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
 762         /*
 763          * Prevent irq alloc/free while the dying cpu reorganizes the
 764          * interrupt affinities.
 765          */
 766         irq_lock_sparse();
 767
 768         /*
 769          * So now all preempt/rcu users must observe !cpu_active().
 770          */
 771         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 772         if (err) {
 773                 /* CPU refused to die */
 774                 irq_unlock_sparse();
 775                 /* Unpark the hotplug thread so we can rollback there */
 776                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 777                 return err;
 778         }
 779         BUG_ON(cpu_online(cpu));
 780
 781         /*
 782          * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
 783          * runnable tasks from the cpu, there's only the idle task left now
 784          * that the migration thread is done doing the stop_machine thing.
 785          *
 786          * Wait for the stop thread to go away.
 787          */
 788         wait_for_ap_thread(st, false);
 789         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 790
 791         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 792         irq_unlock_sparse();
 793
 794         hotplug_cpu__broadcast_tick_pull(cpu);
 795         /* This actually kills the CPU. */
 796         __cpu_die(cpu);
 797
 798         tick_cleanup_dead_cpu(cpu);
 799         rcutree_migrate_callbacks(cpu);
 800         return 0;
 801 }
 802
 803 static void cpuhp_complete_idle_dead(void *arg)
 804 {
 805         struct cpuhp_cpu_state *st = arg;
 806
 807         complete_ap_thread(st, false);
 808 }
 809
 810 void cpuhp_report_idle_dead(void)
 811 {
 812         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 813
 814         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 815         rcu_report_dead(smp_processor_id());
 816         st->state = CPUHP_AP_IDLE_DEAD;
 817         /*
 818          * We cannot call complete after rcu_report_dead() so we delegate it
 819          * to an online cpu.
 820          */
 821         smp_call_function_single(cpumask_first(cpu_online_mask),
 822                                  cpuhp_complete_idle_dead, st, 0);
 823 }
 824
 825 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 826 {
 827         for (st->state++; st->state < st->target; st->state++) {
 828                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 829
 830                 if (!step->skip_onerr)
 831                         cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 832         }
 833 }
 834
 835 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 836                                 enum cpuhp_state target)
 837 {
 838         enum cpuhp_state prev_state = st->state;
 839         int ret = 0;
 840
 841         for (; st->state > target; st->state--) {
 842                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 843                 if (ret) {
 844                         st->target = prev_state;
 845                         undo_cpu_down(cpu, st);
 846                         break;
 847                 }
 848         }
 849         return ret;
 850 }
 851
 852 /* Requires cpu_add_remove_lock to be held */
 853 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 854                            enum cpuhp_state target)
 855 {
 856         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 857         int prev_state, ret = 0;
 858
 859         if (num_online_cpus() == 1)
 860                 return -EBUSY;
 861
 862         if (!cpu_present(cpu))
 863                 return -EINVAL;
 864
 865         cpus_write_lock();
 866
 867         cpuhp_tasks_frozen = tasks_frozen;
 868
 869         prev_state = cpuhp_set_state(st, target);
 870         /*
 871          * If the current CPU state is in the range of the AP hotplug thread,
 872          * then we need to kick the thread.
 873          */
 874         if (st->state > CPUHP_TEARDOWN_CPU) {
 875                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 876                 ret = cpuhp_kick_ap_work(cpu);
 877                 /*
 878                  * The AP side has done the error rollback already. Just
 879                  * return the error code..
 880                  */
 881                 if (ret)
 882                         goto out;
 883
 884                 /*
 885                  * We might have stopped still in the range of the AP hotplug
 886                  * thread. Nothing to do anymore.
 887                  */
 888                 if (st->state > CPUHP_TEARDOWN_CPU)
 889                         goto out;
 890
 891                 st->target = target;
 892         }
 893         /*
 894          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 895          * to do the further cleanups.
 896          */
 897         ret = cpuhp_down_callbacks(cpu, st, target);
 898         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 899                 cpuhp_reset_state(st, prev_state);
 900                 __cpuhp_kick_ap(st);
 901         }
 902
 903 out:
 904         cpus_write_unlock();
 905         /*
 906          * Do post unplug cleanup. This is still protected against
 907          * concurrent CPU hotplug via cpu_add_remove_lock.
 908          */
 909         lockup_detector_cleanup();
 910         return ret;
 911 }
 912
 913 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 914 {
 915         int err;
 916
 917         cpu_maps_update_begin();
 918
 919         if (cpu_hotplug_disabled) {
 920                 err = -EBUSY;
 921                 goto out;
 922         }
 923
 924         err = _cpu_down(cpu, 0, target);
 925
 926 out:
 927         cpu_maps_update_done();
 928         return err;
 929 }
 930
 931 int cpu_down(unsigned int cpu)
 932 {
 933         return do_cpu_down(cpu, CPUHP_OFFLINE);
 934 }
 935 EXPORT_SYMBOL(cpu_down);
 936
 937 #else
 938 #define takedown_cpu            NULL
 939 #endif /*CONFIG_HOTPLUG_CPU*/
 940
 941 /**
 942  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
 943  * @cpu: cpu that just started
 944  *
 945  * It must be called by the arch code on the new cpu, before the new cpu
 946  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 947  */
 948 void notify_cpu_starting(unsigned int cpu)
 949 {
 950         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 951         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 952         int ret;
 953
 954         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
 955         while (st->state < target) {
 956                 st->state++;
 957                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 958                 /*
 959                  * STARTING must not fail!
