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  * @cb_stat:    The state for a single callback (install/uninstall)
  40  * @cb:         Single callback function (install/uninstall)
  41  * @result:     Result of the operation
  42  * @done:       Signal completion to the issuer of the task
  43  */
  44 struct cpuhp_cpu_state {
  45         enum cpuhp_state        state;
  46         enum cpuhp_state        target;
  47 #ifdef CONFIG_SMP
  48         struct task_struct      *thread;
  49         bool                    should_run;
  50         enum cpuhp_state        cb_state;
  51         int                     (*cb)(unsigned int cpu);
  52         int                     result;
  53         struct completion       done;
  54 #endif
  55 };
  56
  57 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
  58
  59 /**
  60  * cpuhp_step - Hotplug state machine step
  61  * @name:       Name of the step
  62  * @startup:    Startup function of the step
  63  * @teardown:   Teardown function of the step
  64  * @skip_onerr: Do not invoke the functions on error rollback
  65  *              Will go away once the notifiers are gone
  66  * @cant_stop:  Bringup/teardown can't be stopped at this step
  67  */
  68 struct cpuhp_step {
  69         const char      *name;
  70         int             (*startup)(unsigned int cpu);
  71         int             (*teardown)(unsigned int cpu);
  72         bool            skip_onerr;
  73         bool            cant_stop;
  74 };
  75
  76 static DEFINE_MUTEX(cpuhp_state_mutex);
  77 static struct cpuhp_step cpuhp_bp_states[];
  78 static struct cpuhp_step cpuhp_ap_states[];
  79
  80 /**
  81  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
  82  * @cpu:        The cpu for which the callback should be invoked
  83  * @step:       The step in the state machine
  84  * @cb:         The callback function to invoke
  85  *
  86  * Called from cpu hotplug and from the state register machinery
  87  */
  88 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state step,
  89                                  int (*cb)(unsigned int))
  90 {
  91         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
  92         int ret = 0;
  93
  94         if (cb) {
  95                 trace_cpuhp_enter(cpu, st->target, step, cb);
  96                 ret = cb(cpu);
  97                 trace_cpuhp_exit(cpu, st->state, step, ret);
  98         }
  99         return ret;
 100 }
 101
 102 #ifdef CONFIG_SMP
 103 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 104 static DEFINE_MUTEX(cpu_add_remove_lock);
 105 bool cpuhp_tasks_frozen;
 106 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 107
 108 /*
 109  * The following two APIs (cpu_maps_update_begin/done) must be used when
 110  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 111  * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
 112  * hotplug callback (un)registration performed using __register_cpu_notifier()
 113  * or __unregister_cpu_notifier().
 114  */
 115 void cpu_maps_update_begin(void)
 116 {
 117         mutex_lock(&cpu_add_remove_lock);
 118 }
 119 EXPORT_SYMBOL(cpu_notifier_register_begin);
 120
 121 void cpu_maps_update_done(void)
 122 {
 123         mutex_unlock(&cpu_add_remove_lock);
 124 }
 125 EXPORT_SYMBOL(cpu_notifier_register_done);
 126
 127 static RAW_NOTIFIER_HEAD(cpu_chain);
 128
 129 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 130  * Should always be manipulated under cpu_add_remove_lock
 131  */
 132 static int cpu_hotplug_disabled;
 133
 134 #ifdef CONFIG_HOTPLUG_CPU
 135
 136 static struct {
 137         struct task_struct *active_writer;
 138         /* wait queue to wake up the active_writer */
 139         wait_queue_head_t wq;
 140         /* verifies that no writer will get active while readers are active */
 141         struct mutex lock;
 142         /*
 143          * Also blocks the new readers during
 144          * an ongoing cpu hotplug operation.
 145          */
 146         atomic_t refcount;
 147
 148 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 149         struct lockdep_map dep_map;
 150 #endif
 151 } cpu_hotplug = {
 152         .active_writer = NULL,
 153         .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
 154         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
 155 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 156         .dep_map = {.name = "cpu_hotplug.lock" },
 157 #endif
 158 };
 159
 160 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
 161 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
 162 #define cpuhp_lock_acquire_tryread() \
 163                                   lock_map_acquire_tryread(&cpu_hotplug.dep_map)
 164 #define cpuhp_lock_acquire()      lock_map_acquire(&cpu_hotplug.dep_map)
 165 #define cpuhp_lock_release()      lock_map_release(&cpu_hotplug.dep_map)
 166
 167
 168 void get_online_cpus(void)
 169 {
 170         might_sleep();
 171         if (cpu_hotplug.active_writer == current)
 172                 return;
 173         cpuhp_lock_acquire_read();
 174         mutex_lock(&cpu_hotplug.lock);
 175         atomic_inc(&cpu_hotplug.refcount);
 176         mutex_unlock(&cpu_hotplug.lock);
 177 }
 178 EXPORT_SYMBOL_GPL(get_online_cpus);
 179
 180 void put_online_cpus(void)
 181 {
 182         int refcount;
 183
 184         if (cpu_hotplug.active_writer == current)
 185                 return;
 186
 187         refcount = atomic_dec_return(&cpu_hotplug.refcount);
 188         if (WARN_ON(refcount < 0)) /* try to fix things up */
 189                 atomic_inc(&cpu_hotplug.refcount);
 190
 191         if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
 192                 wake_up(&cpu_hotplug.wq);
 193
 194         cpuhp_lock_release();
 195
 196 }
 197 EXPORT_SYMBOL_GPL(put_online_cpus);
 198
 199 /*
 200  * This ensures that the hotplug operation can begin only when the
 201  * refcount goes to zero.
 202  *
 203  * Note that during a cpu-hotplug operation, the new readers, if any,
 204  * will be blocked by the cpu_hotplug.lock
 205  *
 206  * Since cpu_hotplug_begin() is always called after invoking
 207  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 208  *
 209  * Note that theoretically, there is a possibility of a livelock:
 210  * - Refcount goes to zero, last reader wakes up the sleeping
 211  *   writer.
 212  * - Last reader unlocks the cpu_hotplug.lock.
 213  * - A new reader arrives at this moment, bumps up the refcount.
 214  * - The writer acquires the cpu_hotplug.lock finds the refcount
 215  *   non zero and goes to sleep again.
 216  *
 217  * However, this is very difficult to achieve in practice since
 218  * get_online_cpus() not an api which is called all that often.
