ifdef CONFIG_HOTPLUG_SMT
 395 enum cpuhp_smt_control cpu_smt_control __read_mostly = CPU_SMT_ENABLED;
 396
 397 void __init cpu_smt_disable(bool force)
 398 {
 399         if (!cpu_smt_possible())
 400                 return;
 401
 402         if (force) {
 403                 pr_info("SMT: Force disabled\n");
 404                 cpu_smt_control = CPU_SMT_FORCE_DISABLED;
 405         } else {
 406                 pr_info("SMT: disabled\n");
 407                 cpu_smt_control = CPU_SMT_DISABLED;
 408         }
 409 }
 410
 411 /*
 412  * The decision whether SMT is supported can only be done after the full
 413  * CPU identification. Called from architecture code.
 414  */
 415 void __init cpu_smt_check_topology(void)
 416 {
 417         if (!topology_smt_supported())
 418                 cpu_smt_control = CPU_SMT_NOT_SUPPORTED;
 419 }
 420
 421 static int __init smt_cmdline_disable(char *str)
 422 {
 423         cpu_smt_disable(str && !strcmp(str, "force"));
 424         return 0;
 425 }
 426 early_param("nosmt", smt_cmdline_disable);
 427
 428 static inline bool cpu_smt_allowed(unsigned int cpu)
 429 {
 430         if (cpu_smt_control == CPU_SMT_ENABLED)
 431                 return true;
 432
 433         if (topology_is_primary_thread(cpu))
 434                 return true;
 435
 436         /*
 437          * On x86 it's required to boot all logical CPUs at least once so
 438          * that the init code can get a chance to set CR4.MCE on each
 439          * CPU. Otherwise, a broadcasted MCE observing CR4.MCE=0b on any
 440          * core will shutdown the machine.
 441          */
 442         return !cpumask_test_cpu(cpu, &cpus_booted_once_mask);
 443 }
 444
 445 /* Returns true if SMT is not supported of forcefully (irreversibly) disabled */
 446 bool cpu_smt_possible(void)
 447 {
 448         return cpu_smt_control != CPU_SMT_FORCE_DISABLED &&
 449                 cpu_smt_control != CPU_SMT_NOT_SUPPORTED;
 450 }
 451 EXPORT_SYMBOL_GPL(cpu_smt_possible);
 452 #else
 453 static inline bool cpu_smt_allowed(unsigned int cpu) { return true; }
 454 #endif
 455
 456 static inline enum cpuhp_state
 457 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 458 {
 459         enum cpuhp_state prev_state = st->state;
 460
 461         st->rollback = false;
 462         st->last = NULL;
 463
 464         st->target = target;
 465         st->single = false;
 466         st->bringup = st->state < target;
 467
 468         return prev_state;
 469 }
 470
 471 static inline void
 472 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 473 {
 474         st->rollback = true;
 475
 476         /*
 477          * If we have st->last we need to undo partial multi_instance of this
 478          * state first. Otherwise start undo at the previous state.
 479          */
 480         if (!st->last) {
 481                 if (st->bringup)
 482                         st->state--;
 483                 else
 484                         st->state++;
 485         }
 486
 487         st->target = prev_state;
 488         st->bringup = !st->bringup;
 489 }
 490
 491 /* Regular hotplug invocation of the AP hotplug thread */
 492 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 493 {
 494         if (!st->single && st->state == st->target)
 495                 return;
 496
 497         st->result = 0;
 498         /*
 499          * Make sure the above stores are visible before should_run becomes
 500          * true. Paired with the mb() above in cpuhp_thread_fun()
 501          */
 502         smp_mb();
 503         st->should_run = true;
 504         wake_up_process(st->thread);
 505         wait_for_ap_thread(st, st->bringup);
 506 }
 507
 508 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 509 {
 510         enum cpuhp_state prev_state;
 511         int ret;
 512
 513         prev_state = cpuhp_set_state(st, target);
 514         __cpuhp_kick_ap(st);
 515         if ((ret = st->result)) {
 516                 cpuhp_reset_state(st, prev_state);
 517                 __cpuhp_kick_ap(st);
 518         }
 519
 520         return ret;
 521 }
 522
 523 static int bringup_wait_for_ap(unsigned int cpu)
 524 {
 525         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 526
 527         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 528         wait_for_ap_thread(st, true);
 529         if (WARN_ON_ONCE((!cpu_online(cpu))))
 530                 return -ECANCELED;
 531
 532         /* Unpark the hotplug thread of the target cpu */
 533         kthread_unpark(st->thread);
 534
 535         /*
 536          * SMT soft disabling on X86 requires to bring the CPU out of the
 537          * BIOS 'wait for SIPI' state in order to set the CR4.MCE bit.  The
 538          * CPU marked itself as booted_once in notify_cpu_starting() so the
 539          * cpu_smt_allowed() check will now return false if this is not the
 540          * primary sibling.
 541          */
 542         if (!cpu_smt_allowed(cpu))
 543                 return -ECANCELED;
 544
 545         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 546                 return 0;
 547
 548         return cpuhp_kick_ap(st, st->target);
 549 }
 550
 551 static int bringup_cpu(unsigned int cpu)
 552 {
 553         struct task_struct *idle = idle_thread_get(cpu);
 554         int ret;
 555
 556         /*
 557          * Reset stale stack state from the last time this CPU was online.
 558          */
 559         scs_task_reset(idle);
 560         kasan_unpoison_task_stack(idle);
 561
 562         /*
 563          * Some architectures have to walk the irq descriptors to
 564          * setup the vector space for the cpu which comes online.
 565          * Prevent irq alloc/free across the bringup.
 566          */
 567         irq_lock_sparse();
 568
 569         /* Arch-specific enabling code. */
 570         ret = __cpu_up(cpu, idle);
 571         irq_unlock_sparse();
 572         if (ret)
 573                 return ret;
 574         return bringup_wait_for_ap(cpu);
 575 }
 576
 577 static int finish_cpu(unsigned int cpu)
 578 {
 579         struct task_struct *idle = idle_thread_get(cpu);
 580         struct mm_struct *mm = idle->active_mm;
 581
 582         /*
 583          * idle_task_exit() will have switched to &init_mm, now
 584          * clean up any remaining active_mm state.
 585          */
 586         if (mm != &init_mm)
 587                 idle->active_mm = &init_mm;
 588         mmdrop(mm);
 589         return 0;
 590 }
 591
 592 /*
 593  * Hotplug state machine related functions
 594  */
 595
 596 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 597 {
 598         for (st->state--; st->state > st->target; st->state--)
 599                 cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 600 }
 601
 602 static inline bool can_rollback_cpu(struct cpuhp_cpu_state *st)
 603 {
 604         if (IS_ENABLED(CONFIG_HOTPLUG_CPU))
 605                 return true;
 606         /*
 607          * When CPU hotplug is disabled, then taking the CPU down is not
 608          * possible because takedown_cpu() and the architecture and
 609          * subsystem specific mechanisms are not available. So the CPU
 610          * which would be completely unplugged again needs to stay around
 611          * in the current state.
 612          */
 613         return st->state <= CPUHP_BRINGUP_CPU;
 614 }
 615
 616 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 617                               enum cpuhp_state target)
 618 {
 619         enum cpuhp_state prev_state = st->state;
 620         int ret = 0;
 621
 622         while (st->state < target) {
 623                 st->state++;
 624                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 625                 if (ret) {
 626                         if (can_rollback_cpu(st)) {
 627                                 st->target = prev_state;
 628                                 undo_cpu_up(cpu, st);
 629                         }
 630                         break;
 631                 }
 632         }
 633         return ret;
 634 }
 635
 636 /*
 637  * The cpu hotplug threads manage the bringup and teardown of the cpus
 638  */
 639 static void cpuhp_create(unsigned int cpu)
 640 {
 641         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 642
 643         init_completion(&st->done_up);
 644         init_completion(&st->done_down);
 645 }
 646
 647 static int cpuhp_should_run(unsigned int cpu)
 648 {
 649         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 650
 651         return st->should_run;
 652 }
 653
 654 /*
 655  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 656  * callbacks when a state gets [un]installed at runtime.
 657  *
 658  * Each invocation of this function by the smpboot thread does a single AP
 659  * state callback.
 660  *
 661  * It has 3 modes of operation:
 662  *  - single: runs st->cb_state
 663  *  - up:     runs ++st->state, while st->state < st->target
 664  *  - down:   runs st->state--, while st->state > st->target
 665  *
 666  * When complete or on error, should_run is cleared and the completion is fired.
 667  */
 668 static void cpuhp_thread_fun(unsigned int cpu)
 669 {
 670         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 671         bool bringup = st->bringup;
 672         enum cpuhp_state state;
 673
 674         if (WARN_ON_ONCE(!st->should_run))
 675                 return;
 676
 677         /*
 678          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 679          * that if we see ->should_run we also see the rest of the state.
 680          */
 681         smp_mb();
 682
 683         /*
 684          * The BP holds the hotplug lock, but we're now running on the AP,
 685          * ensure that anybody asserting the lock is held, will actually find
 686          * it so.