 960                  */
 961                 WARN_ON_ONCE(ret);
 962         }
 963 }
 964
 965 /*
 966  * Called from the idle task. Wake up the controlling task which brings the
 967  * stopper and the hotplug thread of the upcoming CPU up and then delegates
 968  * the rest of the online bringup to the hotplug thread.
 969  */
 970 void cpuhp_online_idle(enum cpuhp_state state)
 971 {
 972         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 973
 974         /* Happens for the boot cpu */
 975         if (state != CPUHP_AP_ONLINE_IDLE)
 976                 return;
 977
 978         st->state = CPUHP_AP_ONLINE_IDLE;
 979         complete_ap_thread(st, true);
 980 }
 981
 982 /* Requires cpu_add_remove_lock to be held */
 983 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 984 {
 985         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 986         struct task_struct *idle;
 987         int ret = 0;
 988
 989         cpus_write_lock();
 990
 991         if (!cpu_present(cpu)) {
 992                 ret = -EINVAL;
 993                 goto out;
 994         }
 995
 996         /*
 997          * The caller of do_cpu_up might have raced with another
 998          * caller. Ignore it for now.
 999          */
1000         if (st->state >= target)
1001                 goto out;
1002
1003         if (st->state == CPUHP_OFFLINE) {
1004                 /* Let it fail before we try to bring the cpu up */
1005                 idle = idle_thread_get(cpu);
1006                 if (IS_ERR(idle)) {
1007                         ret = PTR_ERR(idle);
1008                         goto out;
1009                 }
1010         }
1011
1012         cpuhp_tasks_frozen = tasks_frozen;
1013
1014         cpuhp_set_state(st, target);
1015         /*
1016          * If the current CPU state is in the range of the AP hotplug thread,
1017          * then we need to kick the thread once more.
1018          */
1019         if (st->state > CPUHP_BRINGUP_CPU) {
1020                 ret = cpuhp_kick_ap_work(cpu);
1021                 /*
1022                  * The AP side has done the error rollback already. Just
1023                  * return the error code..
1024                  */
1025                 if (ret)
1026                         goto out;
1027         }
1028
1029         /*
1030          * Try to reach the target state. We max out on the BP at
1031          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1032          * responsible for bringing it up to the target state.
1033          */
1034         target = min((int)target, CPUHP_BRINGUP_CPU);
1035         ret = cpuhp_up_callbacks(cpu, st, target);
1036 out:
1037         cpus_write_unlock();
1038         return ret;
1039 }
1040
1041 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1042 {
1043         int err = 0;
1044
1045         if (!cpu_possible(cpu)) {
1046                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1047                        cpu);
1048 #if defined(CONFIG_IA64)
1049                 pr_err("please check additional_cpus= boot parameter\n");
1050 #endif
1051                 return -EINVAL;
1052         }
1053
1054         err = try_online_node(cpu_to_node(cpu));
1055         if (err)
1056                 return err;
1057
1058         cpu_maps_update_begin();
1059
1060         if (cpu_hotplug_disabled) {
1061                 err = -EBUSY;
1062                 goto out;
1063         }
1064
1065         err = _cpu_up(cpu, 0, target);
1066 out:
1067         cpu_maps_update_done();
1068         return err;
1069 }
1070
1071 int cpu_up(unsigned int cpu)
1072 {
1073         return do_cpu_up(cpu, CPUHP_ONLINE);
1074 }
1075 EXPORT_SYMBOL_GPL(cpu_up);
1076
1077 #ifdef CONFIG_PM_SLEEP_SMP
1078 static cpumask_var_t frozen_cpus;
1079
1080 int freeze_secondary_cpus(int primary)
1081 {
1082         int cpu, error = 0;
1083
1084         cpu_maps_update_begin();
1085         if (!cpu_online(primary))
1086                 primary = cpumask_first(cpu_online_mask);
1087         /*
1088          * We take down all of the non-boot CPUs in one shot to avoid races
1089          * with the userspace trying to use the CPU hotplug at the same time
1090          */
1091         cpumask_clear(frozen_cpus);
1092
1093         pr_info("Disabling non-boot CPUs ...\n");
1094         for_each_online_cpu(cpu) {
1095                 if (cpu == primary)
1096                         continue;
1097                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1098                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1099                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1100                 if (!error)
1101                         cpumask_set_cpu(cpu, frozen_cpus);
1102                 else {
1103                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1104                         break;
1105                 }
1106         }
1107
1108         if (!error)
1109                 BUG_ON(num_online_cpus() > 1);
1110         else
1111                 pr_err("Non-boot CPUs are not disabled\n");
1112
1113         /*
1114          * Make sure the CPUs won't be enabled by someone else. We need to do
1115          * this even in case of failure as all disable_nonboot_cpus() users are
1116          * supposed to do enable_nonboot_cpus() on the failure path.
