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         return ret;
 636 }
 637
 638 static int cpuhp_kick_ap_work(unsigned int cpu)
 639 {
 640         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 641         enum cpuhp_state prev_state = st->state;
 642         int ret;
 643
 644         cpuhp_lock_acquire(false);
 645         cpuhp_lock_release(false);
 646
 647         cpuhp_lock_acquire(true);
 648         cpuhp_lock_release(true);
 649
 650         trace_cpuhp_enter(cpu, st->target, prev_state, cpuhp_kick_ap_work);
 651         ret = cpuhp_kick_ap(st, st->target);
 652         trace_cpuhp_exit(cpu, st->state, prev_state, ret);
 653
 654         return ret;
 655 }
 656
 657 static struct smp_hotplug_thread cpuhp_threads = {
 658         .store                  = &cpuhp_state.thread,
 659         .create                 = &cpuhp_create,
 660         .thread_should_run      = cpuhp_should_run,
 661         .thread_fn              = cpuhp_thread_fun,
 662         .thread_comm            = "cpuhp/%u",
 663         .selfparking            = true,
 664 };
 665
 666 void __init cpuhp_threads_init(void)
 667 {
 668         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 669         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 670 }
 671
 672 #ifdef CONFIG_HOTPLUG_CPU
 673 /**
 674  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 675  * @cpu: a CPU id
 676  *
 677  * This function walks all processes, finds a valid mm struct for each one and
 678  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 679  * trivial, there are various non-obvious corner cases, which this function
 680  * tries to solve in a safe manner.
 681  *
 682  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 683  * be called only for an already offlined CPU.
 684  */
 685 void clear_tasks_mm_cpumask(int cpu)
 686 {
 687         struct task_struct *p;
 688
 689         /*
 690          * This function is called after the cpu is taken down and marked
 691          * offline, so its not like new tasks will ever get this cpu set in
 692          * their mm mask. -- Peter Zijlstra
 693          * Thus, we may use rcu_read_lock() here, instead of grabbing
 694          * full-fledged tasklist_lock.
 695          */
 696         WARN_ON(cpu_online(cpu));
 697         rcu_read_lock();
 698         for_each_process(p) {
 699                 struct task_struct *t;
 700
 701                 /*
 702                  * Main thread might exit, but other threads may still have
 703                  * a valid mm. Find one.
 704                  */
 705                 t = find_lock_task_mm(p);
 706                 if (!t)
 707                         continue;
 708                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 709                 task_unlock(t);
 710         }
 711         rcu_read_unlock();
 712 }
 713
 714 /* Take this CPU down. */
 715 static int take_cpu_down(void *_param)
 716 {
 717         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 718         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 719         int err, cpu = smp_processor_id();
 720         int ret;
 721
 722         /* Ensure this CPU doesn't handle any more interrupts. */
 723         err = __cpu_disable();
 724         if (err < 0)
 725                 return err;
 726
 727         /*
 728          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 729          * do this step again.
 730          */
 731         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 732         st->state--;
 733         /* Invoke the former CPU_DYING callbacks */
 734         for (; st->state > target; st->state--) {
 735                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 736                 /*
 737                  * DYING must not fail!
 738                  */
 739                 WARN_ON_ONCE(ret);
 740         }
 741
 742         /* Give up timekeeping duties */
 743         tick_handover_do_timer();
 744         /* Park the stopper thread */
 745         stop_machine_park(cpu);
 746         return 0;
 747 }
 748
 749 static int takedown_cpu(unsigned int cpu)
 750 {
 751         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 752         int err;
 753
 754         /* Park the smpboot threads */
 755         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 756         smpboot_park_threads(cpu);
 757
 758         /*
 759          * Prevent irq alloc/free while the dying cpu reorganizes the
 760          * interrupt affinities.
 761          */
 762         irq_lock_sparse();
 763
 764         /*
 765          * So now all preempt/rcu users must observe !cpu_active().
 766          */
 767         err = stop_machine_cpuslocked(take_cpu_down, NULL, cpumask_of(cpu));
 768         if (err) {
 769                 /* CPU refused to die */
 770                 irq_unlock_sparse();
 771                 /* Unpark the hotplug thread so we can rollback there */
 772                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 773                 return err;
 774         }
 775         BUG_ON(cpu_online(cpu));
 776
 777         /*
 778          * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
 779          * runnable tasks from the cpu, there's only the idle task left now
 780          * that the migration thread is done doing the stop_machine thing.
 781          *
 782          * Wait for the stop thread to go away.
 783          */
 784         wait_for_ap_thread(st, false);
 785         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 786
 787         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 788         irq_unlock_sparse();
 789
 790         hotplug_cpu__broadcast_tick_pull(cpu);
 791         /* This actually kills the CPU. */
 792         __cpu_die(cpu);
 793
 794         tick_cleanup_dead_cpu(cpu);
 795         rcutree_migrate_callbacks(cpu);
 796         return 0;
 797 }
 798
 799 static void cpuhp_complete_idle_dead(void *arg)
 800 {
 801         struct cpuhp_cpu_state *st = arg;
 802
 803         complete_ap_thread(st, false);
 804 }
 805
 806 void cpuhp_report_idle_dead(void)
 807 {
 808         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 809
 810         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 811         rcu_report_dead(smp_processor_id());
 812         st->state = CPUHP_AP_IDLE_DEAD;
 813         /*
 814          * We cannot call complete after rcu_report_dead() so we delegate it
 815          * to an online cpu.