 219  *
 220  */
 221 void cpu_hotplug_begin(void)
 222 {
 223         DEFINE_WAIT(wait);
 224
 225         cpu_hotplug.active_writer = current;
 226         cpuhp_lock_acquire();
 227
 228         for (;;) {
 229                 mutex_lock(&cpu_hotplug.lock);
 230                 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
 231                 if (likely(!atomic_read(&cpu_hotplug.refcount)))
 232                                 break;
 233                 mutex_unlock(&cpu_hotplug.lock);
 234                 schedule();
 235         }
 236         finish_wait(&cpu_hotplug.wq, &wait);
 237 }
 238
 239 void cpu_hotplug_done(void)
 240 {
 241         cpu_hotplug.active_writer = NULL;
 242         mutex_unlock(&cpu_hotplug.lock);
 243         cpuhp_lock_release();
 244 }
 245
 246 /*
 247  * Wait for currently running CPU hotplug operations to complete (if any) and
 248  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 249  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 250  * hotplug path before performing hotplug operations. So acquiring that lock
 251  * guarantees mutual exclusion from any currently running hotplug operations.
 252  */
 253 void cpu_hotplug_disable(void)
 254 {
 255         cpu_maps_update_begin();
 256         cpu_hotplug_disabled++;
 257         cpu_maps_update_done();
 258 }
 259 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 260
 261 void cpu_hotplug_enable(void)
 262 {
 263         cpu_maps_update_begin();
 264         WARN_ON(--cpu_hotplug_disabled < 0);
 265         cpu_maps_update_done();
 266 }
 267 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 268 #endif  /* CONFIG_HOTPLUG_CPU */
 269
 270 /* Need to know about CPUs going up/down? */
 271 int register_cpu_notifier(struct notifier_block *nb)
 272 {
 273         int ret;
 274         cpu_maps_update_begin();
 275         ret = raw_notifier_chain_register(&cpu_chain, nb);
 276         cpu_maps_update_done();
 277         return ret;
 278 }
 279
 280 int __register_cpu_notifier(struct notifier_block *nb)
 281 {
 282         return raw_notifier_chain_register(&cpu_chain, nb);
 283 }
 284
 285 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
 286                         int *nr_calls)
 287 {
 288         unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
 289         void *hcpu = (void *)(long)cpu;
 290
 291         int ret;
 292
 293         ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
 294                                         nr_calls);
 295
 296         return notifier_to_errno(ret);
 297 }
 298
 299 static int cpu_notify(unsigned long val, unsigned int cpu)
 300 {
 301         return __cpu_notify(val, cpu, -1, NULL);
 302 }
 303
 304 /* Notifier wrappers for transitioning to state machine */
 305 static int notify_prepare(unsigned int cpu)
 306 {
 307         int nr_calls = 0;
 308         int ret;
 309
 310         ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
 311         if (ret) {
 312                 nr_calls--;
 313                 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
 314                                 __func__, cpu);
 315                 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
 316         }
 317         return ret;
 318 }
 319
 320 static int notify_online(unsigned int cpu)
 321 {
 322         cpu_notify(CPU_ONLINE, cpu);
 323         return 0;
 324 }
 325
 326 static int notify_starting(unsigned int cpu)
 327 {
 328         cpu_notify(CPU_STARTING, cpu);
 329         return 0;
 330 }
 331
 332 static int bringup_wait_for_ap(unsigned int cpu)
 333 {
 334         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 335
 336         wait_for_completion(&st->done);
 337         return st->result;
 338 }
 339
 340 static int bringup_cpu(unsigned int cpu)
 341 {
 342         struct task_struct *idle = idle_thread_get(cpu);
 343         int ret;
 344
 345         /* Arch-specific enabling code. */
 346         ret = __cpu_up(cpu, idle);
 347         if (ret) {
 348                 cpu_notify(CPU_UP_CANCELED, cpu);
 349                 return ret;
 350         }
 351         ret = bringup_wait_for_ap(cpu);
 352         BUG_ON(!cpu_online(cpu));
 353         return ret;
 354 }
 355
 356 /*
 357  * Hotplug state machine related functions
 358  */
 359 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st,
 360                           struct cpuhp_step *steps)
 361 {
 362         for (st->state++; st->state < st->target; st->state++) {
 363                 struct cpuhp_step *step = steps + st->state;
 364
 365                 if (!step->skip_onerr)
 366                         cpuhp_invoke_callback(cpu, st->state, step->startup);
 367         }
 368 }
 369
 370 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 371                                 struct cpuhp_step *steps, enum cpuhp_state target)
 372 {
 373         enum cpuhp_state prev_state = st->state;
 374         int ret = 0;
 375
 376         for (; st->state > target; st->state--) {
 377                 struct cpuhp_step *step = steps + st->state;
 378
 379                 ret = cpuhp_invoke_callback(cpu, st->state, step->teardown);
 380                 if (ret) {
 381                         st->target = prev_state;
 382                         undo_cpu_down(cpu, st, steps);
 383                         break;
 384                 }
 385         }
 386         return ret;
 387 }
 388
 389 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st,
 390                         struct cpuhp_step *steps)
 391 {
 392         for (st->state--; st->state > st->target; st->state--) {
 393                 struct cpuhp_step *step = steps + st->state;
 394
 395                 if (!step->skip_onerr)
 396                         cpuhp_invoke_callback(cpu, st->state, step->teardown);
 397         }
 398 }
 399
 400 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 401                               struct cpuhp_step *steps, enum cpuhp_state target)
 402 {
 403         enum cpuhp_state prev_state = st->state;
 404         int ret = 0;
 405
 406         while (st->state < target) {
 407                 struct cpuhp_step *step;
 408
 409                 st->state++;
 410                 step = steps + st->state;
 411                 ret = cpuhp_invoke_callback(cpu, st->state, step->startup);
 412                 if (ret) {
 413                         st->target = prev_state;
 414                         undo_cpu_up(cpu, st, steps);
 415                         break;
 416                 }
 417         }
 418         return ret;
 419 }
 420
 421 /*
 422  * The cpu hotplug threads manage the bringup and teardown of the cpus
 423  */
 424 static void cpuhp_create(unsigned int cpu)
 425 {
 426         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 427
 428         init_completion(&st->done);
 429 }
 430
 431 static int cpuhp_should_run(unsigned int cpu)
 432 {
 433         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 434
 435         return st->should_run;
 436 }
 437
 438 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
 439 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
 440 {
 441         enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
 442
 443         return cpuhp_down_callbacks(cpu, st, cpuhp_ap_states, target);
 444 }
 445
 446 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
 447 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
 448 {
 449         return cpuhp_up_callbacks(cpu, st, cpuhp_ap_states, st->target);
 450 }
 451
 452 /*
 453  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 454  * callbacks when a state gets [un]installed at runtime.
