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/task.h>
  13 #include <linux/unistd.h>
  14 #include <linux/cpu.h>
  15 #include <linux/oom.h>
  16 #include <linux/rcupdate.h>
  17 #include <linux/export.h>
  18 #include <linux/bug.h>
  19 #include <linux/kthread.h>
  20 #include <linux/stop_machine.h>
  21 #include <linux/mutex.h>
  22 #include <linux/gfp.h>
  23 #include <linux/suspend.h>
  24 #include <linux/lockdep.h>
  25 #include <linux/tick.h>
  26 #include <linux/irq.h>
  27 #include <linux/nmi.h>
  28 #include <linux/smpboot.h>
  29 #include <linux/relay.h>
  30 #include <linux/slab.h>
  31 #include <linux/percpu-rwsem.h>
  32
  33 #include <trace/events/power.h>
  34 #define CREATE_TRACE_POINTS
  35 #include <trace/events/cpuhp.h>
  36
  37 #include "smpboot.h"
  38
  39 /**
  40  * cpuhp_cpu_state - Per cpu hotplug state storage
  41  * @state:      The current cpu state
  42  * @target:     The target state
  43  * @thread:     Pointer to the hotplug thread
  44  * @should_run: Thread should execute
  45  * @rollback:   Perform a rollback
  46  * @single:     Single callback invocation
  47  * @bringup:    Single callback bringup or teardown selector
  48  * @cb_state:   The state for a single callback (install/uninstall)
  49  * @result:     Result of the operation
  50  * @done_up:    Signal completion to the issuer of the task for cpu-up
  51  * @done_down:  Signal completion to the issuer of the task for cpu-down
  52  */
  53 struct cpuhp_cpu_state {
  54         enum cpuhp_state        state;
  55         enum cpuhp_state        target;
  56         enum cpuhp_state        fail;
  57 #ifdef CONFIG_SMP
  58         struct task_struct      *thread;
  59         bool                    should_run;
  60         bool                    rollback;
  61         bool                    single;
  62         bool                    bringup;
  63         struct hlist_node       *node;
  64         struct hlist_node       *last;
  65         enum cpuhp_state        cb_state;
  66         int                     result;
  67         struct completion       done_up;
  68         struct completion       done_down;
  69 #endif
  70 };
  71
  72 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state) = {
  73         .fail = CPUHP_INVALID,
  74 };
  75
  76 #if defined(CONFIG_LOCKDEP) && defined(CONFIG_SMP)
  77 static struct lockdep_map cpuhp_state_up_map =
  78         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-up", &cpuhp_state_up_map);
  79 static struct lockdep_map cpuhp_state_down_map =
  80         STATIC_LOCKDEP_MAP_INIT("cpuhp_state-down", &cpuhp_state_down_map);
  81
  82
  83 static void inline cpuhp_lock_acquire(bool bringup)
  84 {
  85         lock_map_acquire(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  86 }
  87
  88 static void inline cpuhp_lock_release(bool bringup)
  89 {
  90         lock_map_release(bringup ? &cpuhp_state_up_map : &cpuhp_state_down_map);
  91 }
  92 #else
  93
  94 static void inline cpuhp_lock_acquire(bool bringup) { }
  95 static void inline cpuhp_lock_release(bool bringup) { }
  96
  97 #endif
  98
  99 /**
 100  * cpuhp_step - Hotplug state machine step
 101  * @name:       Name of the step
 102  * @startup:    Startup function of the step
 103  * @teardown:   Teardown function of the step
 104  * @skip_onerr: Do not invoke the functions on error rollback
 105  *              Will go away once the notifiers are gone
 106  * @cant_stop:  Bringup/teardown can't be stopped at this step
 107  */
 108 struct cpuhp_step {
 109         const char              *name;
 110         union {
 111                 int             (*single)(unsigned int cpu);
 112                 int             (*multi)(unsigned int cpu,
 113                                          struct hlist_node *node);
 114         } startup;
 115         union {
 116                 int             (*single)(unsigned int cpu);
 117                 int             (*multi)(unsigned int cpu,
 118                                          struct hlist_node *node);
 119         } teardown;
 120         struct hlist_head       list;
 121         bool                    skip_onerr;
 122         bool                    cant_stop;
 123         bool                    multi_instance;
 124 };
 125
 126 static DEFINE_MUTEX(cpuhp_state_mutex);
 127 static struct cpuhp_step cpuhp_bp_states[];
 128 static struct cpuhp_step cpuhp_ap_states[];
 129
 130 static bool cpuhp_is_ap_state(enum cpuhp_state state)
 131 {
 132         /*
 133          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 134          * purposes as that state is handled explicitly in cpu_down.
 135          */
 136         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 137 }
 138
 139 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 140 {
 141         struct cpuhp_step *sp;
 142
 143         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
 144         return sp + state;
 145 }
 146
 147 /**
 148  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 149  * @cpu:        The cpu for which the callback should be invoked
 150  * @state:      The state to do callbacks for
 151  * @bringup:    True if the bringup callback should be invoked
 152  * @node:       For multi-instance, do a single entry callback for install/remove
 153  * @lastp:      For multi-instance rollback, remember how far we got
 154  *
 155  * Called from cpu hotplug and from the state register machinery.
 156  */
 157 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 158                                  bool bringup, struct hlist_node *node,
 159                                  struct hlist_node **lastp)
 160 {
 161         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 162         struct cpuhp_step *step = cpuhp_get_step(state);
 163         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 164         int (*cb)(unsigned int cpu);
 165         int ret, cnt;
 166
 167         if (st->fail == state) {
 168                 st->fail = CPUHP_INVALID;
 169
 170                 if (!(bringup ? step->startup.single : step->teardown.single))
 171                         return 0;
 172
 173                 return -EAGAIN;
 174         }
 175
 176         if (!step->multi_instance) {
 177                 WARN_ON_ONCE(lastp && *lastp);
 178                 cb = bringup ? step->startup.single : step->teardown.single;
 179                 if (!cb)
 180                         return 0;
 181                 trace_cpuhp_enter(cpu, st->target, state, cb);
 182                 ret = cb(cpu);
 183                 trace_cpuhp_exit(cpu, st->state, state, ret);
 184                 return ret;
 185         }
 186         cbm = bringup ? step->startup.multi : step->teardown.multi;
 187         if (!cbm)
 188                 return 0;
 189
 190         /* Single invocation for instance add/remove */
 191         if (node) {
 192                 WARN_ON_ONCE(lastp && *lastp);
 193                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 194                 ret = cbm(cpu, node);
 195                 trace_cpuhp_exit(cpu, st->state, state, ret);
 196                 return ret;
 197         }
 198
 199         /* State transition. Invoke on all instances */
 200         cnt = 0;
 201         hlist_for_each(node, &step->list) {
 202                 if (lastp && node == *lastp)
 203                         break;
 204
 205                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 206                 ret = cbm(cpu, node);
 207                 trace_cpuhp_exit(cpu, st->state, state, ret);
 208                 if (ret) {
 209                         if (!lastp)
 210                                 goto err;
 211
 212                         *lastp = node;
 213                         return ret;
 214                 }
 215                 cnt++;
 216         }
 217         if (lastp)
 218                 *lastp = NULL;
 219         return 0;
 220 err:
 221         /* Rollback the instances if one failed */
 222         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 223         if (!cbm)
 224                 return ret;
 225
 226         hlist_for_each(node, &step->list) {
 227                 if (!cnt--)
 228                         break;
 229
 230                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 231                 ret = cbm(cpu, node);
 232                 trace_cpuhp_exit(cpu, st->state, state, ret);
 233                 /*
 234                  * Rollback must not fail,
 235                  */
 236                 WARN_ON_ONCE(ret);
 237         }
 238         return ret;
 239 }
 240
 241 #ifdef CONFIG_SMP
 242 static inline void wait_for_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 243 {
 244         struct completion *done = bringup ? &st->done_up : &st->done_down;
 245         wait_for_completion(done);
 246 }
 247
 248 static inline void complete_ap_thread(struct cpuhp_cpu_state *st, bool bringup)
 249 {
 250         struct completion *done = bringup ? &st->done_up : &st->done_down;
 251         complete(done);
 252 }
 253
 254 /*
 255  * The former STARTING/DYING states, ran with IRQs disabled and must not fail.