 816          */
 817         smp_call_function_single(cpumask_first(cpu_online_mask),
 818                                  cpuhp_complete_idle_dead, st, 0);
 819 }
 820
 821 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 822 {
 823         for (st->state++; st->state < st->target; st->state++) {
 824                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 825
 826                 if (!step->skip_onerr)
 827                         cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 828         }
 829 }
 830
 831 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 832                                 enum cpuhp_state target)
 833 {
 834         enum cpuhp_state prev_state = st->state;
 835         int ret = 0;
 836
 837         for (; st->state > target; st->state--) {
 838                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL, NULL);
 839                 if (ret) {
 840                         st->target = prev_state;
 841                         undo_cpu_down(cpu, st);
 842                         break;
 843                 }
 844         }
 845         return ret;
 846 }
 847
 848 /* Requires cpu_add_remove_lock to be held */
 849 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 850                            enum cpuhp_state target)
 851 {
 852         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 853         int prev_state, ret = 0;
 854
 855         if (num_online_cpus() == 1)
 856                 return -EBUSY;
 857
 858         if (!cpu_present(cpu))
 859                 return -EINVAL;
 860
 861         cpus_write_lock();
 862
 863         cpuhp_tasks_frozen = tasks_frozen;
 864
 865         prev_state = cpuhp_set_state(st, target);
 866         /*
 867          * If the current CPU state is in the range of the AP hotplug thread,
 868          * then we need to kick the thread.
 869          */
 870         if (st->state > CPUHP_TEARDOWN_CPU) {
 871                 st->target = max((int)target, CPUHP_TEARDOWN_CPU);
 872                 ret = cpuhp_kick_ap_work(cpu);
 873                 /*
 874                  * The AP side has done the error rollback already. Just
 875                  * return the error code..
 876                  */
 877                 if (ret)
 878                         goto out;
 879
 880                 /*
 881                  * We might have stopped still in the range of the AP hotplug
 882                  * thread. Nothing to do anymore.
 883                  */
 884                 if (st->state > CPUHP_TEARDOWN_CPU)
 885                         goto out;
 886
 887                 st->target = target;
 888         }
 889         /*
 890          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 891          * to do the further cleanups.
 892          */
 893         ret = cpuhp_down_callbacks(cpu, st, target);
 894         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 895                 cpuhp_reset_state(st, prev_state);
 896                 __cpuhp_kick_ap(st);
 897         }
 898
 899 out:
 900         cpus_write_unlock();
 901         /*
 902          * Do post unplug cleanup. This is still protected against
 903          * concurrent CPU hotplug via cpu_add_remove_lock.
 904          */
 905         lockup_detector_cleanup();
 906         return ret;
 907 }
 908
 909 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 910 {
 911         int err;
 912
 913         cpu_maps_update_begin();
 914
 915         if (cpu_hotplug_disabled) {
 916                 err = -EBUSY;
 917                 goto out;
 918         }
 919
 920         err = _cpu_down(cpu, 0, target);
 921
 922 out:
 923         cpu_maps_update_done();
 924         return err;
 925 }
 926
 927 int cpu_down(unsigned int cpu)
 928 {
 929         return do_cpu_down(cpu, CPUHP_OFFLINE);
 930 }
 931 EXPORT_SYMBOL(cpu_down);
 932
 933 #else
 934 #define takedown_cpu            NULL
 935 #endif /*CONFIG_HOTPLUG_CPU*/
 936
 937 /**
 938  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
 939  * @cpu: cpu that just started
 940  *
 941  * It must be called by the arch code on the new cpu, before the new cpu
 942  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 943  */
 944 void notify_cpu_starting(unsigned int cpu)
 945 {
 946         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 947         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 948         int ret;
 949
 950         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
 951         while (st->state < target) {
 952                 st->state++;
 953                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL, NULL);
 954                 /*
 955                  * STARTING must not fail!
 956                  */
 957                 WARN_ON_ONCE(ret);
 958         }
 959 }
 960
 961 /*
 962  * Called from the idle task. Wake up the controlling task which brings the
 963  * stopper and the hotplug thread of the upcoming CPU up and then delegates
 964  * the rest of the online bringup to the hotplug thread.
 965  */
 966 void cpuhp_online_idle(enum cpuhp_state state)
 967 {
 968         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 969
 970         /* Happens for the boot cpu */
 971         if (state != CPUHP_AP_ONLINE_IDLE)
 972                 return;
 973
 974         st->state = CPUHP_AP_ONLINE_IDLE;
 975         complete_ap_thread(st, true);
 976 }
 977
 978 /* Requires cpu_add_remove_lock to be held */
 979 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 980 {
 981         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 982         struct task_struct *idle;
 983         int ret = 0;
 984
 985         cpus_write_lock();
 986
 987         if (!cpu_present(cpu)) {
 988                 ret = -EINVAL;
 989                 goto out;
 990         }
 991
 992         /*
 993          * The caller of do_cpu_up might have raced with another
 994          * caller. Ignore it for now.
 995          */
 996         if (st->state >= target)
 997                 goto out;
 998
 999         if (st->state == CPUHP_OFFLINE) {
1000                 /* Let it fail before we try to bring the cpu up */
1001                 idle = idle_thread_get(cpu);
1002                 if (IS_ERR(idle)) {
1003                         ret = PTR_ERR(idle);
1004                         goto out;
1005                 }
1006         }
1007
1008         cpuhp_tasks_frozen = tasks_frozen;
1009
1010         cpuhp_set_state(st, target);
1011         /*
1012          * If the current CPU state is in the range of the AP hotplug thread,
1013          * then we need to kick the thread once more.
