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         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
 893         while (st->state < target) {
 894                 struct cpuhp_step *step;
 895
 896                 st->state++;
 897                 step = cpuhp_ap_states + st->state;
 898                 cpuhp_invoke_callback(cpu, st->state, step->startup);
 899         }
 900 }
 901
 902 /*
 903  * Called from the idle task. We need to set active here, so we can kick off
 904  * the stopper thread and unpark the smpboot threads. If the target state is
 905  * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
 906  * cpu further.
 907  */
 908 void cpuhp_online_idle(enum cpuhp_state state)
 909 {
 910         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 911         unsigned int cpu = smp_processor_id();
 912
 913         /* Happens for the boot cpu */
 914         if (state != CPUHP_AP_ONLINE_IDLE)
 915                 return;
 916
 917         st->state = CPUHP_AP_ONLINE_IDLE;
 918
 919         /* Unpark the stopper thread and the hotplug thread of this cpu */
 920         stop_machine_unpark(cpu);
 921         kthread_unpark(st->thread);
 922
 923         /* Should we go further up ? */
 924         if (st->target > CPUHP_AP_ONLINE_IDLE)
 925                 __cpuhp_kick_ap_work(st);
 926         else
 927                 complete(&st->done);
 928 }
 929
 930 /* Requires cpu_add_remove_lock to be held */
 931 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 932 {
 933         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 934         struct task_struct *idle;
 935         int ret = 0;
 936
 937         cpu_hotplug_begin();
 938
 939         if (!cpu_present(cpu)) {
 940                 ret = -EINVAL;
 941                 goto out;
 942         }
 943
 944         /*
 945          * The caller of do_cpu_up might have raced with another
 946          * caller. Ignore it for now.
 947          */
 948         if (st->state >= target)
 949                 goto out;
 950
 951         if (st->state == CPUHP_OFFLINE) {
 952                 /* Let it fail before we try to bring the cpu up */
 953                 idle = idle_thread_get(cpu);
 954                 if (IS_ERR(idle)) {
 955                         ret = PTR_ERR(idle);
 956                         goto out;
 957                 }
 958         }
 959
 960         cpuhp_tasks_frozen = tasks_frozen;
 961
 962         st->target = target;
 963         /*
 964          * If the current CPU state is in the range of the AP hotplug thread,
 965          * then we need to kick the thread once more.
 966          */
 967         if (st->state > CPUHP_BRINGUP_CPU) {
 968                 ret = cpuhp_kick_ap_work(cpu);
 969                 /*
 970                  * The AP side has done the error rollback already. Just
 971                  * return the error code..
 972                  */
 973                 if (ret)
 974                         goto out;
 975         }
 976
 977         /*
 978          * Try to reach the target state. We max out on the BP at
 979          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
 980          * responsible for bringing it up to the target state.
 981          */
 982         target = min((int)target, CPUHP_BRINGUP_CPU);
 983         ret = cpuhp_up_callbacks(cpu, st, cpuhp_bp_states, target);
 984 out:
 985         cpu_hotplug_done();
 986         return ret;
 987 }
 988
 989 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
 990 {
 991         int err = 0;
 992
 993         if (!cpu_possible(cpu)) {
 994                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
 995                        cpu);
 996 #if defined(CONFIG_IA64)
 997                 pr_err("please check additional_cpus= boot parameter\n");
 998 #endif
 999                 return -EINVAL;
1000         }
1001
1002         err = try_online_node(cpu_to_node(cpu));
1003         if (err)
1004                 return err;
1005
1006         cpu_maps_update_begin();
1007
1008         if (cpu_hotplug_disabled) {
1009                 err = -EBUSY;
1010                 goto out;
1011         }
1012
1013         err = _cpu_up(cpu, 0, target);
1014 out:
1015         cpu_maps_update_done();
1016         return err;
1017 }
1018
1019 int cpu_up(unsigned int cpu)
1020 {
1021         return do_cpu_up(cpu, CPUHP_ONLINE);
1022 }
1023 EXPORT_SYMBOL_GPL(cpu_up);
1024
1025 #ifdef CONFIG_PM_SLEEP_SMP
1026 static cpumask_var_t frozen_cpus;
1027
1028 int freeze_secondary_cpus(int primary)
1029 {