 687          */
 688         lockdep_acquire_cpus_lock();
 689         cpuhp_lock_acquire(bringup);
 690
 691         if (st->single) {
 692                 state = st->cb_state;
 693                 st->should_run = false;
 694         } else {
 695                 if (bringup) {
 696                         st->state++;
 697                         state = st->state;
 698                         st->should_run = (st->state < st->target);
 699                         WARN_ON_ONCE(st->state > st->target);
 700                 } else {
 701                         state = st->state;
 702                         st->state--;
 703                         st->should_run = (st->state > st->target);
 704                         WARN_ON_ONCE(st->state < st->target);
 705                 }
 706         }
 707
 708         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 709
 710         if (cpuhp_is_atomic_state(state)) {
 711                 local_irq_disable();
 712                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 713                 local_irq_enable();
 714
 715                 /*
 716                  * STARTING/DYING must not fail!
 717                  */
 718                 WARN_ON_ONCE(st->result);
 719         } else {
 720                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 721         }
 722
 723         if (st->result) {
 724                 /*
 725                  * If we fail on a rollback, we're up a creek without no
 726                  * paddle, no way forward, no way back. We loose, thanks for
 727                  * playing.
 728                  */
 729                 WARN_ON_ONCE(st->rollback);
 730                 st->should_run = false;
 731         }
 732
 733         cpuhp_lock_release(bringup);
 734         lockdep_release_cpus_lock();
 735
 736         if (!st->should_run)
 737                 complete_ap_thread(st, bringup);
 738 }
 739
 740 /* Invoke a single callback on a remote cpu */
 741 static int
 742 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 743                          struct hlist_node *node)
 744 {
 745         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 746         int ret;
 747
 748         if (!cpu_online(cpu))
 749                 return 0;
 750
 751         cpuhp_lock_acquire(false);
 752         cpuhp_lock_release(false);
 753
 754         cpuhp_lock_acquire(true);
 755         cpuhp_lock_release(true);
 756
 757         /*
 758          * If we are up and running, use the hotplug thread. For early calls
 759          * we invoke the thread function directly.
 760          */
 761         if (!st->thread)
 762                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 763
 764         st->rollback = false;
 765         st->last = NULL;
 766
 767         st->node = node;
 768         st->bringup = bringup;
 769         st->cb_state = state;
 770         st->single = true;
 771
 772         __cpuhp_kick_ap(st);
 773
 774         /*
 775          * If we failed and did a partial, do a rollback.
 776          */
 777         if ((ret = st->result) && st->last) {
 778                 st->rollback = true;
 779                 st->bringup = !bringup;
 780
 781                 __cpuhp_kick_ap(st);
 782         }
 783
 784         /*
 785          * Clean up the leftovers so the next hotplug operation wont use stale
 786          * data.
 787          */
 788         st->node = st->last = NULL;
 789         return ret;
 790 }
 791
 792 static int cpuhp_kick_ap_work(unsigned int cpu)
 793 {
 794         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 795         enum cpuhp_state prev_state = st->state;
 796         int ret;
 797
 798         cpuhp_lock_acquire(false);
 799         cpuhp_lock_release(false);
 800
 801         cpuhp_lock_acquire(true);
 802         cpuhp_lock_release(true);
 803
 804         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 805         ret = cpuhp_kick_ap(st, st->target);
 806         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 807
 808         return ret;
 809 }
 810
 811 static struct smp_hotplug_thread cpuhp_threads = {
 812         .store                  = &cpuhp_state.thread,
 813         .create                 = &cpuhp_create,
 814         .thread_should_run      = cpuhp_should_run,
 815         .thread_fn              = cpuhp_thread_fun,
 816         .thread_comm            = "cpuhp/%u",
 817         .selfparking            = true,
 818 };
 819
 820 void __init cpuhp_threads_init(void)
 821 {
 822         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 823         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 824 }
 825
 826 /*
 827  *
 828  * Serialize hotplug trainwrecks outside of the cpu_hotplug_lock
 829  * protected region.
 830  *
 831  * The operation is still serialized against concurrent CPU hotplug via
 832  * cpu_add_remove_lock, i.e. CPU map protection.  But it is _not_
 833  * serialized against other hotplug related activity like adding or
 834  * removing of state callbacks and state instances, which invoke either the
 835  * startup or the teardown callback of the affected state.
 836  *
 837  * This is required for subsystems which are unfixable vs. CPU hotplug and
 838  * evade lock inversion problems by scheduling work which has to be
 839  * completed _before_ cpu_up()/_cpu_down() returns.
 840  *
 841  * Don't even think about adding anything to this for any new code or even
 842  * drivers. It's only purpose is to keep existing lock order trainwrecks
 843  * working.
 844  *
 845  * For cpu_down() there might be valid reasons to finish cleanups which are
 846  * not required to be done under cpu_hotplug_lock, but that's a different
 847  * story and would be not invoked via this.
 848  */
 849 static void cpu_up_down_serialize_trainwrecks(bool tasks_frozen)
 850 {
 851         /*
 852          * cpusets delegate hotplug operations to a worker to "solve" the
 853          * lock order problems. Wait for the worker, but only if tasks are
 854          * _not_ frozen (suspend, hibernate) as that would wait forever.
 855          *
 856          * The wait is required because otherwise the hotplug operation
 857          * returns with inconsistent state, which could even be observed in
 858          * user space when a new CPU is brought up. The CPU plug uevent
 859          * would be delivered and user space reacting on it would fail to
 860          * move tasks to the newly plugged CPU up to the point where the
 861          * work has finished because up to that point the newly plugged CPU
 862          * is not assignable in cpusets/cgroups. On unplug that's not
 863          * necessarily a visible issue, but it is still inconsistent state,
 864          * which is the real problem which needs to be "fixed". This can't
 865          * prevent the transient state between scheduling the work and
 866          * returning from waiting for it.
 867          */
 868         if (!tasks_frozen)
 869                 cpuset_wait_for_hotplug();
 870 }
 871
 872 #ifdef CONFIG_HOTPLUG_CPU
 873 #ifndef arch_clear_mm_cpumask_cpu
 874 #define arch_clear_mm_cpumask_cpu(cpu, mm) cpumask_clear_cpu(cpu, mm_cpumask(mm))
 875 #endif
 876
 877 /**
 878  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 879  * @cpu: a CPU id
 880  *
 881  * This function walks all processes, finds a valid mm struct for each one and
 882  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 883  * trivial, there are various non-obvious corner cases, which this function
 884  * tries to solve in a safe manner.
 885  *
 886  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 887  * be called only for an already offlined CPU.
 888  */
 889 void clear_tasks_mm_cpumask(int cpu)
 890 {
 891         struct task_struct *p;
 892
 893         /*
 894          * This function is called after the cpu is taken down and marked
 895          * offline, so its not like new tasks will ever get this cpu set in
 896          * their mm mask. -- Peter Zijlstra
 897          * Thus, we may use rcu_read_lock() here, instead of grabbing
 898          * full-fledged tasklist_lock.
 899          */
 900         WARN_ON(cpu_online(cpu));
 901         rcu_read_lock();
 902         for_each_process(p) {
 903                 struct task_struct *t;
 904
 905                 /*
 906                  * Main thread might exit, but other threads may still have
 907                  * a valid mm. Find one.
 908                  */
 909                 t = find_lock_task_mm(p);
 910                 if (!t)
 911                         continue;
 912                 arch_clear_mm_cpumask_cpu(cpu, t->mm);
 913                 task_unlock(t);
 914         }
 915         rcu_read_unlock();
 916 }
 917
 918 /* Take this CPU down. */
 919 static int take_cpu_down(void *_param)
 920 {
 921         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 922         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 923         int err, cpu = smp_processor_id();
 924         int ret;
 925
 926         /* Ensure this CPU doesn't handle any more interrupts. */
 927         err = __cpu_disable();
 928         if (err < 0)
 929                 return err;
 930
 931         /*
 932          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 933          * do this step again.
 934          */
 935         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 936         st->state--;
 937         /* Invoke the former CPU_DYING callbacks */
 938         for (; st->state > target; st->state--) {
 939                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 940                 /*
 941                  * DYING must not fail!
 942                  */
 943                 WARN_ON_ONCE(ret);
 944         }
 945
 946         /* Give up timekeeping duties */
 947         tick_handover_do_timer();
 948         /* Remove CPU from timer broadcasting */
 949         tick_offline_cpu(cpu);
 950         /* Park the stopper thread */
 951         stop_machine_park(cpu);
 952         return 0;
 953 }
 954
 955 static int takedown_cpu(unsigned int cpu)
 956 {
 957         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 958         int err;
 959
 960         /* Park the smpboot threads */
 961         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 962
 963         /*
 964          * Prevent irq alloc/free while the dying cpu reorganizes the
 965          * interrupt affinities.
 966          */
 967         irq_lock_sparse();
 968
 969         /*
 970          * So now all preempt/rcu users must observe !cpu_active().