1014          */
1015         if (st->state > CPUHP_BRINGUP_CPU) {
1016                 ret = cpuhp_kick_ap_work(cpu);
1017                 /*
1018                  * The AP side has done the error rollback already. Just
1019                  * return the error code..
1020                  */
1021                 if (ret)
1022                         goto out;
1023         }
1024
1025         /*
1026          * Try to reach the target state. We max out on the BP at
1027          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1028          * responsible for bringing it up to the target state.
1029          */
1030         target = min((int)target, CPUHP_BRINGUP_CPU);
1031         ret = cpuhp_up_callbacks(cpu, st, target);
1032 out:
1033         cpus_write_unlock();
1034         return ret;
1035 }
1036
1037 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1038 {
1039         int err = 0;
1040
1041         if (!cpu_possible(cpu)) {
1042                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1043                        cpu);
1044 #if defined(CONFIG_IA64)
1045                 pr_err("please check additional_cpus= boot parameter\n");
1046 #endif
1047                 return -EINVAL;
1048         }
1049
1050         err = try_online_node(cpu_to_node(cpu));
1051         if (err)
1052                 return err;
1053
1054         cpu_maps_update_begin();
1055
1056         if (cpu_hotplug_disabled) {
1057                 err = -EBUSY;
1058                 goto out;
1059         }
1060
1061         err = _cpu_up(cpu, 0, target);
1062 out:
1063         cpu_maps_update_done();
1064         return err;
1065 }
1066
1067 int cpu_up(unsigned int cpu)
1068 {
1069         return do_cpu_up(cpu, CPUHP_ONLINE);
1070 }
1071 EXPORT_SYMBOL_GPL(cpu_up);
1072
1073 #ifdef CONFIG_PM_SLEEP_SMP
1074 static cpumask_var_t frozen_cpus;
1075
1076 int freeze_secondary_cpus(int primary)
1077 {
1078         int cpu, error = 0;
1079
1080         cpu_maps_update_begin();
1081         if (!cpu_online(primary))
1082                 primary = cpumask_first(cpu_online_mask);
1083         /*
1084          * We take down all of the non-boot CPUs in one shot to avoid races
1085          * with the userspace trying to use the CPU hotplug at the same time
1086          */
1087         cpumask_clear(frozen_cpus);
1088
1089         pr_info("Disabling non-boot CPUs ...\n");
1090         for_each_online_cpu(cpu) {
1091                 if (cpu == primary)
1092                         continue;
1093                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1094                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1095                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1096                 if (!error)
1097                         cpumask_set_cpu(cpu, frozen_cpus);
1098                 else {
1099                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1100                         break;
1101                 }
1102         }
1103
1104         if (!error)
1105                 BUG_ON(num_online_cpus() > 1);
1106         else
1107                 pr_err("Non-boot CPUs are not disabled\n");
1108
1109         /*
1110          * Make sure the CPUs won't be enabled by someone else. We need to do
1111          * this even in case of failure as all disable_nonboot_cpus() users are
1112          * supposed to do enable_nonboot_cpus() on the failure path.
1113          */
1114         cpu_hotplug_disabled++;
1115
1116         cpu_maps_update_done();
1117         return error;
1118 }
1119
1120 void __weak arch_enable_nonboot_cpus_begin(void)
1121 {
1122 }
1123
1124 void __weak arch_enable_nonboot_cpus_end(void)
1125 {
1126 }
1127
1128 void enable_nonboot_cpus(void)
1129 {
1130         int cpu, error;
1131
1132         /* Allow everyone to use the CPU hotplug again */
1133         cpu_maps_update_begin();
1134         __cpu_hotplug_enable();
1135         if (cpumask_empty(frozen_cpus))
1136                 goto out;
1137
1138         pr_info("Enabling non-boot CPUs ...\n");
1139
1140         arch_enable_nonboot_cpus_begin();
1141
1142         for_each_cpu(cpu, frozen_cpus) {
1143                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1144                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1145                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1146                 if (!error) {
1147                         pr_info("CPU%d is up\n", cpu);
1148                         continue;
1149                 }
1150                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1151         }
1152
1153         arch_enable_nonboot_cpus_end();
1154
1155         cpumask_clear(frozen_cpus);
1156 out:
1157         cpu_maps_update_done();
1158 }
1159
1160 static int __init alloc_frozen_cpus(void)
1161 {
1162         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1163                 return -ENOMEM;
1164         return 0;
1165 }
1166 core_initcall(alloc_frozen_cpus);
1167
1168 /*
1169  * When callbacks for CPU hotplug notifications are being executed, we must
1170  * ensure that the state of the system with respect to the tasks being frozen
1171  * or not, as reported by the notification, remains unchanged *throughout the
1172  * duration* of the execution of the callbacks.
1173  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1174  *
1175  * This synchronization is implemented by mutually excluding regular CPU
1176  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1177  * Hibernate notifications.
1178  */
1179 static int
1180 cpu_hotplug_pm_callback(struct notifier_block *nb,
1181                         unsigned long action, void *ptr)
1182 {
1183         switch (action) {
1184
1185         case PM_SUSPEND_PREPARE:
1186         case PM_HIBERNATION_PREPARE:
1187                 cpu_hotplug_disable();
1188                 break;
1189
1190         case PM_POST_SUSPEND:
1191         case PM_POST_HIBERNATION:
1192                 cpu_hotplug_enable();
1193                 break;
1194
1195         default:
1196                 return NOTIFY_DONE;
1197         }
1198
1199         return NOTIFY_OK;
1200 }
1201
1202
1203 static int __init cpu_hotplug_pm_sync_init(void)
1204 {
1205         /*
1206          * cpu_hotplug_pm_callback has higher priority than x86
1207          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1208          * to disable cpu hotplug to avoid cpu hotplug race.