1030         int cpu, error = 0;
1031
1032         cpu_maps_update_begin();
1033         if (!cpu_online(primary))
1034                 primary = cpumask_first(cpu_online_mask);
1035         /*
1036          * We take down all of the non-boot CPUs in one shot to avoid races
1037          * with the userspace trying to use the CPU hotplug at the same time
1038          */
1039         cpumask_clear(frozen_cpus);
1040
1041         pr_info("Disabling non-boot CPUs ...\n");
1042         for_each_online_cpu(cpu) {
1043                 if (cpu == primary)
1044                         continue;
1045                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1046                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1047                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1048                 if (!error)
1049                         cpumask_set_cpu(cpu, frozen_cpus);
1050                 else {
1051                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1052                         break;
1053                 }
1054         }
1055
1056         if (!error)
1057                 BUG_ON(num_online_cpus() > 1);
1058         else
1059                 pr_err("Non-boot CPUs are not disabled\n");
1060
1061         /*
1062          * Make sure the CPUs won't be enabled by someone else. We need to do
1063          * this even in case of failure as all disable_nonboot_cpus() users are
1064          * supposed to do enable_nonboot_cpus() on the failure path.
1065          */
1066         cpu_hotplug_disabled++;
1067
1068         cpu_maps_update_done();
1069         return error;
1070 }
1071
1072 void __weak arch_enable_nonboot_cpus_begin(void)
1073 {
1074 }
1075
1076 void __weak arch_enable_nonboot_cpus_end(void)
1077 {
1078 }
1079
1080 void enable_nonboot_cpus(void)
1081 {
1082         int cpu, error;
1083
1084         /* Allow everyone to use the CPU hotplug again */
1085         cpu_maps_update_begin();
1086         WARN_ON(--cpu_hotplug_disabled < 0);
1087         if (cpumask_empty(frozen_cpus))
1088                 goto out;
1089
1090         pr_info("Enabling non-boot CPUs ...\n");
1091
1092         arch_enable_nonboot_cpus_begin();
1093
1094         for_each_cpu(cpu, frozen_cpus) {
1095                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1096                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1097                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1098                 if (!error) {
1099                         pr_info("CPU%d is up\n", cpu);
1100                         continue;
1101                 }
1102                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1103         }
1104
1105         arch_enable_nonboot_cpus_end();
1106
1107         cpumask_clear(frozen_cpus);
1108 out:
1109         cpu_maps_update_done();
1110 }
1111
1112 static int __init alloc_frozen_cpus(void)
1113 {
1114         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1115                 return -ENOMEM;
1116         return 0;
1117 }
1118 core_initcall(alloc_frozen_cpus);
1119
1120 /*
1121  * When callbacks for CPU hotplug notifications are being executed, we must
1122  * ensure that the state of the system with respect to the tasks being frozen
1123  * or not, as reported by the notification, remains unchanged *throughout the
1124  * duration* of the execution of the callbacks.
1125  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1126  *
1127  * This synchronization is implemented by mutually excluding regular CPU
1128  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1129  * Hibernate notifications.
1130  */
1131 static int
1132 cpu_hotplug_pm_callback(struct notifier_block *nb,
1133                         unsigned long action, void *ptr)
1134 {
1135         switch (action) {
1136
1137         case PM_SUSPEND_PREPARE:
1138         case PM_HIBERNATION_PREPARE:
1139                 cpu_hotplug_disable();
1140                 break;
1141
1142         case PM_POST_SUSPEND:
1143         case PM_POST_HIBERNATION:
1144                 cpu_hotplug_enable();
1145                 break;
1146
1147         default:
1148                 return NOTIFY_DONE;
1149         }
1150
1151         return NOTIFY_OK;
1152 }
1153
1154
1155 static int __init cpu_hotplug_pm_sync_init(void)
1156 {
1157         /*
1158          * cpu_hotplug_pm_callback has higher priority than x86
1159          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1160          * to disable cpu hotplug to avoid cpu hotplug race.
