kernel/cpu.c

   1 /* CPU control.
   2  * (C) 2001, 2002, 2003, 2004 Rusty Russell
   3  *
   4  * This code is licenced under the GPL.
   5  */
   6 #include <linux/proc_fs.h>
   7 #include <linux/smp.h>
   8 #include <linux/init.h>
   9 #include <linux/notifier.h>
  10 #include <linux/sched.h>
  11 #include <linux/unistd.h>
  12 #include <linux/cpu.h>
  13 #include <linux/oom.h>
  14 #include <linux/rcupdate.h>
  15 #include <linux/export.h>
  16 #include <linux/bug.h>
  17 #include <linux/kthread.h>
  18 #include <linux/stop_machine.h>
  19 #include <linux/mutex.h>
  20 #include <linux/gfp.h>
  21 #include <linux/suspend.h>
  22 #include <linux/lockdep.h>
  23 #include <linux/tick.h>
  24 #include <linux/irq.h>
  25 #include <linux/smpboot.h>
  26
  27 #include <trace/events/power.h>
  28 #define CREATE_TRACE_POINTS
  29 #include <trace/events/cpuhp.h>
  30
  31 #include "smpboot.h"
  32
  33 /**
  34  * cpuhp_cpu_state - Per cpu hotplug state storage
  35  * @state:      The current cpu state
  36  * @target:     The target state
  37  * @thread:     Pointer to the hotplug thread
  38  * @should_run: Thread should execute
  39  * @cb_stat:    The state for a single callback (install/uninstall)
  40  * @cb:         Single callback function (install/uninstall)
  41  * @result:     Result of the operation
  42  * @done:       Signal completion to the issuer of the task
  43  */
  44 struct cpuhp_cpu_state {
  45         enum cpuhp_state        state;
  46         enum cpuhp_state        target;
  47 #ifdef CONFIG_SMP
  48         struct task_struct      *thread;
  49         bool                    should_run;
  50         enum cpuhp_state        cb_state;
  51         int                     (*cb)(unsigned int cpu);
  52         int                     result;
  53         struct completion       done;
  54 #endif
  55 };
  56
  57 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
  58
  59 /**
  60  * cpuhp_step - Hotplug state machine step
  61  * @name:       Name of the step
  62  * @startup:    Startup function of the step
  63  * @teardown:   Teardown function of the step
  64  * @skip_onerr: Do not invoke the functions on error rollback
  65  *              Will go away once the notifiers are gone
  66  * @cant_stop:  Bringup/teardown can't be stopped at this step
  67  */
  68 struct cpuhp_step {
  69         const char      *name;
  70         int             (*startup)(unsigned int cpu);
  71         int             (*teardown)(unsigned int cpu);
  72         bool            skip_onerr;
  73         bool            cant_stop;
  74 };
  75
  76 static DEFINE_MUTEX(cpuhp_state_mutex);
  77 static struct cpuhp_step cpuhp_bp_states[];
  78 static struct cpuhp_step cpuhp_ap_states[];
  79
  80 /**
  81  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
  82  * @cpu:        The cpu for which the callback should be invoked
  83  * @step:       The step in the state machine
  84  * @cb:         The callback function to invoke
  85  *
  86  * Called from cpu hotplug and from the state register machinery
  87  */
  88 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state step,
  89                                  int (*cb)(unsigned int))
  90 {
  91         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
  92         int ret = 0;
  93
  94         if (cb) {
  95                 trace_cpuhp_enter(cpu, st->target, step, cb);
  96                 ret = cb(cpu);
  97                 trace_cpuhp_exit(cpu, st->state, step, ret);
  98         }
  99         return ret;
 100 }
 101
 102 #ifdef CONFIG_SMP
 103 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 104 static DEFINE_MUTEX(cpu_add_remove_lock);
 105 bool cpuhp_tasks_frozen;
 106 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 107
 108 /*
 109  * The following two APIs (cpu_maps_update_begin/done) must be used when
 110  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 111  * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
 112  * hotplug callback (un)registration performed using __register_cpu_notifier()
 113  * or __unregister_cpu_notifier().
 114  */
 115 void cpu_maps_update_begin(void)
 116 {
 117         mutex_lock(&cpu_add_remove_lock);
 118 }
 119 EXPORT_SYMBOL(cpu_notifier_register_begin);
 120
 121 void cpu_maps_update_done(void)
 122 {
 123         mutex_unlock(&cpu_add_remove_lock);
 124 }
 125 EXPORT_SYMBOL(cpu_notifier_register_done);
 126
 127 static RAW_NOTIFIER_HEAD(cpu_chain);
 128
 129 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 130  * Should always be manipulated under cpu_add_remove_lock
 131  */
 132 static int cpu_hotplug_disabled;
 133
 134 #ifdef CONFIG_HOTPLUG_CPU
 135
 136 static struct {
 137         struct task_struct *active_writer;
 138         /* wait queue to wake up the active_writer */
 139         wait_queue_head_t wq;
 140         /* verifies that no writer will get active while readers are active */
 141         struct mutex lock;
 142         /*
 143          * Also blocks the new readers during
 144          * an ongoing cpu hotplug operation.
 145          */
 146         atomic_t refcount;
 147
 148 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 149         struct lockdep_map dep_map;
 150 #endif
 151 } cpu_hotplug = {
 152         .active_writer = NULL,
 153         .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
 154         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
 155 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 156         .dep_map = {.name = "cpu_hotplug.lock" },
 157 #endif
 158 };
 159
 160 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
 161 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
 162 #define cpuhp_lock_acquire_tryread() \
 163                                   lock_map_acquire_tryread(&cpu_hotplug.dep_map)
 164 #define cpuhp_lock_acquire()      lock_map_acquire(&cpu_hotplug.dep_map)
 165 #define cpuhp_lock_release()      lock_map_release(&cpu_hotplug.dep_map)
 166
 167
 168 void get_online_cpus(void)
 169 {
 170         might_sleep();
 171         if (cpu_hotplug.active_writer == current)
 172                 return;
 173         cpuhp_lock_acquire_read();
 174         mutex_lock(&cpu_hotplug.lock);
 175         atomic_inc(&cpu_hotplug.refcount);
 176         mutex_unlock(&cpu_hotplug.lock);
 177 }
 178 EXPORT_SYMBOL_GPL(get_online_cpus);
 179
 180 void put_online_cpus(void)
 181 {
 182         int refcount;
 183
 184         if (cpu_hotplug.active_writer == current)
 185                 return;
 186
 187         refcount = atomic_dec_return(&cpu_hotplug.refcount);
 188         if (WARN_ON(refcount < 0)) /* try to fix things up */
 189                 atomic_inc(&cpu_hotplug.refcount);
 190
 191         if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
 192                 wake_up(&cpu_hotplug.wq);
 193
 194         cpuhp_lock_release();
 195
 196 }
 197 EXPORT_SYMBOL_GPL(put_online_cpus);
 198
 199 /*
 200  * This ensures that the hotplug operation can begin only when the
 201  * refcount goes to zero.