 256  */
 257 static bool cpuhp_is_atomic_state(enum cpuhp_state state)
 258 {
 259         return CPUHP_AP_IDLE_DEAD <= state && state < CPUHP_AP_ONLINE;
 260 }
 261
 262 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 263 static DEFINE_MUTEX(cpu_add_remove_lock);
 264 bool cpuhp_tasks_frozen;
 265 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 266
 267 /*
 268  * The following two APIs (cpu_maps_update_begin/done) must be used when
 269  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 270  */
 271 void cpu_maps_update_begin(void)
 272 {
 273         mutex_lock(&cpu_add_remove_lock);
 274 }
 275
 276 void cpu_maps_update_done(void)
 277 {
 278         mutex_unlock(&cpu_add_remove_lock);
 279 }
 280
 281 /*
 282  * If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 283  * Should always be manipulated under cpu_add_remove_lock
 284  */
 285 static int cpu_hotplug_disabled;
 286
 287 #ifdef CONFIG_HOTPLUG_CPU
 288
 289 DEFINE_STATIC_PERCPU_RWSEM(cpu_hotplug_lock);
 290
 291 void cpus_read_lock(void)
 292 {
 293         percpu_down_read(&cpu_hotplug_lock);
 294 }
 295 EXPORT_SYMBOL_GPL(cpus_read_lock);
 296
 297 void cpus_read_unlock(void)
 298 {
 299         percpu_up_read(&cpu_hotplug_lock);
 300 }
 301 EXPORT_SYMBOL_GPL(cpus_read_unlock);
 302
 303 void cpus_write_lock(void)
 304 {
 305         percpu_down_write(&cpu_hotplug_lock);
 306 }
 307
 308 void cpus_write_unlock(void)
 309 {
 310         percpu_up_write(&cpu_hotplug_lock);
 311 }
 312
 313 void lockdep_assert_cpus_held(void)
 314 {
 315         percpu_rwsem_assert_held(&cpu_hotplug_lock);
 316 }
 317
 318 /*
 319  * Wait for currently running CPU hotplug operations to complete (if any) and
 320  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 321  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 322  * hotplug path before performing hotplug operations. So acquiring that lock
 323  * guarantees mutual exclusion from any currently running hotplug operations.
 324  */
 325 void cpu_hotplug_disable(void)
 326 {
 327         cpu_maps_update_begin();
 328         cpu_hotplug_disabled++;
 329         cpu_maps_update_done();
 330 }
 331 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 332
 333 static void __cpu_hotplug_enable(void)
 334 {
 335         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 336                 return;
 337         cpu_hotplug_disabled--;
 338 }
 339
 340 void cpu_hotplug_enable(void)
 341 {
 342         cpu_maps_update_begin();
 343         __cpu_hotplug_enable();
 344         cpu_maps_update_done();
 345 }
 346 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 347 #endif  /* CONFIG_HOTPLUG_CPU */
 348
 349 static inline enum cpuhp_state
 350 cpuhp_set_state(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 351 {
 352         enum cpuhp_state prev_state = st->state;
 353
 354         st->rollback = false;
 355         st->last = NULL;
 356
 357         st->target = target;
 358         st->single = false;
 359         st->bringup = st->state < target;
 360
 361         return prev_state;
 362 }
 363
 364 static inline void
 365 cpuhp_reset_state(struct cpuhp_cpu_state *st, enum cpuhp_state prev_state)
 366 {
 367         st->rollback = true;
 368
 369         /*
 370          * If we have st->last we need to undo partial multi_instance of this
 371          * state first. Otherwise start undo at the previous state.
 372          */
 373         if (!st->last) {
 374                 if (st->bringup)
 375                         st->state--;
 376                 else
 377                         st->state++;
 378         }
 379
 380         st->target = prev_state;
 381         st->bringup = !st->bringup;
 382 }
 383
 384 /* Regular hotplug invocation of the AP hotplug thread */
 385 static void __cpuhp_kick_ap(struct cpuhp_cpu_state *st)
 386 {
 387         if (!st->single && st->state == st->target)
 388                 return;
 389
 390         st->result = 0;
 391         /*
 392          * Make sure the above stores are visible before should_run becomes
 393          * true. Paired with the mb() above in cpuhp_thread_fun()
 394          */
 395         smp_mb();
 396         st->should_run = true;
 397         wake_up_process(st->thread);
 398         wait_for_ap_thread(st, st->bringup);
 399 }
 400
 401 static int cpuhp_kick_ap(struct cpuhp_cpu_state *st, enum cpuhp_state target)
 402 {
 403         enum cpuhp_state prev_state;
 404         int ret;
 405
 406         prev_state = cpuhp_set_state(st, target);
 407         __cpuhp_kick_ap(st);
 408         if ((ret = st->result)) {
 409                 cpuhp_reset_state(st, prev_state);
 410                 __cpuhp_kick_ap(st);
 411         }
 412
 413         return ret;
 414 }
 415
 416 static int bringup_wait_for_ap(unsigned int cpu)
 417 {
 418         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 419
 420         /* Wait for the CPU to reach CPUHP_AP_ONLINE_IDLE */
 421         wait_for_ap_thread(st, true);
 422         if (WARN_ON_ONCE((!cpu_online(cpu))))
 423                 return -ECANCELED;
 424
 425         /* Unpark the stopper thread and the hotplug thread of the target cpu */
 426         stop_machine_unpark(cpu);
 427         kthread_unpark(st->thread);
 428
 429         if (st->target <= CPUHP_AP_ONLINE_IDLE)
 430                 return 0;
 431
 432         return cpuhp_kick_ap(st, st->target);
 433 }
 434
 435 static int bringup_cpu(unsigned int cpu)
 436 {
 437         struct task_struct *idle = idle_thread_get(cpu);
 438         int ret;
 439
 440         /*
 441          * Some architectures have to walk the irq descriptors to
 442          * setup the vector space for the cpu which comes online.
 443          * Prevent irq alloc/free across the bringup.
 444          */
 445         irq_lock_sparse();
 446
 447         /* Arch-specific enabling code. */
 448         ret = __cpu_up(cpu, idle);
 449         irq_unlock_sparse();
 450         if (ret)
 451                 return ret;
 452         return bringup_wait_for_ap(cpu);
 453 }
 454
 455 /*
 456  * Hotplug state machine related functions
 457  */
 458
 459 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 460 {
 461         for (st->state--; st->state > st->target; st->state--) {
 462                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 463
 464                 if (!step->skip_onerr)
 465                         cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 466         }
 467 }
 468
 469 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 470                               enum cpuhp_state target)
 471 {
 472         enum cpuhp_state prev_state = st->state;
 473         int ret = 0;
 474
 475         while (st->state < target) {
 476                 st->state++;
 477                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 478                 if (ret) {
 479                         st->target = prev_state;
 480                         undo_cpu_up(cpu, st);
 481                         break;
 482                 }
 483         }
 484         return ret;
 485 }
 486
 487 /*
 488  * The cpu hotplug threads manage the bringup and teardown of the cpus
 489  */
 490 static void cpuhp_create(unsigned int cpu)
 491 {
 492         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 493
 494         init_completion(&st->done_up);
 495         init_completion(&st->done_down);
 496 }
 497
 498 static int cpuhp_should_run(unsigned int cpu)
 499 {
 500         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 501
 502         return st->should_run;
 503 }
 504
 505 /*
 506  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 507  * callbacks when a state gets [un]installed at runtime.