1161          */
1162         pm_notifier(cpu_hotplug_pm_callback, 0);
1163         return 0;
1164 }
1165 core_initcall(cpu_hotplug_pm_sync_init);
1166
1167 #endif /* CONFIG_PM_SLEEP_SMP */
1168
1169 #endif /* CONFIG_SMP */
1170
1171 /* Boot processor state steps */
1172 static struct cpuhp_step cpuhp_bp_states[] = {
1173         [CPUHP_OFFLINE] = {
1174                 .name                   = "offline",
1175                 .startup                = NULL,
1176                 .teardown               = NULL,
1177         },
1178 #ifdef CONFIG_SMP
1179         [CPUHP_CREATE_THREADS]= {
1180                 .name                   = "threads:create",
1181                 .startup                = smpboot_create_threads,
1182                 .teardown               = NULL,
1183                 .cant_stop              = true,
1184         },
1185         [CPUHP_PERF_PREPARE] = {
1186                 .name = "perf prepare",
1187                 .startup = perf_event_init_cpu,
1188                 .teardown = perf_event_exit_cpu,
1189         },
1190         [CPUHP_WORKQUEUE_PREP] = {
1191                 .name = "workqueue prepare",
1192                 .startup = workqueue_prepare_cpu,
1193                 .teardown = NULL,
1194         },
1195         [CPUHP_HRTIMERS_PREPARE] = {
1196                 .name = "hrtimers prepare",
1197                 .startup = hrtimers_prepare_cpu,
1198                 .teardown = hrtimers_dead_cpu,
1199         },
1200         [CPUHP_SMPCFD_PREPARE] = {
1201                 .name = "SMPCFD prepare",
1202                 .startup = smpcfd_prepare_cpu,
1203                 .teardown = smpcfd_dead_cpu,
1204         },
1205         [CPUHP_RCUTREE_PREP] = {
1206                 .name = "RCU-tree prepare",
1207                 .startup = rcutree_prepare_cpu,
1208                 .teardown = rcutree_dead_cpu,
1209         },
1210         /*
1211          * Preparatory and dead notifiers. Will be replaced once the notifiers
1212          * are converted to states.
1213          */
1214         [CPUHP_NOTIFY_PREPARE] = {
1215                 .name                   = "notify:prepare",
1216                 .startup                = notify_prepare,
1217                 .teardown               = notify_dead,
1218                 .skip_onerr             = true,
1219                 .cant_stop              = true,
1220         },
1221         /*
1222          * On the tear-down path, timers_dead_cpu() must be invoked
1223          * before blk_mq_queue_reinit_notify() from notify_dead(),
1224          * otherwise a RCU stall occurs.
1225          */
1226         [CPUHP_TIMERS_DEAD] = {
1227                 .name = "timers dead",
1228                 .startup = NULL,
1229                 .teardown = timers_dead_cpu,
1230         },
1231         /* Kicks the plugged cpu into life */
1232         [CPUHP_BRINGUP_CPU] = {
1233                 .name                   = "cpu:bringup",
1234                 .startup                = bringup_cpu,
1235                 .teardown               = NULL,
1236                 .cant_stop              = true,
1237         },
1238         [CPUHP_AP_SMPCFD_DYING] = {
1239                 .startup = NULL,
1240                 .teardown = smpcfd_dying_cpu,
1241         },
1242         /*
1243          * Handled on controll processor until the plugged processor manages
1244          * this itself.
1245          */
1246         [CPUHP_TEARDOWN_CPU] = {
1247                 .name                   = "cpu:teardown",
1248                 .startup                = NULL,
1249                 .teardown               = takedown_cpu,
1250                 .cant_stop              = true,
1251         },
1252 #else
1253         [CPUHP_BRINGUP_CPU] = { },
1254 #endif
1255 };
1256
1257 /* Application processor state steps */
1258 static struct cpuhp_step cpuhp_ap_states[] = {
1259 #ifdef CONFIG_SMP
1260         /* Final state before CPU kills itself */
1261         [CPUHP_AP_IDLE_DEAD] = {
1262                 .name                   = "idle:dead",
1263         },
1264         /*
1265          * Last state before CPU enters the idle loop to die. Transient state
1266          * for synchronization.