1117          */
1118         cpu_hotplug_disabled++;
1119
1120         cpu_maps_update_done();
1121         return error;
1122 }
1123
1124 void __weak arch_enable_nonboot_cpus_begin(void)
1125 {
1126 }
1127
1128 void __weak arch_enable_nonboot_cpus_end(void)
1129 {
1130 }
1131
1132 void enable_nonboot_cpus(void)
1133 {
1134         int cpu, error;
1135
1136         /* Allow everyone to use the CPU hotplug again */
1137         cpu_maps_update_begin();
1138         __cpu_hotplug_enable();
1139         if (cpumask_empty(frozen_cpus))
1140                 goto out;
1141
1142         pr_info("Enabling non-boot CPUs ...\n");
1143
1144         arch_enable_nonboot_cpus_begin();
1145
1146         for_each_cpu(cpu, frozen_cpus) {
1147                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1148                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1149                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1150                 if (!error) {
1151                         pr_info("CPU%d is up\n", cpu);
1152                         continue;
1153                 }
1154                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1155         }
1156
1157         arch_enable_nonboot_cpus_end();
1158
1159         cpumask_clear(frozen_cpus);
1160 out:
1161         cpu_maps_update_done();
1162 }
1163
1164 static int __init alloc_frozen_cpus(void)
1165 {
1166         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1167                 return -ENOMEM;
1168         return 0;
1169 }
1170 core_initcall(alloc_frozen_cpus);
1171
1172 /*
1173  * When callbacks for CPU hotplug notifications are being executed, we must
1174  * ensure that the state of the system with respect to the tasks being frozen
1175  * or not, as reported by the notification, remains unchanged *throughout the
1176  * duration* of the execution of the callbacks.
1177  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1178  *
1179  * This synchronization is implemented by mutually excluding regular CPU
1180  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1181  * Hibernate notifications.
1182  */
1183 static int
1184 cpu_hotplug_pm_callback(struct notifier_block *nb,
1185                         unsigned long action, void *ptr)
1186 {
1187         switch (action) {
1188
1189         case PM_SUSPEND_PREPARE:
1190         case PM_HIBERNATION_PREPARE:
1191                 cpu_hotplug_disable();
1192                 break;
1193
1194         case PM_POST_SUSPEND:
1195         case PM_POST_HIBERNATION:
1196                 cpu_hotplug_enable();
1197                 break;
1198
1199         default:
1200                 return NOTIFY_DONE;
1201         }
1202
1203         return NOTIFY_OK;
1204 }
1205
1206
1207 static int __init cpu_hotplug_pm_sync_init(void)
1208 {
1209         /*
1210          * cpu_hotplug_pm_callback has higher priority than x86
1211          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1212          * to disable cpu hotplug to avoid cpu hotplug race.
1213          */
1214         pm_notifier(cpu_hotplug_pm_callback, 0);
1215         return 0;
1216 }
1217 core_initcall(cpu_hotplug_pm_sync_init);
1218
1219 #endif /* CONFIG_PM_SLEEP_SMP */
1220
1221 int __boot_cpu_id;
1222
1223 #endif /* CONFIG_SMP */
1224
1225 /* Boot processor state steps */
1226 static struct cpuhp_step cpuhp_bp_states[] = {
1227         [CPUHP_OFFLINE] = {
1228                 .name                   = "offline",
1229                 .startup.single         = NULL,
1230                 .teardown.single        = NULL,
1231         },
1232 #ifdef CONFIG_SMP
1233         [CPUHP_CREATE_THREADS]= {
1234                 .name                   = "threads:prepare",
1235                 .startup.single         = smpboot_create_threads,
1236                 .teardown.single        = NULL,
1237                 .cant_stop              = true,
1238         },
1239         [CPUHP_PERF_PREPARE] = {
1240                 .name                   = "perf:prepare",
1241                 .startup.single         = perf_event_init_cpu,
1242                 .teardown.single        = perf_event_exit_cpu,
1243         },
1244         [CPUHP_WORKQUEUE_PREP] = {
1245                 .name                   = "workqueue:prepare",
1246                 .startup.single         = workqueue_prepare_cpu,
1247                 .teardown.single        = NULL,
1248         },
1249         [CPUHP_HRTIMERS_PREPARE] = {
1250                 .name                   = "hrtimers:prepare",
1251                 .startup.single         = hrtimers_prepare_cpu,
1252                 .teardown.single        = hrtimers_dead_cpu,
1253         },
1254         [CPUHP_SMPCFD_PREPARE] = {
1255                 .name                   = "smpcfd:prepare",
1256                 .startup.single         = smpcfd_prepare_cpu,
1257                 .teardown.single        = smpcfd_dead_cpu,
1258         },
1259         [CPUHP_RELAY_PREPARE] = {
1260                 .name                   = "relay:prepare",
1261                 .startup.single         = relay_prepare_cpu,
1262                 .teardown.single        = NULL,
1263         },
1264         [CPUHP_SLAB_PREPARE] = {
1265                 .name                   = "slab:prepare",
1266                 .startup.single         = slab_prepare_cpu,
1267                 .teardown.single        = slab_dead_cpu,
1268         },
1269         [CPUHP_RCUTREE_PREP] = {
1270                 .name                   = "RCU/tree:prepare",
1271                 .startup.single         = rcutree_prepare_cpu,
1272                 .teardown.single        = rcutree_dead_cpu,
1273         },
1274         /*
1275          * On the tear-down path, timers_dead_cpu() must be invoked
1276          * before blk_mq_queue_reinit_notify() from notify_dead(),
1277          * otherwise a RCU stall occurs.