1209          */
1210         pm_notifier(cpu_hotplug_pm_callback, 0);
1211         return 0;
1212 }
1213 core_initcall(cpu_hotplug_pm_sync_init);
1214
1215 #endif /* CONFIG_PM_SLEEP_SMP */
1216
1217 int __boot_cpu_id;
1218
1219 #endif /* CONFIG_SMP */
1220
1221 /* Boot processor state steps */
1222 static struct cpuhp_step cpuhp_bp_states[] = {
1223         [CPUHP_OFFLINE] = {
1224                 .name                   = "offline",
1225                 .startup.single         = NULL,
1226                 .teardown.single        = NULL,
1227         },
1228 #ifdef CONFIG_SMP
1229         [CPUHP_CREATE_THREADS]= {
1230                 .name                   = "threads:prepare",
1231                 .startup.single         = smpboot_create_threads,
1232                 .teardown.single        = NULL,
1233                 .cant_stop              = true,
1234         },
1235         [CPUHP_PERF_PREPARE] = {
1236                 .name                   = "perf:prepare",
1237                 .startup.single         = perf_event_init_cpu,
1238                 .teardown.single        = perf_event_exit_cpu,
1239         },
1240         [CPUHP_WORKQUEUE_PREP] = {
1241                 .name                   = "workqueue:prepare",
1242                 .startup.single         = workqueue_prepare_cpu,
1243                 .teardown.single        = NULL,
1244         },
1245         [CPUHP_HRTIMERS_PREPARE] = {
1246                 .name                   = "hrtimers:prepare",
1247                 .startup.single         = hrtimers_prepare_cpu,
1248                 .teardown.single        = hrtimers_dead_cpu,
1249         },
1250         [CPUHP_SMPCFD_PREPARE] = {
1251                 .name                   = "smpcfd:prepare",
1252                 .startup.single         = smpcfd_prepare_cpu,
1253                 .teardown.single        = smpcfd_dead_cpu,
1254         },
1255         [CPUHP_RELAY_PREPARE] = {
1256                 .name                   = "relay:prepare",
1257                 .startup.single         = relay_prepare_cpu,
1258                 .teardown.single        = NULL,
1259         },
1260         [CPUHP_SLAB_PREPARE] = {
1261                 .name                   = "slab:prepare",
1262                 .startup.single         = slab_prepare_cpu,
1263                 .teardown.single        = slab_dead_cpu,
1264         },
1265         [CPUHP_RCUTREE_PREP] = {
1266                 .name                   = "RCU/tree:prepare",
1267                 .startup.single         = rcutree_prepare_cpu,
1268                 .teardown.single        = rcutree_dead_cpu,
1269         },
1270         /*
1271          * On the tear-down path, timers_dead_cpu() must be invoked
1272          * before blk_mq_queue_reinit_notify() from notify_dead(),
1273          * otherwise a RCU stall occurs.
1274          */
1275         [CPUHP_TIMERS_DEAD] = {
1276                 .name                   = "timers:dead",
1277                 .startup.single         = NULL,
1278                 .teardown.single        = timers_dead_cpu,
1279         },
1280         /* Kicks the plugged cpu into life */
1281         [CPUHP_BRINGUP_CPU] = {
1282                 .name                   = "cpu:bringup",
1283                 .startup.single         = bringup_cpu,
1284                 .teardown.single        = NULL,
1285                 .cant_stop              = true,
1286         },
1287         [CPUHP_AP_SMPCFD_DYING] = {
1288                 .name                   = "smpcfd:dying",
1289                 .startup.single         = NULL,
1290                 .teardown.single        = smpcfd_dying_cpu,
1291         },
1292         /*
1293          * Handled on controll processor until the plugged processor manages
1294          * this itself.
1295          */
1296         [CPUHP_TEARDOWN_CPU] = {
1297                 .name                   = "cpu:teardown",
1298                 .startup.single         = NULL,
1299                 .teardown.single        = takedown_cpu,
1300                 .cant_stop              = true,
1301         },
1302 #else
1303         [CPUHP_BRINGUP_CPU] = { },
1304 #endif
1305 };
1306
1307 /* Application processor state steps */
1308 static struct cpuhp_step cpuhp_ap_states[] = {
1309 #ifdef CONFIG_SMP
1310         /* Final state before CPU kills itself */
1311         [CPUHP_AP_IDLE_DEAD] = {
1312                 .name                   = "idle:dead",
1313         },
1314         /*
1315          * Last state before CPU enters the idle loop to die. Transient state
1316          * for synchronization.