 202  *
 203  * Note that during a cpu-hotplug operation, the new readers, if any,
 204  * will be blocked by the cpu_hotplug.lock
 205  *
 206  * Since cpu_hotplug_begin() is always called after invoking
 207  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 208  *
 209  * Note that theoretically, there is a possibility of a livelock:
 210  * - Refcount goes to zero, last reader wakes up the sleeping
 211  *   writer.
 212  * - Last reader unlocks the cpu_hotplug.lock.
 213  * - A new reader arrives at this moment, bumps up the refcount.
 214  * - The writer acquires the cpu_hotplug.lock finds the refcount
 215  *   non zero and goes to sleep again.
 216  *
 217  * However, this is very difficult to achieve in practice since
 218  * get_online_cpus() not an api which is called all that often.
 219  *
 220  */
 221 void cpu_hotplug_begin(void)
 222 {
 223         DEFINE_WAIT(wait);
 224
 225         cpu_hotplug.active_writer = current;
 226         cpuhp_lock_acquire();
 227
 228         for (;;) {
 229                 mutex_lock(&cpu_hotplug.lock);
 230                 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
 231                 if (likely(!atomic_read(&cpu_hotplug.refcount)))
 232                                 break;
 233                 mutex_unlock(&cpu_hotplug.lock);
 234                 schedule();
 235         }
 236         finish_wait(&cpu_hotplug.wq, &wait);
 237 }
 238
 239 void cpu_hotplug_done(void)
 240 {
 241         cpu_hotplug.active_writer = NULL;
 242         mutex_unlock(&cpu_hotplug.lock);
 243         cpuhp_lock_release();
 244 }
 245
 246 /*
 247  * Wait for currently running CPU hotplug operations to complete (if any) and
 248  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 249  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 250  * hotplug path before performing hotplug operations. So acquiring that lock
 251  * guarantees mutual exclusion from any currently running hotplug operations.
 252  */
 253 void cpu_hotplug_disable(void)
 254 {
 255         cpu_maps_update_begin();
 256         cpu_hotplug_disabled++;
 257         cpu_maps_update_done();
 258 }
 259 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 260
 261 void cpu_hotplug_enable(void)
 262 {
 263         cpu_maps_update_begin();
 264         WARN_ON(--cpu_hotplug_disabled < 0);
 265         cpu_maps_update_done();
 266 }
 267 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 268 #endif  /* CONFIG_HOTPLUG_CPU */
 269
 270 /* Need to know about CPUs going up/down? */
 271 int register_cpu_notifier(struct notifier_block *nb)
 272 {
 273         int ret;
 274         cpu_maps_update_begin();
 275         ret = raw_notifier_chain_register(&cpu_chain, nb);
 276         cpu_maps_update_done();
 277         return ret;
 278 }
 279
 280 int __register_cpu_notifier(struct notifier_block *nb)
 281 {
 282         return raw_notifier_chain_register(&cpu_chain, nb);
 283 }
 284
 285 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
 286                         int *nr_calls)
 287 {
 288         unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
 289         void *hcpu = (void *)(long)cpu;
 290
 291         int ret;
 292
 293         ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
 294                                         nr_calls);
 295
 296         return notifier_to_errno(ret);
 297 }
 298
 299 static int cpu_notify(unsigned long val, unsigned int cpu)
 300 {
 301         return __cpu_notify(val, cpu, -1, NULL);
 302 }
 303
 304 /* Notifier wrappers for transitioning to state machine */
 305 static int notify_prepare(unsigned int cpu)
 306 {
 307         int nr_calls = 0;
 308         int ret;
 309
 310         ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
 311         if (ret) {
 312                 nr_calls--;
 313                 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
 314                                 __func__, cpu);
 315                 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
 316         }
 317         return ret;
 318 }
 319
 320 static int notify_online(unsigned int cpu)
 321 {
 322         cpu_notify(CPU_ONLINE, cpu);
 323         return 0;
 324 }
 325
 326 static int notify_starting(unsigned int cpu)
 327 {
 328         cpu_notify(CPU_STARTING, cpu);
 329         return 0;
 330 }
 331
 332 static int bringup_cpu(unsigned int cpu)
 333 {
 334         struct task_struct *idle = idle_thread_get(cpu);
 335         int ret;
 336
 337         /* Arch-specific enabling code. */
 338         ret = __cpu_up(cpu, idle);
 339         if (ret) {
 340                 cpu_notify(CPU_UP_CANCELED, cpu);
 341                 return ret;
 342         }
 343         BUG_ON(!cpu_online(cpu));
 344         return 0;
 345 }
 346
 347 /*
 348  * Hotplug state machine related functions
 349  */
 350 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st,
 351                           struct cpuhp_step *steps)
 352 {
 353         for (st->state++; st->state < st->target; st->state++) {
 354                 struct cpuhp_step *step = steps + st->state;
 355
 356                 if (!step->skip_onerr)
 357                         cpuhp_invoke_callback(cpu, st->state, step->startup);
 358         }
 359 }
 360
 361 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 362                                 struct cpuhp_step *steps, enum cpuhp_state target)
 363 {
 364         enum cpuhp_state prev_state = st->state;
 365         int ret = 0;
 366
 367         for (; st->state > target; st->state--) {
 368                 struct cpuhp_step *step = steps + st->state;
 369
 370                 ret = cpuhp_invoke_callback(cpu, st->state, step->teardown);
 371                 if (ret) {
 372                         st->target = prev_state;
 373                         undo_cpu_down(cpu, st, steps);
 374                         break;
 375                 }
 376         }
 377         return ret;
 378 }
 379
 380 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st,
 381                         struct cpuhp_step *steps)
 382 {
 383         for (st->state--; st->state > st->target; st->state--) {
 384                 struct cpuhp_step *step = steps + st->state;
 385
 386                 if (!step->skip_onerr)
 387                         cpuhp_invoke_callback(cpu, st->state, step->teardown);
 388         }
 389 }
 390
 391 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 392                               struct cpuhp_step *steps, enum cpuhp_state target)
 393 {
 394         enum cpuhp_state prev_state = st->state;
 395         int ret = 0;
 396
 397         while (st->state < target) {
 398                 struct cpuhp_step *step;
 399
 400                 st->state++;
 401                 step = steps + st->state;
 402                 ret = cpuhp_invoke_callback(cpu, st->state, step->startup);
 403                 if (ret) {
 404                         st->target = prev_state;
 405                         undo_cpu_up(cpu, st, steps);
 406                         break;
 407                 }
 408         }
 409         return ret;
 410 }
 411
 412 /*
 413  * The cpu hotplug threads manage the bringup and teardown of the cpus
 414  */
 415 static void cpuhp_create(unsigned int cpu)
 416 {
 417         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 418
 419         init_completion(&st->done);
 420 }
 421
 422 static int cpuhp_should_run(unsigned int cpu)
 423 {
 424         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 425
 426         return st->should_run;
 427 }
 428
 429 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
 430 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
 431 {
 432         enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
 433
 434         return cpuhp_down_callbacks(cpu, st, cpuhp_ap_states, target);
 435 }
 436
 437 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
 438 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
 439 {
 440         return cpuhp_up_callbacks(cpu, st, cpuhp_ap_states, st->target);
 441 }
 442
 443 /*
 444  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 445  * callbacks when a state gets [un]installed at runtime.
