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         st->cb_state = state;
 521         st->cb = cb;
 522         /*
 523          * Make sure the above stores are visible before should_run becomes
 524          * true. Paired with the mb() above in cpuhp_thread_fun()
 525          */
 526         smp_mb();
 527         st->should_run = true;
 528         wake_up_process(st->thread);
 529         wait_for_completion(&st->done);
 530         return st->result;
 531 }
 532
 533 /* Regular hotplug invocation of the AP hotplug thread */
 534 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
 535 {
 536         st->result = 0;
 537         st->cb = NULL;
 538         /*
 539          * Make sure the above stores are visible before should_run becomes
 540          * true. Paired with the mb() above in cpuhp_thread_fun()
 541          */
 542         smp_mb();
 543         st->should_run = true;
 544         wake_up_process(st->thread);
 545 }
 546
 547 static int cpuhp_kick_ap_work(unsigned int cpu)
 548 {
 549         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 550         enum cpuhp_state state = st->state;
 551
 552         trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
 553         __cpuhp_kick_ap_work(st);
 554         wait_for_completion(&st->done);
 555         trace_cpuhp_exit(cpu, st->state, state, st->result);
 556         return st->result;
 557 }
 558
 559 static struct smp_hotplug_thread cpuhp_threads = {
 560         .store                  = &cpuhp_state.thread,
 561         .create                 = &cpuhp_create,
 562         .thread_should_run      = cpuhp_should_run,
 563         .thread_fn              = cpuhp_thread_fun,
 564         .thread_comm            = "cpuhp/%u",
 565         .selfparking            = true,
 566 };
 567
 568 void __init cpuhp_threads_init(void)
 569 {
 570         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 571         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 572 }
 573
 574 #ifdef CONFIG_HOTPLUG_CPU
 575 EXPORT_SYMBOL(register_cpu_notifier);
 576 EXPORT_SYMBOL(__register_cpu_notifier);
 577 void unregister_cpu_notifier(struct notifier_block *nb)
 578 {
 579         cpu_maps_update_begin();
 580         raw_notifier_chain_unregister(&cpu_chain, nb);
 581         cpu_maps_update_done();
 582 }
 583 EXPORT_SYMBOL(unregister_cpu_notifier);
 584
 585 void __unregister_cpu_notifier(struct notifier_block *nb)
 586 {
 587         raw_notifier_chain_unregister(&cpu_chain, nb);
 588 }
 589 EXPORT_SYMBOL(__unregister_cpu_notifier);
 590
 591 /**
 592  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 593  * @cpu: a CPU id
 594  *
 595  * This function walks all processes, finds a valid mm struct for each one and
 596  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 597  * trivial, there are various non-obvious corner cases, which this function
 598  * tries to solve in a safe manner.
 599  *
 600  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 601  * be called only for an already offlined CPU.
 602  */
 603 void clear_tasks_mm_cpumask(int cpu)
 604 {
 605         struct task_struct *p;
 606
 607         /*
 608          * This function is called after the cpu is taken down and marked
 609          * offline, so its not like new tasks will ever get this cpu set in
 610          * their mm mask. -- Peter Zijlstra
 611          * Thus, we may use rcu_read_lock() here, instead of grabbing
 612          * full-fledged tasklist_lock.
 613          */
 614         WARN_ON(cpu_online(cpu));
 615         rcu_read_lock();
 616         for_each_process(p) {
 617                 struct task_struct *t;
 618
 619                 /*
 620                  * Main thread might exit, but other threads may still have
 621                  * a valid mm. Find one.
 622                  */
 623                 t = find_lock_task_mm(p);
 624                 if (!t)
 625                         continue;
 626                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 627                 task_unlock(t);
 628         }
 629         rcu_read_unlock();
 630 }
 631
 632 static inline void check_for_tasks(int dead_cpu)
 633 {
 634         struct task_struct *g, *p;
 635
 636         read_lock(&tasklist_lock);
 637         for_each_process_thread(g, p) {
 638                 if (!p->on_rq)
 639                         continue;
 640                 /*
 641                  * We do the check with unlocked task_rq(p)->lock.
 642                  * Order the reading to do not warn about a task,
 643                  * which was running on this cpu in the past, and
 644                  * it's just been woken on another cpu.