 455  */
 456 static void cpuhp_thread_fun(unsigned int cpu)
 457 {
 458         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 459         int ret = 0;
 460
 461         /*
 462          * Paired with the mb() in cpuhp_kick_ap_work and
 463          * cpuhp_invoke_ap_callback, so the work set is consistent visible.
 464          */
 465         smp_mb();
 466         if (!st->should_run)
 467                 return;
 468
 469         st->should_run = false;
 470
 471         /* Single callback invocation for [un]install ? */
 472         if (st->cb) {
 473                 if (st->cb_state < CPUHP_AP_ONLINE) {
 474                         local_irq_disable();
 475                         ret = cpuhp_invoke_callback(cpu, st->cb_state, st->cb);
 476                         local_irq_enable();
 477                 } else {
 478                         ret = cpuhp_invoke_callback(cpu, st->cb_state, st->cb);
 479                 }
 480         } else {
 481                 /* Cannot happen .... */
 482                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 483
 484                 /* Regular hotplug work */
 485                 if (st->state < st->target)
 486                         ret = cpuhp_ap_online(cpu, st);
 487                 else if (st->state > st->target)
 488                         ret = cpuhp_ap_offline(cpu, st);
 489         }
 490         st->result = ret;
 491         complete(&st->done);
 492 }
 493
 494 /* Invoke a single callback on a remote cpu */
 495 static int cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state,
 496                                     int (*cb)(unsigned int))
 497 {
 498         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 499
 500         if (!cpu_online(cpu))
 501                 return 0;
 502
 503         st->cb_state = state;
 504         st->cb = cb;
 505         /*
 506          * Make sure the above stores are visible before should_run becomes
 507          * true. Paired with the mb() above in cpuhp_thread_fun()
 508          */
 509         smp_mb();
 510         st->should_run = true;
 511         wake_up_process(st->thread);
 512         wait_for_completion(&st->done);
 513         return st->result;
 514 }
 515
 516 /* Regular hotplug invocation of the AP hotplug thread */
 517 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
 518 {
 519         st->result = 0;
 520         st->cb = NULL;
 521         /*
 522          * Make sure the above stores are visible before should_run becomes
 523          * true. Paired with the mb() above in cpuhp_thread_fun()
 524          */
 525         smp_mb();
 526         st->should_run = true;
 527         wake_up_process(st->thread);
 528 }
 529
 530 static int cpuhp_kick_ap_work(unsigned int cpu)
 531 {
 532         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 533         enum cpuhp_state state = st->state;
 534
 535         trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
 536         __cpuhp_kick_ap_work(st);
 537         wait_for_completion(&st->done);
 538         trace_cpuhp_exit(cpu, st->state, state, st->result);
 539         return st->result;
 540 }
 541
 542 static struct smp_hotplug_thread cpuhp_threads = {
 543         .store                  = &cpuhp_state.thread,
 544         .create                 = &cpuhp_create,
 545         .thread_should_run      = cpuhp_should_run,
 546         .thread_fn              = cpuhp_thread_fun,
 547         .thread_comm            = "cpuhp/%u",
 548         .selfparking            = true,
 549 };
 550
 551 void __init cpuhp_threads_init(void)
 552 {
 553         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 554         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 555 }
 556
 557 #ifdef CONFIG_HOTPLUG_CPU
 558 EXPORT_SYMBOL(register_cpu_notifier);
 559 EXPORT_SYMBOL(__register_cpu_notifier);
 560 void unregister_cpu_notifier(struct notifier_block *nb)
 561 {
 562         cpu_maps_update_begin();
 563         raw_notifier_chain_unregister(&cpu_chain, nb);
 564         cpu_maps_update_done();
 565 }
 566 EXPORT_SYMBOL(unregister_cpu_notifier);
 567
 568 void __unregister_cpu_notifier(struct notifier_block *nb)
 569 {
 570         raw_notifier_chain_unregister(&cpu_chain, nb);
 571 }
 572 EXPORT_SYMBOL(__unregister_cpu_notifier);
 573
 574 /**
 575  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 576  * @cpu: a CPU id
 577  *
 578  * This function walks all processes, finds a valid mm struct for each one and
 579  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 580  * trivial, there are various non-obvious corner cases, which this function
 581  * tries to solve in a safe manner.
 582  *
 583  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 584  * be called only for an already offlined CPU.
 585  */
 586 void clear_tasks_mm_cpumask(int cpu)
 587 {
 588         struct task_struct *p;
 589
 590         /*
 591          * This function is called after the cpu is taken down and marked
 592          * offline, so its not like new tasks will ever get this cpu set in
 593          * their mm mask. -- Peter Zijlstra
 594          * Thus, we may use rcu_read_lock() here, instead of grabbing
 595          * full-fledged tasklist_lock.
 596          */
 597         WARN_ON(cpu_online(cpu));
 598         rcu_read_lock();
 599         for_each_process(p) {
 600                 struct task_struct *t;
 601
 602                 /*
 603                  * Main thread might exit, but other threads may still have
 604                  * a valid mm. Find one.
 605                  */
 606                 t = find_lock_task_mm(p);
 607                 if (!t)
 608                         continue;
 609                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 610                 task_unlock(t);
 611         }
 612         rcu_read_unlock();
 613 }
 614
 615 static inline void check_for_tasks(int dead_cpu)
 616 {
 617         struct task_struct *g, *p;
 618
 619         read_lock(&tasklist_lock);
 620         for_each_process_thread(g, p) {
 621                 if (!p->on_rq)
 622                         continue;
 623                 /*
 624                  * We do the check with unlocked task_rq(p)->lock.
 625                  * Order the reading to do not warn about a task,
 626                  * which was running on this cpu in the past, and
 627                  * it's just been woken on another cpu.