 446  */
 447 static void cpuhp_thread_fun(unsigned int cpu)
 448 {
 449         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 450         int ret = 0;
 451
 452         /*
 453          * Paired with the mb() in cpuhp_kick_ap_work and
 454          * cpuhp_invoke_ap_callback, so the work set is consistent visible.
 455          */
 456         smp_mb();
 457         if (!st->should_run)
 458                 return;
 459
 460         st->should_run = false;
 461
 462         /* Single callback invocation for [un]install ? */
 463         if (st->cb) {
 464                 if (st->cb_state < CPUHP_AP_ONLINE) {
 465                         local_irq_disable();
 466                         ret = cpuhp_invoke_callback(cpu, st->cb_state, st->cb);
 467                         local_irq_enable();
 468                 } else {
 469                         ret = cpuhp_invoke_callback(cpu, st->cb_state, st->cb);
 470                 }
 471         } else {
 472                 /* Cannot happen .... */
 473                 BUG_ON(st->state < CPUHP_KICK_AP_THREAD);
 474
 475                 /* Regular hotplug work */
 476                 if (st->state < st->target)
 477                         ret = cpuhp_ap_online(cpu, st);
 478                 else if (st->state > st->target)
 479                         ret = cpuhp_ap_offline(cpu, st);
 480         }
 481         st->result = ret;
 482         complete(&st->done);
 483 }
 484
 485 /* Invoke a single callback on a remote cpu */
 486 static int cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state,
 487                                     int (*cb)(unsigned int))
 488 {
 489         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 490
 491         if (!cpu_online(cpu))
 492                 return 0;
 493
 494         st->cb_state = state;
 495         st->cb = cb;
 496         /*
 497          * Make sure the above stores are visible before should_run becomes
 498          * true. Paired with the mb() above in cpuhp_thread_fun()
 499          */
 500         smp_mb();
 501         st->should_run = true;
 502         wake_up_process(st->thread);
 503         wait_for_completion(&st->done);
 504         return st->result;
 505 }
 506
 507 /* Regular hotplug invocation of the AP hotplug thread */
 508 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
 509 {
 510         st->result = 0;
 511         st->cb = NULL;
 512         /*
 513          * Make sure the above stores are visible before should_run becomes
 514          * true. Paired with the mb() above in cpuhp_thread_fun()
 515          */
 516         smp_mb();
 517         st->should_run = true;
 518         wake_up_process(st->thread);
 519 }
 520
 521 static int cpuhp_kick_ap_work(unsigned int cpu)
 522 {
 523         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 524         enum cpuhp_state state = st->state;
 525
 526         trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
 527         __cpuhp_kick_ap_work(st);
 528         wait_for_completion(&st->done);
 529         trace_cpuhp_exit(cpu, st->state, state, st->result);
 530         return st->result;
 531 }
 532
 533 static struct smp_hotplug_thread cpuhp_threads = {
 534         .store                  = &cpuhp_state.thread,
 535         .create                 = &cpuhp_create,
 536         .thread_should_run      = cpuhp_should_run,
 537         .thread_fn              = cpuhp_thread_fun,
 538         .thread_comm            = "cpuhp/%u",
 539         .selfparking            = true,
 540 };
 541
 542 void __init cpuhp_threads_init(void)
 543 {
 544         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 545         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 546 }
 547
 548 #ifdef CONFIG_HOTPLUG_CPU
 549 EXPORT_SYMBOL(register_cpu_notifier);
 550 EXPORT_SYMBOL(__register_cpu_notifier);
 551 void unregister_cpu_notifier(struct notifier_block *nb)
 552 {
 553         cpu_maps_update_begin();
 554         raw_notifier_chain_unregister(&cpu_chain, nb);
 555         cpu_maps_update_done();
 556 }
 557 EXPORT_SYMBOL(unregister_cpu_notifier);
 558
 559 void __unregister_cpu_notifier(struct notifier_block *nb)
 560 {
 561         raw_notifier_chain_unregister(&cpu_chain, nb);
 562 }
 563 EXPORT_SYMBOL(__unregister_cpu_notifier);
 564
 565 /**
 566  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 567  * @cpu: a CPU id
 568  *
 569  * This function walks all processes, finds a valid mm struct for each one and
 570  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 571  * trivial, there are various non-obvious corner cases, which this function
 572  * tries to solve in a safe manner.
 573  *
 574  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 575  * be called only for an already offlined CPU.
 576  */
 577 void clear_tasks_mm_cpumask(int cpu)
 578 {
 579         struct task_struct *p;
 580
 581         /*
 582          * This function is called after the cpu is taken down and marked
 583          * offline, so its not like new tasks will ever get this cpu set in
 584          * their mm mask. -- Peter Zijlstra
 585          * Thus, we may use rcu_read_lock() here, instead of grabbing
 586          * full-fledged tasklist_lock.