1278          */
1279         [CPUHP_TIMERS_PREPARE] = {
1280                 .name                   = "timers:dead",
1281                 .startup.single         = timers_prepare_cpu,
1282                 .teardown.single        = timers_dead_cpu,
1283         },
1284         /* Kicks the plugged cpu into life */
1285         [CPUHP_BRINGUP_CPU] = {
1286                 .name                   = "cpu:bringup",
1287                 .startup.single         = bringup_cpu,
1288                 .teardown.single        = NULL,
1289                 .cant_stop              = true,
1290         },
1291         /*
1292          * Handled on controll processor until the plugged processor manages
1293          * this itself.
1294          */
1295         [CPUHP_TEARDOWN_CPU] = {
1296                 .name                   = "cpu:teardown",
1297                 .startup.single         = NULL,
1298                 .teardown.single        = takedown_cpu,
1299                 .cant_stop              = true,
1300         },
1301 #else
1302         [CPUHP_BRINGUP_CPU] = { },
1303 #endif
1304 };
1305
1306 /* Application processor state steps */
1307 static struct cpuhp_step cpuhp_ap_states[] = {
1308 #ifdef CONFIG_SMP
1309         /* Final state before CPU kills itself */
1310         [CPUHP_AP_IDLE_DEAD] = {
1311                 .name                   = "idle:dead",
1312         },
1313         /*
1314          * Last state before CPU enters the idle loop to die. Transient state
1315          * for synchronization.
1316          */
1317         [CPUHP_AP_OFFLINE] = {
1318                 .name                   = "ap:offline",
1319                 .cant_stop              = true,
1320         },
1321         /* First state is scheduler control. Interrupts are disabled */
1322         [CPUHP_AP_SCHED_STARTING] = {
1323                 .name                   = "sched:starting",
1324                 .startup.single         = sched_cpu_starting,
1325                 .teardown.single        = sched_cpu_dying,
1326         },
1327         [CPUHP_AP_RCUTREE_DYING] = {
1328                 .name                   = "RCU/tree:dying",
1329                 .startup.single         = NULL,
1330                 .teardown.single        = rcutree_dying_cpu,
1331         },
1332         [CPUHP_AP_SMPCFD_DYING] = {
1333                 .name                   = "smpcfd:dying",
1334                 .startup.single         = NULL,
1335                 .teardown.single        = smpcfd_dying_cpu,
1336         },
1337         /* Entry state on starting. Interrupts enabled from here on. Transient
1338          * state for synchronsization */
1339         [CPUHP_AP_ONLINE] = {
1340                 .name                   = "ap:online",
1341         },
1342         /* Handle smpboot threads park/unpark */
1343         [CPUHP_AP_SMPBOOT_THREADS] = {
1344                 .name                   = "smpboot/threads:online",
1345                 .startup.single         = smpboot_unpark_threads,
1346                 .teardown.single        = smpboot_park_threads,
1347         },
1348         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1349                 .name                   = "irq/affinity:online",
1350                 .startup.single         = irq_affinity_online_cpu,
1351                 .teardown.single        = NULL,
1352         },
1353         [CPUHP_AP_PERF_ONLINE] = {
1354                 .name                   = "perf:online",
1355                 .startup.single         = perf_event_init_cpu,
1356                 .teardown.single        = perf_event_exit_cpu,
1357         },
1358         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1359                 .name                   = "workqueue:online",
1360                 .startup.single         = workqueue_online_cpu,
1361                 .teardown.single        = workqueue_offline_cpu,
1362         },
1363         [CPUHP_AP_RCUTREE_ONLINE] = {
1364                 .name                   = "RCU/tree:online",
1365                 .startup.single         = rcutree_online_cpu,
1366                 .teardown.single        = rcutree_offline_cpu,
1367         },
1368 #endif
1369         /*
1370          * The dynamically registered state space is here
1371          */
1372
1373 #ifdef CONFIG_SMP
1374         /* Last state is scheduler control setting the cpu active */
1375         [CPUHP_AP_ACTIVE] = {
1376                 .name                   = "sched:active",
1377                 .startup.single         = sched_cpu_activate,
1378                 .teardown.single        = sched_cpu_deactivate,
1379         },
1380 #endif
1381
1382         /* CPU is fully up and running. */
1383         [CPUHP_ONLINE] = {
1384                 .name                   = "online",
1385                 .startup.single         = NULL,
1386                 .teardown.single        = NULL,
1387         },
1388 };
1389
1390 /* Sanity check for callbacks */
1391 static int cpuhp_cb_check(enum cpuhp_state state)
1392 {
1393         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1394                 return -EINVAL;
1395         return 0;
1396 }
1397
1398 /*
1399  * Returns a free for dynamic slot assignment of the Online state. The states
1400  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1401  * by having no name assigned.