 971          */
 972         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 973         if (err) {
 974                 /* CPU refused to die */
 975                 irq_unlock_sparse();
 976                 /* Unpark the hotplug thread so we can rollback there */
 977                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 978                 return err;
 979         }
 980         BUG_ON(cpu_online(cpu));
 981
 982         /*
 983          * The teardown callback for CPUHP_AP_SCHED_STARTING will have removed
 984          * all runnable tasks from the CPU, there's only the idle task left now
 985          * that the migration thread is done doing the stop_machine thing.
 986          *
 987          * Wait for the stop thread to go away.
 988          */
 989         wait_for_ap_thread(st, false);
 990         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 991
 992         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 993         irq_unlock_sparse();
 994
 995         hotplug_cpu__broadcast_tick_pull(cpu);
 996         /* This actually kills the CPU. */
 997         __cpu_die(cpu);
 998
 999         tick_cleanup_dead_cpu(cpu);
1000         rcutree_migrate_callbacks(cpu);
1001         return 0;
1002 }
1003
1004 static void cpuhp_complete_idle_dead(void *arg)
1005 {
1006         struct cpuhp_cpu_state *st = arg;
1007
1008         complete_ap_thread(st, false);
1009 }
1010
1011 void cpuhp_report_idle_dead(void)
1012 {
1013         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1014
1015         BUG_ON(st->state != CPUHP_AP_OFFLINE);
1016         rcu_report_dead(smp_processor_id());
1017         st->state = CPUHP_AP_IDLE_DEAD;
1018         /*
1019          * We cannot call complete after rcu_report_dead() so we delegate it
1020          * to an online cpu.
1021          */
1022         smp_call_function_single(cpumask_first(cpu_online_mask),
1023                                  cpuhp_complete_idle_dead, st, 0);
1024 }
1025
1026 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
1027 {
1028         for (st->state++; st->state < st->target; st->state++)
1029                 cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1030 }
1031
1032 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
1033                                 enum cpuhp_state target)
1034 {
1035         enum cpuhp_state prev_state = st->state;
1036         int ret = 0;
1037
1038         for (; st->state > target; st->state--) {
1039                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
1040                 if (ret) {
1041                         st->target = prev_state;
1042                         if (st->state < prev_state)
1043                                 undo_cpu_down(cpu, st);
1044                         break;
1045                 }
1046         }
1047         return ret;
1048 }
1049
1050 /* Requires cpu_add_remove_lock to be held */
1051 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
1052                            enum cpuhp_state target)
1053 {
1054         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1055         int prev_state, ret = 0;
1056
1057         if (num_online_cpus() == 1)
1058                 return -EBUSY;
1059
1060         if (!cpu_present(cpu))
1061                 return -EINVAL;
1062
1063         cpus_write_lock();
1064
1065         cpuhp_tasks_frozen = tasks_frozen;
1066
1067         prev_state = cpuhp_set_state(st, target);
1068         /*
1069          * If the current CPU state is in the range of the AP hotplug thread,
1070          * then we need to kick the thread.
1071          */
1072         if (st->state > CPUHP_TEARDOWN_CPU) {
1073                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
1074                 ret = cpuhp_kick_ap_work(cpu);
1075                 /*
1076                  * The AP side has done the error rollback already. Just
1077                  * return the error code..
1078                  */
1079                 if (ret)
1080                         goto out;
1081
1082                 /*
1083                  * We might have stopped still in the range of the AP hotplug
1084                  * thread. Nothing to do anymore.
1085                  */
1086                 if (st->state > CPUHP_TEARDOWN_CPU)
1087                         goto out;
1088
1089                 st->target = target;
1090         }
1091         /*
1092          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
1093          * to do the further cleanups.
1094          */
1095         ret = cpuhp_down_callbacks(cpu, st, target);
1096         if (ret && st->state == CPUHP_TEARDOWN_CPU && st->state < prev_state) {
1097                 cpuhp_reset_state(st, prev_state);
1098                 __cpuhp_kick_ap(st);
1099         }
1100
1101 out:
1102         cpus_write_unlock();
1103         /*
1104          * Do post unplug cleanup. This is still protected against
1105          * concurrent CPU hotplug via cpu_add_remove_lock.
1106          */
1107         lockup_detector_cleanup();
1108         arch_smt_update();
1109         cpu_up_down_serialize_trainwrecks(tasks_frozen);
1110         return ret;
1111 }
1112
1113 static int cpu_down_maps_locked(unsigned int cpu, enum cpuhp_state target)
1114 {
1115         if (cpu_hotplug_disabled)
1116                 return -EBUSY;
1117         return _cpu_down(cpu, 0, target);
1118 }
1119
1120 static int cpu_down(unsigned int cpu, enum cpuhp_state target)
1121 {
1122         int err;
1123
1124         cpu_maps_update_begin();
1125         err = cpu_down_maps_locked(cpu, target);
1126         cpu_maps_update_done();
1127         return err;
1128 }
1129
1130 /**
1131  * cpu_device_down - Bring down a cpu device
1132  * @dev: Pointer to the cpu device to offline
1133  *
1134  * This function is meant to be used by device core cpu subsystem only.
1135  *
1136  * Other subsystems should use remove_cpu() instead.
1137  */
1138 int cpu_device_down(struct device *dev)
1139 {
1140         return cpu_down(dev->id, CPUHP_OFFLINE);
1141 }
1142
1143 int remove_cpu(unsigned int cpu)
1144 {
1145         int ret;
1146
1147         lock_device_hotplug();
1148         ret = device_offline(get_cpu_device(cpu));
1149         unlock_device_hotplug();
1150
1151         return ret;
1152 }
1153 EXPORT_SYMBOL_GPL(remove_cpu);
1154
1155 void smp_shutdown_nonboot_cpus(unsigned int primary_cpu)
1156 {
1157         unsigned int cpu;
1158         int error;
1159
1160         cpu_maps_update_begin();
1161
1162         /*
1163          * Make certain the cpu I'm about to reboot on is online.
1164          *
1165          * This is inline to what migrate_to_reboot_cpu() already do.
1166          */
1167         if (!cpu_online(primary_cpu))
1168                 primary_cpu = cpumask_first(cpu_online_mask);
1169
1170         for_each_online_cpu(cpu) {
1171                 if (cpu == primary_cpu)
1172                         continue;
1173
1174                 error = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
1175                 if (error) {
1176                         pr_err("Failed to offline CPU%d - error=%d",
1177                                 cpu, error);
1178                         break;
1179                 }
1180         }
1181
1182         /*
1183          * Ensure all but the reboot CPU are offline.
1184          */
1185         BUG_ON(num_online_cpus() > 1);
1186
1187         /*
1188          * Make sure the CPUs won't be enabled by someone else after this
1189          * point. Kexec will reboot to a new kernel shortly resetting
1190          * everything along the way.
1191          */
1192         cpu_hotplug_disabled++;
1193
1194         cpu_maps_update_done();
1195 }
1196
1197 #else
1198 #define takedown_cpu            NULL
1199 #endif /*CONFIG_HOTPLUG_CPU*/
1200
1201 /**
1202  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
1203  * @cpu: cpu that just started
1204  *
1205  * It must be called by the arch code on the new cpu, before the new cpu
1206  * enables interrupts and before the "boot" cpu returns from __cpu_up().
1207  */
1208 void notify_cpu_starting(unsigned int cpu)
1209 {
1210         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1211         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
1212         int ret;
1213
1214         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
1215         cpumask_set_cpu(cpu, &cpus_booted_once_mask);
1216         while (st->state < target) {
1217                 st->state++;
1218                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
1219                 /*
1220                  * STARTING must not fail!
1221                  */
1222                 WARN_ON_ONCE(ret);
1223         }
1224 }
1225
1226 /*
1227  * Called from the idle task. Wake up the controlling task which brings the
1228  * hotplug thread of the upcoming CPU up and then delegates the rest of the
1229  * online bringup to the hotplug thread.
1230  */
1231 void cpuhp_online_idle(enum cpuhp_state state)
1232 {
1233         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
1234
1235         /* Happens for the boot cpu */
1236         if (state != CPUHP_AP_ONLINE_IDLE)
1237                 return;
1238
1239         /*
1240          * Unpart the stopper thread before we start the idle loop (and start
1241          * scheduling); this ensures the stopper task is always available.
1242          */
1243         stop_machine_unpark(smp_processor_id());
1244
1245         st->state = CPUHP_AP_ONLINE_IDLE;
1246         complete_ap_thread(st, true);
1247 }
1248
1249 /* Requires cpu_add_remove_lock to be held */
1250 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1251 {
1252         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1253         struct task_struct *idle;
1254         int ret = 0;
1255
1256         cpus_write_lock();
1257
1258         if (!cpu_present(cpu)) {
1259                 ret = -EINVAL;
1260                 goto out;
1261         }
1262
1263         /*
1264          * The caller of cpu_up() might have raced with another
1265          * caller. Nothing to do.