 628                  */
 629                 rmb();
 630                 if (task_cpu(p) != dead_cpu)
 631                         continue;
 632
 633                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 634                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 635         }
 636         read_unlock(&tasklist_lock);
 637 }
 638
 639 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
 640 {
 641         BUG_ON(cpu_notify(val, cpu));
 642 }
 643
 644 static int notify_down_prepare(unsigned int cpu)
 645 {
 646         int err, nr_calls = 0;
 647
 648         err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
 649         if (err) {
 650                 nr_calls--;
 651                 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
 652                 pr_warn("%s: attempt to take down CPU %u failed\n",
 653                                 __func__, cpu);
 654         }
 655         return err;
 656 }
 657
 658 static int notify_dying(unsigned int cpu)
 659 {
 660         cpu_notify(CPU_DYING, cpu);
 661         return 0;
 662 }
 663
 664 /* Take this CPU down. */
 665 static int take_cpu_down(void *_param)
 666 {
 667         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 668         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 669         int err, cpu = smp_processor_id();
 670
 671         /* Ensure this CPU doesn't handle any more interrupts. */
 672         err = __cpu_disable();
 673         if (err < 0)
 674                 return err;
 675
 676         /* Invoke the former CPU_DYING callbacks */
 677         for (; st->state > target; st->state--) {
 678                 struct cpuhp_step *step = cpuhp_ap_states + st->state;
 679
 680                 cpuhp_invoke_callback(cpu, st->state, step->teardown);
 681         }
 682         /* Give up timekeeping duties */
 683         tick_handover_do_timer();
 684         /* Park the stopper thread */
 685         stop_machine_park(cpu);
 686         return 0;
 687 }
 688
 689 static int takedown_cpu(unsigned int cpu)
 690 {
 691         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 692         int err;
 693
 694         /*
 695          * By now we've cleared cpu_active_mask, wait for all preempt-disabled
 696          * and RCU users of this state to go away such that all new such users
 697          * will observe it.
 698          *
 699          * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
 700          * not imply sync_sched(), so wait for both.
 701          *
 702          * Do sync before park smpboot threads to take care the rcu boost case.
 703          */
 704         if (IS_ENABLED(CONFIG_PREEMPT))
 705                 synchronize_rcu_mult(call_rcu, call_rcu_sched);
 706         else
 707                 synchronize_rcu();
 708
 709         /* Park the smpboot threads */
 710         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 711         smpboot_park_threads(cpu);
 712
 713         /*
 714          * Prevent irq alloc/free while the dying cpu reorganizes the
 715          * interrupt affinities.
 716          */
 717         irq_lock_sparse();
 718
 719         /*
 720          * So now all preempt/rcu users must observe !cpu_active().
 721          */
 722         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 723         if (err) {
 724                 /* CPU didn't die: tell everyone.  Can't complain. */
 725                 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
 726                 irq_unlock_sparse();
 727                 return err;
 728         }
 729         BUG_ON(cpu_online(cpu));
 730
 731         /*
 732          * The migration_call() CPU_DYING callback will have removed all
 733          * runnable tasks from the cpu, there's only the idle task left now
 734          * that the migration thread is done doing the stop_machine thing.
 735          *
 736          * Wait for the stop thread to go away.
 737          */
 738         wait_for_completion(&st->done);
 739         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 740
 741         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 742         irq_unlock_sparse();
 743
 744         hotplug_cpu__broadcast_tick_pull(cpu);
 745         /* This actually kills the CPU. */
 746         __cpu_die(cpu);
 747
 748         tick_cleanup_dead_cpu(cpu);
 749         return 0;
 750 }
 751
 752 static int notify_dead(unsigned int cpu)
 753 {
 754         cpu_notify_nofail(CPU_DEAD, cpu);
 755         check_for_tasks(cpu);
 756         return 0;
 757 }
 758
 759 static void cpuhp_complete_idle_dead(void *arg)
 760 {
 761         struct cpuhp_cpu_state *st = arg;
 762
 763         complete(&st->done);
 764 }
 765
 766 void cpuhp_report_idle_dead(void)
 767 {
 768         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 769
 770         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 771         rcu_report_dead(smp_processor_id());
 772         st->state = CPUHP_AP_IDLE_DEAD;
 773         /*
 774          * We cannot call complete after rcu_report_dead() so we delegate it
 775          * to an online cpu.
 776          */
 777         smp_call_function_single(cpumask_first(cpu_online_mask),
 778                                  cpuhp_complete_idle_dead, st, 0);
 779 }
 780
 781 #else
 782 #define notify_down_prepare     NULL
 783 #define takedown_cpu            NULL
 784 #define notify_dead             NULL
 785 #define notify_dying            NULL
 786 #endif
 787
 788 #ifdef CONFIG_HOTPLUG_CPU
 789
 790 /* Requires cpu_add_remove_lock to be held */
 791 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 792                            enum cpuhp_state target)
 793 {
 794         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 795         int prev_state, ret = 0;
 796         bool hasdied = false;
 797
 798         if (num_online_cpus() == 1)
 799                 return -EBUSY;
 800
 801         if (!cpu_present(cpu))
 802                 return -EINVAL;
 803
 804         cpu_hotplug_begin();
 805
 806         cpuhp_tasks_frozen = tasks_frozen;
 807
 808         prev_state = st->state;
 809         st->target = target;
 810         /*
 811          * If the current CPU state is in the range of the AP hotplug thread,
 812          * then we need to kick the thread.
 813          */
 814         if (st->state > CPUHP_TEARDOWN_CPU) {
 815                 ret = cpuhp_kick_ap_work(cpu);
 816                 /*
 817                  * The AP side has done the error rollback already. Just
 818                  * return the error code..
 819                  */
 820                 if (ret)
 821                         goto out;
 822
 823                 /*
 824                  * We might have stopped still in the range of the AP hotplug
 825                  * thread. Nothing to do anymore.
 826                  */
 827                 if (st->state > CPUHP_TEARDOWN_CPU)
 828                         goto out;
 829         }
 830         /*
 831          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 832          * to do the further cleanups.
 833          */
 834         ret = cpuhp_down_callbacks(cpu, st, cpuhp_bp_states, target);
 835
 836         hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
 837 out:
 838         cpu_hotplug_done();
 839         /* This post dead nonsense must die */
 840         if (!ret && hasdied)
 841                 cpu_notify_nofail(CPU_POST_DEAD, cpu);
 842         return ret;
 843 }
 844
 845 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 846 {
 847         int err;
 848
 849         cpu_maps_update_begin();
 850
 851         if (cpu_hotplug_disabled) {
 852                 err = -EBUSY;
 853                 goto out;
 854         }
 855
 856         err = _cpu_down(cpu, 0, target);
 857
 858 out:
 859         cpu_maps_update_done();
 860         return err;
 861 }
 862 int cpu_down(unsigned int cpu)
 863 {
 864         return do_cpu_down(cpu, CPUHP_OFFLINE);
 865 }
 866 EXPORT_SYMBOL(cpu_down);
 867 #endif /*CONFIG_HOTPLUG_CPU*/
 868
 869 /**
 870  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 871  * @cpu: cpu that just started
 872  *
 873  * This function calls the cpu_chain notifiers with CPU_STARTING.