 645                  */
 646                 rmb();
 647                 if (task_cpu(p) != dead_cpu)
 648                         continue;
 649
 650                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 651                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 652         }
 653         read_unlock(&tasklist_lock);
 654 }
 655
 656 static int notify_down_prepare(unsigned int cpu)
 657 {
 658         int err, nr_calls = 0;
 659
 660         err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
 661         if (err) {
 662                 nr_calls--;
 663                 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
 664                 pr_warn("%s: attempt to take down CPU %u failed\n",
 665                                 __func__, cpu);
 666         }
 667         return err;
 668 }
 669
 670 static int notify_dying(unsigned int cpu)
 671 {
 672         cpu_notify(CPU_DYING, cpu);
 673         return 0;
 674 }
 675
 676 /* Take this CPU down. */
 677 static int take_cpu_down(void *_param)
 678 {
 679         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 680         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 681         int err, cpu = smp_processor_id();
 682
 683         /* Ensure this CPU doesn't handle any more interrupts. */
 684         err = __cpu_disable();
 685         if (err < 0)
 686                 return err;
 687
 688         /* Invoke the former CPU_DYING callbacks */
 689         for (; st->state > target; st->state--) {
 690                 struct cpuhp_step *step = cpuhp_ap_states + st->state;
 691
 692                 cpuhp_invoke_callback(cpu, st->state, step->teardown);
 693         }
 694         /* Give up timekeeping duties */
 695         tick_handover_do_timer();
 696         /* Park the stopper thread */
 697         stop_machine_park(cpu);
 698         return 0;
 699 }
 700
 701 static int takedown_cpu(unsigned int cpu)
 702 {
 703         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 704         int err;
 705
 706         /*
 707          * By now we've cleared cpu_active_mask, wait for all preempt-disabled
 708          * and RCU users of this state to go away such that all new such users
 709          * will observe it.
 710          *
 711          * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
 712          * not imply sync_sched(), so wait for both.
 713          *
 714          * Do sync before park smpboot threads to take care the rcu boost case.
 715          */
 716         if (IS_ENABLED(CONFIG_PREEMPT))
 717                 synchronize_rcu_mult(call_rcu, call_rcu_sched);
 718         else
 719                 synchronize_rcu();
 720
 721         /* Park the smpboot threads */
 722         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 723         smpboot_park_threads(cpu);
 724
 725         /*
 726          * Prevent irq alloc/free while the dying cpu reorganizes the
 727          * interrupt affinities.
 728          */
 729         irq_lock_sparse();
 730
 731         /*
 732          * So now all preempt/rcu users must observe !cpu_active().
 733          */
 734         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 735         if (err) {
 736                 /* CPU refused to die */
 737                 irq_unlock_sparse();
 738                 /* Unpark the hotplug thread so we can rollback there */
 739                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 740                 return err;
 741         }
 742         BUG_ON(cpu_online(cpu));
 743
 744         /*
 745          * The migration_call() CPU_DYING callback will have removed all
 746          * runnable tasks from the cpu, there's only the idle task left now
 747          * that the migration thread is done doing the stop_machine thing.
 748          *
 749          * Wait for the stop thread to go away.
 750          */
 751         wait_for_completion(&st->done);
 752         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 753
 754         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 755         irq_unlock_sparse();
 756
 757         hotplug_cpu__broadcast_tick_pull(cpu);
 758         /* This actually kills the CPU. */
 759         __cpu_die(cpu);
 760
 761         tick_cleanup_dead_cpu(cpu);
 762         return 0;
 763 }
 764
 765 static int notify_dead(unsigned int cpu)
 766 {
 767         cpu_notify_nofail(CPU_DEAD, cpu);
 768         check_for_tasks(cpu);
 769         return 0;
 770 }
 771
 772 static void cpuhp_complete_idle_dead(void *arg)
 773 {
 774         struct cpuhp_cpu_state *st = arg;
 775
 776         complete(&st->done);
 777 }
 778
 779 void cpuhp_report_idle_dead(void)
 780 {
 781         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 782
 783         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 784         rcu_report_dead(smp_processor_id());
 785         st->state = CPUHP_AP_IDLE_DEAD;
 786         /*
 787          * We cannot call complete after rcu_report_dead() so we delegate it
 788          * to an online cpu.
 789          */
 790         smp_call_function_single(cpumask_first(cpu_online_mask),
 791                                  cpuhp_complete_idle_dead, st, 0);
 792 }
 793
 794 #else
 795 #define notify_down_prepare     NULL
 796 #define takedown_cpu            NULL
 797 #define notify_dead             NULL
 798 #define notify_dying            NULL
 799 #endif
 800
 801 #ifdef CONFIG_HOTPLUG_CPU
 802
 803 /* Requires cpu_add_remove_lock to be held */
 804 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 805                            enum cpuhp_state target)
 806 {
 807         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 808         int prev_state, ret = 0;
 809         bool hasdied = false;
 810
 811         if (num_online_cpus() == 1)
 812                 return -EBUSY;
 813
 814         if (!cpu_present(cpu))
 815                 return -EINVAL;
 816
 817         cpu_hotplug_begin();
 818
 819         cpuhp_tasks_frozen = tasks_frozen;
 820
 821         prev_state = st->state;
 822         st->target = target;
 823         /*
 824          * If the current CPU state is in the range of the AP hotplug thread,
 825          * then we need to kick the thread.