1266          */
1267         if (st->state >= target)
1268                 goto out;
1269
1270         if (st->state == CPUHP_OFFLINE) {
1271                 /* Let it fail before we try to bring the cpu up */
1272                 idle = idle_thread_get(cpu);
1273                 if (IS_ERR(idle)) {
1274                         ret = PTR_ERR(idle);
1275                         goto out;
1276                 }
1277         }
1278
1279         cpuhp_tasks_frozen = tasks_frozen;
1280
1281         cpuhp_set_state(st, target);
1282         /*
1283          * If the current CPU state is in the range of the AP hotplug thread,
1284          * then we need to kick the thread once more.
1285          */
1286         if (st->state > CPUHP_BRINGUP_CPU) {
1287                 ret = cpuhp_kick_ap_work(cpu);
1288                 /*
1289                  * The AP side has done the error rollback already. Just
1290                  * return the error code..
1291                  */
1292                 if (ret)
1293                         goto out;
1294         }
1295
1296         /*
1297          * Try to reach the target state. We max out on the BP at
1298          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1299          * responsible for bringing it up to the target state.
1300          */
1301         target = min((int)target, CPUHP_BRINGUP_CPU);
1302         ret = cpuhp_up_callbacks(cpu, st, target);
1303 out:
1304         cpus_write_unlock();
1305         arch_smt_update();
1306         cpu_up_down_serialize_trainwrecks(tasks_frozen);
1307         return ret;
1308 }
1309
1310 static int cpu_up(unsigned int cpu, enum cpuhp_state target)
1311 {
1312         int err = 0;
1313
1314         if (!cpu_possible(cpu)) {
1315                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1316                        cpu);
1317 #if defined(CONFIG_IA64)
1318                 pr_err("please check additional_cpus= boot parameter\n");
1319 #endif
1320                 return -EINVAL;
1321         }
1322
1323         err = try_online_node(cpu_to_node(cpu));
1324         if (err)
1325                 return err;
1326
1327         cpu_maps_update_begin();
1328
1329         if (cpu_hotplug_disabled) {
1330                 err = -EBUSY;
1331                 goto out;
1332         }
1333         if (!cpu_smt_allowed(cpu)) {
1334                 err = -EPERM;
1335                 goto out;
1336         }
1337
1338         err = _cpu_up(cpu, 0, target);
1339 out:
1340         cpu_maps_update_done();
1341         return err;
1342 }
1343
1344 /**
1345  * cpu_device_up - Bring up a cpu device
1346  * @dev: Pointer to the cpu device to online
1347  *
1348  * This function is meant to be used by device core cpu subsystem only.
1349  *
1350  * Other subsystems should use add_cpu() instead.
1351  */
1352 int cpu_device_up(struct device *dev)
1353 {
1354         return cpu_up(dev->id, CPUHP_ONLINE);
1355 }
1356
1357 int add_cpu(unsigned int cpu)
1358 {
1359         int ret;
1360
1361         lock_device_hotplug();
1362         ret = device_online(get_cpu_device(cpu));
1363         unlock_device_hotplug();
1364
1365         return ret;
1366 }
1367 EXPORT_SYMBOL_GPL(add_cpu);
1368
1369 /**
1370  * bringup_hibernate_cpu - Bring up the CPU that we hibernated on
1371  * @sleep_cpu: The cpu we hibernated on and should be brought up.
1372  *
1373  * On some architectures like arm64, we can hibernate on any CPU, but on
1374  * wake up the CPU we hibernated on might be offline as a side effect of
1375  * using maxcpus= for example.
1376  */
1377 int bringup_hibernate_cpu(unsigned int sleep_cpu)
1378 {
1379         int ret;
1380
1381         if (!cpu_online(sleep_cpu)) {
1382                 pr_info("Hibernated on a CPU that is offline! Bringing CPU up.\n");
1383                 ret = cpu_up(sleep_cpu, CPUHP_ONLINE);
1384                 if (ret) {
1385                         pr_err("Failed to bring hibernate-CPU up!\n");
1386                         return ret;
1387                 }
1388         }
1389         return 0;
1390 }
1391
1392 void bringup_nonboot_cpus(unsigned int setup_max_cpus)
1393 {
1394         unsigned int cpu;
1395
1396         for_each_present_cpu(cpu) {
1397                 if (num_online_cpus() >= setup_max_cpus)
1398                         break;
1399                 if (!cpu_online(cpu))
1400                         cpu_up(cpu, CPUHP_ONLINE);
1401         }
1402 }
1403
1404 #ifdef CONFIG_PM_SLEEP_SMP
1405 static cpumask_var_t frozen_cpus;
1406
1407 int freeze_secondary_cpus(int primary)
1408 {
1409         int cpu, error = 0;
1410
1411         cpu_maps_update_begin();
1412         if (primary == -1) {
1413                 primary = cpumask_first(cpu_online_mask);
1414                 if (!housekeeping_cpu(primary, HK_FLAG_TIMER))
1415                         primary = housekeeping_any_cpu(HK_FLAG_TIMER);
1416         } else {
1417                 if (!cpu_online(primary))
1418                         primary = cpumask_first(cpu_online_mask);
1419         }
1420
1421         /*
1422          * We take down all of the non-boot CPUs in one shot to avoid races
1423          * with the userspace trying to use the CPU hotplug at the same time
1424          */
1425         cpumask_clear(frozen_cpus);
1426
1427         pr_info("Disabling non-boot CPUs ...\n");
1428         for_each_online_cpu(cpu) {
1429                 if (cpu == primary)
1430                         continue;
1431
1432                 if (pm_wakeup_pending()) {
1433                         pr_info("Wakeup pending. Abort CPU freeze\n");
1434                         error = -EBUSY;
1435                         break;
1436                 }
1437
1438                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1439                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1440                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1441                 if (!error)
1442                         cpumask_set_cpu(cpu, frozen_cpus);
1443                 else {
1444                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1445                         break;
1446                 }
1447         }
1448
1449         if (!error)
1450                 BUG_ON(num_online_cpus() > 1);
1451         else
1452                 pr_err("Non-boot CPUs are not disabled\n");
1453
1454         /*
1455          * Make sure the CPUs won't be enabled by someone else. We need to do
1456          * this even in case of failure as all freeze_secondary_cpus() users are
1457          * supposed to do thaw_secondary_cpus() on the failure path.
1458          */
1459         cpu_hotplug_disabled++;
1460
1461         cpu_maps_update_done();
1462         return error;
1463 }
1464
1465 void __weak arch_thaw_secondary_cpus_begin(void)
1466 {
1467 }
1468
1469 void __weak arch_thaw_secondary_cpus_end(void)
1470 {
1471 }
1472
1473 void thaw_secondary_cpus(void)
1474 {
1475         int cpu, error;
1476
1477         /* Allow everyone to use the CPU hotplug again */
1478         cpu_maps_update_begin();
1479         __cpu_hotplug_enable();
1480         if (cpumask_empty(frozen_cpus))
1481                 goto out;
1482
1483         pr_info("Enabling non-boot CPUs ...\n");
1484
1485         arch_thaw_secondary_cpus_begin();
1486
1487         for_each_cpu(cpu, frozen_cpus) {
1488                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1489                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1490                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1491                 if (!error) {
1492                         pr_info("CPU%d is up\n", cpu);
1493                         continue;
1494                 }
1495                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1496         }
1497
1498         arch_thaw_secondary_cpus_end();
1499
1500         cpumask_clear(frozen_cpus);
1501 out:
1502         cpu_maps_update_done();
1503 }
1504
1505 static int __init alloc_frozen_cpus(void)
1506 {
1507         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1508                 return -ENOMEM;
1509         return 0;
1510 }
1511 core_initcall(alloc_frozen_cpus);
1512
1513 /*
1514  * When callbacks for CPU hotplug notifications are being executed, we must
1515  * ensure that the state of the system with respect to the tasks being frozen
1516  * or not, as reported by the notification, remains unchanged *throughout the
1517  * duration* of the execution of the callbacks.
1518  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1519  *
1520  * This synchronization is implemented by mutually excluding regular CPU
1521  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1522  * Hibernate notifications.
1523  */
1524 static int
1525 cpu_hotplug_pm_callback(struct notifier_block *nb,
1526                         unsigned long action, void *ptr)
1527 {
1528         switch (action) {
1529
1530         case PM_SUSPEND_PREPARE:
1531         case PM_HIBERNATION_PREPARE:
1532                 cpu_hotplug_disable();
1533                 break;
1534
1535         case PM_POST_SUSPEND:
1536         case PM_POST_HIBERNATION:
1537                 cpu_hotplug_enable();
1538                 break;
1539
1540         default:
1541                 return NOTIFY_DONE;
1542         }
1543
1544         return NOTIFY_OK;
1545 }
1546
1547
1548 static int __init cpu_hotplug_pm_sync_init(void)
1549 {
1550         /*
1551          * cpu_hotplug_pm_callback has higher priority than x86
1552          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1553          * to disable cpu hotplug to avoid cpu hotplug race.