1317          */
1318         [CPUHP_AP_OFFLINE] = {
1319                 .name                   = "ap:offline",
1320                 .cant_stop              = true,
1321         },
1322         /* First state is scheduler control. Interrupts are disabled */
1323         [CPUHP_AP_SCHED_STARTING] = {
1324                 .name                   = "sched:starting",
1325                 .startup.single         = sched_cpu_starting,
1326                 .teardown.single        = sched_cpu_dying,
1327         },
1328         [CPUHP_AP_RCUTREE_DYING] = {
1329                 .name                   = "RCU/tree:dying",
1330                 .startup.single         = NULL,
1331                 .teardown.single        = rcutree_dying_cpu,
1332         },
1333         /* Entry state on starting. Interrupts enabled from here on. Transient
1334          * state for synchronsization */
1335         [CPUHP_AP_ONLINE] = {
1336                 .name                   = "ap:online",
1337         },
1338         /* Handle smpboot threads park/unpark */
1339         [CPUHP_AP_SMPBOOT_THREADS] = {
1340                 .name                   = "smpboot/threads:online",
1341                 .startup.single         = smpboot_unpark_threads,
1342                 .teardown.single        = NULL,
1343         },
1344         [CPUHP_AP_IRQ_AFFINITY_ONLINE] = {
1345                 .name                   = "irq/affinity:online",
1346                 .startup.single         = irq_affinity_online_cpu,
1347                 .teardown.single        = NULL,
1348         },
1349         [CPUHP_AP_PERF_ONLINE] = {
1350                 .name                   = "perf:online",
1351                 .startup.single         = perf_event_init_cpu,
1352                 .teardown.single        = perf_event_exit_cpu,
1353         },
1354         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1355                 .name                   = "workqueue:online",
1356                 .startup.single         = workqueue_online_cpu,
1357                 .teardown.single        = workqueue_offline_cpu,
1358         },
1359         [CPUHP_AP_RCUTREE_ONLINE] = {
1360                 .name                   = "RCU/tree:online",
1361                 .startup.single         = rcutree_online_cpu,
1362                 .teardown.single        = rcutree_offline_cpu,
1363         },
1364 #endif
1365         /*
1366          * The dynamically registered state space is here
1367          */
1368
1369 #ifdef CONFIG_SMP
1370         /* Last state is scheduler control setting the cpu active */
1371         [CPUHP_AP_ACTIVE] = {
1372                 .name                   = "sched:active",
1373                 .startup.single         = sched_cpu_activate,
1374                 .teardown.single        = sched_cpu_deactivate,
1375         },
1376 #endif
1377
1378         /* CPU is fully up and running. */
1379         [CPUHP_ONLINE] = {
1380                 .name                   = "online",
1381                 .startup.single         = NULL,
1382                 .teardown.single        = NULL,
1383         },
1384 };
1385
1386 /* Sanity check for callbacks */
1387 static int cpuhp_cb_check(enum cpuhp_state state)
1388 {
1389         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1390                 return -EINVAL;
1391         return 0;
1392 }
1393
1394 /*
1395  * Returns a free for dynamic slot assignment of the Online state. The states
1396  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1397  * by having no name assigned.
1398  */
1399 static int cpuhp_reserve_state(enum cpuhp_state state)
1400 {
1401         enum cpuhp_state i, end;
1402         struct cpuhp_step *step;
1403
1404         switch (state) {
1405         case CPUHP_AP_ONLINE_DYN:
1406                 step = cpuhp_ap_states + CPUHP_AP_ONLINE_DYN;
1407                 end = CPUHP_AP_ONLINE_DYN_END;
1408                 break;
1409         case CPUHP_BP_PREPARE_DYN:
1410                 step = cpuhp_bp_states + CPUHP_BP_PREPARE_DYN;
1411                 end = CPUHP_BP_PREPARE_DYN_END;
1412                 break;
1413         default:
1414                 return -EINVAL;
1415         }
1416
1417         for (i = state; i <= end; i++, step++) {
1418                 if (!step->name)
1419                         return i;
1420         }
1421         WARN(1, "No more dynamic states available for CPU hotplug\n");
1422         return -ENOSPC;
1423 }
1424
1425 static int cpuhp_store_callbacks(enum cpuhp_state state, const char *name,
1426                                  int (*startup)(unsigned int cpu),
1427                                  int (*teardown)(unsigned int cpu),
1428                                  bool multi_instance)
1429 {
1430         /* (Un)Install the callbacks for further cpu hotplug operations */
1431         struct cpuhp_step *sp;
1432         int ret = 0;
1433
1434         /*
1435          * If name is NULL, then the state gets removed.
1436          *
1437          * CPUHP_AP_ONLINE_DYN and CPUHP_BP_PREPARE_DYN are handed out on
1438          * the first allocation from these dynamic ranges, so the removal
1439          * would trigger a new allocation and clear the wrong (already
1440          * empty) state, leaving the callbacks of the to be cleared state
1441          * dangling, which causes wreckage on the next hotplug operation.
1442          */
1443         if (name && (state == CPUHP_AP_ONLINE_DYN ||
1444                      state == CPUHP_BP_PREPARE_DYN)) {
1445                 ret = cpuhp_reserve_state(state);
1446                 if (ret < 0)
1447                         return ret;
1448                 state = ret;
1449         }
1450         sp = cpuhp_get_step(state);
1451         if (name && sp->name)
1452                 return -EBUSY;
1453
1454         sp->startup.single = startup;
1455         sp->teardown.single = teardown;
1456         sp->name = name;
1457         sp->multi_instance = multi_instance;
1458         INIT_HLIST_HEAD(&sp->list);
1459         return ret;
1460 }
1461
1462 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1463 {
1464         return cpuhp_get_step(state)->teardown.single;
1465 }
1466
1467 /*
1468  * Call the startup/teardown function for a step either on the AP or
1469  * on the current CPU.
1470  */
1471 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1472                             struct hlist_node *node)
1473 {
1474         struct cpuhp_step *sp = cpuhp_get_step(state);
1475         int ret;
1476
1477         /*
1478          * If there's nothing to do, we done.
1479          * Relies on the union for multi_instance.