1402  */
1403 static int cpuhp_reserve_state(enum cpuhp_state state)
1404 {
1405         enum cpuhp_state i, end;
1406         struct cpuhp_step *step;
1407
1408         switch (state) {
1409         case CPUHP_AP_ONLINE_DYN:
1410                 step = cpuhp_ap_states + CPUHP_AP_ONLINE_DYN;
1411                 end = CPUHP_AP_ONLINE_DYN_END;
1412                 break;
1413         case CPUHP_BP_PREPARE_DYN:
1414                 step = cpuhp_bp_states + CPUHP_BP_PREPARE_DYN;
1415                 end = CPUHP_BP_PREPARE_DYN_END;
1416                 break;
1417         default:
1418                 return -EINVAL;
1419         }
1420
1421         for (i = state; i <= end; i++, step++) {
1422                 if (!step->name)
1423                         return i;
1424         }
1425         WARN(1, "No more dynamic states available for CPU hotplug\n");
1426         return -ENOSPC;
1427 }
1428
1429 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1430                                  int (*startup)(unsigned int cpu),
1431                                  int (*teardown)(unsigned int cpu),
1432                                  bool multi_instance)
1433 {
1434         /* (Un)Install the callbacks for further cpu hotplug operations */
1435         struct cpuhp_step *sp;
1436         int ret = 0;
1437
1438         /*
1439          * If name is NULL, then the state gets removed.
1440          *
1441          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1442          * the first allocation from these dynamic ranges, so the removal
1443          * would trigger a new allocation and clear the wrong (already
1444          * empty) state, leaving the callbacks of the to be cleared state
1445          * dangling, which causes wreckage on the next hotplug operation.
1446          */
1447         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1448                      state == CPUHP_BP_PREPARE_DYN)) {
1449                 ret = cpuhp_reserve_state(state);
1450                 if (ret < 0)
1451                         return ret;
1452                 state = ret;
1453         }
1454         sp = cpuhp_get_step(state);
1455         if (name && sp->name)
1456                 return -EBUSY;
1457
1458         sp->startup.single = startup;
1459         sp->teardown.single = teardown;
1460         sp->name = name;
1461         sp->multi_instance = multi_instance;
1462         INIT_HLIST_HEAD(&sp->list);
1463         return ret;
1464 }
1465
1466 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1467 {
1468         return cpuhp_get_step(state)->teardown.single;
1469 }
1470
1471 /*
1472  * Call the startup/teardown function for a step either on the AP or
1473  * on the current CPU.
1474  */
1475 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1476                             struct hlist_node *node)
1477 {
1478         struct cpuhp_step *sp = cpuhp_get_step(state);
1479         int ret;
1480
1481         /*
1482          * If there's nothing to do, we done.
1483          * Relies on the union for multi_instance.
1484          */
1485         if ((bringup && !sp->startup.single) ||
1486             (!bringup && !sp->teardown.single))
1487                 return 0;
1488         /*
1489          * The non AP bound callbacks can fail on bringup. On teardown
1490          * e.g. module removal we crash for now.
1491          */
1492 #ifdef CONFIG_SMP
1493         if (cpuhp_is_ap_state(state))
1494                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1495         else
1496                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1497 #else
1498         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1499 #endif
1500         BUG_ON(ret && !bringup);
1501         return ret;
1502 }
1503
1504 /*
1505  * Called from __cpuhp_setup_state on a recoverable failure.
1506  *
1507  * Note: The teardown callbacks for rollback are not allowed to fail!