1267          */
1268         [CPUHP_AP_OFFLINE] = {
1269                 .name                   = "ap:offline",
1270                 .cant_stop              = true,
1271         },
1272         /* First state is scheduler control. Interrupts are disabled */
1273         [CPUHP_AP_SCHED_STARTING] = {
1274                 .name                   = "sched:starting",
1275                 .startup                = sched_cpu_starting,
1276                 .teardown               = sched_cpu_dying,
1277         },
1278         [CPUHP_AP_RCUTREE_DYING] = {
1279                 .startup = NULL,
1280                 .teardown = rcutree_dying_cpu,
1281         },
1282         /*
1283          * Low level startup/teardown notifiers. Run with interrupts
1284          * disabled. Will be removed once the notifiers are converted to
1285          * states.
1286          */
1287         [CPUHP_AP_NOTIFY_STARTING] = {
1288                 .name                   = "notify:starting",
1289                 .startup                = notify_starting,
1290                 .teardown               = notify_dying,
1291                 .skip_onerr             = true,
1292                 .cant_stop              = true,
1293         },
1294         /* Entry state on starting. Interrupts enabled from here on. Transient
1295          * state for synchronsization */
1296         [CPUHP_AP_ONLINE] = {
1297                 .name                   = "ap:online",
1298         },
1299         /* Handle smpboot threads park/unpark */
1300         [CPUHP_AP_SMPBOOT_THREADS] = {
1301                 .name                   = "smpboot:threads",
1302                 .startup                = smpboot_unpark_threads,
1303                 .teardown               = NULL,
1304         },
1305         [CPUHP_AP_PERF_ONLINE] = {
1306                 .name = "perf online",
1307                 .startup = perf_event_init_cpu,
1308                 .teardown = perf_event_exit_cpu,
1309         },
1310         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1311                 .name = "workqueue online",
1312                 .startup = workqueue_online_cpu,
1313                 .teardown = workqueue_offline_cpu,
1314         },
1315         [CPUHP_AP_RCUTREE_ONLINE] = {
1316                 .name = "RCU-tree online",
1317                 .startup = rcutree_online_cpu,
1318                 .teardown = rcutree_offline_cpu,
1319         },
1320
1321         /*
1322          * Online/down_prepare notifiers. Will be removed once the notifiers
1323          * are converted to states.
1324          */
1325         [CPUHP_AP_NOTIFY_ONLINE] = {
1326                 .name                   = "notify:online",
1327                 .startup                = notify_online,
1328                 .teardown               = notify_down_prepare,
1329                 .skip_onerr             = true,
1330         },
1331 #endif
1332         /*
1333          * The dynamically registered state space is here
1334          */
1335
1336 #ifdef CONFIG_SMP
1337         /* Last state is scheduler control setting the cpu active */
1338         [CPUHP_AP_ACTIVE] = {
1339                 .name                   = "sched:active",
1340                 .startup                = sched_cpu_activate,
1341                 .teardown               = sched_cpu_deactivate,
1342         },
1343 #endif
1344
1345         /* CPU is fully up and running. */
1346         [CPUHP_ONLINE] = {
1347                 .name                   = "online",
1348                 .startup                = NULL,
1349                 .teardown               = NULL,
1350         },
1351 };
1352
1353 /* Sanity check for callbacks */
1354 static int cpuhp_cb_check(enum cpuhp_state state)
1355 {
1356         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1357                 return -EINVAL;
1358         return 0;
1359 }
1360
1361 static bool cpuhp_is_ap_state(enum cpuhp_state state)
1362 {
1363         /*
1364          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
1365          * purposes as that state is handled explicitely in cpu_down.