1480          */
1481         if ((bringup && !sp->startup.single) ||
1482             (!bringup && !sp->teardown.single))
1483                 return 0;
1484         /*
1485          * The non AP bound callbacks can fail on bringup. On teardown
1486          * e.g. module removal we crash for now.
1487          */
1488 #ifdef CONFIG_SMP
1489         if (cpuhp_is_ap_state(state))
1490                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1491         else
1492                 ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1493 #else
1494         ret = cpuhp_invoke_callback(cpu, state, bringup, node, NULL);
1495 #endif
1496         BUG_ON(ret && !bringup);
1497         return ret;
1498 }
1499
1500 /*
1501  * Called from __cpuhp_setup_state on a recoverable failure.
1502  *
1503  * Note: The teardown callbacks for rollback are not allowed to fail!
1504  */
1505 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1506                                    struct hlist_node *node)
1507 {
1508         int cpu;
1509
1510         /* Roll back the already executed steps on the other cpus */
1511         for_each_present_cpu(cpu) {
1512                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1513                 int cpustate = st->state;
1514
1515                 if (cpu >= failedcpu)
1516                         break;
1517
1518                 /* Did we invoke the startup call on that cpu ? */
1519                 if (cpustate >= state)
1520                         cpuhp_issue_call(cpu, state, false, node);
1521         }
1522 }
1523
1524 int __cpuhp_state_add_instance_cpuslocked(enum cpuhp_state state,
1525                                           struct hlist_node *node,
1526                                           bool invoke)
1527 {
1528         struct cpuhp_step *sp;
1529         int cpu;
1530         int ret;
1531
1532         lockdep_assert_cpus_held();
1533
1534         sp = cpuhp_get_step(state);
1535         if (sp->multi_instance == false)
1536                 return -EINVAL;
1537
1538         mutex_lock(&cpuhp_state_mutex);
1539
1540         if (!invoke || !sp->startup.multi)
1541                 goto add_node;
1542
1543         /*
1544          * Try to call the startup callback for each present cpu
1545          * depending on the hotplug state of the cpu.
1546          */
1547         for_each_present_cpu(cpu) {
1548                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1549                 int cpustate = st->state;
1550
1551                 if (cpustate < state)
1552                         continue;
1553
1554                 ret = cpuhp_issue_call(cpu, state, true, node);
1555                 if (ret) {
1556                         if (sp->teardown.multi)
1557                                 cpuhp_rollback_install(cpu, state, node);
1558                         goto unlock;
1559                 }
1560         }
1561 add_node:
1562         ret = 0;
1563         hlist_add_head(node, &sp->list);
1564 unlock:
1565         mutex_unlock(&cpuhp_state_mutex);
1566         return ret;
1567 }
1568
1569 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1570                                bool invoke)
1571 {
1572         int ret;
1573
1574         cpus_read_lock();
1575         ret = __cpuhp_state_add_instance_cpuslocked(state, node, invoke);
1576         cpus_read_unlock();
1577         return ret;
1578 }
1579 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1580
1581 /**
1582  * __cpuhp_setup_state_cpuslocked - Setup the callbacks for an hotplug machine state
1583  * @state:              The state to setup
1584  * @invoke:             If true, the startup function is invoked for cpus where
1585  *                      cpu state >= @state
1586  * @startup:            startup callback function
1587  * @teardown:           teardown callback function
1588  * @multi_instance:     State is set up for multiple instances which get
1589  *                      added afterwards.
1590  *
1591  * The caller needs to hold cpus read locked while calling this function.
1592  * Returns:
1593  *   On success:
1594  *      Positive state number if @state is CPUHP_AP_ONLINE_DYN
1595  *      0 for all other states
1596  *   On failure: proper (negative) error code
1597  */
1598 int __cpuhp_setup_state_cpuslocked(enum cpuhp_state state,
1599                                    const char *name, bool invoke,
1600                                    int (*startup)(unsigned int cpu),
1601                                    int (*teardown)(unsigned int cpu),
1602                                    bool multi_instance)
1603 {
1604         int cpu, ret = 0;
1605         bool dynstate;
1606
1607         lockdep_assert_cpus_held();
1608
1609         if (cpuhp_cb_check(state) || !name)
1610                 return -EINVAL;
1611
1612         mutex_lock(&cpuhp_state_mutex);
1613
1614         ret = cpuhp_store_callbacks(state, name, startup, teardown,
1615                                     multi_instance);
1616
1617         dynstate = state == CPUHP_AP_ONLINE_DYN;
1618         if (ret > 0 && dynstate) {
1619                 state = ret;
1620                 ret = 0;
1621         }
1622
1623         if (ret || !invoke || !startup)
1624                 goto out;
1625
1626         /*
1627          * Try to call the startup callback for each present cpu
1628          * depending on the hotplug state of the cpu.
1629          */
1630         for_each_present_cpu(cpu) {
1631                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1632                 int cpustate = st->state;
1633
1634                 if (cpustate < state)
1635                         continue;
1636
1637                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1638                 if (ret) {
1639                         if (teardown)
1640                                 cpuhp_rollback_install(cpu, state, NULL);
1641                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1642                         goto out;
1643                 }
1644         }
1645 out:
1646         mutex_unlock(&cpuhp_state_mutex);
1647         /*
1648          * If the requested state is CPUHP_AP_ONLINE_DYN, return the
1649          * dynamically allocated state in case of success.