 874  * It must be called by the arch code on the new cpu, before the new cpu
 875  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 876  */
 877 void notify_cpu_starting(unsigned int cpu)
 878 {
 879         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 880         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 881
 882         while (st->state < target) {
 883                 struct cpuhp_step *step;
 884
 885                 st->state++;
 886                 step = cpuhp_ap_states + st->state;
 887                 cpuhp_invoke_callback(cpu, st->state, step->startup);
 888         }
 889 }
 890
 891 /*
 892  * Called from the idle task. We need to set active here, so we can kick off
 893  * the stopper thread and unpark the smpboot threads. If the target state is
 894  * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
 895  * cpu further.
 896  */
 897 void cpuhp_online_idle(enum cpuhp_state state)
 898 {
 899         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 900         unsigned int cpu = smp_processor_id();
 901
 902         /* Happens for the boot cpu */
 903         if (state != CPUHP_AP_ONLINE_IDLE)
 904                 return;
 905
 906         st->state = CPUHP_AP_ONLINE_IDLE;
 907
 908         /* The cpu is marked online, set it active now */
 909         set_cpu_active(cpu, true);
 910         /* Unpark the stopper thread and the hotplug thread of this cpu */
 911         stop_machine_unpark(cpu);
 912         kthread_unpark(st->thread);
 913
 914         /* Should we go further up ? */
 915         if (st->target > CPUHP_AP_ONLINE_IDLE)
 916                 __cpuhp_kick_ap_work(st);
 917         else
 918                 complete(&st->done);
 919 }
 920
 921 /* Requires cpu_add_remove_lock to be held */
 922 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 923 {
 924         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 925         struct task_struct *idle;
 926         int ret = 0;
 927
 928         cpu_hotplug_begin();
 929
 930         if (!cpu_present(cpu)) {
 931                 ret = -EINVAL;
 932                 goto out;
 933         }
 934
 935         /*
 936          * The caller of do_cpu_up might have raced with another
 937          * caller. Ignore it for now.
 938          */
 939         if (st->state >= target)
 940                 goto out;
 941
 942         if (st->state == CPUHP_OFFLINE) {
 943                 /* Let it fail before we try to bring the cpu up */
 944                 idle = idle_thread_get(cpu);
 945                 if (IS_ERR(idle)) {
 946                         ret = PTR_ERR(idle);
 947                         goto out;
 948                 }
 949         }
 950
 951         cpuhp_tasks_frozen = tasks_frozen;
 952
 953         st->target = target;
 954         /*
 955          * If the current CPU state is in the range of the AP hotplug thread,
 956          * then we need to kick the thread once more.
 957          */
 958         if (st->state > CPUHP_BRINGUP_CPU) {
 959                 ret = cpuhp_kick_ap_work(cpu);
 960                 /*
 961                  * The AP side has done the error rollback already. Just
 962                  * return the error code..
 963                  */
 964                 if (ret)
 965                         goto out;
 966         }
 967
 968         /*
 969          * Try to reach the target state. We max out on the BP at
 970          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
 971          * responsible for bringing it up to the target state.
 972          */
 973         target = min((int)target, CPUHP_BRINGUP_CPU);
 974         ret = cpuhp_up_callbacks(cpu, st, cpuhp_bp_states, target);
 975 out:
 976         cpu_hotplug_done();
 977         return ret;
 978 }
 979
 980 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
 981 {
 982         int err = 0;
 983
 984         if (!cpu_possible(cpu)) {
 985                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
 986                        cpu);
 987 #if defined(CONFIG_IA64)
 988                 pr_err("please check additional_cpus= boot parameter\n");
 989 #endif
 990                 return -EINVAL;
 991         }
 992
 993         err = try_online_node(cpu_to_node(cpu));
 994         if (err)
 995                 return err;
 996
 997         cpu_maps_update_begin();
 998
 999         if (cpu_hotplug_disabled) {
1000                 err = -EBUSY;
1001                 goto out;
1002         }
1003
1004         err = _cpu_up(cpu, 0, target);
1005 out:
1006         cpu_maps_update_done();
1007         return err;
1008 }
1009
1010 int cpu_up(unsigned int cpu)
1011 {
1012         return do_cpu_up(cpu, CPUHP_ONLINE);
1013 }
1014 EXPORT_SYMBOL_GPL(cpu_up);
1015
1016 #ifdef CONFIG_PM_SLEEP_SMP
1017 static cpumask_var_t frozen_cpus;
1018
1019 int disable_nonboot_cpus(void)
1020 {
1021         int cpu, first_cpu, error = 0;
1022
1023         cpu_maps_update_begin();
1024         first_cpu = cpumask_first(cpu_online_mask);
1025         /*
1026          * We take down all of the non-boot CPUs in one shot to avoid races
1027          * with the userspace trying to use the CPU hotplug at the same time
1028          */
1029         cpumask_clear(frozen_cpus);
1030
1031         pr_info("Disabling non-boot CPUs ...\n");
1032         for_each_online_cpu(cpu) {
1033                 if (cpu == first_cpu)
1034                         continue;
1035                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1036                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1037                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1038                 if (!error)
1039                         cpumask_set_cpu(cpu, frozen_cpus);
1040                 else {
1041                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1042                         break;
1043                 }
1044         }
1045
1046         if (!error)
1047                 BUG_ON(num_online_cpus() > 1);
1048         else
1049                 pr_err("Non-boot CPUs are not disabled\n");
1050
1051         /*
1052          * Make sure the CPUs won't be enabled by someone else. We need to do
1053          * this even in case of failure as all disable_nonboot_cpus() users are
1054          * supposed to do enable_nonboot_cpus() on the failure path.