 826          */
 827         if (st->state > CPUHP_TEARDOWN_CPU) {
 828                 ret = cpuhp_kick_ap_work(cpu);
 829                 /*
 830                  * The AP side has done the error rollback already. Just
 831                  * return the error code..
 832                  */
 833                 if (ret)
 834                         goto out;
 835
 836                 /*
 837                  * We might have stopped still in the range of the AP hotplug
 838                  * thread. Nothing to do anymore.
 839                  */
 840                 if (st->state > CPUHP_TEARDOWN_CPU)
 841                         goto out;
 842         }
 843         /*
 844          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 845          * to do the further cleanups.
 846          */
 847         ret = cpuhp_down_callbacks(cpu, st, cpuhp_bp_states, target);
 848         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 849                 st->target = prev_state;
 850                 st->rollback = true;
 851                 cpuhp_kick_ap_work(cpu);
 852         }
 853
 854         hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
 855 out:
 856         cpu_hotplug_done();
 857         /* This post dead nonsense must die */
 858         if (!ret && hasdied)
 859                 cpu_notify_nofail(CPU_POST_DEAD, cpu);
 860         return ret;
 861 }
 862
 863 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 864 {
 865         int err;
 866
 867         cpu_maps_update_begin();
 868
 869         if (cpu_hotplug_disabled) {
 870                 err = -EBUSY;
 871                 goto out;
 872         }
 873
 874         err = _cpu_down(cpu, 0, target);
 875
 876 out:
 877         cpu_maps_update_done();
 878         return err;
 879 }
 880 int cpu_down(unsigned int cpu)
 881 {
 882         return do_cpu_down(cpu, CPUHP_OFFLINE);
 883 }
 884 EXPORT_SYMBOL(cpu_down);
 885 #endif /*CONFIG_HOTPLUG_CPU*/
 886
 887 /**
 888  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 889  * @cpu: cpu that just started
 890  *
 891  * This function calls the cpu_chain notifiers with CPU_STARTING.
 892  * It must be called by the arch code on the new cpu, before the new cpu
 893  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 894  */
 895 void notify_cpu_starting(unsigned int cpu)
 896 {
 897         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 898         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 899
 900         while (st->state < target) {
 901                 struct cpuhp_step *step;
 902
 903                 st->state++;
 904                 step = cpuhp_ap_states + st->state;
 905                 cpuhp_invoke_callback(cpu, st->state, step->startup);
 906         }
 907 }
 908
 909 /*
 910  * Called from the idle task. We need to set active here, so we can kick off
 911  * the stopper thread and unpark the smpboot threads. If the target state is
 912  * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
 913  * cpu further.
 914  */
 915 void cpuhp_online_idle(enum cpuhp_state state)
 916 {
 917         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 918         unsigned int cpu = smp_processor_id();
 919
 920         /* Happens for the boot cpu */
 921         if (state != CPUHP_AP_ONLINE_IDLE)
 922                 return;
 923
 924         st->state = CPUHP_AP_ONLINE_IDLE;
 925
 926         /* Unpark the stopper thread and the hotplug thread of this cpu */
 927         stop_machine_unpark(cpu);
 928         kthread_unpark(st->thread);
 929
 930         /* Should we go further up ? */
 931         if (st->target > CPUHP_AP_ONLINE_IDLE)
 932                 __cpuhp_kick_ap_work(st);
 933         else
 934                 complete(&st->done);
 935 }
 936
 937 /* Requires cpu_add_remove_lock to be held */
 938 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 939 {
 940         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 941         struct task_struct *idle;
 942         int ret = 0;
 943
 944         cpu_hotplug_begin();
 945
 946         if (!cpu_present(cpu)) {
 947                 ret = -EINVAL;
 948                 goto out;
 949         }
 950
 951         /*
 952          * The caller of do_cpu_up might have raced with another
 953          * caller. Ignore it for now.
 954          */
 955         if (st->state >= target)
 956                 goto out;
 957
 958         if (st->state == CPUHP_OFFLINE) {
 959                 /* Let it fail before we try to bring the cpu up */
 960                 idle = idle_thread_get(cpu);
 961                 if (IS_ERR(idle)) {
 962                         ret = PTR_ERR(idle);
 963                         goto out;
 964                 }
 965         }
 966
 967         cpuhp_tasks_frozen = tasks_frozen;
 968
 969         st->target = target;
 970         /*
 971          * If the current CPU state is in the range of the AP hotplug thread,
 972          * then we need to kick the thread once more.