 508  *
 509  * Each invocation of this function by the smpboot thread does a single AP
 510  * state callback.
 511  *
 512  * It has 3 modes of operation:
 513  *  - single: runs st->cb_state
 514  *  - up:     runs ++st->state, while st->state < st->target
 515  *  - down:   runs st->state--, while st->state > st->target
 516  *
 517  * When complete or on error, should_run is cleared and the completion is fired.
 518  */
 519 static void cpuhp_thread_fun(unsigned int cpu)
 520 {
 521         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 522         bool bringup = st->bringup;
 523         enum cpuhp_state state;
 524
 525         /*
 526          * ACQUIRE for the cpuhp_should_run() load of ->should_run. Ensures
 527          * that if we see ->should_run we also see the rest of the state.
 528          */
 529         smp_mb();
 530
 531         if (WARN_ON_ONCE(!st->should_run))
 532                 return;
 533
 534         cpuhp_lock_acquire(bringup);
 535
 536         if (st->single) {
 537                 state = st->cb_state;
 538                 st->should_run = false;
 539         } else {
 540                 if (bringup) {
 541                         st->state++;
 542                         state = st->state;
 543                         st->should_run = (st->state < st->target);
 544                         WARN_ON_ONCE(st->state > st->target);
 545                 } else {
 546                         state = st->state;
 547                         st->state--;
 548                         st->should_run = (st->state > st->target);
 549                         WARN_ON_ONCE(st->state < st->target);
 550                 }
 551         }
 552
 553         WARN_ON_ONCE(!cpuhp_is_ap_state(state));
 554
 555         if (st->rollback) {
 556                 struct cpuhp_step *step = cpuhp_get_step(state);
 557                 if (step->skip_onerr)
 558                         goto next;
 559         }
 560
 561         if (cpuhp_is_atomic_state(state)) {
 562                 local_irq_disable();
 563                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 564                 local_irq_enable();
 565
 566                 /*
 567                  * STARTING/DYING must not fail!
 568                  */
 569                 WARN_ON_ONCE(st->result);
 570         } else {
 571                 st->result = cpuhp_invoke_callback(cpu, state, bringup, st->node, &st->last);
 572         }
 573
 574         if (st->result) {
 575                 /*
 576                  * If we fail on a rollback, we're up a creek without no
 577                  * paddle, no way forward, no way back. We loose, thanks for
 578                  * playing.
 579                  */
 580                 WARN_ON_ONCE(st->rollback);
 581                 st->should_run = false;
 582         }
 583
 584 next:
 585         cpuhp_lock_release(bringup);
 586
 587         if (!st->should_run)
 588                 complete_ap_thread(st, bringup);
 589 }
 590
 591 /* Invoke a single callback on a remote cpu */
 592 static int
 593 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 594                          struct hlist_node *node)
 595 {
 596         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 597         int ret;
 598
 599         if (!cpu_online(cpu))
 600                 return 0;
 601
 602         cpuhp_lock_acquire(false);
 603         cpuhp_lock_release(false);
 604
 605         cpuhp_lock_acquire(true);
 606         cpuhp_lock_release(true);
 607
 608         /*
 609          * If we are up and running, use the hotplug thread. For early calls
 610          * we invoke the thread function directly.
 611          */
 612         if (!st->thread)
 613                 return cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
 614
 615         st->rollback = false;
 616         st->last = NULL;
 617
 618         st->node = node;
 619         st->bringup = bringup;
 620         st->cb_state = state;
 621         st->single = true;
 622
 623         __cpuhp_kick_ap(st);
 624
 625         /*
 626          * If we failed and did a partial, do a rollback.
 627          */
 628         if ((ret = st->result) && st->last) {
 629                 st->rollback = true;
 630                 st->bringup = !bringup;
 631
 632                 __cpuhp_kick_ap(st);
 633         }
 634
 635         /*
 636          * Clean up the leftovers so the next hotplug operation wont use stale
 637          * data.
 638          */
 639         st->node = st->last = NULL;
 640         return ret;
 641 }
 642
 643 static int cpuhp_kick_ap_work(unsigned int cpu)
 644 {
 645         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 646         enum cpuhp_state prev_state = st->state;
 647         int ret;
 648
 649         cpuhp_lock_acquire(false);
 650         cpuhp_lock_release(false);
 651
 652         cpuhp_lock_acquire(true);
 653         cpuhp_lock_release(true);
 654
 655         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 656         ret = cpuhp_kick_ap(st, st->target);
 657         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 658
 659         return ret;
 660 }
 661
 662 static struct smp_hotplug_thread cpuhp_threads = {
 663         .store                  = &cpuhp_state.thread,
 664         .create                 = &cpuhp_create,
 665         .thread_should_run      = cpuhp_should_run,
 666         .thread_fn              = cpuhp_thread_fun,
 667         .thread_comm            = "cpuhp/%u",
 668         .selfparking            = true,
 669 };
 670
 671 void __init cpuhp_threads_init(void)
 672 {
 673         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 674         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 675 }
 676
 677 #ifdef CONFIG_HOTPLUG_CPU
 678 /**
 679  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 680  * @cpu: a CPU id
 681  *
 682  * This function walks all processes, finds a valid mm struct for each one and
 683  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 684  * trivial, there are various non-obvious corner cases, which this function
 685  * tries to solve in a safe manner.
 686  *
 687  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 688  * be called only for an already offlined CPU.
 689  */
 690 void clear_tasks_mm_cpumask(int cpu)
 691 {
 692         struct task_struct *p;
 693
 694         /*
 695          * This function is called after the cpu is taken down and marked
 696          * offline, so its not like new tasks will ever get this cpu set in
 697          * their mm mask. -- Peter Zijlstra
 698          * Thus, we may use rcu_read_lock() here, instead of grabbing
 699          * full-fledged tasklist_lock.
 700          */
 701         WARN_ON(cpu_online(cpu));
 702         rcu_read_lock();
 703         for_each_process(p) {
 704                 struct task_struct *t;
 705
 706                 /*
 707                  * Main thread might exit, but other threads may still have
 708                  * a valid mm. Find one.
 709                  */
 710                 t = find_lock_task_mm(p);
 711                 if (!t)
 712                         continue;
 713                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 714                 task_unlock(t);
 715         }
 716         rcu_read_unlock();
 717 }
 718
 719 /* Take this CPU down. */
 720 static int take_cpu_down(void *_param)
 721 {
 722         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 723         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 724         int err, cpu = smp_processor_id();
 725         int ret;
 726
 727         /* Ensure this CPU doesn't handle any more interrupts. */
 728         err = __cpu_disable();
 729         if (err < 0)
 730                 return err;
 731
 732         /*
 733          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 734          * do this step again.
 735          */
 736         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 737         st->state--;
 738         /* Invoke the former CPU_DYING callbacks */
 739         for (; st->state > target; st->state--) {
 740                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 741                 /*
 742                  * DYING must not fail!
 743                  */
 744                 WARN_ON_ONCE(ret);
 745         }
 746
 747         /* Give up timekeeping duties */
 748         tick_handover_do_timer();
 749         /* Park the stopper thread */
 750         stop_machine_park(cpu);
 751         return 0;
 752 }
 753
 754 static int takedown_cpu(unsigned int cpu)
 755 {
 756         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 757         int err;
 758
 759         /* Park the smpboot threads */
 760         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 761         smpboot_park_threads(cpu);
 762
 763         /*
 764          * Prevent irq alloc/free while the dying cpu reorganizes the
 765          * interrupt affinities.