1366          */
1367         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
1368 }
1369
1370 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
1371 {
1372         struct cpuhp_step *sp;
1373
1374         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
1375         return sp + state;
1376 }
1377
1378 static void cpuhp_store_callbacks(enum cpuhp_state state,
1379                                   const char *name,
1380                                   int (*startup)(unsigned int cpu),
1381                                   int (*teardown)(unsigned int cpu))
1382 {
1383         /* (Un)Install the callbacks for further cpu hotplug operations */
1384         struct cpuhp_step *sp;
1385
1386         mutex_lock(&cpuhp_state_mutex);
1387         sp = cpuhp_get_step(state);
1388         sp->startup = startup;
1389         sp->teardown = teardown;
1390         sp->name = name;
1391         mutex_unlock(&cpuhp_state_mutex);
1392 }
1393
1394 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1395 {
1396         return cpuhp_get_step(state)->teardown;
1397 }
1398
1399 /*
1400  * Call the startup/teardown function for a step either on the AP or
1401  * on the current CPU.
1402  */
1403 static int cpuhp_issue_call(int cpu, enum cpuhp_state state,
1404                             int (*cb)(unsigned int), bool bringup)
1405 {
1406         int ret;
1407
1408         if (!cb)
1409                 return 0;
1410         /*
1411          * The non AP bound callbacks can fail on bringup. On teardown
1412          * e.g. module removal we crash for now.
1413          */
1414 #ifdef CONFIG_SMP
1415         if (cpuhp_is_ap_state(state))
1416                 ret = cpuhp_invoke_ap_callback(cpu, state, cb);
1417         else
1418                 ret = cpuhp_invoke_callback(cpu, state, cb);
1419 #else
1420         ret = cpuhp_invoke_callback(cpu, state, cb);
1421 #endif
1422         BUG_ON(ret && !bringup);
1423         return ret;
1424 }
1425
1426 /*
1427  * Called from __cpuhp_setup_state on a recoverable failure.
1428  *
1429  * Note: The teardown callbacks for rollback are not allowed to fail!
1430  */
1431 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1432                                    int (*teardown)(unsigned int cpu))
1433 {
1434         int cpu;
1435
1436         if (!teardown)
1437                 return;
1438
1439         /* Roll back the already executed steps on the other cpus */
1440         for_each_present_cpu(cpu) {
1441                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1442                 int cpustate = st->state;
1443
1444                 if (cpu >= failedcpu)
1445                         break;
1446
1447                 /* Did we invoke the startup call on that cpu ? */
1448                 if (cpustate >= state)
1449                         cpuhp_issue_call(cpu, state, teardown, false);
1450         }
1451 }
1452
1453 /*
1454  * Returns a free for dynamic slot assignment of the Online state. The states
1455  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1456  * by having no name assigned.
1457  */
1458 static int cpuhp_reserve_state(enum cpuhp_state state)
1459 {
1460         enum cpuhp_state i;
1461
1462         mutex_lock(&cpuhp_state_mutex);
1463         for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1464                 if (cpuhp_ap_states[i].name)
1465                         continue;
1466
1467                 cpuhp_ap_states[i].name = "Reserved";
1468                 mutex_unlock(&cpuhp_state_mutex);
1469                 return i;
1470         }
1471         mutex_unlock(&cpuhp_state_mutex);
1472         WARN(1, "No more dynamic states available for CPU hotplug\n");
1473         return -ENOSPC;
1474 }
1475
1476 /**
1477  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1478  * @state:      The state to setup
1479  * @invoke:     If true, the startup function is invoked for cpus where
1480  *              cpu state >= @state
1481  * @startup:    startup callback function
1482  * @teardown:   teardown callback function
1483  *
1484  * Returns 0 if successful, otherwise a proper error code
1485  */
1486 int __cpuhp_setup_state(enum cpuhp_state state,
1487                         const char *name, bool invoke,
1488                         int (*startup)(unsigned int cpu),
1489                         int (*teardown)(unsigned int cpu))
1490 {
1491         int cpu, ret = 0;
1492         int dyn_state = 0;
1493
1494         if (cpuhp_cb_check(state) || !name)
1495                 return -EINVAL;
1496
1497         get_online_cpus();
1498
1499         /* currently assignments for the ONLINE state are possible */
1500         if (state == CPUHP_AP_ONLINE_DYN) {
1501                 dyn_state = 1;
1502                 ret = cpuhp_reserve_state(state);
1503                 if (ret < 0)
1504                         goto out;
1505                 state = ret;
1506         }
1507
1508         cpuhp_store_callbacks(state, name, startup, teardown);
1509
1510         if (!invoke || !startup)
1511                 goto out;
1512
1513         /*
1514          * Try to call the startup callback for each present cpu
1515          * depending on the hotplug state of the cpu.