 973          */
 974         if (st->state > CPUHP_BRINGUP_CPU) {
 975                 ret = cpuhp_kick_ap_work(cpu);
 976                 /*
 977                  * The AP side has done the error rollback already. Just
 978                  * return the error code..
 979                  */
 980                 if (ret)
 981                         goto out;
 982         }
 983
 984         /*
 985          * Try to reach the target state. We max out on the BP at
 986          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
 987          * responsible for bringing it up to the target state.
 988          */
 989         target = min((int)target, CPUHP_BRINGUP_CPU);
 990         ret = cpuhp_up_callbacks(cpu, st, cpuhp_bp_states, target);
 991 out:
 992         cpu_hotplug_done();
 993         return ret;
 994 }
 995
 996 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
 997 {
 998         int err = 0;
 999
1000         if (!cpu_possible(cpu)) {
1001                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1002                        cpu);
1003 #if defined(CONFIG_IA64)
1004                 pr_err("please check additional_cpus= boot parameter\n");
1005 #endif
1006                 return -EINVAL;
1007         }
1008
1009         err = try_online_node(cpu_to_node(cpu));
1010         if (err)
1011                 return err;
1012
1013         cpu_maps_update_begin();
1014
1015         if (cpu_hotplug_disabled) {
1016                 err = -EBUSY;
1017                 goto out;
1018         }
1019
1020         err = _cpu_up(cpu, 0, target);
1021 out:
1022         cpu_maps_update_done();
1023         return err;
1024 }
1025
1026 int cpu_up(unsigned int cpu)
1027 {
1028         return do_cpu_up(cpu, CPUHP_ONLINE);
1029 }
1030 EXPORT_SYMBOL_GPL(cpu_up);
1031
1032 #ifdef CONFIG_PM_SLEEP_SMP
1033 static cpumask_var_t frozen_cpus;
1034
1035 int disable_nonboot_cpus(void)
1036 {
1037         int cpu, first_cpu, error = 0;
1038
1039         cpu_maps_update_begin();
1040         first_cpu = cpumask_first(cpu_online_mask);
1041         /*
1042          * We take down all of the non-boot CPUs in one shot to avoid races
1043          * with the userspace trying to use the CPU hotplug at the same time
1044          */
1045         cpumask_clear(frozen_cpus);
1046
1047         pr_info("Disabling non-boot CPUs ...\n");
1048         for_each_online_cpu(cpu) {
1049                 if (cpu == first_cpu)
1050                         continue;
1051                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1052                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1053                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1054                 if (!error)
1055                         cpumask_set_cpu(cpu, frozen_cpus);
1056                 else {
1057                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1058                         break;
1059                 }
1060         }
1061
1062         if (!error)
1063                 BUG_ON(num_online_cpus() > 1);
1064         else
1065                 pr_err("Non-boot CPUs are not disabled\n");
1066
1067         /*
1068          * Make sure the CPUs won't be enabled by someone else. We need to do
1069          * this even in case of failure as all disable_nonboot_cpus() users are
1070          * supposed to do enable_nonboot_cpus() on the failure path.
1071          */
1072         cpu_hotplug_disabled++;
1073
1074         cpu_maps_update_done();
1075         return error;
1076 }
1077
1078 void __weak arch_enable_nonboot_cpus_begin(void)
1079 {
1080 }
1081
1082 void __weak arch_enable_nonboot_cpus_end(void)
1083 {
1084 }
1085
1086 void enable_nonboot_cpus(void)
1087 {
1088         int cpu, error;
1089
1090         /* Allow everyone to use the CPU hotplug again */
1091         cpu_maps_update_begin();
1092         WARN_ON(--cpu_hotplug_disabled < 0);
1093         if (cpumask_empty(frozen_cpus))
1094                 goto out;
1095
1096         pr_info("Enabling non-boot CPUs ...\n");
1097
1098         arch_enable_nonboot_cpus_begin();
1099
1100         for_each_cpu(cpu, frozen_cpus) {
1101                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1102                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1103                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1104                 if (!error) {
1105                         pr_info("CPU%d is up\n", cpu);
1106                         continue;
1107                 }
1108                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1109         }
1110
1111         arch_enable_nonboot_cpus_end();
1112
1113         cpumask_clear(frozen_cpus);
1114 out:
1115         cpu_maps_update_done();
1116 }
1117
1118 static int __init alloc_frozen_cpus(void)
1119 {
1120         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1121                 return -ENOMEM;
1122         return 0;
1123 }
1124 core_initcall(alloc_frozen_cpus);
1125
1126 /*
1127  * When callbacks for CPU hotplug notifications are being executed, we must
1128  * ensure that the state of the system with respect to the tasks being frozen
1129  * or not, as reported by the notification, remains unchanged *throughout the
1130  * duration* of the execution of the callbacks.