1508  */
1509 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1510                                    struct hlist_node *node)
1511 {
1512         int cpu;
1513
1514         /* Roll back the already executed steps on the other cpus */
1515         for_each_present_cpu(cpu) {
1516                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1517                 int cpustate = st->state;
1518
1519                 if (cpu >= failedcpu)
1520                         break;
1521
1522                 /* Did we invoke the startup call on that cpu ? */
1523                 if (cpustate >= state)
1524                         cpuhp_issue_call(cpu, state, false, node);
1525         }
1526 }
1527
1528 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1529                                           struct hlist_node *node,
1530                                           bool invoke)
1531 {
1532         struct cpuhp_step *sp;
1533         int cpu;
1534         int ret;
1535
1536         lockdep_assert_cpus_held();
1537
1538         sp = cpuhp_get_step(state);
1539         if (sp->multi_instance == false)
1540                 return -EINVAL;
1541
1542         mutex_lock(&cpuhp_state_mutex);
1543
1544         if (!invoke || !sp->startup.multi)
1545                 goto add_node;
1546
1547         /*
1548          * Try to call the startup callback for each present cpu
1549          * depending on the hotplug state of the cpu.
1550          */
1551         for_each_present_cpu(cpu) {
1552                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1553                 int cpustate = st->state;
1554
1555                 if (cpustate < state)
1556                         continue;
1557
1558                 ret = cpuhp_issue_call(cpu, state, true, node);
1559                 if (ret) {
1560                         if (sp->teardown.multi)
1561                                 cpuhp_rollback_install(cpu, state, node);
1562                         goto unlock;
1563                 }
1564         }
1565 add_node:
1566         ret = 0;
1567         hlist_add_head(node, &sp->list);
1568 unlock:
1569         mutex_unlock(&cpuhp_state_mutex);
1570         return ret;
1571 }
1572
1573 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1574                                bool invoke)
1575 {
1576         int ret;
1577
1578         cpus_read_lock();
1579         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1580         cpus_read_unlock();
1581         return ret;
1582 }
1583 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1584
1585 /**
1586  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1587  * @state:              The state to setup
1588  * @invoke:             If true, the startup function is invoked for cpus where
1589  *                      cpu state >= @state
1590  * @startup:            startup callback function
1591  * @teardown:           teardown callback function
1592  * @multi_instance:     State is set up for multiple instances which get
1593  *                      added afterwards.
1594  *
1595  * The caller needs to hold cpus read locked while calling this function.
1596  * Returns:
1597  *   On success:
1598  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1599  *      0 for all other states
1600  *   On failure: proper (negative) error code
1601  */
1602 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1603                                    const char *name, bool invoke,
1604                                    int (*startup)(unsigned int cpu),
1605                                    int (*teardown)(unsigned int cpu),
1606                                    bool multi_instance)
1607 {
1608         int cpu, ret = 0;
1609         bool dynstate;
1610
1611         lockdep_assert_cpus_held();
1612
1613         if (cpuhp_cb_check(state) || !name)
1614                 return -EINVAL;
1615
1616         mutex_lock(&cpuhp_state_mutex);
1617
1618         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1619                                     multi_instance);
1620
1621         dynstate = state == CPUHP_AP_ONLINE_DYN;
1622         if (ret > 0 && dynstate) {
1623                 state = ret;
1624                 ret = 0;
1625         }
1626
1627         if (ret || !invoke || !startup)
1628                 goto out;
1629
1630         /*
1631          * Try to call the startup callback for each present cpu
1632          * depending on the hotplug state of the cpu.
1633          */
1634         for_each_present_cpu(cpu) {
1635                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1636                 int cpustate = st->state;
1637
1638                 if (cpustate < state)
1639                         continue;
1640
1641                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1642                 if (ret) {
1643                         if (teardown)
1644                                 cpuhp_rollback_install(cpu, state, NULL);
1645                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1646                         goto out;
1647                 }
1648         }
1649 out:
1650         mutex_unlock(&cpuhp_state_mutex);
1651         /*
1652          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1653          * dynamically allocated state in case of success.
1654          */
1655         if (!ret && dynstate)
1656                 return state;
1657         return ret;
1658 }
1659 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1660
1661 int __cpuhp_setup_state(enum cpuhp_state state,
1662                         const char *name, bool invoke,
1663                         int (*startup)(unsigned int cpu),
1664                         int (*teardown)(unsigned int cpu),
1665                         bool multi_instance)
1666 {
1667         int ret;
1668
1669         cpus_read_lock();
1670         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1671                                              teardown, multi_instance);
1672         cpus_read_unlock();
1673         return ret;
1674 }
1675 EXPORT_SYMBOL(__cpuhp_setup_state);
1676
1677 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1678                                   struct hlist_node *node, bool invoke)
1679 {
1680         struct cpuhp_step *sp = cpuhp_get_step(state);
1681         int cpu;
1682
1683         BUG_ON(cpuhp_cb_check(state));
1684
1685         if (!sp->multi_instance)
1686                 return -EINVAL;
1687
1688         cpus_read_lock();
1689         mutex_lock(&cpuhp_state_mutex);
1690
1691         if (!invoke || !cpuhp_get_teardown_cb(state))
1692                 goto remove;
1693         /*
1694          * Call the teardown callback for each present cpu depending
1695          * on the hotplug state of the cpu. This function is not
1696          * allowed to fail currently!