1516          */
1517         for_each_present_cpu(cpu) {
1518                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1519                 int cpustate = st->state;
1520
1521                 if (cpustate < state)
1522                         continue;
1523
1524                 ret = cpuhp_issue_call(cpu, state, startup, true);
1525                 if (ret) {
1526                         cpuhp_rollback_install(cpu, state, teardown);
1527                         cpuhp_store_callbacks(state, NULL, NULL, NULL);
1528                         goto out;
1529                 }
1530         }
1531 out:
1532         put_online_cpus();
1533         if (!ret && dyn_state)
1534                 return state;
1535         return ret;
1536 }
1537 EXPORT_SYMBOL(__cpuhp_setup_state);
1538
1539 /**
1540  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1541  * @state:      The state to remove
1542  * @invoke:     If true, the teardown function is invoked for cpus where
1543  *              cpu state >= @state
1544  *
1545  * The teardown callback is currently not allowed to fail. Think
1546  * about module removal!
1547  */
1548 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1549 {
1550         int (*teardown)(unsigned int cpu) = cpuhp_get_teardown_cb(state);
1551         int cpu;
1552
1553         BUG_ON(cpuhp_cb_check(state));
1554
1555         get_online_cpus();
1556
1557         if (!invoke || !teardown)
1558                 goto remove;
1559
1560         /*
1561          * Call the teardown callback for each present cpu depending
1562          * on the hotplug state of the cpu. This function is not
1563          * allowed to fail currently!
1564          */
1565         for_each_present_cpu(cpu) {
1566                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1567                 int cpustate = st->state;
1568
1569                 if (cpustate >= state)
1570                         cpuhp_issue_call(cpu, state, teardown, false);
1571         }
1572 remove:
1573         cpuhp_store_callbacks(state, NULL, NULL, NULL);
1574         put_online_cpus();
1575 }
1576 EXPORT_SYMBOL(__cpuhp_remove_state);
1577
1578 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1579 static ssize_t show_cpuhp_state(struct device *dev,
1580                                 struct device_attribute *attr, char *buf)
1581 {
1582         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1583
1584         return sprintf(buf, "%d\n", st->state);
1585 }
1586 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1587
1588 static ssize_t write_cpuhp_target(struct device *dev,
1589                                   struct device_attribute *attr,
1590                                   const char *buf, size_t count)
1591 {
1592         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1593         struct cpuhp_step *sp;
1594         int target, ret;
1595
1596         ret = kstrtoint(buf, 10, &target);
1597         if (ret)
1598                 return ret;
1599
1600 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1601         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1602                 return -EINVAL;
1603 #else
1604         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1605                 return -EINVAL;
1606 #endif
1607
1608         ret = lock_device_hotplug_sysfs();
1609         if (ret)
1610                 return ret;
1611
1612         mutex_lock(&cpuhp_state_mutex);
1613         sp = cpuhp_get_step(target);
1614         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1615         mutex_unlock(&cpuhp_state_mutex);
1616         if (ret)
1617                 return ret;
1618
1619         if (st->state < target)
1620                 ret = do_cpu_up(dev->id, target);
1621         else
1622                 ret = do_cpu_down(dev->id, target);
1623
1624         unlock_device_hotplug();
1625         return ret ? ret : count;
1626 }
1627
1628 static ssize_t show_cpuhp_target(struct device *dev,
1629                                  struct device_attribute *attr, char *buf)
1630 {
1631         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1632
1633         return sprintf(buf, "%d\n", st->target);
1634 }
1635 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1636
1637 static struct attribute *cpuhp_cpu_attrs[] = {
1638         &dev_attr_state.attr,
1639         &dev_attr_target.attr,
1640         NULL
1641 };
1642
1643 static struct attribute_group cpuhp_cpu_attr_group = {