1554          */
1555         pm_notifier(cpu_hotplug_pm_callback, 0);
1556         return 0;
1557 }
1558 core_initcall(cpu_hotplug_pm_sync_init);
1559
1560 #endif /* CONFIG_PM_SLEEP_SMP */
1561
1562 int __boot_cpu_id;
1563
1564 #endif /* CONFIG_SMP */
1565
1566 /* Boot processor state steps */
1567 static struct cpuhp_step cpuhp_hp_states[] = {
1568         [CPUHP_OFFLINE] = {
1569                 .name                   = "offline",
1570                 .startup.single         = NULL,
1571                 .teardown.single        = NULL,
1572         },
1573 #ifdef CONFIG_SMP
1574         [CPUHP_CREATE_THREADS]= {
1575                 .name                   = "threads:prepare",
1576                 .startup.single         = smpboot_create_threads,
1577                 .teardown.single        = NULL,
1578                 .cant_stop              = true,
1579         },
1580         [CPUHP_PERF_PREPARE] = {
1581                 .name                   = "perf:prepare",
1582                 .startup.single         = perf_event_init_cpu,
1583                 .teardown.single        = perf_event_exit_cpu,
1584         },
1585         [CPUHP_RANDOM_PREPARE] = {
1586                 .name                   = "random:prepare",
1587                 .startup.single         = random_prepare_cpu,
1588                 .teardown.single        = NULL,
1589         },
1590         [CPUHP_WORKQUEUE_PREP] = {
1591                 .name                   = "workqueue:prepare",
1592                 .startup.single         = workqueue_prepare_cpu,
1593                 .teardown.single        = NULL,
1594         },
1595         [CPUHP_HRTIMERS_PREPARE] = {
1596                 .name                   = "hrtimers:prepare",
1597                 .startup.single         = hrtimers_prepare_cpu,
1598                 .teardown.single        = NULL,
1599         },
1600         [CPUHP_SMPCFD_PREPARE] = {
1601                 .name                   = "smpcfd:prepare",
1602                 .startup.single         = smpcfd_prepare_cpu,
1603                 .teardown.single        = smpcfd_dead_cpu,
1604         },
1605         [CPUHP_RELAY_PREPARE] = {
1606                 .name                   = "relay:prepare",
1607                 .startup.single         = relay_prepare_cpu,
1608                 .teardown.single        = NULL,
1609         },
1610         [CPUHP_SLAB_PREPARE] = {
1611                 .name                   = "slab:prepare",
1612                 .startup.single         = slab_prepare_cpu,
1613                 .teardown.single        = slab_dead_cpu,
1614         },
1615         [CPUHP_RCUTREE_PREP] = {
1616                 .name                   = "RCU/tree:prepare",
1617                 .startup.single         = rcutree_prepare_cpu,
1618                 .teardown.single        = rcutree_dead_cpu,
1619         },
1620         /*
1621          * On the tear-down path, timers_dead_cpu() must be invoked
1622          * before blk_mq_queue_reinit_notify() from notify_dead(),
1623          * otherwise a RCU stall occurs.
1624          */
1625         [CPUHP_TIMERS_PREPARE] = {
1626                 .name                   = "timers:prepare",
1627                 .startup.single         = timers_prepare_cpu,
1628                 .teardown.single        = timers_dead_cpu,
1629         },
1630         /* Kicks the plugged cpu into life */
1631         [CPUHP_BRINGUP_CPU] = {
1632                 .name                   = "cpu:bringup",
1633                 .startup.single         = bringup_cpu,
1634                 .teardown.single        = finish_cpu,
1635                 .cant_stop              = true,
1636         },
1637         /* Final state before CPU kills itself */
1638         [CPUHP_AP_IDLE_DEAD] = {
1639                 .name                   = "idle:dead",
1640         },
1641         /*
1642          * Last state before CPU enters the idle loop to die. Transient state
1643          * for synchronization.
1644          */
1645         [CPUHP_AP_OFFLINE] = {
1646                 .name                   = "ap:offline",
1647                 .cant_stop              = true,
1648         },
1649         /* First state is scheduler control. Interrupts are disabled */
1650         [CPUHP_AP_SCHED_STARTING] = {
1651                 .name                   = "sched:starting",
1652                 .startup.single         = sched_cpu_starting,
1653                 .teardown.single        = sched_cpu_dying,
1654         },
1655         [CPUHP_AP_RCUTREE_DYING] = {
1656                 .name                   = "RCU/tree:dying",
1657                 .startup.single         = NULL,
1658                 .teardown.single        = rcutree_dying_cpu,
1659         },
1660         [CPUHP_AP_SMPCFD_DYING] = {
1661                 .name                   = "smpcfd:dying",
1662                 .startup.single         = NULL,
1663                 .teardown.single        = smpcfd_dying_cpu,
1664         },
1665         [CPUHP_AP_HRTIMERS_DYING] = {
1666                 .name                   = "hrtimers:dying",
1667                 .startup.single         = NULL,
1668                 .teardown.single        = hrtimers_cpu_dying,
1669         },
1670
1671         /* Entry state on starting. Interrupts enabled from here on. Transient
1672          * state for synchronsization */
1673         [CPUHP_AP_ONLINE] = {
1674                 .name                   = "ap:online",
1675         },
1676         /*
1677          * Handled on controll processor until the plugged processor manages
1678          * this itself.
1679          */
1680         [CPUHP_TEARDOWN_CPU] = {
1681                 .name                   = "cpu:teardown",
1682                 .startup.single         = NULL,
1683                 .teardown.single        = takedown_cpu,
1684                 .cant_stop              = true,
1685         },
1686         /* Handle smpboot threads park/unpark */
1687         [CPUHP_AP_SMPBOOT_THREADS] = {
1688                 .name                   = "smpboot/threads:online",
1689                 .startup.single         = smpboot_unpark_threads,
1690                 .teardown.single        = smpboot_park_threads,
1691         },
1692         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1693                 .name                   = "irq/affinity:online",
1694                 .startup.single         = irq_affinity_online_cpu,
1695                 .teardown.single        = NULL,
1696         },
1697         [CPUHP_AP_PERF_ONLINE] = {
1698                 .name                   = "perf:online",
1699                 .startup.single         = perf_event_init_cpu,
1700                 .teardown.single        = perf_event_exit_cpu,
1701         },
1702         [CPUHP_AP_WATCHDOG_ONLINE] = {
1703                 .name                   = "lockup_detector:online",
1704                 .startup.single         = lockup_detector_online_cpu,
1705                 .teardown.single        = lockup_detector_offline_cpu,
1706         },
1707         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1708                 .name                   = "workqueue:online",
1709                 .startup.single         = workqueue_online_cpu,
1710                 .teardown.single        = workqueue_offline_cpu,
1711         },
1712         [CPUHP_AP_RANDOM_ONLINE] = {
1713                 .name                   = "random:online",
1714                 .startup.single         = random_online_cpu,
1715                 .teardown.single        = NULL,
1716         },
1717         [CPUHP_AP_RCUTREE_ONLINE] = {
1718                 .name                   = "RCU/tree:online",
1719                 .startup.single         = rcutree_online_cpu,
1720                 .teardown.single        = rcutree_offline_cpu,
1721         },
1722 #endif
1723         /*
1724          * The dynamically registered state space is here
1725          */
1726
1727 #ifdef CONFIG_SMP
1728         /* Last state is scheduler control setting the cpu active */
1729         [CPUHP_AP_ACTIVE] = {
1730                 .name                   = "sched:active",
1731                 .startup.single         = sched_cpu_activate,
1732                 .teardown.single        = sched_cpu_deactivate,
1733         },
1734 #endif
1735
1736         /* CPU is fully up and running. */
1737         [CPUHP_ONLINE] = {
1738                 .name                   = "online",
1739                 .startup.single         = NULL,
1740                 .teardown.single        = NULL,
1741         },
1742 };
1743
1744 /* Sanity check for callbacks */
1745 static int cpuhp_cb_check(enum cpuhp_state state)
1746 {
1747         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1748                 return -EINVAL;
1749         return 0;
1750 }
1751
1752 /*
1753  * Returns a free for dynamic slot assignment of the Online state. The states
1754  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1755  * by having no name assigned.
1756  */
1757 static int cpuhp_reserve_state(enum cpuhp_state state)
1758 {
1759         enum cpuhp_state i, end;
1760         struct cpuhp_step *step;
1761
1762         switch (state) {
1763         case CPUHP_AP_ONLINE_DYN:
1764                 step = cpuhp_hp_states + CPUHP_AP_ONLINE_DYN;
1765                 end = CPUHP_AP_ONLINE_DYN_END;
1766                 break;
1767         case CPUHP_BP_PREPARE_DYN:
1768                 step = cpuhp_hp_states + CPUHP_BP_PREPARE_DYN;
1769                 end = CPUHP_BP_PREPARE_DYN_END;
1770                 break;
1771         default:
1772                 return -EINVAL;
1773         }
1774
1775         for (i = state; i <= end; i++, step++) {
1776                 if (!step->name)
1777                         return i;
1778         }
1779         WARN(1, "No more dynamic states available for CPU hotplug\n");
1780         return -ENOSPC;
1781 }
1782
1783 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1784                                  int (*startup)(unsigned int cpu),
1785                                  int (*teardown)(unsigned int cpu),
1786                                  bool multi_instance)
1787 {
1788         /* (Un)Install the callbacks for further cpu hotplug operations */
1789         struct cpuhp_step *sp;
1790         int ret = 0;
1791
1792         /*
1793          * If name is NULL, then the state gets removed.