1131  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1132  *
1133  * This synchronization is implemented by mutually excluding regular CPU
1134  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1135  * Hibernate notifications.
1136  */
1137 static int
1138 cpu_hotplug_pm_callback(struct notifier_block *nb,
1139                         unsigned long action, void *ptr)
1140 {
1141         switch (action) {
1142
1143         case PM_SUSPEND_PREPARE:
1144         case PM_HIBERNATION_PREPARE:
1145                 cpu_hotplug_disable();
1146                 break;
1147
1148         case PM_POST_SUSPEND:
1149         case PM_POST_HIBERNATION:
1150                 cpu_hotplug_enable();
1151                 break;
1152
1153         default:
1154                 return NOTIFY_DONE;
1155         }
1156
1157         return NOTIFY_OK;
1158 }
1159
1160
1161 static int __init cpu_hotplug_pm_sync_init(void)
1162 {
1163         /*
1164          * cpu_hotplug_pm_callback has higher priority than x86
1165          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1166          * to disable cpu hotplug to avoid cpu hotplug race.
1167          */
1168         pm_notifier(cpu_hotplug_pm_callback, 0);
1169         return 0;
1170 }
1171 core_initcall(cpu_hotplug_pm_sync_init);
1172
1173 #endif /* CONFIG_PM_SLEEP_SMP */
1174
1175 #endif /* CONFIG_SMP */
1176
1177 /* Boot processor state steps */
1178 static struct cpuhp_step cpuhp_bp_states[] = {
1179         [CPUHP_OFFLINE] = {
1180                 .name                   = "offline",
1181                 .startup                = NULL,
1182                 .teardown               = NULL,
1183         },
1184 #ifdef CONFIG_SMP
1185         [CPUHP_CREATE_THREADS]= {
1186                 .name                   = "threads:create",
1187                 .startup                = smpboot_create_threads,
1188                 .teardown               = NULL,
1189                 .cant_stop              = true,
1190         },
1191         /*
1192          * Preparatory and dead notifiers. Will be replaced once the notifiers
1193          * are converted to states.
1194          */
1195         [CPUHP_NOTIFY_PREPARE] = {
1196                 .name                   = "notify:prepare",
1197                 .startup                = notify_prepare,
1198                 .teardown               = notify_dead,
1199                 .skip_onerr             = true,
1200                 .cant_stop              = true,
1201         },
1202         /* Kicks the plugged cpu into life */
1203         [CPUHP_BRINGUP_CPU] = {
1204                 .name                   = "cpu:bringup",
1205                 .startup                = bringup_cpu,
1206                 .teardown               = NULL,
1207                 .cant_stop              = true,
1208         },
1209         /*
1210          * Handled on controll processor until the plugged processor manages
1211          * this itself.
1212          */
1213         [CPUHP_TEARDOWN_CPU] = {
1214                 .name                   = "cpu:teardown",
1215                 .startup                = NULL,
1216                 .teardown               = takedown_cpu,
1217                 .cant_stop              = true,
1218         },
1219 #endif
1220 };
1221
1222 /* Application processor state steps */
1223 static struct cpuhp_step cpuhp_ap_states[] = {
1224 #ifdef CONFIG_SMP
1225         /* Final state before CPU kills itself */
1226         [CPUHP_AP_IDLE_DEAD] = {
1227                 .name                   = "idle:dead",
1228         },
1229         /*
1230          * Last state before CPU enters the idle loop to die. Transient state
1231          * for synchronization.
1232          */
1233         [CPUHP_AP_OFFLINE] = {
1234                 .name                   = "ap:offline",
1235                 .cant_stop              = true,
1236         },
1237         /* First state is scheduler control. Interrupts are disabled */
1238         [CPUHP_AP_SCHED_STARTING] = {
1239                 .name                   = "sched:starting",
1240                 .startup                = sched_cpu_starting,
1241                 .teardown               = NULL,
1242         },
1243         /*
1244          * Low level startup/teardown notifiers. Run with interrupts
1245          * disabled. Will be removed once the notifiers are converted to
1246          * states.
