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  * @single:     Single callback invocation
  41  * @bringup:    Single callback bringup or teardown selector
  42  * @cb_state:   The state for a single callback (install/uninstall)
  43  * @result:     Result of the operation
  44  * @done:       Signal completion to the issuer of the task
  45  */
  46 struct cpuhp_cpu_state {
  47         enum cpuhp_state        state;
  48         enum cpuhp_state        target;
  49 #ifdef CONFIG_SMP
  50         struct task_struct      *thread;
  51         bool                    should_run;
  52         bool                    rollback;
  53         bool                    single;
  54         bool                    bringup;
  55         struct hlist_node       *node;
  56         enum cpuhp_state        cb_state;
  57         int                     result;
  58         struct completion       done;
  59 #endif
  60 };
  61
  62 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
  63
  64 /**
  65  * cpuhp_step - Hotplug state machine step
  66  * @name:       Name of the step
  67  * @startup:    Startup function of the step
  68  * @teardown:   Teardown function of the step
  69  * @skip_onerr: Do not invoke the functions on error rollback
  70  *              Will go away once the notifiers are gone
  71  * @cant_stop:  Bringup/teardown can't be stopped at this step
  72  */
  73 struct cpuhp_step {
  74         const char              *name;
  75         union {
  76                 int             (*startup)(unsigned int cpu);
  77                 int             (*startup_multi)(unsigned int cpu,
  78                                                  struct hlist_node *node);
  79         };
  80         union {
  81                 int             (*teardown)(unsigned int cpu);
  82                 int             (*teardown_multi)(unsigned int cpu,
  83                                                   struct hlist_node *node);
  84         };
  85         struct hlist_head       list;
  86         bool                    skip_onerr;
  87         bool                    cant_stop;
  88         bool                    multi_instance;
  89 };
  90
  91 static DEFINE_MUTEX(cpuhp_state_mutex);
  92 static struct cpuhp_step cpuhp_bp_states[];
  93 static struct cpuhp_step cpuhp_ap_states[];
  94
  95 static bool cpuhp_is_ap_state(enum cpuhp_state state)
  96 {
  97         /*
  98          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
  99          * purposes as that state is handled explicitly in cpu_down.
 100          */
 101         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 102 }
 103
 104 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 105 {
 106         struct cpuhp_step *sp;
 107
 108         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
 109         return sp + state;
 110 }
 111
 112 /**
 113  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 114  * @cpu:        The cpu for which the callback should be invoked
 115  * @step:       The step in the state machine
 116  * @bringup:    True if the bringup callback should be invoked
 117  *
 118  * Called from cpu hotplug and from the state register machinery.
 119  */
 120 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 121                                  bool bringup, struct hlist_node *node)
 122 {
 123         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 124         struct cpuhp_step *step = cpuhp_get_step(state);
 125         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 126         int (*cb)(unsigned int cpu);
 127         int ret, cnt;
 128
 129         if (!step->multi_instance) {
 130                 cb = bringup ? step->startup : step->teardown;
 131                 if (!cb)
 132                         return 0;
 133                 trace_cpuhp_enter(cpu, st->target, state, cb);
 134                 ret = cb(cpu);
 135                 trace_cpuhp_exit(cpu, st->state, state, ret);
 136                 return ret;
 137         }
 138         cbm = bringup ? step->startup_multi : step->teardown_multi;
 139         if (!cbm)
 140                 return 0;
 141
 142         /* Single invocation for instance add/remove */
 143         if (node) {
 144                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 145                 ret = cbm(cpu, node);
 146                 trace_cpuhp_exit(cpu, st->state, state, ret);
 147                 return ret;
 148         }
 149
 150         /* State transition. Invoke on all instances */
 151         cnt = 0;
 152         hlist_for_each(node, &step->list) {
 153                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 154                 ret = cbm(cpu, node);
 155                 trace_cpuhp_exit(cpu, st->state, state, ret);
 156                 if (ret)
 157                         goto err;
 158                 cnt++;
 159         }
 160         return 0;
 161 err:
 162         /* Rollback the instances if one failed */
 163         cbm = !bringup ? step->startup_multi : step->teardown_multi;
 164         if (!cbm)
 165                 return ret;
 166
 167         hlist_for_each(node, &step->list) {
 168                 if (!cnt--)
 169                         break;
 170                 cbm(cpu, node);
 171         }
 172         return ret;
 173 }
 174
 175 #ifdef CONFIG_SMP
 176 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 177 static DEFINE_MUTEX(cpu_add_remove_lock);
 178 bool cpuhp_tasks_frozen;
 179 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 180
 181 /*
 182  * The following two APIs (cpu_maps_update_begin/done) must be used when
 183  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 184  * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
 185  * hotplug callback (un)registration performed using __register_cpu_notifier()
 186  * or __unregister_cpu_notifier().