 766          */
 767         irq_lock_sparse();
 768
 769         /*
 770          * So now all preempt/rcu users must observe !cpu_active().
 771          */
 772         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 773         if (err) {
 774                 /* CPU refused to die */
 775                 irq_unlock_sparse();
 776                 /* Unpark the hotplug thread so we can rollback there */
 777                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 778                 return err;
 779         }
 780         BUG_ON(cpu_online(cpu));
 781
 782         /*
 783          * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
 784          * runnable tasks from the cpu, there's only the idle task left now
 785          * that the migration thread is done doing the stop_machine thing.
 786          *
 787          * Wait for the stop thread to go away.
 788          */
 789         wait_for_ap_thread(st, false);
 790         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 791
 792         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 793         irq_unlock_sparse();
 794
 795         hotplug_cpu__broadcast_tick_pull(cpu);
 796         /* This actually kills the CPU. */
 797         __cpu_die(cpu);
 798
 799         tick_cleanup_dead_cpu(cpu);
 800         rcutree_migrate_callbacks(cpu);
 801         return 0;
 802 }
 803
 804 static void cpuhp_complete_idle_dead(void *arg)
 805 {
 806         struct cpuhp_cpu_state *st = arg;
 807
 808         complete_ap_thread(st, false);
 809 }
 810
 811 void cpuhp_report_idle_dead(void)
 812 {
 813         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 814
 815         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 816         rcu_report_dead(smp_processor_id());
 817         st->state = CPUHP_AP_IDLE_DEAD;
 818         /*
 819          * We cannot call complete after rcu_report_dead() so we delegate it
 820          * to an online cpu.
 821          */
 822         smp_call_function_single(cpumask_first(cpu_online_mask),
 823                                  cpuhp_complete_idle_dead, st, 0);
 824 }
 825
 826 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 827 {
 828         for (st->state++; st->state < st->target; st->state++) {
 829                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 830
 831                 if (!step->skip_onerr)
 832                         cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 833         }
 834 }
 835
 836 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 837                                 enum cpuhp_state target)
 838 {
 839         enum cpuhp_state prev_state = st->state;
 840         int ret = 0;
 841
 842         for (; st->state > target; st->state--) {
 843                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 844                 if (ret) {
 845                         st->target = prev_state;
 846                         undo_cpu_down(cpu, st);
 847                         break;
 848                 }
 849         }
 850         return ret;
 851 }
 852
 853 /* Requires cpu_add_remove_lock to be held */
 854 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 855                            enum cpuhp_state target)
 856 {
 857         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 858         int prev_state, ret = 0;
 859
 860         if (num_online_cpus() == 1)
 861                 return -EBUSY;
 862
 863         if (!cpu_present(cpu))
 864                 return -EINVAL;
 865
 866         cpus_write_lock();
 867
 868         cpuhp_tasks_frozen = tasks_frozen;
 869
 870         prev_state = cpuhp_set_state(st, target);
 871         /*
 872          * If the current CPU state is in the range of the AP hotplug thread,
 873          * then we need to kick the thread.
 874          */
 875         if (st->state > CPUHP_TEARDOWN_CPU) {
 876                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 877                 ret = cpuhp_kick_ap_work(cpu);
 878                 /*
 879                  * The AP side has done the error rollback already. Just
 880                  * return the error code..
 881                  */
 882                 if (ret)
 883                         goto out;
 884
 885                 /*
 886                  * We might have stopped still in the range of the AP hotplug
 887                  * thread. Nothing to do anymore.
 888                  */
 889                 if (st->state > CPUHP_TEARDOWN_CPU)
 890                         goto out;
 891
 892                 st->target = target;
 893         }
 894         /*
 895          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 896          * to do the further cleanups.
 897          */
 898         ret = cpuhp_down_callbacks(cpu, st, target);
 899         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 900                 cpuhp_reset_state(st, prev_state);
 901                 __cpuhp_kick_ap(st);
 902         }
 903
 904 out:
 905         cpus_write_unlock();
 906         /*
 907          * Do post unplug cleanup. This is still protected against
 908          * concurrent CPU hotplug via cpu_add_remove_lock.
 909          */
 910         lockup_detector_cleanup();
 911         return ret;
 912 }
 913
 914 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 915 {
 916         int err;
 917
 918         cpu_maps_update_begin();
 919
 920         if (cpu_hotplug_disabled) {
 921                 err = -EBUSY;
 922                 goto out;
 923         }
 924
 925         err = _cpu_down(cpu, 0, target);
 926
 927 out:
 928         cpu_maps_update_done();
 929         return err;
 930 }
 931
 932 int cpu_down(unsigned int cpu)
 933 {
 934         return do_cpu_down(cpu, CPUHP_OFFLINE);
 935 }
 936 EXPORT_SYMBOL(cpu_down);
 937
 938 #else
 939 #define takedown_cpu            NULL
 940 #endif /*CONFIG_HOTPLUG_CPU*/
 941
 942 /**
 943  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
 944  * @cpu: cpu that just started
 945  *
 946  * It must be called by the arch code on the new cpu, before the new cpu
 947  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 948  */
 949 void notify_cpu_starting(unsigned int cpu)
 950 {
 951         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 952         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 953         int ret;
 954
 955         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
 956         while (st->state < target) {
 957                 st->state++;
 958                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 959                 /*
 960                  * STARTING must not fail!
 961                  */
 962                 WARN_ON_ONCE(ret);
 963         }
 964 }
 965
 966 /*
 967  * Called from the idle task. Wake up the controlling task which brings the
 968  * stopper and the hotplug thread of the upcoming CPU up and then delegates
 969  * the rest of the online bringup to the hotplug thread.
 970  */
 971 void cpuhp_online_idle(enum cpuhp_state state)
 972 {
 973         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 974
 975         /* Happens for the boot cpu */
 976         if (state != CPUHP_AP_ONLINE_IDLE)
 977                 return;
 978
 979         st->state = CPUHP_AP_ONLINE_IDLE;
 980         complete_ap_thread(st, true);
 981 }
 982
 983 /* Requires cpu_add_remove_lock to be held */
 984 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 985 {
 986         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 987         struct task_struct *idle;
 988         int ret = 0;
 989
 990         cpus_write_lock();
 991
 992         if (!cpu_present(cpu)) {
 993                 ret = -EINVAL;
 994                 goto out;
 995         }
 996
 997         /*
 998          * The caller of do_cpu_up might have raced with another
 999          * caller. Ignore it for now.
1000          */
1001         if (st->state >= target)
1002                 goto out;
1003
1004         if (st->state == CPUHP_OFFLINE) {
1005                 /* Let it fail before we try to bring the cpu up */
1006                 idle = idle_thread_get(cpu);
1007                 if (IS_ERR(idle)) {
1008                         ret = PTR_ERR(idle);
1009                         goto out;
1010                 }
1011         }
1012
1013         cpuhp_tasks_frozen = tasks_frozen;
1014
1015         cpuhp_set_state(st, target);
1016         /*
1017          * If the current CPU state is in the range of the AP hotplug thread,
1018          * then we need to kick the thread once more.
1019          */
1020         if (st->state > CPUHP_BRINGUP_CPU) {
1021                 ret = cpuhp_kick_ap_work(cpu);
1022                 /*
1023                  * The AP side has done the error rollback already. Just
1024                  * return the error code..
1025                  */
1026                 if (ret)
1027                         goto out;
1028         }
1029
1030         /*
1031          * Try to reach the target state. We max out on the BP at
1032          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1033          * responsible for bringing it up to the target state.