1247          */
1248         [CPUHP_AP_NOTIFY_STARTING] = {
1249                 .name                   = "notify:starting",
1250                 .startup                = notify_starting,
1251                 .teardown               = notify_dying,
1252                 .skip_onerr             = true,
1253                 .cant_stop              = true,
1254         },
1255         /* Entry state on starting. Interrupts enabled from here on. Transient
1256          * state for synchronsization */
1257         [CPUHP_AP_ONLINE] = {
1258                 .name                   = "ap:online",
1259         },
1260         /* First state is scheduler control. Interrupts are enabled */
1261         [CPUHP_AP_ACTIVE] = {
1262                 .name                   = "sched:active",
1263                 .startup                = sched_cpu_activate,
1264                 .teardown               = sched_cpu_deactivate,
1265         },
1266         /* Handle smpboot threads park/unpark */
1267         [CPUHP_AP_SMPBOOT_THREADS] = {
1268                 .name                   = "smpboot:threads",
1269                 .startup                = smpboot_unpark_threads,
1270                 .teardown               = NULL,
1271         },
1272         /*
1273          * Online/down_prepare notifiers. Will be removed once the notifiers
1274          * are converted to states.
1275          */
1276         [CPUHP_AP_NOTIFY_ONLINE] = {
1277                 .name                   = "notify:online",
1278                 .startup                = notify_online,
1279                 .teardown               = notify_down_prepare,
1280                 .skip_onerr             = true,
1281         },
1282 #endif
1283         /*
1284          * The dynamically registered state space is here
1285          */
1286
1287         /* CPU is fully up and running. */
1288         [CPUHP_ONLINE] = {
1289                 .name                   = "online",
1290                 .startup                = NULL,
1291                 .teardown               = NULL,
1292         },
1293 };
1294
1295 /* Sanity check for callbacks */
1296 static int cpuhp_cb_check(enum cpuhp_state state)
1297 {
1298         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1299                 return -EINVAL;
1300         return 0;
1301 }
1302
1303 static bool cpuhp_is_ap_state(enum cpuhp_state state)
1304 {
1305         /*
1306          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
1307          * purposes as that state is handled explicitely in cpu_down.
1308          */
1309         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
1310 }
1311
1312 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
1313 {
1314         struct cpuhp_step *sp;
1315
1316         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
1317         return sp + state;
1318 }
1319
1320 static void cpuhp_store_callbacks(enum cpuhp_state state,
1321                                   const char *name,
1322                                   int (*startup)(unsigned int cpu),
1323                                   int (*teardown)(unsigned int cpu))
1324 {
1325         /* (Un)Install the callbacks for further cpu hotplug operations */
1326         struct cpuhp_step *sp;
1327
1328         mutex_lock(&cpuhp_state_mutex);
1329         sp = cpuhp_get_step(state);
1330         sp->startup = startup;
1331         sp->teardown = teardown;
1332         sp->name = name;
1333         mutex_unlock(&cpuhp_state_mutex);
1334 }
1335
1336 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1337 {
1338         return cpuhp_get_step(state)->teardown;
1339 }
1340
1341 /*
1342  * Call the startup/teardown function for a step either on the AP or
1343  * on the current CPU.
1344  */
1345 static int cpuhp_issue_call(int cpu, enum cpuhp_state state,
1346                             int (*cb)(unsigned int), bool bringup)
1347 {
1348         int ret;
1349
1350         if (!cb)
1351                 return 0;
1352         /*
1353          * The non AP bound callbacks can fail on bringup. On teardown
1354          * e.g. module removal we crash for now.
1355          */
1356 #ifdef CONFIG_SMP
1357         if (cpuhp_is_ap_state(state))
1358                 ret = cpuhp_invoke_ap_callback(cpu, state, cb);
1359         else
1360                 ret = cpuhp_invoke_callback(cpu, state, cb);
1361 #else
1362         ret = cpuhp_invoke_callback(cpu, state, cb);
1363 #endif
1364         BUG_ON(ret && !bringup);
1365         return ret;
1366 }
1367
1368 /*
1369  * Called from __cpuhp_setup_state on a recoverable failure.
1370  *
1371  * Note: The teardown callbacks for rollback are not allowed to fail!
1372  */
1373 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1374                                    int (*teardown)(unsigned int cpu))
1375 {
1376         int cpu;
1377
1378         if (!teardown)
1379                 return;
1380
1381         /* Roll back the already executed steps on the other cpus */
1382         for_each_present_cpu(cpu) {
1383                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1384                 int cpustate = st->state;
1385
1386                 if (cpu >= failedcpu)
1387                         break;
1388
1389                 /* Did we invoke the startup call on that cpu ? */
1390                 if (cpustate >= state)
1391                         cpuhp_issue_call(cpu, state, teardown, false);
1392         }
1393 }
1394
1395 /*
1396  * Returns a free for dynamic slot assignment of the Online state. The states
1397  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1398  * by having no name assigned.
