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