 587          */
 588         WARN_ON(cpu_online(cpu));
 589         rcu_read_lock();
 590         for_each_process(p) {
 591                 struct task_struct *t;
 592
 593                 /*
 594                  * Main thread might exit, but other threads may still have
 595                  * a valid mm. Find one.
 596                  */
 597                 t = find_lock_task_mm(p);
 598                 if (!t)
 599                         continue;
 600                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 601                 task_unlock(t);
 602         }
 603         rcu_read_unlock();
 604 }
 605
 606 static inline void check_for_tasks(int dead_cpu)
 607 {
 608         struct task_struct *g, *p;
 609
 610         read_lock(&tasklist_lock);
 611         for_each_process_thread(g, p) {
 612                 if (!p->on_rq)
 613                         continue;
 614                 /*
 615                  * We do the check with unlocked task_rq(p)->lock.
 616                  * Order the reading to do not warn about a task,
 617                  * which was running on this cpu in the past, and
 618                  * it's just been woken on another cpu.
 619                  */
 620                 rmb();
 621                 if (task_cpu(p) != dead_cpu)
 622                         continue;
 623
 624                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 625                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 626         }
 627         read_unlock(&tasklist_lock);
 628 }
 629
 630 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
 631 {
 632         BUG_ON(cpu_notify(val, cpu));
 633 }
 634
 635 static int notify_down_prepare(unsigned int cpu)
 636 {
 637         int err, nr_calls = 0;
 638
 639         err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
 640         if (err) {
 641                 nr_calls--;
 642                 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
 643                 pr_warn("%s: attempt to take down CPU %u failed\n",
 644                                 __func__, cpu);
 645         }
 646         return err;
 647 }
 648
 649 static int notify_dying(unsigned int cpu)
 650 {
 651         cpu_notify(CPU_DYING, cpu);
 652         return 0;
 653 }
 654
 655 /* Take this CPU down. */
 656 static int take_cpu_down(void *_param)
 657 {
 658         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 659         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 660         int err, cpu = smp_processor_id();
 661
 662         /* Ensure this CPU doesn't handle any more interrupts. */
 663         err = __cpu_disable();
 664         if (err < 0)
 665                 return err;
 666
 667         /* Invoke the former CPU_DYING callbacks */
 668         for (; st->state > target; st->state--) {
 669                 struct cpuhp_step *step = cpuhp_ap_states + st->state;
 670
 671                 cpuhp_invoke_callback(cpu, st->state, step->teardown);
 672         }
 673         /* Give up timekeeping duties */
 674         tick_handover_do_timer();
 675         /* Park the stopper thread */
 676         stop_machine_park(cpu);
 677         return 0;
 678 }
 679
 680 static int takedown_cpu(unsigned int cpu)
 681 {
 682         int err;
 683
 684         /*
 685          * By now we've cleared cpu_active_mask, wait for all preempt-disabled
 686          * and RCU users of this state to go away such that all new such users
 687          * will observe it.
 688          *
 689          * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
 690          * not imply sync_sched(), so wait for both.
 691          *
 692          * Do sync before park smpboot threads to take care the rcu boost case.
 693          */
 694         if (IS_ENABLED(CONFIG_PREEMPT))
 695                 synchronize_rcu_mult(call_rcu, call_rcu_sched);
 696         else
 697                 synchronize_rcu();
 698
 699         /* Park the hotplug thread */
 700         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 701
 702         /*
 703          * Prevent irq alloc/free while the dying cpu reorganizes the
 704          * interrupt affinities.
 705          */
 706         irq_lock_sparse();
 707
 708         /*
 709          * So now all preempt/rcu users must observe !cpu_active().
 710          */
 711         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 712         if (err) {
 713                 /* CPU didn't die: tell everyone.  Can't complain. */
 714                 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
 715                 irq_unlock_sparse();
 716                 return err;
 717         }
 718         BUG_ON(cpu_online(cpu));
 719
 720         /*
 721          * The migration_call() CPU_DYING callback will have removed all
 722          * runnable tasks from the cpu, there's only the idle task left now
 723          * that the migration thread is done doing the stop_machine thing.
 724          *
 725          * Wait for the stop thread to go away.
 726          */
 727         while (!per_cpu(cpu_dead_idle, cpu))
 728                 cpu_relax();
 729         smp_mb(); /* Read from cpu_dead_idle before __cpu_die(). */
 730         per_cpu(cpu_dead_idle, cpu) = false;
 731
 732         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 733         irq_unlock_sparse();
 734
 735         hotplug_cpu__broadcast_tick_pull(cpu);
 736         /* This actually kills the CPU. */
 737         __cpu_die(cpu);
 738
 739         tick_cleanup_dead_cpu(cpu);
 740         return 0;
 741 }
 742
 743 static int notify_dead(unsigned int cpu)
 744 {
 745         cpu_notify_nofail(CPU_DEAD, cpu);
 746         check_for_tasks(cpu);
 747         return 0;
 748 }
 749
 750 #else
 751 #define notify_down_prepare     NULL
 752 #define takedown_cpu            NULL
 753 #define notify_dead             NULL
 754 #define notify_dying            NULL
 755 #endif
 756
 757 #ifdef CONFIG_HOTPLUG_CPU
 758
 759 /* Requires cpu_add_remove_lock to be held */
 760 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 761                            enum cpuhp_state target)
 762 {
 763         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 764         int prev_state, ret = 0;
 765         bool hasdied = false;
 766
 767         if (num_online_cpus() == 1)
 768                 return -EBUSY;
 769
 770         if (!cpu_present(cpu))
 771                 return -EINVAL;
 772
 773         cpu_hotplug_begin();
 774
 775         cpuhp_tasks_frozen = tasks_frozen;
 776
 777         prev_state = st->state;
 778         st->target = target;
 779         /*
 780          * If the current CPU state is in the range of the AP hotplug thread,
 781          * then we need to kick the thread.
 782          */
 783         if (st->state >= CPUHP_KICK_AP_THREAD) {
 784                 ret = cpuhp_kick_ap_work(cpu);
 785                 /*
 786                  * The AP side has done the error rollback already. Just
 787                  * return the error code..
 788                  */
 789                 if (ret)
 790                         goto out;
 791
 792                 /*
 793                  * We might have stopped still in the range of the AP hotplug
 794                  * thread. Nothing to do anymore.