1697          */
1698         for_each_present_cpu(cpu) {
1699                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1700                 int cpustate = st->state;
1701
1702                 if (cpustate >= state)
1703                         cpuhp_issue_call(cpu, state, false, node);
1704         }
1705
1706 remove:
1707         hlist_del(node);
1708         mutex_unlock(&cpuhp_state_mutex);
1709         cpus_read_unlock();
1710
1711         return 0;
1712 }
1713 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1714
1715 /**
1716  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1717  * @state:      The state to remove
1718  * @invoke:     If true, the teardown function is invoked for cpus where
1719  *              cpu state >= @state
1720  *
1721  * The caller needs to hold cpus read locked while calling this function.
1722  * The teardown callback is currently not allowed to fail. Think
1723  * about module removal!
1724  */
1725 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1726 {
1727         struct cpuhp_step *sp = cpuhp_get_step(state);
1728         int cpu;
1729
1730         BUG_ON(cpuhp_cb_check(state));
1731
1732         lockdep_assert_cpus_held();
1733
1734         mutex_lock(&cpuhp_state_mutex);
1735         if (sp->multi_instance) {
1736                 WARN(!hlist_empty(&sp->list),
1737                      "Error: Removing state %d which has instances left.\n",
1738                      state);
1739                 goto remove;
1740         }
1741
1742         if (!invoke || !cpuhp_get_teardown_cb(state))
1743                 goto remove;
1744
1745         /*
1746          * Call the teardown callback for each present cpu depending
1747          * on the hotplug state of the cpu. This function is not
1748          * allowed to fail currently!
1749          */
1750         for_each_present_cpu(cpu) {
1751                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1752                 int cpustate = st->state;
1753
1754                 if (cpustate >= state)
1755                         cpuhp_issue_call(cpu, state, false, NULL);
1756         }
1757 remove:
1758         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1759         mutex_unlock(&cpuhp_state_mutex);
1760 }
1761 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1762
1763 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1764 {
1765         cpus_read_lock();
1766         __cpuhp_remove_state_cpuslocked(state, invoke);
1767         cpus_read_unlock();
1768 }
1769 EXPORT_SYMBOL(__cpuhp_remove_state);
1770
1771 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1772 static ssize_t show_cpuhp_state(struct device *dev,
1773                                 struct device_attribute *attr, char *buf)
1774 {
1775         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1776
1777         return sprintf(buf, "%d\n", st->state);
1778 }
1779 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1780
1781 static ssize_t write_cpuhp_target(struct device *dev,
1782                                   struct device_attribute *attr,
1783                                   const char *buf, size_t count)
1784 {
1785         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1786         struct cpuhp_step *sp;
1787         int target, ret;
1788
1789         ret = kstrtoint(buf, 10, &target);
1790         if (ret)
1791                 return ret;
1792
1793 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1794         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1795                 return -EINVAL;
1796 #else
1797         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1798                 return -EINVAL;
1799 #endif
1800
1801         ret = lock_device_hotplug_sysfs();
1802         if (ret)
1803                 return ret;
1804
1805         mutex_lock(&cpuhp_state_mutex);
1806         sp = cpuhp_get_step(target);
1807         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1808         mutex_unlock(&cpuhp_state_mutex);
1809         if (ret)
1810                 goto out;
1811
1812         if (st->state < target)
1813                 ret = do_cpu_up(dev->id, target);
1814         else
1815                 ret = do_cpu_down(dev->id, target);
1816 out:
1817         unlock_device_hotplug();
1818         return ret ? ret : count;
1819 }
1820
1821 static ssize_t show_cpuhp_target(struct device *dev,
1822                                  struct device_attribute *attr, char *buf)
1823 {
1824         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1825
1826         return sprintf(buf, "%d\n", st->target);
1827 }
1828 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1829
1830
1831 static ssize_t write_cpuhp_fail(struct device *dev,
1832                                 struct device_attribute *attr,
1833                                 const char *buf, size_t count)
1834 {
1835         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1836         struct cpuhp_step *sp;
1837         int fail, ret;
1838
1839         ret = kstrtoint(buf, 10, &fail);
1840         if (ret)
1841                 return ret;
1842
1843         /*
1844          * Cannot fail STARTING/DYING callbacks.
1845          */
1846         if (cpuhp_is_atomic_state(fail))
1847                 return -EINVAL;
1848
1849         /*
1850          * Cannot fail anything that doesn't have callbacks.
