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