1794          *
1795          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1796          * the first allocation from these dynamic ranges, so the removal
1797          * would trigger a new allocation and clear the wrong (already
1798          * empty) state, leaving the callbacks of the to be cleared state
1799          * dangling, which causes wreckage on the next hotplug operation.
1800          */
1801         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1802                      state == CPUHP_BP_PREPARE_DYN)) {
1803                 ret = cpuhp_reserve_state(state);
1804                 if (ret < 0)
1805                         return ret;
1806                 state = ret;
1807         }
1808         sp = cpuhp_get_step(state);
1809         if (name && sp->name)
1810                 return -EBUSY;
1811
1812         sp->startup.single = startup;
1813         sp->teardown.single = teardown;
1814         sp->name = name;
1815         sp->multi_instance = multi_instance;
1816         INIT_HLIST_HEAD(&sp->list);
1817         return ret;
1818 }
1819
1820 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1821 {
1822         return cpuhp_get_step(state)->teardown.single;
1823 }
1824
1825 /*
1826  * Call the startup/teardown function for a step either on the AP or
1827  * on the current CPU.
1828  */
1829 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1830                             struct hlist_node *node)
1831 {
1832         struct cpuhp_step *sp = cpuhp_get_step(state);
1833         int ret;
1834
1835         /*
1836          * If there's nothing to do, we done.
1837          * Relies on the union for multi_instance.
1838          */
1839         if ((bringup && !sp->startup.single) ||
1840             (!bringup && !sp->teardown.single))
1841                 return 0;
1842         /*
1843          * The non AP bound callbacks can fail on bringup. On teardown
1844          * e.g. module removal we crash for now.
1845          */
1846 #ifdef CONFIG_SMP
1847         if (cpuhp_is_ap_state(state))
1848                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1849         else
1850                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1851 #else
1852         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1853 #endif
1854         BUG_ON(ret && !bringup);
1855         return ret;
1856 }
1857
1858 /*
1859  * Called from __cpuhp_setup_state on a recoverable failure.
1860  *
1861  * Note: The teardown callbacks for rollback are not allowed to fail!
1862  */
1863 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1864                                    struct hlist_node *node)
1865 {
1866         int cpu;
1867
1868         /* Roll back the already executed steps on the other cpus */
1869         for_each_present_cpu(cpu) {
1870                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1871                 int cpustate = st->state;
1872
1873                 if (cpu >= failedcpu)
1874                         break;
1875
1876                 /* Did we invoke the startup call on that cpu ? */
1877                 if (cpustate >= state)
1878                         cpuhp_issue_call(cpu, state, false, node);
1879         }
1880 }
1881
1882 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1883                                           struct hlist_node *node,
1884                                           bool invoke)
1885 {
1886         struct cpuhp_step *sp;
1887         int cpu;
1888         int ret;
1889
1890         lockdep_assert_cpus_held();
1891
1892         sp = cpuhp_get_step(state);
1893         if (sp->multi_instance == false)
1894                 return -EINVAL;
1895
1896         mutex_lock(&cpuhp_state_mutex);
1897
1898         if (!invoke || !sp->startup.multi)
1899                 goto add_node;
1900
1901         /*
1902          * Try to call the startup callback for each present cpu
1903          * depending on the hotplug state of the cpu.
1904          */
1905         for_each_present_cpu(cpu) {
1906                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1907                 int cpustate = st->state;
1908
1909                 if (cpustate < state)
1910                         continue;
1911
1912                 ret = cpuhp_issue_call(cpu, state, true, node);
1913                 if (ret) {
1914                         if (sp->teardown.multi)
1915                                 cpuhp_rollback_install(cpu, state, node);
1916                         goto unlock;
1917                 }
1918         }
1919 add_node:
1920         ret = 0;
1921         hlist_add_head(node, &sp->list);
1922 unlock:
1923         mutex_unlock(&cpuhp_state_mutex);
1924         return ret;
1925 }
1926
1927 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1928                                bool invoke)
1929 {
1930         int ret;
1931
1932         cpus_read_lock();
1933         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1934         cpus_read_unlock();
1935         return ret;
1936 }
1937 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1938
1939 /**
1940  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1941  * @state:              The state to setup
1942  * @invoke:             If true, the startup function is invoked for cpus where
1943  *                      cpu state >= @state
1944  * @startup:            startup callback function
1945  * @teardown:           teardown callback function
1946  * @multi_instance:     State is set up for multiple instances which get
1947  *                      added afterwards.
1948  *
1949  * The caller needs to hold cpus read locked while calling this function.
1950  * Returns:
1951  *   On success:
1952  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1953  *      0 for all other states
1954  *   On failure: proper (negative) error code
1955  */
1956 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1957                                    const char *name, bool invoke,
1958                                    int (*startup)(unsigned int cpu),
1959                                    int (*teardown)(unsigned int cpu),
1960                                    bool multi_instance)
1961 {
1962         int cpu, ret = 0;
1963         bool dynstate;
1964
1965         lockdep_assert_cpus_held();
1966
1967         if (cpuhp_cb_check(state) || !name)
1968                 return -EINVAL;
1969
1970         mutex_lock(&cpuhp_state_mutex);
1971
1972         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1973                                     multi_instance);
1974
1975         dynstate = state == CPUHP_AP_ONLINE_DYN;
1976         if (ret > 0 && dynstate) {
1977                 state = ret;
1978                 ret = 0;
1979         }
1980
1981         if (ret || !invoke || !startup)
1982                 goto out;
1983
1984         /*
1985          * Try to call the startup callback for each present cpu
1986          * depending on the hotplug state of the cpu.
1987          */
1988         for_each_present_cpu(cpu) {
1989                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1990                 int cpustate = st->state;
1991
1992                 if (cpustate < state)
1993                         continue;
1994
1995                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1996                 if (ret) {
1997                         if (teardown)
1998                                 cpuhp_rollback_install(cpu, state, NULL);
1999                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2000                         goto out;
2001                 }
2002         }
2003 out:
2004         mutex_unlock(&cpuhp_state_mutex);
2005         /*
2006          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
2007          * dynamically allocated state in case of success.
2008          */
2009         if (!ret && dynstate)
2010                 return state;
2011         return ret;
2012 }
2013 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
2014
2015 int __cpuhp_setup_state(enum cpuhp_state state,
2016                         const char *name, bool invoke,
2017                         int (*startup)(unsigned int cpu),
2018                         int (*teardown)(unsigned int cpu),
2019                         bool multi_instance)
2020 {
2021         int ret;
2022
2023         cpus_read_lock();
2024         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
2025                                              teardown, multi_instance);
2026         cpus_read_unlock();
2027         return ret;
2028 }
2029 EXPORT_SYMBOL(__cpuhp_setup_state);
2030
2031 int __cpuhp_state_remove_instance(enum cpuhp_state state,
2032                                   struct hlist_node *node, bool invoke)
2033 {
2034         struct cpuhp_step *sp = cpuhp_get_step(state);
2035         int cpu;
2036
2037         BUG_ON(cpuhp_cb_check(state));
2038
2039         if (!sp->multi_instance)
2040                 return -EINVAL;
2041
2042         cpus_read_lock();
2043         mutex_lock(&cpuhp_state_mutex);
2044
2045         if (!invoke || !cpuhp_get_teardown_cb(state))
2046                 goto remove;
2047         /*
2048          * Call the teardown callback for each present cpu depending
2049          * on the hotplug state of the cpu. This function is not
2050          * allowed to fail currently!
2051          */
2052         for_each_present_cpu(cpu) {
2053                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2054                 int cpustate = st->state;
2055
2056                 if (cpustate >= state)
2057                         cpuhp_issue_call(cpu, state, false, node);
2058         }
2059
2060 remove:
2061         hlist_del(node);
2062         mutex_unlock(&cpuhp_state_mutex);
2063         cpus_read_unlock();
2064
2065         return 0;
2066 }
2067 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
2068
2069 /**
2070  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
2071  * @state:      The state to remove
2072  * @invoke:     If true, the teardown function is invoked for cpus where
2073  *              cpu state >= @state
2074  *
2075  * The caller needs to hold cpus read locked while calling this function.
2076  * The teardown callback is currently not allowed to fail. Think
2077  * about module removal!
2078  */
2079 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
2080 {
2081         struct cpuhp_step *sp = cpuhp_get_step(state);
2082         int cpu;
2083
2084         BUG_ON(cpuhp_cb_check(state));
2085
2086         lockdep_assert_cpus_held();
2087
2088         mutex_lock(&cpuhp_state_mutex);
2089         if (sp->multi_instance) {
2090                 WARN(!hlist_empty(&sp->list),
2091                      "Error: Removing state %d which has instances left.\n",
2092                      state);
2093                 goto remove;
2094         }
2095
2096         if (!invoke || !cpuhp_get_teardown_cb(state))
2097                 goto remove;
2098
2099         /*
2100          * Call the teardown callback for each present cpu depending
2101          * on the hotplug state of the cpu. This function is not
2102          * allowed to fail currently!