 795                  */
 796                 if (st->state >= CPUHP_KICK_AP_THREAD)
 797                         goto out;
 798         }
 799         /*
 800          * The AP brought itself down below CPUHP_KICK_AP_THREAD. So we need
 801          * to do the further cleanups.
 802          */
 803         ret = cpuhp_down_callbacks(cpu, st, cpuhp_bp_states, target);
 804
 805         hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
 806 out:
 807         cpu_hotplug_done();
 808         /* This post dead nonsense must die */
 809         if (!ret && hasdied)
 810                 cpu_notify_nofail(CPU_POST_DEAD, cpu);
 811         return ret;
 812 }
 813
 814 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 815 {
 816         int err;
 817
 818         cpu_maps_update_begin();
 819
 820         if (cpu_hotplug_disabled) {
 821                 err = -EBUSY;
 822                 goto out;
 823         }
 824
 825         err = _cpu_down(cpu, 0, target);
 826
 827 out:
 828         cpu_maps_update_done();
 829         return err;
 830 }
 831 int cpu_down(unsigned int cpu)
 832 {
 833         return do_cpu_down(cpu, CPUHP_OFFLINE);
 834 }
 835 EXPORT_SYMBOL(cpu_down);
 836 #endif /*CONFIG_HOTPLUG_CPU*/
 837
 838 /**
 839  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 840  * @cpu: cpu that just started
 841  *
 842  * This function calls the cpu_chain notifiers with CPU_STARTING.
 843  * It must be called by the arch code on the new cpu, before the new cpu
 844  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 845  */
 846 void notify_cpu_starting(unsigned int cpu)
 847 {
 848         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 849         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 850
 851         while (st->state < target) {
 852                 struct cpuhp_step *step;
 853
 854                 st->state++;
 855                 step = cpuhp_ap_states + st->state;
 856                 cpuhp_invoke_callback(cpu, st->state, step->startup);
 857         }
 858 }
 859
 860 /*
 861  * Called from the idle task. We need to set active here, so we can kick off
 862  * the stopper thread.
 863  */
 864 static int cpuhp_set_cpu_active(unsigned int cpu)
 865 {
 866         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 867
 868         /* The cpu is marked online, set it active now */
 869         set_cpu_active(cpu, true);
 870         /* Unpark the stopper thread and the hotplug thread */
 871         stop_machine_unpark(cpu);
 872         kthread_unpark(st->thread);
 873         return 0;
 874 }
 875
 876 /* Requires cpu_add_remove_lock to be held */
 877 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 878 {
 879         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 880         struct task_struct *idle;
 881         int ret = 0;
 882
 883         cpu_hotplug_begin();
 884
 885         if (!cpu_present(cpu)) {
 886                 ret = -EINVAL;
 887                 goto out;
 888         }
 889
 890         /*
 891          * The caller of do_cpu_up might have raced with another
 892          * caller. Ignore it for now.
 893          */
 894         if (st->state >= target)
 895                 goto out;
 896
 897         if (st->state == CPUHP_OFFLINE) {
 898                 /* Let it fail before we try to bring the cpu up */
 899                 idle = idle_thread_get(cpu);
 900                 if (IS_ERR(idle)) {
 901                         ret = PTR_ERR(idle);
 902                         goto out;
 903                 }
 904         }
 905
 906         cpuhp_tasks_frozen = tasks_frozen;
 907
 908         st->target = target;
 909         /*
 910          * If the current CPU state is in the range of the AP hotplug thread,
 911          * then we need to kick the thread once more.
 912          */
 913         if (st->state >= CPUHP_KICK_AP_THREAD) {
 914                 ret = cpuhp_kick_ap_work(cpu);
 915                 /*
 916                  * The AP side has done the error rollback already. Just
 917                  * return the error code..
 918                  */
 919                 if (ret)
 920                         goto out;
 921         }
 922
 923         /*
 924          * Try to reach the target state. We max out on the BP at
 925          * CPUHP_KICK_AP_THREAD. After that the AP hotplug thread is
 926          * responsible for bringing it up to the target state.
 927          */
 928         target = min((int)target, CPUHP_KICK_AP_THREAD);
 929         ret = cpuhp_up_callbacks(cpu, st, cpuhp_bp_states, target);
 930 out:
 931         cpu_hotplug_done();
 932         return ret;
 933 }
 934
 935 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
 936 {
 937         int err = 0;
 938
 939         if (!cpu_possible(cpu)) {
 940                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
 941                        cpu);
 942 #if defined(CONFIG_IA64)
 943                 pr_err("please check additional_cpus= boot parameter\n");
 944 #endif
 945                 return -EINVAL;
 946         }
 947
 948         err = try_online_node(cpu_to_node(cpu));
 949         if (err)
 950                 return err;
 951
 952         cpu_maps_update_begin();
 953
 954         if (cpu_hotplug_disabled) {
 955                 err = -EBUSY;
 956                 goto out;
 957         }
 958
 959         err = _cpu_up(cpu, 0, target);
 960 out:
 961         cpu_maps_update_done();
 962         return err;
 963 }
 964
 965 int cpu_up(unsigned int cpu)
 966 {
 967         return do_cpu_up(cpu, CPUHP_ONLINE);
 968 }
 969 EXPORT_SYMBOL_GPL(cpu_up);
 970
 971 #ifdef CONFIG_PM_SLEEP_SMP
 972 static cpumask_var_t frozen_cpus;
 973
 974 int disable_nonboot_cpus(void)
 975 {
 976         int cpu, first_cpu, error = 0;
 977
 978         cpu_maps_update_begin();
 979         first_cpu = cpumask_first(cpu_online_mask);
 980         /*
 981          * We take down all of the non-boot CPUs in one shot to avoid races
 982          * with the userspace trying to use the CPU hotplug at the same time
 983          */
 984         cpumask_clear(frozen_cpus);
 985
 986         pr_info("Disabling non-boot CPUs ...\n");
 987         for_each_online_cpu(cpu) {
 988                 if (cpu == first_cpu)
 989                         continue;
 990                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
 991                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
 992                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
 993                 if (!error)
 994                         cpumask_set_cpu(cpu, frozen_cpus);
 995                 else {
 996                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
 997                         break;
 998                 }
 999         }
1000
1001         if (!error)
1002                 BUG_ON(num_online_cpus() > 1);
1003         else
1004                 pr_err("Non-boot CPUs are not disabled\n");
1005
1006         /*
1007          * Make sure the CPUs won't be enabled by someone else. We need to do
1008          * this even in case of failure as all disable_nonboot_cpus() users are
1009          * supposed to do enable_nonboot_cpus() on the failure path.