1034          */
1035         target = min((int)target, CPUHP_BRINGUP_CPU);
1036         ret = cpuhp_up_callbacks(cpu, st, target);
1037 out:
1038         cpus_write_unlock();
1039         return ret;
1040 }
1041
1042 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1043 {
1044         int err = 0;
1045
1046         if (!cpu_possible(cpu)) {
1047                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1048                        cpu);
1049 #if defined(CONFIG_IA64)
1050                 pr_err("please check additional_cpus= boot parameter\n");
1051 #endif
1052                 return -EINVAL;
1053         }
1054
1055         err = try_online_node(cpu_to_node(cpu));
1056         if (err)
1057                 return err;
1058
1059         cpu_maps_update_begin();
1060
1061         if (cpu_hotplug_disabled) {
1062                 err = -EBUSY;
1063                 goto out;
1064         }
1065
1066         err = _cpu_up(cpu, 0, target);
1067 out:
1068         cpu_maps_update_done();
1069         return err;
1070 }
1071
1072 int cpu_up(unsigned int cpu)
1073 {
1074         return do_cpu_up(cpu, CPUHP_ONLINE);
1075 }
1076 EXPORT_SYMBOL_GPL(cpu_up);
1077
1078 #ifdef CONFIG_PM_SLEEP_SMP
1079 static cpumask_var_t frozen_cpus;
1080
1081 int freeze_secondary_cpus(int primary)
1082 {
1083         int cpu, error = 0;
1084
1085         cpu_maps_update_begin();
1086         if (!cpu_online(primary))
1087                 primary = cpumask_first(cpu_online_mask);
1088         /*
1089          * We take down all of the non-boot CPUs in one shot to avoid races
1090          * with the userspace trying to use the CPU hotplug at the same time
1091          */
1092         cpumask_clear(frozen_cpus);
1093
1094         pr_info("Disabling non-boot CPUs ...\n");
1095         for_each_online_cpu(cpu) {
1096                 if (cpu == primary)
1097                         continue;
1098                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1099                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1100                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1101                 if (!error)
1102                         cpumask_set_cpu(cpu, frozen_cpus);
1103                 else {
1104                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1105                         break;
1106                 }
1107         }
1108
1109         if (!error)
1110                 BUG_ON(num_online_cpus() > 1);
1111         else
1112                 pr_err("Non-boot CPUs are not disabled\n");
1113
1114         /*
1115          * Make sure the CPUs won't be enabled by someone else. We need to do
1116          * this even in case of failure as all disable_nonboot_cpus() users are
1117          * supposed to do enable_nonboot_cpus() on the failure path.
1118          */
1119         cpu_hotplug_disabled++;
1120
1121         cpu_maps_update_done();
1122         return error;
1123 }
1124
1125 void __weak arch_enable_nonboot_cpus_begin(void)
1126 {
1127 }
1128
1129 void __weak arch_enable_nonboot_cpus_end(void)
1130 {
1131 }
1132
1133 void enable_nonboot_cpus(void)
1134 {
1135         int cpu, error;
1136
1137         /* Allow everyone to use the CPU hotplug again */
1138         cpu_maps_update_begin();
1139         __cpu_hotplug_enable();
1140         if (cpumask_empty(frozen_cpus))
1141                 goto out;
1142
1143         pr_info("Enabling non-boot CPUs ...\n");
1144
1145         arch_enable_nonboot_cpus_begin();
1146
1147         for_each_cpu(cpu, frozen_cpus) {
1148                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1149                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1150                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1151                 if (!error) {
1152                         pr_info("CPU%d is up\n", cpu);
1153                         continue;
1154                 }
1155                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1156         }
1157
1158         arch_enable_nonboot_cpus_end();
1159
1160         cpumask_clear(frozen_cpus);
1161 out:
1162         cpu_maps_update_done();
1163 }
1164
1165 static int __init alloc_frozen_cpus(void)
1166 {
1167         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1168                 return -ENOMEM;
1169         return 0;
1170 }
1171 core_initcall(alloc_frozen_cpus);
1172
1173 /*
1174  * When callbacks for CPU hotplug notifications are being executed, we must
1175  * ensure that the state of the system with respect to the tasks being frozen
1176  * or not, as reported by the notification, remains unchanged *throughout the
1177  * duration* of the execution of the callbacks.
1178  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1179  *
1180  * This synchronization is implemented by mutually excluding regular CPU
1181  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1182  * Hibernate notifications.
1183  */
1184 static int
1185 cpu_hotplug_pm_callback(struct notifier_block *nb,
1186                         unsigned long action, void *ptr)
1187 {
1188         switch (action) {
1189
1190         case PM_SUSPEND_PREPARE:
1191         case PM_HIBERNATION_PREPARE:
1192                 cpu_hotplug_disable();
1193                 break;
1194
1195         case PM_POST_SUSPEND:
1196         case PM_POST_HIBERNATION:
1197                 cpu_hotplug_enable();
1198                 break;
1199
1200         default:
1201                 return NOTIFY_DONE;
1202         }
1203
1204         return NOTIFY_OK;
1205 }
1206
1207
1208 static int __init cpu_hotplug_pm_sync_init(void)
1209 {
1210         /*
1211          * cpu_hotplug_pm_callback has higher priority than x86
1212          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1213          * to disable cpu hotplug to avoid cpu hotplug race.
1214          */
1215         pm_notifier(cpu_hotplug_pm_callback, 0);
1216         return 0;
1217 }
1218 core_initcall(cpu_hotplug_pm_sync_init);
1219
1220 #endif /* CONFIG_PM_SLEEP_SMP */
1221
1222 int __boot_cpu_id;
1223
1224 #endif /* CONFIG_SMP */
1225
1226 /* Boot processor state steps */
1227 static struct cpuhp_step cpuhp_bp_states[] = {
1228         [CPUHP_OFFLINE] = {
1229                 .name                   = "offline",
1230                 .startup.single         = NULL,
1231                 .teardown.single        = NULL,
1232         },
1233 #ifdef CONFIG_SMP
1234         [CPUHP_CREATE_THREADS]= {
1235                 .name                   = "threads:prepare",
1236                 .startup.single         = smpboot_create_threads,
1237                 .teardown.single        = NULL,
1238                 .cant_stop              = true,
1239         },
1240         [CPUHP_PERF_PREPARE] = {
1241                 .name                   = "perf:prepare",
1242                 .startup.single         = perf_event_init_cpu,
1243                 .teardown.single        = perf_event_exit_cpu,
1244         },
1245         [CPUHP_WORKQUEUE_PREP] = {
1246                 .name                   = "workqueue:prepare",
1247                 .startup.single         = workqueue_prepare_cpu,
1248                 .teardown.single        = NULL,
1249         },
1250         [CPUHP_HRTIMERS_PREPARE] = {
1251                 .name                   = "hrtimers:prepare",
1252                 .startup.single         = hrtimers_prepare_cpu,
1253                 .teardown.single        = hrtimers_dead_cpu,
1254         },
1255         [CPUHP_SMPCFD_PREPARE] = {
1256                 .name                   = "smpcfd:prepare",
1257                 .startup.single         = smpcfd_prepare_cpu,
1258                 .teardown.single        = smpcfd_dead_cpu,
1259         },
1260         [CPUHP_RELAY_PREPARE] = {
1261                 .name                   = "relay:prepare",
1262                 .startup.single         = relay_prepare_cpu,
1263                 .teardown.single        = NULL,
1264         },
1265         [CPUHP_SLAB_PREPARE] = {
1266                 .name                   = "slab:prepare",
1267                 .startup.single         = slab_prepare_cpu,
1268                 .teardown.single        = slab_dead_cpu,
1269         },
1270         [CPUHP_RCUTREE_PREP] = {
1271                 .name                   = "RCU/tree:prepare",
1272                 .startup.single         = rcutree_prepare_cpu,
1273                 .teardown.single        = rcutree_dead_cpu,
1274         },
1275         /*
1276          * On the tear-down path, timers_dead_cpu() must be invoked
1277          * before blk_mq_queue_reinit_notify() from notify_dead(),
1278          * otherwise a RCU stall occurs.