2103          */
2104         for_each_present_cpu(cpu) {
2105                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
2106                 int cpustate = st->state;
2107
2108                 if (cpustate >= state)
2109                         cpuhp_issue_call(cpu, state, false, NULL);
2110         }
2111 remove:
2112         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
2113         mutex_unlock(&cpuhp_state_mutex);
2114 }
2115 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
2116
2117 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
2118 {
2119         cpus_read_lock();
2120         __cpuhp_remove_state_cpuslocked(state, invoke);
2121         cpus_read_unlock();
2122 }
2123 EXPORT_SYMBOL(__cpuhp_remove_state);
2124
2125 #ifdef CONFIG_HOTPLUG_SMT
2126 static void cpuhp_offline_cpu_device(unsigned int cpu)
2127 {
2128         struct device *dev = get_cpu_device(cpu);
2129
2130         dev->offline = true;
2131         /* Tell user space about the state change */
2132         kobject_uevent(&dev->kobj, KOBJ_OFFLINE);
2133 }
2134
2135 static void cpuhp_online_cpu_device(unsigned int cpu)
2136 {
2137         struct device *dev = get_cpu_device(cpu);
2138
2139         dev->offline = false;
2140         /* Tell user space about the state change */
2141         kobject_uevent(&dev->kobj, KOBJ_ONLINE);
2142 }
2143
2144 int cpuhp_smt_disable(enum cpuhp_smt_control ctrlval)
2145 {
2146         int cpu, ret = 0;
2147
2148         cpu_maps_update_begin();
2149         for_each_online_cpu(cpu) {
2150                 if (topology_is_primary_thread(cpu))
2151                         continue;
2152                 ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
2153                 if (ret)
2154                         break;
2155                 /*
2156                  * As this needs to hold the cpu maps lock it's impossible
2157                  * to call device_offline() because that ends up calling
2158                  * cpu_down() which takes cpu maps lock. cpu maps lock
2159                  * needs to be held as this might race against in kernel
2160                  * abusers of the hotplug machinery (thermal management).
2161                  *
2162                  * So nothing would update device:offline state. That would
2163                  * leave the sysfs entry stale and prevent onlining after
2164                  * smt control has been changed to 'off' again. This is
2165                  * called under the sysfs hotplug lock, so it is properly
2166                  * serialized against the regular offline usage.
2167                  */
2168                 cpuhp_offline_cpu_device(cpu);
2169         }
2170         if (!ret)
2171                 cpu_smt_control = ctrlval;
2172         cpu_maps_update_done();
2173         return ret;
2174 }
2175
2176 int cpuhp_smt_enable(void)
2177 {
2178         int cpu, ret = 0;
2179
2180         cpu_maps_update_begin();
2181         cpu_smt_control = CPU_SMT_ENABLED;
2182         for_each_present_cpu(cpu) {
2183                 /* Skip online CPUs and CPUs on offline nodes */
2184                 if (cpu_online(cpu) || !node_online(cpu_to_node(cpu)))
2185                         continue;
2186                 ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
2187                 if (ret)
2188                         break;
2189                 /* See comment in cpuhp_smt_disable() */
2190                 cpuhp_online_cpu_device(cpu);
2191         }
2192         cpu_maps_update_done();
2193         return ret;
2194 }
2195 #endif
2196
2197 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
2198 static ssize_t show_cpuhp_state(struct device *dev,
2199                                 struct device_attribute *attr, char *buf)
2200 {
2201         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2202
2203         return sprintf(buf, "%d\n", st->state);
2204 }
2205 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
2206
2207 static ssize_t write_cpuhp_target(struct device *dev,
2208                                   struct device_attribute *attr,
2209                                   const char *buf, size_t count)
2210 {
2211         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2212         struct cpuhp_step *sp;
2213         int target, ret;
2214
2215         ret = kstrtoint(buf, 10, &target);
2216         if (ret)
2217                 return ret;
2218
2219 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
2220         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
2221                 return -EINVAL;
2222 #else
2223         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
2224                 return -EINVAL;
2225 #endif
2226
2227         ret = lock_device_hotplug_sysfs();
2228         if (ret)
2229                 return ret;
2230
2231         mutex_lock(&cpuhp_state_mutex);
2232         sp = cpuhp_get_step(target);
2233         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
2234         mutex_unlock(&cpuhp_state_mutex);
2235         if (ret)
2236                 goto out;
2237
2238         if (st->state < target)
2239                 ret = cpu_up(dev->id, target);
2240         else if (st->state > target)
2241                 ret = cpu_down(dev->id, target);
2242         else if (WARN_ON(st->target != target))
2243                 st->target = target;
2244 out:
2245         unlock_device_hotplug();
2246         return ret ? ret : count;
2247 }
2248
2249 static ssize_t show_cpuhp_target(struct device *dev,
2250                                  struct device_attribute *attr, char *buf)
2251 {
2252         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2253
2254         return sprintf(buf, "%d\n", st->target);
2255 }
2256 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
2257
2258
2259 static ssize_t write_cpuhp_fail(struct device *dev,
2260                                 struct device_attribute *attr,
2261                                 const char *buf, size_t count)
2262 {
2263         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2264         struct cpuhp_step *sp;
2265         int fail, ret;
2266
2267         ret = kstrtoint(buf, 10, &fail);
2268         if (ret)
2269                 return ret;
2270
2271         if (fail < CPUHP_OFFLINE || fail > CPUHP_ONLINE)
2272                 return -EINVAL;
2273
2274         /*
2275          * Cannot fail STARTING/DYING callbacks.
2276          */
2277         if (cpuhp_is_atomic_state(fail))
2278                 return -EINVAL;
2279
2280         /*
2281          * Cannot fail anything that doesn't have callbacks.
2282          */
2283         mutex_lock(&cpuhp_state_mutex);
2284         sp = cpuhp_get_step(fail);
2285         if (!sp->startup.single && !sp->teardown.single)
2286                 ret = -EINVAL;
2287         mutex_unlock(&cpuhp_state_mutex);
2288         if (ret)
2289                 return ret;
2290
2291         st->fail = fail;
2292
2293         return count;
2294 }
2295
2296 static ssize_t show_cpuhp_fail(struct device *dev,
2297                                struct device_attribute *attr, char *buf)
2298 {
2299         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
2300
2301         return sprintf(buf, "%d\n", st->fail);
2302 }
2303
2304 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
2305
2306 static struct attribute *cpuhp_cpu_attrs[] = {
2307         &dev_attr_state.attr,
2308         &dev_attr_target.attr,
2309         &dev_attr_fail.attr,
2310         NULL
2311 };
2312
2313 static const struct attribute_group cpuhp_cpu_attr_group = {
2314         .attrs = cpuhp_cpu_attrs,
2315         .name = "hotplug",
2316         NULL
2317 };
2318
2319 static ssize_t show_cpuhp_states(struct device *dev,
2320                                  struct device_attribute *attr, char *buf)
2321 {
2322         ssize_t cur, res = 0;
2323         int i;
2324
2325         mutex_lock(&cpuhp_state_mutex);
2326         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
2327                 struct cpuhp_step *sp = cpuhp_get_step(i);
2328
2329                 if (sp->name) {
2330                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
2331                         buf += cur;
2332                         res += cur;
2333                 }
2334         }
2335         mutex_unlock(&cpuhp_state_mutex);
2336         return res;
2337 }
2338 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
2339
2340 static struct attribute *cpuhp_cpu_root_attrs[] = {
2341         &dev_attr_states.attr,
2342         NULL
2343 };
2344
2345 static const struct attribute_group cpuhp_cpu_root_attr_group = {
2346         .attrs = cpuhp_cpu_root_attrs,
2347         .name = "hotplug",
2348         NULL
2349 };
2350
2351 #ifdef CONFIG_HOTPLUG_SMT
2352
2353 static ssize_t
2354 __store_smt_control(struct device *dev, struct device_attribute *attr,
2355                     const char *buf, size_t count)
2356 {
2357         int ctrlval, ret;
2358
2359         if (sysfs_streq(buf, "on"))
2360                 ctrlval = CPU_SMT_ENABLED;
2361         else if (sysfs_streq(buf, "off"))