1010          */
1011         cpu_hotplug_disabled++;
1012
1013         cpu_maps_update_done();
1014         return error;
1015 }
1016
1017 void __weak arch_enable_nonboot_cpus_begin(void)
1018 {
1019 }
1020
1021 void __weak arch_enable_nonboot_cpus_end(void)
1022 {
1023 }
1024
1025 void enable_nonboot_cpus(void)
1026 {
1027         int cpu, error;
1028
1029         /* Allow everyone to use the CPU hotplug again */
1030         cpu_maps_update_begin();
1031         WARN_ON(--cpu_hotplug_disabled < 0);
1032         if (cpumask_empty(frozen_cpus))
1033                 goto out;
1034
1035         pr_info("Enabling non-boot CPUs ...\n");
1036
1037         arch_enable_nonboot_cpus_begin();
1038
1039         for_each_cpu(cpu, frozen_cpus) {
1040                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1041                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1042                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1043                 if (!error) {
1044                         pr_info("CPU%d is up\n", cpu);
1045                         continue;
1046                 }
1047                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1048         }
1049
1050         arch_enable_nonboot_cpus_end();
1051
1052         cpumask_clear(frozen_cpus);
1053 out:
1054         cpu_maps_update_done();
1055 }
1056
1057 static int __init alloc_frozen_cpus(void)
1058 {
1059         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1060                 return -ENOMEM;
1061         return 0;
1062 }
1063 core_initcall(alloc_frozen_cpus);
1064
1065 /*
1066  * When callbacks for CPU hotplug notifications are being executed, we must
1067  * ensure that the state of the system with respect to the tasks being frozen
1068  * or not, as reported by the notification, remains unchanged *throughout the
1069  * duration* of the execution of the callbacks.
1070  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1071  *
1072  * This synchronization is implemented by mutually excluding regular CPU
1073  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1074  * Hibernate notifications.
1075  */
1076 static int
1077 cpu_hotplug_pm_callback(struct notifier_block *nb,
1078                         unsigned long action, void *ptr)
1079 {
1080         switch (action) {
1081
1082         case PM_SUSPEND_PREPARE:
1083         case PM_HIBERNATION_PREPARE:
1084                 cpu_hotplug_disable();
1085                 break;
1086
1087         case PM_POST_SUSPEND:
1088         case PM_POST_HIBERNATION:
1089                 cpu_hotplug_enable();
1090                 break;
1091
1092         default:
1093                 return NOTIFY_DONE;
1094         }
1095
1096         return NOTIFY_OK;
1097 }
1098
1099
1100 static int __init cpu_hotplug_pm_sync_init(void)
1101 {
1102         /*
1103          * cpu_hotplug_pm_callback has higher priority than x86
1104          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1105          * to disable cpu hotplug to avoid cpu hotplug race.
1106          */
1107         pm_notifier(cpu_hotplug_pm_callback, 0);
1108         return 0;
1109 }
1110 core_initcall(cpu_hotplug_pm_sync_init);
1111
1112 #endif /* CONFIG_PM_SLEEP_SMP */
1113
1114 #endif /* CONFIG_SMP */
1115
1116 /* Boot processor state steps */
1117 static struct cpuhp_step cpuhp_bp_states[] = {
1118         [CPUHP_OFFLINE] = {
1119                 .name                   = "offline",
1120                 .startup                = NULL,
1121                 .teardown               = NULL,
1122         },
1123 #ifdef CONFIG_SMP
1124         [CPUHP_CREATE_THREADS]= {
1125                 .name                   = "threads:create",
1126                 .startup                = smpboot_create_threads,
1127                 .teardown               = NULL,
1128                 .cant_stop              = true,
1129         },
1130         [CPUHP_NOTIFY_PREPARE] = {
1131                 .name                   = "notify:prepare",
1132                 .startup                = notify_prepare,
1133                 .teardown               = notify_dead,
1134                 .skip_onerr             = true,
1135                 .cant_stop              = true,
1136         },
1137         [CPUHP_BRINGUP_CPU] = {
1138                 .name                   = "cpu:bringup",
1139                 .startup                = bringup_cpu,
1140                 .teardown               = NULL,
1141                 .cant_stop              = true,
1142         },
1143         [CPUHP_TEARDOWN_CPU] = {
1144                 .name                   = "cpu:teardown",
1145                 .startup                = NULL,
1146                 .teardown               = takedown_cpu,
1147                 .cant_stop              = true,
1148         },
1149         [CPUHP_CPU_SET_ACTIVE] = {
1150                 .name                   = "cpu:active",
1151                 .startup                = cpuhp_set_cpu_active,
1152                 .teardown               = NULL,
1153         },
1154         [CPUHP_KICK_AP_THREAD] = {
1155                 .name                   = "cpuhp:kickthread",
1156                 .startup                = cpuhp_kick_ap_work,
1157                 .teardown               = cpuhp_kick_ap_work,
1158         },
1159 #endif
1160         [CPUHP_BP_ONLINE] = {
1161                 .name                   = "online",
1162                 .startup                = NULL,
1163                 .teardown               = NULL,
1164         },
1165 };
1166
1167 /* Application processor state steps */
1168 static struct cpuhp_step cpuhp_ap_states[] = {
1169 #ifdef CONFIG_SMP
1170         [CPUHP_AP_NOTIFY_STARTING] = {
1171                 .name                   = "notify:starting",
1172                 .startup                = notify_starting,
1173                 .teardown               = notify_dying,
1174                 .skip_onerr             = true,
1175                 .cant_stop              = true,
1176         },
1177         [CPUHP_AP_SMPBOOT_THREADS] = {
1178                 .name                   = "smpboot:threads",
1179                 .startup                = smpboot_unpark_threads,
1180                 .teardown               = smpboot_park_threads,
1181         },
1182         [CPUHP_AP_NOTIFY_ONLINE] = {
1183                 .name                   = "notify:online",
1184                 .startup                = notify_online,
1185                 .teardown               = notify_down_prepare,
1186         },
1187 #endif
1188         [CPUHP_ONLINE] = {
1189                 .name                   = "online",
1190                 .startup                = NULL,
1191                 .teardown               = NULL,
1192         },
1193 };
1194
1195 /* Sanity check for callbacks */
1196 static int cpuhp_cb_check(enum cpuhp_state state)
1197 {
1198         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1199                 return -EINVAL;
1200         return 0;
1201 }
1202
1203 static bool cpuhp_is_ap_state(enum cpuhp_state state)
1204 {
1205         if (state >= CPUHP_AP_OFFLINE && state <= CPUHP_AP_ONLINE)
1206                 return true;
1207         return state > CPUHP_BP_ONLINE;
1208 }
1209
1210 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
1211 {
1212         struct cpuhp_step *sp;
1213
1214         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
1215         return sp + state;
1216 }
1217
1218 static void cpuhp_store_callbacks(enum cpuhp_state state,
1219                                   const char *name,
1220                                   int (*startup)(unsigned int cpu),
1221                                   int (*teardown)(unsigned int cpu))
1222 {
1223         /* (Un)Install the callbacks for further cpu hotplug operations */
1224         struct cpuhp_step *sp;
1225
1226         mutex_lock(&cpuhp_state_mutex);
1227         sp = cpuhp_get_step(state);
1228         sp->startup = startup;
1229         sp->teardown = teardown;
1230         sp->name = name;
1231         mutex_unlock(&cpuhp_state_mutex);
1232 }
1233
1234 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1235 {
1236         return cpuhp_get_step(state)->teardown;
1237 }
1238
1239 /*
1240  * Call the startup/teardown function for a step either on the AP or
1241  * on the current CPU.