1279          */
1280         [CPUHP_TIMERS_DEAD] = {
1281                 .name                   = "timers:dead",
1282                 .startup.single         = NULL,
1283                 .teardown.single        = timers_dead_cpu,
1284         },
1285         /* Kicks the plugged cpu into life */
1286         [CPUHP_BRINGUP_CPU] = {
1287                 .name                   = "cpu:bringup",
1288                 .startup.single         = bringup_cpu,
1289                 .teardown.single        = NULL,
1290                 .cant_stop              = true,
1291         },
1292         [CPUHP_AP_SMPCFD_DYING] = {
1293                 .name                   = "smpcfd:dying",
1294                 .startup.single         = NULL,
1295                 .teardown.single        = smpcfd_dying_cpu,
1296         },
1297         /*
1298          * Handled on controll processor until the plugged processor manages
1299          * this itself.
1300          */
1301         [CPUHP_TEARDOWN_CPU] = {
1302                 .name                   = "cpu:teardown",
1303                 .startup.single         = NULL,
1304                 .teardown.single        = takedown_cpu,
1305                 .cant_stop              = true,
1306         },
1307 #else
1308         [CPUHP_BRINGUP_CPU] = { },
1309 #endif
1310 };
1311
1312 /* Application processor state steps */
1313 static struct cpuhp_step cpuhp_ap_states[] = {
1314 #ifdef CONFIG_SMP
1315         /* Final state before CPU kills itself */
1316         [CPUHP_AP_IDLE_DEAD] = {
1317                 .name                   = "idle:dead",
1318         },
1319         /*
1320          * Last state before CPU enters the idle loop to die. Transient state
1321          * for synchronization.
1322          */
1323         [CPUHP_AP_OFFLINE] = {
1324                 .name                   = "ap:offline",
1325                 .cant_stop              = true,
1326         },
1327         /* First state is scheduler control. Interrupts are disabled */
1328         [CPUHP_AP_SCHED_STARTING] = {
1329                 .name                   = "sched:starting",
1330                 .startup.single         = sched_cpu_starting,
1331                 .teardown.single        = sched_cpu_dying,
1332         },
1333         [CPUHP_AP_RCUTREE_DYING] = {
1334                 .name                   = "RCU/tree:dying",
1335                 .startup.single         = NULL,
1336                 .teardown.single        = rcutree_dying_cpu,
1337         },
1338         /* Entry state on starting. Interrupts enabled from here on. Transient
1339          * state for synchronsization */
1340         [CPUHP_AP_ONLINE] = {
1341                 .name                   = "ap:online",
1342         },
1343         /* Handle smpboot threads park/unpark */
1344         [CPUHP_AP_SMPBOOT_THREADS] = {
1345                 .name                   = "smpboot/threads:online",
1346                 .startup.single         = smpboot_unpark_threads,
1347                 .teardown.single        = NULL,
1348         },
1349         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1350                 .name                   = "irq/affinity:online",
1351                 .startup.single         = irq_affinity_online_cpu,
1352                 .teardown.single        = NULL,
1353         },
1354         [CPUHP_AP_PERF_ONLINE] = {
1355                 .name                   = "perf:online",
1356                 .startup.single         = perf_event_init_cpu,
1357                 .teardown.single        = perf_event_exit_cpu,
1358         },
1359         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1360                 .name                   = "workqueue:online",
1361                 .startup.single         = workqueue_online_cpu,
1362                 .teardown.single        = workqueue_offline_cpu,
1363         },
1364         [CPUHP_AP_RCUTREE_ONLINE] = {
1365                 .name                   = "RCU/tree:online",
1366                 .startup.single         = rcutree_online_cpu,
1367                 .teardown.single        = rcutree_offline_cpu,
1368         },
1369 #endif
1370         /*
1371          * The dynamically registered state space is here
1372          */
1373
1374 #ifdef CONFIG_SMP
1375         /* Last state is scheduler control setting the cpu active */
1376         [CPUHP_AP_ACTIVE] = {
1377                 .name                   = "sched:active",
1378                 .startup.single         = sched_cpu_activate,
1379                 .teardown.single        = sched_cpu_deactivate,
1380         },
1381 #endif
1382
1383         /* CPU is fully up and running. */
1384         [CPUHP_ONLINE] = {
1385                 .name                   = "online",
1386                 .startup.single         = NULL,
1387                 .teardown.single        = NULL,
1388         },
1389 };
1390
1391 /* Sanity check for callbacks */
1392 static int cpuhp_cb_check(enum cpuhp_state state)
1393 {
1394         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1395                 return -EINVAL;
1396         return 0;
1397 }
1398
1399 /*
1400  * Returns a free for dynamic slot assignment of the Online state. The states
1401  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1402  * by having no name assigned.
1403  */
1404 static int cpuhp_reserve_state(enum cpuhp_state state)
1405 {
1406         enum cpuhp_state i, end;
1407         struct cpuhp_step *step;
1408
1409         switch (state) {
1410         case CPUHP_AP_ONLINE_DYN:
1411                 step = cpuhp_ap_states + CPUHP_AP_ONLINE_DYN;
1412                 end = CPUHP_AP_ONLINE_DYN_END;
1413                 break;
1414         case CPUHP_BP_PREPARE_DYN:
1415                 step = cpuhp_bp_states + CPUHP_BP_PREPARE_DYN;
1416                 end = CPUHP_BP_PREPARE_DYN_END;
1417                 break;
1418         default:
1419                 return -EINVAL;
1420         }
1421
1422         for (i = state; i <= end; i++, step++) {
1423                 if (!step->name)
1424                         return i;
1425         }
1426         WARN(1, "No more dynamic states available for CPU hotplug\n");
1427         return -ENOSPC;
1428 }
1429
1430 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1431                                  int (*startup)(unsigned int cpu),
1432                                  int (*teardown)(unsigned int cpu),
1433                                  bool multi_instance)
1434 {
1435         /* (Un)Install the callbacks for further cpu hotplug operations */
1436         struct cpuhp_step *sp;
1437         int ret = 0;
1438
1439         /*
1440          * If name is NULL, then the state gets removed.
1441          *
1442          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1443          * the first allocation from these dynamic ranges, so the removal
1444          * would trigger a new allocation and clear the wrong (already
1445          * empty) state, leaving the callbacks of the to be cleared state
1446          * dangling, which causes wreckage on the next hotplug operation.
1447          */
1448         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1449                      state == CPUHP_BP_PREPARE_DYN)) {
1450                 ret = cpuhp_reserve_state(state);
1451                 if (ret < 0)
1452                         return ret;
1453                 state = ret;
1454         }
1455         sp = cpuhp_get_step(state);
1456         if (name && sp->name)
1457                 return -EBUSY;
1458
1459         sp->startup.single = startup;
1460         sp->teardown.single = teardown;
1461         sp->name = name;
1462         sp->multi_instance = multi_instance;
1463         INIT_HLIST_HEAD(&sp->list);
1464         return ret;
1465 }
1466
1467 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1468 {
1469         return cpuhp_get_step(state)->teardown.single;
1470 }
1471
1472 /*
1473  * Call the startup/teardown function for a step either on the AP or
1474  * on the current CPU.
1475  */
1476 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1477                             struct hlist_node *node)
1478 {
1479         struct cpuhp_step *sp = cpuhp_get_step(state);
1480         int ret;
1481
1482         /*
1483          * If there's nothing to do, we done.
1484          * Relies on the union for multi_instance.
