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