1399  */
1400 static int cpuhp_reserve_state(enum cpuhp_state state)
1401 {
1402         enum cpuhp_state i;
1403
1404         mutex_lock(&cpuhp_state_mutex);
1405         for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1406                 if (cpuhp_ap_states[i].name)
1407                         continue;
1408
1409                 cpuhp_ap_states[i].name = "Reserved";
1410                 mutex_unlock(&cpuhp_state_mutex);
1411                 return i;
1412         }
1413         mutex_unlock(&cpuhp_state_mutex);
1414         WARN(1, "No more dynamic states available for CPU hotplug\n");
1415         return -ENOSPC;
1416 }
1417
1418 /**
1419  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1420  * @state:      The state to setup
1421  * @invoke:     If true, the startup function is invoked for cpus where
1422  *              cpu state >= @state
1423  * @startup:    startup callback function
1424  * @teardown:   teardown callback function
1425  *
1426  * Returns 0 if successful, otherwise a proper error code
1427  */
1428 int __cpuhp_setup_state(enum cpuhp_state state,
1429                         const char *name, bool invoke,
1430                         int (*startup)(unsigned int cpu),
1431                         int (*teardown)(unsigned int cpu))
1432 {
1433         int cpu, ret = 0;
1434         int dyn_state = 0;
1435
1436         if (cpuhp_cb_check(state) || !name)
1437                 return -EINVAL;
1438
1439         get_online_cpus();
1440
1441         /* currently assignments for the ONLINE state are possible */
1442         if (state == CPUHP_AP_ONLINE_DYN) {
1443                 dyn_state = 1;
1444                 ret = cpuhp_reserve_state(state);
1445                 if (ret < 0)
1446                         goto out;
1447                 state = ret;
1448         }
1449
1450         cpuhp_store_callbacks(state, name, startup, teardown);
1451
1452         if (!invoke || !startup)
1453                 goto out;
1454
1455         /*
1456          * Try to call the startup callback for each present cpu
1457          * depending on the hotplug state of the cpu.
1458          */
1459         for_each_present_cpu(cpu) {
1460                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1461                 int cpustate = st->state;
1462
1463                 if (cpustate < state)
1464                         continue;
1465
1466                 ret = cpuhp_issue_call(cpu, state, startup, true);
1467                 if (ret) {
1468                         cpuhp_rollback_install(cpu, state, teardown);
1469                         cpuhp_store_callbacks(state, NULL, NULL, NULL);
1470                         goto out;
1471                 }
1472         }
1473 out:
1474         put_online_cpus();
1475         if (!ret && dyn_state)
1476                 return state;
1477         return ret;
1478 }
1479 EXPORT_SYMBOL(__cpuhp_setup_state);
1480
1481 /**
1482  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1483  * @state:      The state to remove
1484  * @invoke:     If true, the teardown function is invoked for cpus where
1485  *              cpu state >= @state
1486  *
1487  * The teardown callback is currently not allowed to fail. Think
1488  * about module removal!
1489  */
1490 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1491 {
1492         int (*teardown)(unsigned int cpu) = cpuhp_get_teardown_cb(state);
1493         int cpu;
1494
1495         BUG_ON(cpuhp_cb_check(state));
1496
1497         get_online_cpus();
1498
1499         if (!invoke || !teardown)
1500                 goto remove;
1501
1502         /*
1503          * Call the teardown callback for each present cpu depending
1504          * on the hotplug state of the cpu. This function is not
1505          * allowed to fail currently!
1506          */
1507         for_each_present_cpu(cpu) {
1508                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1509                 int cpustate = st->state;
1510
1511                 if (cpustate >= state)
1512                         cpuhp_issue_call(cpu, state, teardown, false);
1513         }
1514 remove:
1515         cpuhp_store_callbacks(state, NULL, NULL, NULL);
1516         put_online_cpus();
1517 }
1518 EXPORT_SYMBOL(__cpuhp_remove_state);
1519
1520 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1521 static ssize_t show_cpuhp_state(struct device *dev,
1522                                 struct device_attribute *attr, char *buf)
1523 {
1524         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1525
1526         return sprintf(buf, "%d\n", st->state);
1527 }
1528 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1529
1530 static ssize_t write_cpuhp_target(struct device *dev,
1531                                   struct device_attribute *attr,
1532                                   const char *buf, size_t count)
1533 {
1534         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1535         struct cpuhp_step *sp;
1536         int target, ret;
1537
1538         ret = kstrtoint(buf, 10, &target);
1539         if (ret)
1540                 return ret;
1541
1542 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1543         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1544                 return -EINVAL;
1545 #else
1546         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1547                 return -EINVAL;
1548 #endif
1549
1550         ret = lock_device_hotplug_sysfs();
1551         if (ret)
1552                 return ret;
1553