2362                 ctrlval = CPU_SMT_DISABLED;
2363         else if (sysfs_streq(buf, "forceoff"))
2364                 ctrlval = CPU_SMT_FORCE_DISABLED;
2365         else
2366                 return -EINVAL;
2367
2368         if (cpu_smt_control == CPU_SMT_FORCE_DISABLED)
2369                 return -EPERM;
2370
2371         if (cpu_smt_control == CPU_SMT_NOT_SUPPORTED)
2372                 return -ENODEV;
2373
2374         ret = lock_device_hotplug_sysfs();
2375         if (ret)
2376                 return ret;
2377
2378         if (ctrlval != cpu_smt_control) {
2379                 switch (ctrlval) {
2380                 case CPU_SMT_ENABLED:
2381                         ret = cpuhp_smt_enable();
2382                         break;
2383                 case CPU_SMT_DISABLED:
2384                 case CPU_SMT_FORCE_DISABLED:
2385                         ret = cpuhp_smt_disable(ctrlval);
2386                         break;
2387                 }
2388         }
2389
2390         unlock_device_hotplug();
2391         return ret ? ret : count;
2392 }
2393
2394 #else /* !CONFIG_HOTPLUG_SMT */
2395 static ssize_t
2396 __store_smt_control(struct device *dev, struct device_attribute *attr,
2397                     const char *buf, size_t count)
2398 {
2399         return -ENODEV;
2400 }
2401 #endif /* CONFIG_HOTPLUG_SMT */
2402
2403 static const char *smt_states[] = {
2404         [CPU_SMT_ENABLED]               = "on",
2405         [CPU_SMT_DISABLED]              = "off",
2406         [CPU_SMT_FORCE_DISABLED]        = "forceoff",
2407         [CPU_SMT_NOT_SUPPORTED]         = "notsupported",
2408         [CPU_SMT_NOT_IMPLEMENTED]       = "notimplemented",
2409 };
2410
2411 static ssize_t
2412 show_smt_control(struct device *dev, struct device_attribute *attr, char *buf)
2413 {
2414         const char *state = smt_states[cpu_smt_control];
2415
2416         return snprintf(buf, PAGE_SIZE - 2, "%s\n", state);
2417 }
2418
2419 static ssize_t
2420 store_smt_control(struct device *dev, struct device_attribute *attr,
2421                   const char *buf, size_t count)
2422 {
2423         return __store_smt_control(dev, attr, buf, count);
2424 }
2425 static DEVICE_ATTR(control, 0644, show_smt_control, store_smt_control);
2426
2427 static ssize_t
2428 show_smt_active(struct device *dev, struct device_attribute *attr, char *buf)
2429 {
2430         return snprintf(buf, PAGE_SIZE - 2, "%d\n", sched_smt_active());
2431 }
2432 static DEVICE_ATTR(active, 0444, show_smt_active, NULL);
2433
2434 static struct attribute *cpuhp_smt_attrs[] = {
2435         &dev_attr_control.attr,
2436         &dev_attr_active.attr,
2437         NULL
2438 };
2439
2440 static const struct attribute_group cpuhp_smt_attr_group = {
2441         .attrs = cpuhp_smt_attrs,
2442         .name = "smt",
2443         NULL
2444 };
2445
2446 static int __init cpu_smt_sysfs_init(void)
2447 {
2448         return sysfs_create_group(&cpu_subsys.dev_root->kobj,
2449                                   &cpuhp_smt_attr_group);
2450 }
2451
2452 static int __init cpuhp_sysfs_init(void)
2453 {
2454         int cpu, ret;
2455
2456         ret = cpu_smt_sysfs_init();
2457         if (ret)
2458                 return ret;
2459
2460         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
2461                                  &cpuhp_cpu_root_attr_group);
2462         if (ret)
2463                 return ret;
2464
2465         for_each_possible_cpu(cpu) {
2466                 struct device *dev = get_cpu_device(cpu);
2467
2468                 if (!dev)
2469                         continue;
2470                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
2471                 if (ret)
2472                         return ret;
2473         }
2474         return 0;
2475 }
2476 device_initcall(cpuhp_sysfs_init);
2477 #endif /* CONFIG_SYSFS && CONFIG_HOTPLUG_CPU */
2478
2479 /*
2480  * cpu_bit_bitmap[] is a special, "compressed" data structure that
2481  * represents all NR_CPUS bits binary values of 1<<nr.
2482  *
2483  * It is used by cpumask_of() to get a constant address to a CPU
2484  * mask value that has a single bit set only.
2485  */
2486
2487 /* cpu_bit_bitmap[0] is empty - so we can back into it */
2488 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
2489 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
2490 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
2491 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
2492
2493 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
2494
2495         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
2496         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
2497 #if BITS_PER_LONG > 32
2498         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
2499         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
2500 #endif
2501 };
2502 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
2503
2504 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
2505 EXPORT_SYMBOL(cpu_all_bits);
2506
2507 #ifdef CONFIG_INIT_ALL_POSSIBLE
2508 struct cpumask __cpu_possible_mask __read_mostly
2509         = {CPU_BITS_ALL};
2510 #else
2511 struct cpumask __cpu_possible_mask __read_mostly;
2512 #endif
2513 EXPORT_SYMBOL(__cpu_possible_mask);
2514
2515 struct cpumask __cpu_online_mask __read_mostly;
2516 EXPORT_SYMBOL(__cpu_online_mask);
2517
2518 struct cpumask __cpu_present_mask __read_mostly;
2519 EXPORT_SYMBOL(__cpu_present_mask);
2520
2521 struct cpumask __cpu_active_mask __read_mostly;
2522 EXPORT_SYMBOL(__cpu_active_mask);
2523
2524 atomic_t __num_online_cpus __read_mostly;
2525 EXPORT_SYMBOL(__num_online_cpus);
2526
2527 void init_cpu_present(const struct cpumask *src)
2528 {
2529         cpumask_copy(&__cpu_present_mask, src);
2530 }
2531
2532 void init_cpu_possible(const struct cpumask *src)
2533 {
2534         cpumask_copy(&__cpu_possible_mask, src);
2535 }
2536
2537 void init_cpu_online(const struct cpumask *src)
2538 {
2539         cpumask_copy(&__cpu_online_mask, src);
2540 }
2541
2542 void set_cpu_online(unsigned int cpu, bool online)
2543 {
2544         /*
2545          * atomic_inc/dec() is required to handle the horrid abuse of this
2546          * function by the reboot and kexec code which invoke it from
2547          * IPI/NMI broadcasts when shutting down CPUs. Invocation from
2548          * regular CPU hotplug is properly serialized.
2549          *
2550          * Note, that the fact that __num_online_cpus is of type atomic_t
2551          * does not protect readers which are not serialized against
2552          * concurrent hotplug operations.
2553          */
2554         if (online) {
2555                 if (!cpumask_test_and_set_cpu(cpu, &__cpu_online_mask))
2556                         atomic_inc(&__num_online_cpus);
2557         } else {
2558                 if (cpumask_test_and_clear_cpu(cpu, &__cpu_online_mask))
2559                         atomic_dec(&__num_online_cpus);
2560         }
2561 }
2562
2563 /*
2564  * Activate the first processor.
2565  */
2566 void __init boot_cpu_init(void)
2567 {
2568         int cpu = smp_processor_id();
2569
2570         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2571         set_cpu_online(cpu, true);
2572         set_cpu_active(cpu, true);
2573         set_cpu_present(cpu, true);
2574         set_cpu_possible(cpu, true);
2575
2576 #ifdef CONFIG_SMP
2577         __boot_cpu_id = cpu;
2578 #endif
2579 }
2580
2581 /*
2582  * Must be called _AFTER_ setting up the per_cpu areas
2583  */
2584 void __init boot_cpu_hotplug_init(void)
2585 {
2586 #ifdef CONFIG_SMP
2587         cpumask_set_cpu(smp_processor_id(), &cpus_booted_once_mask);
2588 #endif
2589         this_cpu_write(cpuhp_state.state, CPUHP_ONLINE);
2590 }
2591
2592 /*
2593  * These are used for a global "mitigations=" cmdline option for toggling
2594  * optional CPU mitigations.
2595  */
2596 enum cpu_mitigations {
2597         CPU_MITIGATIONS_OFF,
2598         CPU_MITIGATIONS_AUTO,
2599         CPU_MITIGATIONS_AUTO_NOSMT,
2600 };
2601
2602 static enum cpu_mitigations cpu_mitigations __ro_after_init =
2603         CPU_MITIGATIONS_AUTO;
2604
2605 static int __init mitigations_parse_cmdline(char *arg)
2606 {
2607         if (!strcmp(arg, "off"))
2608                 cpu_mitigations = CPU_MITIGATIONS_OFF;
2609         else if (!strcmp(arg, "auto"))
2610                 cpu_mitigations = CPU_MITIGATIONS_AUTO;
2611         else if (!strcmp(arg, "auto,nosmt"))
2612                 cpu_mitigations = CPU_MITIGATIONS_AUTO_NOSMT;
2613         else
2614                 pr_crit("Unsupported mitigations=%s, system may still be vulnerable\n",
2615                         arg);
2616
2617         return 0;
2618 }
2619 early_param("mitigations", mitigations_parse_cmdline);
2620
2621 /* mitigations=off */
2622 bool cpu_mitigations_off(void)
2623 {
2624         return cpu_mitigations == CPU_MITIGATIONS_OFF;
2625 }
2626 EXPORT_SYMBOL_GPL(cpu_mitigations_off);
2627
2628 /* mitigations=auto,nosmt */
2629 bool cpu_mitigations_auto_nosmt(void)
2630 {
2631         return cpu_mitigations == CPU_MITIGATIONS_AUTO_NOSMT;
2632 }
2633 EXPORT_SYMBOL_GPL(cpu_mitigations_auto_nosmt);