1055          */
1056         cpu_hotplug_disabled++;
1057
1058         cpu_maps_update_done();
1059         return error;
1060 }
1061
1062 void __weak arch_enable_nonboot_cpus_begin(void)
1063 {
1064 }
1065
1066 void __weak arch_enable_nonboot_cpus_end(void)
1067 {
1068 }
1069
1070 void enable_nonboot_cpus(void)
1071 {
1072         int cpu, error;
1073
1074         /* Allow everyone to use the CPU hotplug again */
1075         cpu_maps_update_begin();
1076         WARN_ON(--cpu_hotplug_disabled < 0);
1077         if (cpumask_empty(frozen_cpus))
1078                 goto out;
1079
1080         pr_info("Enabling non-boot CPUs ...\n");
1081
1082         arch_enable_nonboot_cpus_begin();
1083
1084         for_each_cpu(cpu, frozen_cpus) {
1085                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1086                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1087                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1088                 if (!error) {
1089                         pr_info("CPU%d is up\n", cpu);
1090                         continue;
1091                 }
1092                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1093         }
1094
1095         arch_enable_nonboot_cpus_end();
1096
1097         cpumask_clear(frozen_cpus);
1098 out:
1099         cpu_maps_update_done();
1100 }
1101
1102 static int __init alloc_frozen_cpus(void)
1103 {
1104         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1105                 return -ENOMEM;
1106         return 0;
1107 }
1108 core_initcall(alloc_frozen_cpus);
1109
1110 /*
1111  * When callbacks for CPU hotplug notifications are being executed, we must
1112  * ensure that the state of the system with respect to the tasks being frozen
1113  * or not, as reported by the notification, remains unchanged *throughout the
1114  * duration* of the execution of the callbacks.
1115  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1116  *
1117  * This synchronization is implemented by mutually excluding regular CPU
1118  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1119  * Hibernate notifications.
1120  */
1121 static int
1122 cpu_hotplug_pm_callback(struct notifier_block *nb,
1123                         unsigned long action, void *ptr)
1124 {
1125         switch (action) {
1126
1127         case PM_SUSPEND_PREPARE:
1128         case PM_HIBERNATION_PREPARE:
1129                 cpu_hotplug_disable();
1130                 break;
1131
1132         case PM_POST_SUSPEND:
1133         case PM_POST_HIBERNATION:
1134                 cpu_hotplug_enable();
1135                 break;
1136
1137         default:
1138                 return NOTIFY_DONE;
1139         }
1140
1141         return NOTIFY_OK;
1142 }
1143
1144
1145 static int __init cpu_hotplug_pm_sync_init(void)
1146 {
1147         /*
1148          * cpu_hotplug_pm_callback has higher priority than x86
1149          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1150          * to disable cpu hotplug to avoid cpu hotplug race.
1151          */
1152         pm_notifier(cpu_hotplug_pm_callback, 0);
1153         return 0;
1154 }
1155 core_initcall(cpu_hotplug_pm_sync_init);
1156
1157 #endif /* CONFIG_PM_SLEEP_SMP */
1158
1159 #endif /* CONFIG_SMP */
1160
1161 /* Boot processor state steps */
1162 static struct cpuhp_step cpuhp_bp_states[] = {
1163         [CPUHP_OFFLINE] = {
1164                 .name                   = "offline",
1165                 .startup                = NULL,
1166                 .teardown               = NULL,
1167         },
1168 #ifdef CONFIG_SMP
1169         [CPUHP_CREATE_THREADS]= {
1170                 .name                   = "threads:create",
1171                 .startup                = smpboot_create_threads,
1172                 .teardown               = NULL,
1173                 .cant_stop              = true,
1174         },
1175         /*
1176          * Preparatory and dead notifiers. Will be replaced once the notifiers
1177          * are converted to states.
1178          */
1179         [CPUHP_NOTIFY_PREPARE] = {
1180                 .name                   = "notify:prepare",
1181                 .startup                = notify_prepare,
1182                 .teardown               = notify_dead,
1183                 .skip_onerr             = true,
1184                 .cant_stop              = true,
1185         },
1186         /* Kicks the plugged cpu into life */
1187         [CPUHP_BRINGUP_CPU] = {
1188                 .name                   = "cpu:bringup",
1189                 .startup                = bringup_cpu,
1190                 .teardown               = NULL,
1191                 .cant_stop              = true,
1192         },
1193         /*
1194          * Handled on controll processor until the plugged processor manages
1195          * this itself.
1196          */
1197         [CPUHP_TEARDOWN_CPU] = {
1198                 .name                   = "cpu:teardown",
1199                 .startup                = NULL,
1200                 .teardown               = takedown_cpu,
1201                 .cant_stop              = true,
1202         },
1203 #endif
1204 };
1205
1206 /* Application processor state steps */
1207 static struct cpuhp_step cpuhp_ap_states[] = {
1208 #ifdef CONFIG_SMP
1209         /* Final state before CPU kills itself */
1210         [CPUHP_AP_IDLE_DEAD] = {
1211                 .name                   = "idle:dead",
1212         },
1213         /*
1214          * Last state before CPU enters the idle loop to die. Transient state
1215          * for synchronization.
1216          */
1217         [CPUHP_AP_OFFLINE] = {
1218                 .name                   = "ap:offline",
1219                 .cant_stop              = true,
1220         },
1221         /*
1222          * Low level startup/teardown notifiers. Run with interrupts
1223          * disabled. Will be removed once the notifiers are converted to
1224          * states.
1225          */
1226         [CPUHP_AP_NOTIFY_STARTING] = {
1227                 .name                   = "notify:starting",
1228                 .startup                = notify_starting,
1229                 .teardown               = notify_dying,
1230                 .skip_onerr             = true,
1231                 .cant_stop              = true,
1232         },
1233         /* Entry state on starting. Interrupts enabled from here on. Transient
1234          * state for synchronsization */
1235         [CPUHP_AP_ONLINE] = {
1236                 .name                   = "ap:online",
1237         },
1238         /* Handle smpboot threads park/unpark */
1239         [CPUHP_AP_SMPBOOT_THREADS] = {
1240                 .name                   = "smpboot:threads",
1241                 .startup                = smpboot_unpark_threads,
1242                 .teardown               = NULL,
1243         },
1244         /*
1245          * Online/down_prepare notifiers. Will be removed once the notifiers
1246          * are converted to states.
1247          */
1248         [CPUHP_AP_NOTIFY_ONLINE] = {
1249                 .name                   = "notify:online",
1250                 .startup                = notify_online,
1251                 .teardown               = notify_down_prepare,
1252         },
1253 #endif
1254         /*
1255          * The dynamically registered state space is here
1256          */
1257
1258         /* CPU is fully up and running. */
1259         [CPUHP_ONLINE] = {
1260                 .name                   = "online",
1261                 .startup                = NULL,
1262                 .teardown               = NULL,
1263         },
1264 };
1265
1266 /* Sanity check for callbacks */
1267 static int cpuhp_cb_check(enum cpuhp_state state)
1268 {
1269         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1270                 return -EINVAL;
1271         return 0;
1272 }
1273
1274 static bool cpuhp_is_ap_state(enum cpuhp_state state)
1275 {
1276         /*
1277          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
1278          * purposes as that state is handled explicitely in cpu_down.