1242  */
1243 static int cpuhp_issue_call(int cpu, enum cpuhp_state state,
1244                             int (*cb)(unsigned int), bool bringup)
1245 {
1246         int ret;
1247
1248         if (!cb)
1249                 return 0;
1250         /*
1251          * The non AP bound callbacks can fail on bringup. On teardown
1252          * e.g. module removal we crash for now.
1253          */
1254 #ifdef CONFIG_SMP
1255         if (cpuhp_is_ap_state(state))
1256                 ret = cpuhp_invoke_ap_callback(cpu, state, cb);
1257         else
1258                 ret = cpuhp_invoke_callback(cpu, state, cb);
1259 #else
1260         ret = cpuhp_invoke_callback(cpu, state, cb);
1261 #endif
1262         BUG_ON(ret && !bringup);
1263         return ret;
1264 }
1265
1266 /*
1267  * Called from __cpuhp_setup_state on a recoverable failure.
1268  *
1269  * Note: The teardown callbacks for rollback are not allowed to fail!
1270  */
1271 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1272                                    int (*teardown)(unsigned int cpu))
1273 {
1274         int cpu;
1275
1276         if (!teardown)
1277                 return;
1278
1279         /* Roll back the already executed steps on the other cpus */
1280         for_each_present_cpu(cpu) {
1281                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1282                 int cpustate = st->state;
1283
1284                 if (cpu >= failedcpu)
1285                         break;
1286
1287                 /* Did we invoke the startup call on that cpu ? */
1288                 if (cpustate >= state)
1289                         cpuhp_issue_call(cpu, state, teardown, false);
1290         }
1291 }
1292
1293 /*
1294  * Returns a free for dynamic slot assignment of the Online state. The states
1295  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1296  * by having no name assigned.
1297  */
1298 static int cpuhp_reserve_state(enum cpuhp_state state)
1299 {
1300         enum cpuhp_state i;
1301
1302         mutex_lock(&cpuhp_state_mutex);
1303         for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1304                 if (cpuhp_ap_states[i].name)
1305                         continue;
1306
1307                 cpuhp_ap_states[i].name = "Reserved";
1308                 mutex_unlock(&cpuhp_state_mutex);
1309                 return i;
1310         }
1311         mutex_unlock(&cpuhp_state_mutex);
1312         WARN(1, "No more dynamic states available for CPU hotplug\n");
1313         return -ENOSPC;
1314 }
1315
1316 /**
1317  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1318  * @state:      The state to setup
1319  * @invoke:     If true, the startup function is invoked for cpus where
1320  *              cpu state >= @state
1321  * @startup:    startup callback function
1322  * @teardown:   teardown callback function
1323  *
1324  * Returns 0 if successful, otherwise a proper error code
1325  */
1326 int __cpuhp_setup_state(enum cpuhp_state state,
1327                         const char *name, bool invoke,
1328                         int (*startup)(unsigned int cpu),
1329                         int (*teardown)(unsigned int cpu))
1330 {
1331         int cpu, ret = 0;
1332         int dyn_state = 0;
1333
1334         if (cpuhp_cb_check(state) || !name)
1335                 return -EINVAL;
1336
1337         get_online_cpus();
1338
1339         /* currently assignments for the ONLINE state are possible */
1340         if (state == CPUHP_AP_ONLINE_DYN) {
1341                 dyn_state = 1;
1342                 ret = cpuhp_reserve_state(state);
1343                 if (ret < 0)
1344                         goto out;
1345                 state = ret;
1346         }
1347
1348         cpuhp_store_callbacks(state, name, startup, teardown);
1349
1350         if (!invoke || !startup)
1351                 goto out;
1352
1353         /*
1354          * Try to call the startup callback for each present cpu
1355          * depending on the hotplug state of the cpu.
1356          */
1357         for_each_present_cpu(cpu) {
1358                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1359                 int cpustate = st->state;
1360
1361                 if (cpustate < state)
1362                         continue;
1363
1364                 ret = cpuhp_issue_call(cpu, state, startup, true);
1365                 if (ret) {
1366                         cpuhp_rollback_install(cpu, state, teardown);
1367                         cpuhp_store_callbacks(state, NULL, NULL, NULL);
1368                         goto out;
1369                 }
1370         }
1371 out:
1372         put_online_cpus();
1373         if (!ret && dyn_state)
1374                 return state;
1375         return ret;
1376 }
1377 EXPORT_SYMBOL(__cpuhp_setup_state);
1378
1379 /**
1380  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1381  * @state:      The state to remove
1382  * @invoke:     If true, the teardown function is invoked for cpus where
1383  *              cpu state >= @state
1384  *
1385  * The teardown callback is currently not allowed to fail. Think
1386  * about module removal!
1387  */
1388 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1389 {
1390         int (*teardown)(unsigned int cpu) = cpuhp_get_teardown_cb(state);
1391         int cpu;
1392
1393         BUG_ON(cpuhp_cb_check(state));
1394
1395         get_online_cpus();
1396
1397         if (!invoke || !teardown)
1398                 goto remove;
1399
1400         /*
1401          * Call the teardown callback for each present cpu depending
1402          * on the hotplug state of the cpu. This function is not
1403          * allowed to fail currently!