1554         mutex_lock(&cpuhp_state_mutex);
1555         sp = cpuhp_get_step(target);
1556         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1557         mutex_unlock(&cpuhp_state_mutex);
1558         if (ret)
1559                 return ret;
1560
1561         if (st->state < target)
1562                 ret = do_cpu_up(dev->id, target);
1563         else
1564                 ret = do_cpu_down(dev->id, target);
1565
1566         unlock_device_hotplug();
1567         return ret ? ret : count;
1568 }
1569
1570 static ssize_t show_cpuhp_target(struct device *dev,
1571                                  struct device_attribute *attr, char *buf)
1572 {
1573         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1574
1575         return sprintf(buf, "%d\n", st->target);
1576 }
1577 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1578
1579 static struct attribute *cpuhp_cpu_attrs[] = {
1580         &dev_attr_state.attr,
1581         &dev_attr_target.attr,
1582         NULL
1583 };
1584
1585 static struct attribute_group cpuhp_cpu_attr_group = {
1586         .attrs = cpuhp_cpu_attrs,
1587         .name = "hotplug",
1588         NULL
1589 };
1590
1591 static ssize_t show_cpuhp_states(struct device *dev,
1592                                  struct device_attribute *attr, char *buf)
1593 {
1594         ssize_t cur, res = 0;
1595         int i;
1596
1597         mutex_lock(&cpuhp_state_mutex);
1598         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1599                 struct cpuhp_step *sp = cpuhp_get_step(i);
1600
1601                 if (sp->name) {
1602                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1603                         buf += cur;
1604                         res += cur;
1605                 }
1606         }
1607         mutex_unlock(&cpuhp_state_mutex);
1608         return res;
1609 }
1610 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1611
1612 static struct attribute *cpuhp_cpu_root_attrs[] = {
1613         &dev_attr_states.attr,
1614         NULL
1615 };
1616
1617 static struct attribute_group cpuhp_cpu_root_attr_group = {
1618         .attrs = cpuhp_cpu_root_attrs,
1619         .name = "hotplug",
1620         NULL
1621 };
1622
1623 static int __init cpuhp_sysfs_init(void)
1624 {
1625         int cpu, ret;
1626
1627         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1628                                  &cpuhp_cpu_root_attr_group);
1629         if (ret)
1630                 return ret;
1631
1632         for_each_possible_cpu(cpu) {
1633                 struct device *dev = get_cpu_device(cpu);
1634
1635                 if (!dev)
1636                         continue;
1637                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1638                 if (ret)
1639                         return ret;
1640         }
1641         return 0;
1642 }
1643 device_initcall(cpuhp_sysfs_init);
1644 #endif
1645
1646 /*
1647  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1648  * represents all NR_CPUS bits binary values of 1<<nr.
1649  *
1650  * It is used by cpumask_of() to get a constant address to a CPU
1651  * mask value that has a single bit set only.
1652  */
1653
1654 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1655 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1656 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1657 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1658 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1659
1660 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1661
1662         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1663         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1664 #if BITS_PER_LONG > 32
1665         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1666         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1667 #endif
1668 };
1669 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1670
1671 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1672 EXPORT_SYMBOL(cpu_all_bits);
1673
1674 #ifdef CONFIG_INIT_ALL_POSSIBLE
1675 struct cpumask __cpu_possible_mask __read_mostly
1676         = {CPU_BITS_ALL};
1677 #else
1678 struct cpumask __cpu_possible_mask __read_mostly;
1679 #endif
1680 EXPORT_SYMBOL(__cpu_possible_mask);
1681
1682 struct cpumask __cpu_online_mask __read_mostly;
1683 EXPORT_SYMBOL(__cpu_online_mask);
1684
1685 struct cpumask __cpu_present_mask __read_mostly;
1686 EXPORT_SYMBOL(__cpu_present_mask);
1687
1688 struct cpumask __cpu_active_mask __read_mostly;
1689 EXPORT_SYMBOL(__cpu_active_mask);
1690
1691 void init_cpu_present(const struct cpumask *src)
1692 {
1693         cpumask_copy(&__cpu_present_mask, src);
1694 }
1695
1696 void init_cpu_possible(const struct cpumask *src)
1697 {
1698         cpumask_copy(&__cpu_possible_mask, src);
1699 }
1700
1701 void init_cpu_online(const struct cpumask *src)
1702 {
1703         cpumask_copy(&__cpu_online_mask, src);
1704 }
1705
1706 /*
1707  * Activate the first processor.
1708  */
1709 void __init boot_cpu_init(void)
1710 {
1711         int cpu = smp_processor_id();
1712
1713         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1714         set_cpu_online(cpu, true);
1715         set_cpu_active(cpu, true);
1716         set_cpu_present(cpu, true);
1717         set_cpu_possible(cpu, true);
1718 }
1719
1720 /*
1721  * Must be called _AFTER_ setting up the per_cpu areas
1722  */
1723 void __init boot_cpu_state_init(void)
1724 {
1725         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1726 }