1650          */
1651         if (!ret && dynstate)
1652                 return state;
1653         return ret;
1654 }
1655 EXPORT_SYMBOL(__cpuhp_setup_state_cpuslocked);
1656
1657 int __cpuhp_setup_state(enum cpuhp_state state,
1658                         const char *name, bool invoke,
1659                         int (*startup)(unsigned int cpu),
1660                         int (*teardown)(unsigned int cpu),
1661                         bool multi_instance)
1662 {
1663         int ret;
1664
1665         cpus_read_lock();
1666         ret = __cpuhp_setup_state_cpuslocked(state, name, invoke, startup,
1667                                              teardown, multi_instance);
1668         cpus_read_unlock();
1669         return ret;
1670 }
1671 EXPORT_SYMBOL(__cpuhp_setup_state);
1672
1673 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1674                                   struct hlist_node *node, bool invoke)
1675 {
1676         struct cpuhp_step *sp = cpuhp_get_step(state);
1677         int cpu;
1678
1679         BUG_ON(cpuhp_cb_check(state));
1680
1681         if (!sp->multi_instance)
1682                 return -EINVAL;
1683
1684         cpus_read_lock();
1685         mutex_lock(&cpuhp_state_mutex);
1686
1687         if (!invoke || !cpuhp_get_teardown_cb(state))
1688                 goto remove;
1689         /*
1690          * Call the teardown callback for each present cpu depending
1691          * on the hotplug state of the cpu. This function is not
1692          * allowed to fail currently!
1693          */
1694         for_each_present_cpu(cpu) {
1695                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1696                 int cpustate = st->state;
1697
1698                 if (cpustate >= state)
1699                         cpuhp_issue_call(cpu, state, false, node);
1700         }
1701
1702 remove:
1703         hlist_del(node);
1704         mutex_unlock(&cpuhp_state_mutex);
1705         cpus_read_unlock();
1706
1707         return 0;
1708 }
1709 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1710
1711 /**
1712  * __cpuhp_remove_state_cpuslocked - Remove the callbacks for an hotplug machine state
1713  * @state:      The state to remove
1714  * @invoke:     If true, the teardown function is invoked for cpus where
1715  *              cpu state >= @state
1716  *
1717  * The caller needs to hold cpus read locked while calling this function.
1718  * The teardown callback is currently not allowed to fail. Think
1719  * about module removal!
1720  */
1721 void __cpuhp_remove_state_cpuslocked(enum cpuhp_state state, bool invoke)
1722 {
1723         struct cpuhp_step *sp = cpuhp_get_step(state);
1724         int cpu;
1725
1726         BUG_ON(cpuhp_cb_check(state));
1727
1728         lockdep_assert_cpus_held();
1729
1730         mutex_lock(&cpuhp_state_mutex);
1731         if (sp->multi_instance) {
1732                 WARN(!hlist_empty(&sp->list),
1733                      "Error: Removing state %d which has instances left.\n",
1734                      state);
1735                 goto remove;
1736         }
1737
1738         if (!invoke || !cpuhp_get_teardown_cb(state))
1739                 goto remove;
1740
1741         /*
1742          * Call the teardown callback for each present cpu depending
1743          * on the hotplug state of the cpu. This function is not
1744          * allowed to fail currently!
1745          */
1746         for_each_present_cpu(cpu) {
1747                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1748                 int cpustate = st->state;
1749
1750                 if (cpustate >= state)
1751                         cpuhp_issue_call(cpu, state, false, NULL);
1752         }
1753 remove:
1754         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1755         mutex_unlock(&cpuhp_state_mutex);
1756 }
1757 EXPORT_SYMBOL(__cpuhp_remove_state_cpuslocked);
1758
1759 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1760 {
1761         cpus_read_lock();
1762         __cpuhp_remove_state_cpuslocked(state, invoke);
1763         cpus_read_unlock();
1764 }
1765 EXPORT_SYMBOL(__cpuhp_remove_state);
1766
1767 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1768 static ssize_t show_cpuhp_state(struct device *dev,
1769                                 struct device_attribute *attr, char *buf)
1770 {
1771         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1772
1773         return sprintf(buf, "%d\n", st->state);
1774 }
1775 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1776
1777 static ssize_t write_cpuhp_target(struct device *dev,
1778                                   struct device_attribute *attr,
1779                                   const char *buf, size_t count)
1780 {
1781         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1782         struct cpuhp_step *sp;
1783         int target, ret;
1784
1785         ret = kstrtoint(buf, 10, &target);
1786         if (ret)
1787                 return ret;
1788
1789 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1790         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1791                 return -EINVAL;
1792 #else
1793         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1794                 return -EINVAL;
1795 #endif
1796
1797         ret = lock_device_hotplug_sysfs();
1798         if (ret)
1799                 return ret;
1800
1801         mutex_lock(&cpuhp_state_mutex);
1802         sp = cpuhp_get_step(target);
1803         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1804         mutex_unlock(&cpuhp_state_mutex);
1805         if (ret)
1806                 goto out;
1807
1808         if (st->state < target)
1809                 ret = do_cpu_up(dev->id, target);
1810         else
1811                 ret = do_cpu_down(dev->id, target);
1812 out:
1813         unlock_device_hotplug();
1814         return ret ? ret : count;
1815 }
1816
1817 static ssize_t show_cpuhp_target(struct device *dev,
1818                                  struct device_attribute *attr, char *buf)
1819 {
1820         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1821
1822         return sprintf(buf, "%d\n", st->target);
1823 }
1824 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1825
1826
1827 static ssize_t write_cpuhp_fail(struct device *dev,
1828                                 struct device_attribute *attr,
1829                                 const char *buf, size_t count)
1830 {
1831         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1832         struct cpuhp_step *sp;
1833         int fail, ret;
1834
1835         ret = kstrtoint(buf, 10, &fail);
1836         if (ret)
1837                 return ret;
1838
1839         /*
1840          * Cannot fail STARTING/DYING callbacks.