1279          */
1280         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
1281 }
1282
1283 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
1284 {
1285         struct cpuhp_step *sp;
1286
1287         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
1288         return sp + state;
1289 }
1290
1291 static void cpuhp_store_callbacks(enum cpuhp_state state,
1292                                   const char *name,
1293                                   int (*startup)(unsigned int cpu),
1294                                   int (*teardown)(unsigned int cpu))
1295 {
1296         /* (Un)Install the callbacks for further cpu hotplug operations */
1297         struct cpuhp_step *sp;
1298
1299         mutex_lock(&cpuhp_state_mutex);
1300         sp = cpuhp_get_step(state);
1301         sp->startup = startup;
1302         sp->teardown = teardown;
1303         sp->name = name;
1304         mutex_unlock(&cpuhp_state_mutex);
1305 }
1306
1307 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1308 {
1309         return cpuhp_get_step(state)->teardown;
1310 }
1311
1312 /*
1313  * Call the startup/teardown function for a step either on the AP or
1314  * on the current CPU.
1315  */
1316 static int cpuhp_issue_call(int cpu, enum cpuhp_state state,
1317                             int (*cb)(unsigned int), bool bringup)
1318 {
1319         int ret;
1320
1321         if (!cb)
1322                 return 0;
1323         /*
1324          * The non AP bound callbacks can fail on bringup. On teardown
1325          * e.g. module removal we crash for now.
1326          */
1327 #ifdef CONFIG_SMP
1328         if (cpuhp_is_ap_state(state))
1329                 ret = cpuhp_invoke_ap_callback(cpu, state, cb);
1330         else
1331                 ret = cpuhp_invoke_callback(cpu, state, cb);
1332 #else
1333         ret = cpuhp_invoke_callback(cpu, state, cb);
1334 #endif
1335         BUG_ON(ret && !bringup);
1336         return ret;
1337 }
1338
1339 /*
1340  * Called from __cpuhp_setup_state on a recoverable failure.
1341  *
1342  * Note: The teardown callbacks for rollback are not allowed to fail!
1343  */
1344 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1345                                    int (*teardown)(unsigned int cpu))
1346 {
1347         int cpu;
1348
1349         if (!teardown)
1350                 return;
1351
1352         /* Roll back the already executed steps on the other cpus */
1353         for_each_present_cpu(cpu) {
1354                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1355                 int cpustate = st->state;
1356
1357                 if (cpu >= failedcpu)
1358                         break;
1359
1360                 /* Did we invoke the startup call on that cpu ? */
1361                 if (cpustate >= state)
1362                         cpuhp_issue_call(cpu, state, teardown, false);
1363         }
1364 }
1365
1366 /*
1367  * Returns a free for dynamic slot assignment of the Online state. The states
1368  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1369  * by having no name assigned.
1370  */
1371 static int cpuhp_reserve_state(enum cpuhp_state state)
1372 {
1373         enum cpuhp_state i;
1374
1375         mutex_lock(&cpuhp_state_mutex);
1376         for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1377                 if (cpuhp_ap_states[i].name)
1378                         continue;
1379
1380                 cpuhp_ap_states[i].name = "Reserved";
1381                 mutex_unlock(&cpuhp_state_mutex);
1382                 return i;
1383         }
1384         mutex_unlock(&cpuhp_state_mutex);
1385         WARN(1, "No more dynamic states available for CPU hotplug\n");
1386         return -ENOSPC;
1387 }
1388
1389 /**
1390  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1391  * @state:      The state to setup
1392  * @invoke:     If true, the startup function is invoked for cpus where
1393  *              cpu state >= @state
1394  * @startup:    startup callback function
1395  * @teardown:   teardown callback function
1396  *
1397  * Returns 0 if successful, otherwise a proper error code
1398  */
1399 int __cpuhp_setup_state(enum cpuhp_state state,
1400                         const char *name, bool invoke,
1401                         int (*startup)(unsigned int cpu),
1402                         int (*teardown)(unsigned int cpu))
1403 {
1404         int cpu, ret = 0;
1405         int dyn_state = 0;
1406
1407         if (cpuhp_cb_check(state) || !name)
1408                 return -EINVAL;
1409
1410         get_online_cpus();
1411
1412         /* currently assignments for the ONLINE state are possible */
1413         if (state == CPUHP_AP_ONLINE_DYN) {
1414                 dyn_state = 1;
1415                 ret = cpuhp_reserve_state(state);
1416                 if (ret < 0)
1417                         goto out;
1418                 state = ret;
1419         }
1420
1421         cpuhp_store_callbacks(state, name, startup, teardown);
1422
1423         if (!invoke || !startup)
1424                 goto out;
1425
1426         /*
1427          * Try to call the startup callback for each present cpu
1428          * depending on the hotplug state of the cpu.
1429          */
1430         for_each_present_cpu(cpu) {
1431                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1432                 int cpustate = st->state;
1433
1434                 if (cpustate < state)
1435                         continue;
1436
1437                 ret = cpuhp_issue_call(cpu, state, startup, true);
1438                 if (ret) {
1439                         cpuhp_rollback_install(cpu, state, teardown);
1440                         cpuhp_store_callbacks(state, NULL, NULL, NULL);
1441                         goto out;
1442                 }
1443         }
1444 out:
1445         put_online_cpus();
1446         if (!ret && dyn_state)
1447                 return state;
1448         return ret;
1449 }
1450 EXPORT_SYMBOL(__cpuhp_setup_state);
1451
1452 /**
1453  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1454  * @state:      The state to remove
1455  * @invoke:     If true, the teardown function is invoked for cpus where
1456  *              cpu state >= @state
1457  *
1458  * The teardown callback is currently not allowed to fail. Think
1459  * about module removal!
1460  */
1461 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1462 {
1463         int (*teardown)(unsigned int cpu) = cpuhp_get_teardown_cb(state);
1464         int cpu;
1465
1466         BUG_ON(cpuhp_cb_check(state));
1467
1468         get_online_cpus();
1469
1470         if (!invoke || !teardown)
1471                 goto remove;
1472
1473         /*
1474          * Call the teardown callback for each present cpu depending
1475          * on the hotplug state of the cpu. This function is not
1476          * allowed to fail currently!