1404          */
1405         for_each_present_cpu(cpu) {
1406                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1407                 int cpustate = st->state;
1408
1409                 if (cpustate >= state)
1410                         cpuhp_issue_call(cpu, state, teardown, false);
1411         }
1412 remove:
1413         cpuhp_store_callbacks(state, NULL, NULL, NULL);
1414         put_online_cpus();
1415 }
1416 EXPORT_SYMBOL(__cpuhp_remove_state);
1417
1418 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1419 static ssize_t show_cpuhp_state(struct device *dev,
1420                                 struct device_attribute *attr, char *buf)
1421 {
1422         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1423
1424         return sprintf(buf, "%d\n", st->state);
1425 }
1426 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1427
1428 static ssize_t write_cpuhp_target(struct device *dev,
1429                                   struct device_attribute *attr,
1430                                   const char *buf, size_t count)
1431 {
1432         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1433         struct cpuhp_step *sp;
1434         int target, ret;
1435
1436         ret = kstrtoint(buf, 10, &target);
1437         if (ret)
1438                 return ret;
1439
1440 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1441         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1442                 return -EINVAL;
1443 #else
1444         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1445                 return -EINVAL;
1446 #endif
1447
1448         ret = lock_device_hotplug_sysfs();
1449         if (ret)
1450                 return ret;
1451
1452         mutex_lock(&cpuhp_state_mutex);
1453         sp = cpuhp_get_step(target);
1454         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1455         mutex_unlock(&cpuhp_state_mutex);
1456         if (ret)
1457                 return ret;
1458
1459         if (st->state < target)
1460                 ret = do_cpu_up(dev->id, target);
1461         else
1462                 ret = do_cpu_down(dev->id, target);
1463
1464         unlock_device_hotplug();
1465         return ret ? ret : count;
1466 }
1467
1468 static ssize_t show_cpuhp_target(struct device *dev,
1469                                  struct device_attribute *attr, char *buf)
1470 {
1471         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1472
1473         return sprintf(buf, "%d\n", st->target);
1474 }
1475 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1476
1477 static struct attribute *cpuhp_cpu_attrs[] = {
1478         &dev_attr_state.attr,
1479         &dev_attr_target.attr,
1480         NULL
1481 };
1482
1483 static struct attribute_group cpuhp_cpu_attr_group = {
1484         .attrs = cpuhp_cpu_attrs,
1485         .name = "hotplug",
1486         NULL
1487 };
1488
1489 static ssize_t show_cpuhp_states(struct device *dev,
1490                                  struct device_attribute *attr, char *buf)
1491 {
1492         ssize_t cur, res = 0;
1493         int i;
1494
1495         mutex_lock(&cpuhp_state_mutex);
1496         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1497                 struct cpuhp_step *sp = cpuhp_get_step(i);
1498
1499                 if (sp->name) {
1500                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1501                         buf += cur;
1502                         res += cur;
1503                 }
1504         }
1505         mutex_unlock(&cpuhp_state_mutex);
1506         return res;
1507 }
1508 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1509
1510 static struct attribute *cpuhp_cpu_root_attrs[] = {
1511         &dev_attr_states.attr,
1512         NULL
1513 };
1514
1515 static struct attribute_group cpuhp_cpu_root_attr_group = {
1516         .attrs = cpuhp_cpu_root_attrs,
1517         .name = "hotplug",
1518         NULL
1519 };
1520
1521 static int __init cpuhp_sysfs_init(void)
1522 {
1523         int cpu, ret;
1524
1525         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1526                                  &cpuhp_cpu_root_attr_group);
1527         if (ret)
1528                 return ret;
1529
1530         for_each_possible_cpu(cpu) {
1531                 struct device *dev = get_cpu_device(cpu);
1532
1533                 if (!dev)
1534                         continue;
1535                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1536                 if (ret)
1537                         return ret;
1538         }
1539         return 0;
1540 }
1541 device_initcall(cpuhp_sysfs_init);
1542 #endif
1543
1544 /*
1545  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1546  * represents all NR_CPUS bits binary values of 1<<nr.
1547  *
1548  * It is used by cpumask_of() to get a constant address to a CPU
1549  * mask value that has a single bit set only.
1550  */
1551
1552 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1553 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1554 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1555 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1556 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1557
1558 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1559
1560         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1561         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1562 #if BITS_PER_LONG > 32
1563         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1564         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1565 #endif
1566 };
1567 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1568
1569 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1570 EXPORT_SYMBOL(cpu_all_bits);
1571
1572 #ifdef CONFIG_INIT_ALL_POSSIBLE
1573 struct cpumask __cpu_possible_mask __read_mostly
1574         = {CPU_BITS_ALL};
1575 #else
1576 struct cpumask __cpu_possible_mask __read_mostly;
1577 #endif
1578 EXPORT_SYMBOL(__cpu_possible_mask);
1579
1580 struct cpumask __cpu_online_mask __read_mostly;
1581 EXPORT_SYMBOL(__cpu_online_mask);
1582
1583 struct cpumask __cpu_present_mask __read_mostly;
1584 EXPORT_SYMBOL(__cpu_present_mask);
1585
1586 struct cpumask __cpu_active_mask __read_mostly;
1587 EXPORT_SYMBOL(__cpu_active_mask);
1588
1589 void init_cpu_present(const struct cpumask *src)
1590 {
1591         cpumask_copy(&__cpu_present_mask, src);
1592 }
1593
1594 void init_cpu_possible(const struct cpumask *src)
1595 {
1596         cpumask_copy(&__cpu_possible_mask, src);
1597 }
1598
1599 void init_cpu_online(const struct cpumask *src)
1600 {
1601         cpumask_copy(&__cpu_online_mask, src);
1602 }
1603
1604 /*
1605  * Activate the first processor.
1606  */
1607 void __init boot_cpu_init(void)
1608 {
1609         int cpu = smp_processor_id();
1610
1611         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1612         set_cpu_online(cpu, true);
1613         set_cpu_active(cpu, true);
1614         set_cpu_present(cpu, true);
1615         set_cpu_possible(cpu, true);
1616 }
1617
1618 /*
1619  * Must be called _AFTER_ setting up the per_cpu areas
1620  */
1621 void __init boot_cpu_state_init(void)
1622 {
1623         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1624 }