1841          */
1842         if (cpuhp_is_atomic_state(fail))
1843                 return -EINVAL;
1844
1845         /*
1846          * Cannot fail anything that doesn't have callbacks.
1847          */
1848         mutex_lock(&cpuhp_state_mutex);
1849         sp = cpuhp_get_step(fail);
1850         if (!sp->startup.single && !sp->teardown.single)
1851                 ret = -EINVAL;
1852         mutex_unlock(&cpuhp_state_mutex);
1853         if (ret)
1854                 return ret;
1855
1856         st->fail = fail;
1857
1858         return count;
1859 }
1860
1861 static ssize_t show_cpuhp_fail(struct device *dev,
1862                                struct device_attribute *attr, char *buf)
1863 {
1864         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1865
1866         return sprintf(buf, "%d\n", st->fail);
1867 }
1868
1869 static DEVICE_ATTR(fail, 0644, show_cpuhp_fail, write_cpuhp_fail);
1870
1871 static struct attribute *cpuhp_cpu_attrs[] = {
1872         &dev_attr_state.attr,
1873         &dev_attr_target.attr,
1874         &dev_attr_fail.attr,
1875         NULL
1876 };
1877
1878 static const struct attribute_group cpuhp_cpu_attr_group = {
1879         .attrs = cpuhp_cpu_attrs,
1880         .name = "hotplug",
1881         NULL
1882 };
1883
1884 static ssize_t show_cpuhp_states(struct device *dev,
1885                                  struct device_attribute *attr, char *buf)
1886 {
1887         ssize_t cur, res = 0;
1888         int i;
1889
1890         mutex_lock(&cpuhp_state_mutex);
1891         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1892                 struct cpuhp_step *sp = cpuhp_get_step(i);
1893
1894                 if (sp->name) {
1895                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1896                         buf += cur;
1897                         res += cur;
1898                 }
1899         }
1900         mutex_unlock(&cpuhp_state_mutex);
1901         return res;
1902 }
1903 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1904
1905 static struct attribute *cpuhp_cpu_root_attrs[] = {
1906         &dev_attr_states.attr,
1907         NULL
1908 };
1909
1910 static const struct attribute_group cpuhp_cpu_root_attr_group = {
1911         .attrs = cpuhp_cpu_root_attrs,
1912         .name = "hotplug",
1913         NULL
1914 };
1915
1916 static int __init cpuhp_sysfs_init(void)
1917 {
1918         int cpu, ret;
1919
1920         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1921                                  &cpuhp_cpu_root_attr_group);
1922         if (ret)
1923                 return ret;
1924
1925         for_each_possible_cpu(cpu) {
1926                 struct device *dev = get_cpu_device(cpu);
1927
1928                 if (!dev)
1929                         continue;
1930                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1931                 if (ret)
1932                         return ret;
1933         }
1934         return 0;
1935 }
1936 device_initcall(cpuhp_sysfs_init);
1937 #endif
1938
1939 /*
1940  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1941  * represents all NR_CPUS bits binary values of 1<<nr.
1942  *
1943  * It is used by cpumask_of() to get a constant address to a CPU
1944  * mask value that has a single bit set only.
1945  */
1946
1947 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1948 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1949 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1950 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1951 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1952
1953 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1954
1955         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1956         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1957 #if BITS_PER_LONG > 32
1958         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1959         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1960 #endif
1961 };
1962 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1963
1964 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1965 EXPORT_SYMBOL(cpu_all_bits);
1966
1967 #ifdef CONFIG_INIT_ALL_POSSIBLE
1968 struct cpumask __cpu_possible_mask __read_mostly
1969         = {CPU_BITS_ALL};
1970 #else
1971 struct cpumask __cpu_possible_mask __read_mostly;
1972 #endif
1973 EXPORT_SYMBOL(__cpu_possible_mask);
1974
1975 struct cpumask __cpu_online_mask __read_mostly;
1976 EXPORT_SYMBOL(__cpu_online_mask);
1977
1978 struct cpumask __cpu_present_mask __read_mostly;
1979 EXPORT_SYMBOL(__cpu_present_mask);
1980
1981 struct cpumask __cpu_active_mask __read_mostly;
1982 EXPORT_SYMBOL(__cpu_active_mask);
1983
1984 void init_cpu_present(const struct cpumask *src)
1985 {
1986         cpumask_copy(&__cpu_present_mask, src);
1987 }
1988
1989 void init_cpu_possible(const struct cpumask *src)
1990 {
1991         cpumask_copy(&__cpu_possible_mask, src);
1992 }
1993
1994 void init_cpu_online(const struct cpumask *src)
1995 {
1996         cpumask_copy(&__cpu_online_mask, src);
1997 }
1998
1999 /*
2000  * Activate the first processor.
2001  */
2002 void __init boot_cpu_init(void)
2003 {
2004         int cpu = smp_processor_id();
2005
2006         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
2007         set_cpu_online(cpu, true);
2008         set_cpu_active(cpu, true);
2009         set_cpu_present(cpu, true);
2010         set_cpu_possible(cpu, true);
2011
2012 #ifdef CONFIG_SMP
2013         __boot_cpu_id = cpu;
2014 #endif
2015 }
2016
2017 /*
2018  * Must be called _AFTER_ setting up the per_cpu areas
2019  */
2020 void __init boot_cpu_state_init(void)
2021 {
2022         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
2023 }