1477          */
1478         for_each_present_cpu(cpu) {
1479                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1480                 int cpustate = st->state;
1481
1482                 if (cpustate >= state)
1483                         cpuhp_issue_call(cpu, state, teardown, false);
1484         }
1485 remove:
1486         cpuhp_store_callbacks(state, NULL, NULL, NULL);
1487         put_online_cpus();
1488 }
1489 EXPORT_SYMBOL(__cpuhp_remove_state);
1490
1491 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1492 static ssize_t show_cpuhp_state(struct device *dev,
1493                                 struct device_attribute *attr, char *buf)
1494 {
1495         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1496
1497         return sprintf(buf, "%d\n", st->state);
1498 }
1499 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1500
1501 static ssize_t write_cpuhp_target(struct device *dev,
1502                                   struct device_attribute *attr,
1503                                   const char *buf, size_t count)
1504 {
1505         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1506         struct cpuhp_step *sp;
1507         int target, ret;
1508
1509         ret = kstrtoint(buf, 10, &target);
1510         if (ret)
1511                 return ret;
1512
1513 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1514         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1515                 return -EINVAL;
1516 #else
1517         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1518                 return -EINVAL;
1519 #endif
1520
1521         ret = lock_device_hotplug_sysfs();
1522         if (ret)
1523                 return ret;
1524
1525         mutex_lock(&cpuhp_state_mutex);
1526         sp = cpuhp_get_step(target);
1527         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1528         mutex_unlock(&cpuhp_state_mutex);
1529         if (ret)
1530                 return ret;
1531
1532         if (st->state < target)
1533                 ret = do_cpu_up(dev->id, target);
1534         else
1535                 ret = do_cpu_down(dev->id, target);
1536
1537         unlock_device_hotplug();
1538         return ret ? ret : count;
1539 }
1540
1541 static ssize_t show_cpuhp_target(struct device *dev,
1542                                  struct device_attribute *attr, char *buf)
1543 {
1544         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1545
1546         return sprintf(buf, "%d\n", st->target);
1547 }
1548 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1549
1550 static struct attribute *cpuhp_cpu_attrs[] = {
1551         &dev_attr_state.attr,
1552         &dev_attr_target.attr,
1553         NULL
1554 };
1555
1556 static struct attribute_group cpuhp_cpu_attr_group = {
1557         .attrs = cpuhp_cpu_attrs,
1558         .name = "hotplug",
1559         NULL
1560 };
1561
1562 static ssize_t show_cpuhp_states(struct device *dev,
1563                                  struct device_attribute *attr, char *buf)
1564 {
1565         ssize_t cur, res = 0;
1566         int i;
1567
1568         mutex_lock(&cpuhp_state_mutex);
1569         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1570                 struct cpuhp_step *sp = cpuhp_get_step(i);
1571
1572                 if (sp->name) {
1573                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1574                         buf += cur;
1575                         res += cur;
1576                 }
1577         }
1578         mutex_unlock(&cpuhp_state_mutex);
1579         return res;
1580 }
1581 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1582
1583 static struct attribute *cpuhp_cpu_root_attrs[] = {
1584         &dev_attr_states.attr,
1585         NULL
1586 };
1587
1588 static struct attribute_group cpuhp_cpu_root_attr_group = {
1589         .attrs = cpuhp_cpu_root_attrs,
1590         .name = "hotplug",
1591         NULL
1592 };
1593
1594 static int __init cpuhp_sysfs_init(void)
1595 {
1596         int cpu, ret;
1597
1598         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1599                                  &cpuhp_cpu_root_attr_group);
1600         if (ret)
1601                 return ret;
1602
1603         for_each_possible_cpu(cpu) {
1604                 struct device *dev = get_cpu_device(cpu);
1605
1606                 if (!dev)
1607                         continue;
1608                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1609                 if (ret)
1610                         return ret;
1611         }
1612         return 0;
1613 }
1614 device_initcall(cpuhp_sysfs_init);
1615 #endif
1616
1617 /*
1618  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1619  * represents all NR_CPUS bits binary values of 1<<nr.
1620  *
1621  * It is used by cpumask_of() to get a constant address to a CPU
1622  * mask value that has a single bit set only.
1623  */
1624
1625 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1626 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1627 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1628 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1629 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1630
1631 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1632
1633         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1634         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1635 #if BITS_PER_LONG > 32
1636         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1637         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1638 #endif
1639 };
1640 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1641
1642 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1643 EXPORT_SYMBOL(cpu_all_bits);
1644
1645 #ifdef CONFIG_INIT_ALL_POSSIBLE
1646 struct cpumask __cpu_possible_mask __read_mostly
1647         = {CPU_BITS_ALL};
1648 #else
1649 struct cpumask __cpu_possible_mask __read_mostly;
1650 #endif
1651 EXPORT_SYMBOL(__cpu_possible_mask);
1652
1653 struct cpumask __cpu_online_mask __read_mostly;
1654 EXPORT_SYMBOL(__cpu_online_mask);
1655
1656 struct cpumask __cpu_present_mask __read_mostly;
1657 EXPORT_SYMBOL(__cpu_present_mask);
1658
1659 struct cpumask __cpu_active_mask __read_mostly;
1660 EXPORT_SYMBOL(__cpu_active_mask);
1661
1662 void init_cpu_present(const struct cpumask *src)
1663 {
1664         cpumask_copy(&__cpu_present_mask, src);
1665 }
1666
1667 void init_cpu_possible(const struct cpumask *src)
1668 {
1669         cpumask_copy(&__cpu_possible_mask, src);
1670 }
1671
1672 void init_cpu_online(const struct cpumask *src)
1673 {
1674         cpumask_copy(&__cpu_online_mask, src);
1675 }
1676
1677 /*
1678  * Activate the first processor.
1679  */
1680 void __init boot_cpu_init(void)
1681 {
1682         int cpu = smp_processor_id();
1683
1684         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1685         set_cpu_online(cpu, true);
1686         set_cpu_active(cpu, true);
1687         set_cpu_present(cpu, true);
1688         set_cpu_possible(cpu, true);
1689 }
1690
1691 /*
1692  * Must be called _AFTER_ setting up the per_cpu areas
1693  */
1694 void __init boot_cpu_state_init(void)
1695 {
1696         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1697 }