1485          */
1486         if ((bringup && !sp->startup.single) ||
1487             (!bringup && !sp->teardown.single))
1488                 return 0;
1489         /*
1490          * The non AP bound callbacks can fail on bringup. On teardown
1491          * e.g. module removal we crash for now.
1492          */
1493 #ifdef CONFIG_SMP
1494         if (cpuhp_is_ap_state(state))
1495                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1496         else
1497                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1498 #else
1499         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1500 #endif
1501         BUG_ON(ret && !bringup);
1502         return ret;
1503 }
1504
1505 /*
1506  * Called from __cpuhp_setup_state on a recoverable failure.
1507  *
1508  * Note: The teardown callbacks for rollback are not allowed to fail!
1509  */
1510 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1511                                    struct hlist_node *node)
1512 {
1513         int cpu;
1514
1515         /* Roll back the already executed steps on the other cpus */
1516         for_each_present_cpu(cpu) {
1517                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1518                 int cpustate = st->state;
1519
1520                 if (cpu >= failedcpu)
1521                         break;
1522
1523                 /* Did we invoke the startup call on that cpu ? */
1524                 if (cpustate >= state)
1525                         cpuhp_issue_call(cpu, state, false, node);
1526         }
1527 }
1528
1529 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1530                                           struct hlist_node *node,
1531                                           bool invoke)
1532 {
1533         struct cpuhp_step *sp;
1534         int cpu;
1535         int ret;
1536
1537         lockdep_assert_cpus_held();
1538
1539         sp = cpuhp_get_step(state);
1540         if (sp->multi_instance == false)
1541                 return -EINVAL;
1542
1543         mutex_lock(&cpuhp_state_mutex);
1544
1545         if (!invoke || !sp->startup.multi)
1546                 goto add_node;
1547
1548         /*
1549          * Try to call the startup callback for each present cpu
1550          * depending on the hotplug state of the cpu.
1551          */
1552         for_each_present_cpu(cpu) {
1553                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1554                 int cpustate = st->state;
1555
1556                 if (cpustate < state)
1557                         continue;
1558
1559                 ret = cpuhp_issue_call(cpu, state, true, node);
1560                 if (ret) {
1561                         if (sp->teardown.multi)
1562                                 cpuhp_rollback_install(cpu, state, node);
1563                         goto unlock;
1564                 }
1565         }
1566 add_node:
1567         ret = 0;
1568         hlist_add_head(node, &sp->list);
1569 unlock:
1570         mutex_unlock(&cpuhp_state_mutex);
1571         return ret;
1572 }
1573
1574 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1575                                bool invoke)
1576 {
1577         int ret;
1578
1579         cpus_read_lock();
1580         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1581         cpus_read_unlock();
1582         return ret;
1583 }
1584 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1585
1586 /**
1587  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1588  * @state:              The state to setup
1589  * @invoke:             If true, the startup function is invoked for cpus where
1590  *                      cpu state >= @state
1591  * @startup:            startup callback function
1592  * @teardown:           teardown callback function
1593  * @multi_instance:     State is set up for multiple instances which get
1594  *                      added afterwards.
1595  *
1596  * The caller needs to hold cpus read locked while calling this function.
1597  * Returns:
1598  *   On success:
1599  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1600  *      0 for all other states
1601  *   On failure: proper (negative) error code
1602  */
1603 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1604                                    const char *name, bool invoke,
1605                                    int (*startup)(unsigned int cpu),
1606                                    int (*teardown)(unsigned int cpu),
1607                                    bool multi_instance)
1608 {
1609         int cpu, ret = 0;
1610         bool dynstate;
1611
1612         lockdep_assert_cpus_held();
1613
1614         if (cpuhp_cb_check(state) || !name)
1615                 return -EINVAL;
1616
1617         mutex_lock(&cpuhp_state_mutex);
1618
1619         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1620                                     multi_instance);
1621
1622         dynstate = state == CPUHP_AP_ONLINE_DYN;
1623         if (ret > 0 && dynstate) {
1624                 state = ret;
1625                 ret = 0;
1626         }
1627
1628         if (ret || !invoke || !startup)
1629                 goto out;
1630
1631         /*
1632          * Try to call the startup callback for each present cpu
1633          * depending on the hotplug state of the cpu.
1634          */
1635         for_each_present_cpu(cpu) {
1636                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1637                 int cpustate = st->state;
1638
1639                 if (cpustate < state)
1640                         continue;
1641
1642                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1643                 if (ret) {
1644                         if (teardown)
1645                                 cpuhp_rollback_install(cpu, state, NULL);
1646                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1647                         goto out;
1648                 }
1649         }
1650 out:
1651         mutex_unlock(&cpuhp_state_mutex);
1652         /*
1653          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1654          * dynamically allocated state in case of success.
1655          */
1656         if (!ret && dynstate)
1657                 return state;
1658         return ret;
1659 }
1660 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1661
1662 int __cpuhp_setup_state(enum cpuhp_state state,
1663                         const char *name, bool invoke,
1664                         int (*startup)(unsigned int cpu),
1665                         int (*teardown)(unsigned int cpu),
1666                         bool multi_instance)
1667 {
1668         int ret;
1669
1670         cpus_read_lock();
1671         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1672                                              teardown, multi_instance);
1673         cpus_read_unlock();
1674         return ret;
1675 }
1676 EXPORT_SYMBOL(__cpuhp_setup_state);
1677
1678 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1679                                   struct hlist_node *node, bool invoke)
1680 {
1681         struct cpuhp_step *sp = cpuhp_get_step(state);
1682         int cpu;
1683
1684         BUG_ON(cpuhp_cb_check(state));
1685
1686         if (!sp->multi_instance)
1687                 return -EINVAL;
1688
1689         cpus_read_lock();
1690         mutex_lock(&cpuhp_state_mutex);
1691
1692         if (!invoke || !cpuhp_get_teardown_cb(state))
1693                 goto remove;
1694         /*
1695          * Call the teardown callback for each present cpu depending
1696          * on the hotplug state of the cpu. This function is not
1697          * allowed to fail currently!
1698          */
1699         for_each_present_cpu(cpu) {
1700                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1701                 int cpustate = st->state;
1702
1703                 if (cpustate >= state)
1704                         cpuhp_issue_call(cpu, state, false, node);
1705         }
1706
1707 remove:
1708         hlist_del(node);
1709         mutex_unlock(&cpuhp_state_mutex);
1710         cpus_read_unlock();
1711
1712         return 0;
1713 }
1714 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1715
1716 /**
1717  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1718  * @state:      The state to remove
1719  * @invoke:     If true, the teardown function is invoked for cpus where
1720  *              cpu state >= @state
1721  *
1722  * The caller needs to hold cpus read locked while calling this function.
1723  * The teardown callback is currently not allowed to fail. Think
1724  * about module removal!
1725  */
1726 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1727 {
1728         struct cpuhp_step *sp = cpuhp_get_step(state);
1729         int cpu;
1730
1731         BUG_ON(cpuhp_cb_check(state));
1732
1733         lockdep_assert_cpus_held();
1734
1735         mutex_lock(&cpuhp_state_mutex);
1736         if (sp->multi_instance) {
1737                 WARN(!hlist_empty(&sp->list),
1738                      "Error: Removing state %d which has instances left.\n",
1739                      state);
1740                 goto remove;
1741         }
1742
1743         if (!invoke || !cpuhp_get_teardown_cb(state))
1744                 goto remove;
1745
1746         /*
1747          * Call the teardown callback for each present cpu depending
1748          * on the hotplug state of the cpu. This function is not
1749          * allowed to fail currently!