1644         .attrs = cpuhp_cpu_attrs,
1645         .name = "hotplug",
1646         NULL
1647 };
1648
1649 static ssize_t show_cpuhp_states(struct device *dev,
1650                                  struct device_attribute *attr, char *buf)
1651 {
1652         ssize_t cur, res = 0;
1653         int i;
1654
1655         mutex_lock(&cpuhp_state_mutex);
1656         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1657                 struct cpuhp_step *sp = cpuhp_get_step(i);
1658
1659                 if (sp->name) {
1660                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1661                         buf += cur;
1662                         res += cur;
1663                 }
1664         }
1665         mutex_unlock(&cpuhp_state_mutex);
1666         return res;
1667 }
1668 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1669
1670 static struct attribute *cpuhp_cpu_root_attrs[] = {
1671         &dev_attr_states.attr,
1672         NULL
1673 };
1674
1675 static struct attribute_group cpuhp_cpu_root_attr_group = {
1676         .attrs = cpuhp_cpu_root_attrs,
1677         .name = "hotplug",
1678         NULL
1679 };
1680
1681 static int __init cpuhp_sysfs_init(void)
1682 {
1683         int cpu, ret;
1684
1685         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1686                                  &cpuhp_cpu_root_attr_group);
1687         if (ret)
1688                 return ret;
1689
1690         for_each_possible_cpu(cpu) {
1691                 struct device *dev = get_cpu_device(cpu);
1692
1693                 if (!dev)
1694                         continue;
1695                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1696                 if (ret)
1697                         return ret;
1698         }
1699         return 0;
1700 }
1701 device_initcall(cpuhp_sysfs_init);
1702 #endif
1703
1704 /*
1705  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1706  * represents all NR_CPUS bits binary values of 1<<nr.
1707  *
1708  * It is used by cpumask_of() to get a constant address to a CPU
1709  * mask value that has a single bit set only.
1710  */
1711
1712 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1713 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1714 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1715 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1716 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1717
1718 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1719
1720         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1721         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1722 #if BITS_PER_LONG > 32
1723         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1724         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1725 #endif
1726 };
1727 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1728
1729 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1730 EXPORT_SYMBOL(cpu_all_bits);
1731
1732 #ifdef CONFIG_INIT_ALL_POSSIBLE
1733 struct cpumask __cpu_possible_mask __read_mostly
1734         = {CPU_BITS_ALL};
1735 #else
1736 struct cpumask __cpu_possible_mask __read_mostly;
1737 #endif
1738 EXPORT_SYMBOL(__cpu_possible_mask);
1739
1740 struct cpumask __cpu_online_mask __read_mostly;
1741 EXPORT_SYMBOL(__cpu_online_mask);
1742
1743 struct cpumask __cpu_present_mask __read_mostly;
1744 EXPORT_SYMBOL(__cpu_present_mask);
1745
1746 struct cpumask __cpu_active_mask __read_mostly;
1747 EXPORT_SYMBOL(__cpu_active_mask);
1748
1749 void init_cpu_present(const struct cpumask *src)
1750 {
1751         cpumask_copy(&__cpu_present_mask, src);
1752 }
1753
1754 void init_cpu_possible(const struct cpumask *src)
1755 {
1756         cpumask_copy(&__cpu_possible_mask, src);
1757 }
1758
1759 void init_cpu_online(const struct cpumask *src)
1760 {
1761         cpumask_copy(&__cpu_online_mask, src);
1762 }
1763
1764 /*
1765  * Activate the first processor.
1766  */
1767 void __init boot_cpu_init(void)
1768 {
1769         int cpu = smp_processor_id();
1770
1771         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1772         set_cpu_online(cpu, true);
1773         set_cpu_active(cpu, true);
1774         set_cpu_present(cpu, true);
1775         set_cpu_possible(cpu, true);
1776 }
1777
1778 /*
1779  * Must be called _AFTER_ setting up the per_cpu areas
1780  */
1781 void __init boot_cpu_state_init(void)
1782 {
1783         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1784 }