1851          */
1852         mutex_lock(&cpuhp_state_mutex);
1853         sp = cpuhp_get_step(fail);
1854         if (!sp->startup.single && !sp->teardown.single)
1855                 ret = -EINVAL;
1856         mutex_unlock(&cpuhp_state_mutex);
1857         if (ret)
1858                 return ret;
1859
1860         st->fail = fail;
1861
1862         return count;
1863 }
1864
1865 static ssize_t show_cpuhp_fail(struct device *dev,
1866                                struct device_attribute *attr, char *buf)
1867 {
1868         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1869
1870         return sprintf(buf, "%d\n", st->fail);
1871 }
1872
1873 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1874
1875 static struct attribute *cpuhp_cpu_attrs[] = {
1876         &dev_attr_state.attr,
1877         &dev_attr_target.attr,
1878         &dev_attr_fail.attr,
1879         NULL
1880 };
1881
1882 static const struct attribute_group cpuhp_cpu_attr_group = {
1883         .attrs = cpuhp_cpu_attrs,
1884         .name = "hotplug",
1885         NULL
1886 };
1887
1888 static ssize_t show_cpuhp_states(struct device *dev,
1889                                  struct device_attribute *attr, char *buf)
1890 {
1891         ssize_t cur, res = 0;
1892         int i;
1893
1894         mutex_lock(&cpuhp_state_mutex);
1895         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1896                 struct cpuhp_step *sp = cpuhp_get_step(i);
1897
1898                 if (sp->name) {
1899                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1900                         buf += cur;
1901                         res += cur;
1902                 }
1903         }
1904         mutex_unlock(&cpuhp_state_mutex);
1905         return res;
1906 }
1907 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1908
1909 static struct attribute *cpuhp_cpu_root_attrs[] = {
1910         &dev_attr_states.attr,
1911         NULL
1912 };
1913
1914 static const struct attribute_group cpuhp_cpu_root_attr_group = {
1915         .attrs = cpuhp_cpu_root_attrs,
1916         .name = "hotplug",
1917         NULL
1918 };
1919
1920 static int __init cpuhp_sysfs_init(void)
1921 {
1922         int cpu, ret;
1923
1924         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1925                                  &cpuhp_cpu_root_attr_group);
1926         if (ret)
1927                 return ret;
1928
1929         for_each_possible_cpu(cpu) {
1930                 struct device *dev = get_cpu_device(cpu);
1931
1932                 if (!dev)
1933                         continue;
1934                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1935                 if (ret)
1936                         return ret;
1937         }
1938         return 0;
1939 }
1940 device_initcall(cpuhp_sysfs_init);
1941 #endif
1942
1943 /*
1944  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1945  * represents all NR_CPUS bits binary values of 1<<nr.
1946  *
1947  * It is used by cpumask_of() to get a constant address to a CPU
1948  * mask value that has a single bit set only.
1949  */
1950
1951 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1952 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1953 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1954 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1955 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1956
1957 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1958
1959         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1960         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1961 #if BITS_PER_LONG > 32
1962         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1963         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1964 #endif
1965 };
1966 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1967
1968 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1969 EXPORT_SYMBOL(cpu_all_bits);
1970
1971 #ifdef CONFIG_INIT_ALL_POSSIBLE
1972 struct cpumask __cpu_possible_mask __read_mostly
1973         = {CPU_BITS_ALL};
1974 #else
1975 struct cpumask __cpu_possible_mask __read_mostly;
1976 #endif
1977 EXPORT_SYMBOL(__cpu_possible_mask);
1978
1979 struct cpumask __cpu_online_mask __read_mostly;
1980 EXPORT_SYMBOL(__cpu_online_mask);
1981
1982 struct cpumask __cpu_present_mask __read_mostly;
1983 EXPORT_SYMBOL(__cpu_present_mask);
1984
1985 struct cpumask __cpu_active_mask __read_mostly;
1986 EXPORT_SYMBOL(__cpu_active_mask);
1987
1988 void init_cpu_present(const struct cpumask *src)
1989 {
1990         cpumask_copy(&__cpu_present_mask, src);
1991 }
1992
1993 void init_cpu_possible(const struct cpumask *src)
1994 {
1995         cpumask_copy(&__cpu_possible_mask, src);
1996 }
1997
1998 void init_cpu_online(const struct cpumask *src)
1999 {
2000         cpumask_copy(&__cpu_online_mask, src);
2001 }
2002
2003 /*
2004  * Activate the first processor.
2005  */
2006 void __init boot_cpu_init(void)
2007 {
2008         int cpu = smp_processor_id();
2009
2010         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2011         set_cpu_online(cpu, true);
2012         set_cpu_active(cpu, true);
2013         set_cpu_present(cpu, true);
2014         set_cpu_possible(cpu, true);
2015
2016 #ifdef CONFIG_SMP
2017         __boot_cpu_id = cpu;
2018 #endif
2019 }
2020
2021 /*
2022  * Must be called _AFTER_ setting up the per_cpu areas
2023  */
2024 void __init boot_cpu_hotplug_init(void)
2025 {
2026         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
2027 }