1750          */
1751         for_each_present_cpu(cpu) {
1752                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1753                 int cpustate = st->state;
1754
1755                 if (cpustate >= state)
1756                         cpuhp_issue_call(cpu, state, false, NULL);
1757         }
1758 remove:
1759         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1760         mutex_unlock(&cpuhp_state_mutex);
1761 }
1762 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1763
1764 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1765 {
1766         cpus_read_lock();
1767         __cpuhp_remove_state_cpuslocked(state, invoke);
1768         cpus_read_unlock();
1769 }
1770 EXPORT_SYMBOL(__cpuhp_remove_state);
1771
1772 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1773 static ssize_t show_cpuhp_state(struct device *dev,
1774                                 struct device_attribute *attr, char *buf)
1775 {
1776         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1777
1778         return sprintf(buf, "%d\n", st->state);
1779 }
1780 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1781
1782 static ssize_t write_cpuhp_target(struct device *dev,
1783                                   struct device_attribute *attr,
1784                                   const char *buf, size_t count)
1785 {
1786         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1787         struct cpuhp_step *sp;
1788         int target, ret;
1789
1790         ret = kstrtoint(buf, 10, &target);
1791         if (ret)
1792                 return ret;
1793
1794 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1795         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1796                 return -EINVAL;
1797 #else
1798         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1799                 return -EINVAL;
1800 #endif
1801
1802         ret = lock_device_hotplug_sysfs();
1803         if (ret)
1804                 return ret;
1805
1806         mutex_lock(&cpuhp_state_mutex);
1807         sp = cpuhp_get_step(target);
1808         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1809         mutex_unlock(&cpuhp_state_mutex);
1810         if (ret)
1811                 goto out;
1812
1813         if (st->state < target)
1814                 ret = do_cpu_up(dev->id, target);
1815         else
1816                 ret = do_cpu_down(dev->id, target);
1817 out:
1818         unlock_device_hotplug();
1819         return ret ? ret : count;
1820 }
1821
1822 static ssize_t show_cpuhp_target(struct device *dev,
1823                                  struct device_attribute *attr, char *buf)
1824 {
1825         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1826
1827         return sprintf(buf, "%d\n", st->target);
1828 }
1829 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1830
1831
1832 static ssize_t write_cpuhp_fail(struct device *dev,
1833                                 struct device_attribute *attr,
1834                                 const char *buf, size_t count)
1835 {
1836         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1837         struct cpuhp_step *sp;
1838         int fail, ret;
1839
1840         ret = kstrtoint(buf, 10, &fail);
1841         if (ret)
1842                 return ret;
1843
1844         /*
1845          * Cannot fail STARTING/DYING callbacks.
1846          */
1847         if (cpuhp_is_atomic_state(fail))
1848                 return -EINVAL;
1849
1850         /*
1851          * Cannot fail anything that doesn't have callbacks.
1852          */
1853         mutex_lock(&cpuhp_state_mutex);
1854         sp = cpuhp_get_step(fail);
1855         if (!sp->startup.single && !sp->teardown.single)
1856                 ret = -EINVAL;
1857         mutex_unlock(&cpuhp_state_mutex);
1858         if (ret)
1859                 return ret;
1860
1861         st->fail = fail;
1862
1863         return count;
1864 }
1865
1866 static ssize_t show_cpuhp_fail(struct device *dev,
1867                                struct device_attribute *attr, char *buf)
1868 {
1869         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1870
1871         return sprintf(buf, "%d\n", st->fail);
1872 }
1873
1874 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1875
1876 static struct attribute *cpuhp_cpu_attrs[] = {
1877         &dev_attr_state.attr,
1878         &dev_attr_target.attr,
1879         &dev_attr_fail.attr,
1880         NULL
1881 };
1882
1883 static const struct attribute_group cpuhp_cpu_attr_group = {
1884         .attrs = cpuhp_cpu_attrs,
1885         .name = "hotplug",
1886         NULL
1887 };
1888
1889 static ssize_t show_cpuhp_states(struct device *dev,
1890                                  struct device_attribute *attr, char *buf)
1891 {
1892         ssize_t cur, res = 0;
1893         int i;
1894
1895         mutex_lock(&cpuhp_state_mutex);
1896         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1897                 struct cpuhp_step *sp = cpuhp_get_step(i);
1898
1899                 if (sp->name) {
1900                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1901                         buf += cur;
1902                         res += cur;
1903                 }
1904         }
1905         mutex_unlock(&cpuhp_state_mutex);
1906         return res;
1907 }
1908 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1909
1910 static struct attribute *cpuhp_cpu_root_attrs[] = {
1911         &dev_attr_states.attr,
1912         NULL
1913 };
1914
1915 static const struct attribute_group cpuhp_cpu_root_attr_group = {
1916         .attrs = cpuhp_cpu_root_attrs,
1917         .name = "hotplug",
1918         NULL
1919 };
1920
1921 static int __init cpuhp_sysfs_init(void)
1922 {
1923         int cpu, ret;
1924
1925         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1926                                  &cpuhp_cpu_root_attr_group);
1927         if (ret)
1928                 return ret;
1929
1930         for_each_possible_cpu(cpu) {
1931                 struct device *dev = get_cpu_device(cpu);
1932
1933                 if (!dev)
1934                         continue;
1935                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1936                 if (ret)
1937                         return ret;
1938         }
1939         return 0;
1940 }
1941 device_initcall(cpuhp_sysfs_init);
1942 #endif
1943
1944 /*
1945  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1946  * represents all NR_CPUS bits binary values of 1<<nr.
1947  *
1948  * It is used by cpumask_of() to get a constant address to a CPU
1949  * mask value that has a single bit set only.
1950  */
1951
1952 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1953 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1954 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1955 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1956 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1957
1958 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1959
1960         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1961         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1962 #if BITS_PER_LONG > 32
1963         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1964         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1965 #endif
1966 };
1967 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1968
1969 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1970 EXPORT_SYMBOL(cpu_all_bits);
1971
1972 #ifdef CONFIG_INIT_ALL_POSSIBLE
1973 struct cpumask __cpu_possible_mask __read_mostly
1974         = {CPU_BITS_ALL};
1975 #else
1976 struct cpumask __cpu_possible_mask __read_mostly;
1977 #endif
1978 EXPORT_SYMBOL(__cpu_possible_mask);
1979
1980 struct cpumask __cpu_online_mask __read_mostly;
1981 EXPORT_SYMBOL(__cpu_online_mask);
1982
1983 struct cpumask __cpu_present_mask __read_mostly;
1984 EXPORT_SYMBOL(__cpu_present_mask);
1985
1986 struct cpumask __cpu_active_mask __read_mostly;
1987 EXPORT_SYMBOL(__cpu_active_mask);
1988
1989 void init_cpu_present(const struct cpumask *src)
1990 {
1991         cpumask_copy(&__cpu_present_mask, src);
1992 }
1993
1994 void init_cpu_possible(const struct cpumask *src)
1995 {
1996         cpumask_copy(&__cpu_possible_mask, src);
1997 }
1998
1999 void init_cpu_online(const struct cpumask *src)
2000 {
2001         cpumask_copy(&__cpu_online_mask, src);
2002 }
2003
2004 /*
2005  * Activate the first processor.
2006  */
2007 void __init boot_cpu_init(void)
2008 {
2009         int cpu = smp_processor_id();
2010
2011         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2012         set_cpu_online(cpu, true);
2013         set_cpu_active(cpu, true);
2014         set_cpu_present(cpu, true);
2015         set_cpu_possible(cpu, true);
2016
2017 #ifdef CONFIG_SMP
2018         __boot_cpu_id = cpu;
2019 #endif
2020 }
2021
2022 /*
2023  * Must be called _AFTER_ setting up the per_cpu areas
2024  */
2025 void __init boot_cpu_state_init(void)
2026 {
2027         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
2028 }