 187  */
 188 void cpu_maps_update_begin(void)
 189 {
 190         mutex_lock(&cpu_add_remove_lock);
 191 }
 192 EXPORT_SYMBOL(cpu_notifier_register_begin);
 193
 194 void cpu_maps_update_done(void)
 195 {
 196         mutex_unlock(&cpu_add_remove_lock);
 197 }
 198 EXPORT_SYMBOL(cpu_notifier_register_done);
 199
 200 static RAW_NOTIFIER_HEAD(cpu_chain);
 201
 202 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 203  * Should always be manipulated under cpu_add_remove_lock
 204  */
 205 static int cpu_hotplug_disabled;
 206
 207 #ifdef CONFIG_HOTPLUG_CPU
 208
 209 static struct {
 210         struct task_struct *active_writer;
 211         /* wait queue to wake up the active_writer */
 212         wait_queue_head_t wq;
 213         /* verifies that no writer will get active while readers are active */
 214         struct mutex lock;
 215         /*
 216          * Also blocks the new readers during
 217          * an ongoing cpu hotplug operation.
 218          */
 219         atomic_t refcount;
 220
 221 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 222         struct lockdep_map dep_map;
 223 #endif
 224 } cpu_hotplug = {
 225         .active_writer = NULL,
 226         .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
 227         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
 228 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 229         .dep_map = {.name = "cpu_hotplug.lock" },
 230 #endif
 231 };
 232
 233 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
 234 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
 235 #define cpuhp_lock_acquire_tryread() \
 236                                   lock_map_acquire_tryread(&cpu_hotplug.dep_map)
 237 #define cpuhp_lock_acquire()      lock_map_acquire(&cpu_hotplug.dep_map)
 238 #define cpuhp_lock_release()      lock_map_release(&cpu_hotplug.dep_map)
 239
 240
 241 void get_online_cpus(void)
 242 {
 243         might_sleep();
 244         if (cpu_hotplug.active_writer == current)
 245                 return;
 246         cpuhp_lock_acquire_read();
 247         mutex_lock(&cpu_hotplug.lock);
 248         atomic_inc(&cpu_hotplug.refcount);
 249         mutex_unlock(&cpu_hotplug.lock);
 250 }
 251 EXPORT_SYMBOL_GPL(get_online_cpus);
 252
 253 void put_online_cpus(void)
 254 {
 255         int refcount;
 256
 257         if (cpu_hotplug.active_writer == current)
 258                 return;
 259
 260         refcount = atomic_dec_return(&cpu_hotplug.refcount);
 261         if (WARN_ON(refcount < 0)) /* try to fix things up */
 262                 atomic_inc(&cpu_hotplug.refcount);
 263
 264         if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
 265                 wake_up(&cpu_hotplug.wq);
 266
 267         cpuhp_lock_release();
 268
 269 }
 270 EXPORT_SYMBOL_GPL(put_online_cpus);
 271
 272 /*
 273  * This ensures that the hotplug operation can begin only when the
 274  * refcount goes to zero.
 275  *
 276  * Note that during a cpu-hotplug operation, the new readers, if any,
 277  * will be blocked by the cpu_hotplug.lock
 278  *
 279  * Since cpu_hotplug_begin() is always called after invoking
 280  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 281  *
 282  * Note that theoretically, there is a possibility of a livelock:
 283  * - Refcount goes to zero, last reader wakes up the sleeping
 284  *   writer.
 285  * - Last reader unlocks the cpu_hotplug.lock.
 286  * - A new reader arrives at this moment, bumps up the refcount.
 287  * - The writer acquires the cpu_hotplug.lock finds the refcount
 288  *   non zero and goes to sleep again.
 289  *
 290  * However, this is very difficult to achieve in practice since
 291  * get_online_cpus() not an api which is called all that often.
 292  *
 293  */
 294 void cpu_hotplug_begin(void)
 295 {
 296         DEFINE_WAIT(wait);
 297
 298         cpu_hotplug.active_writer = current;
 299         cpuhp_lock_acquire();
 300
 301         for (;;) {
 302                 mutex_lock(&cpu_hotplug.lock);
 303                 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
 304                 if (likely(!atomic_read(&cpu_hotplug.refcount)))
 305                                 break;
 306                 mutex_unlock(&cpu_hotplug.lock);
 307                 schedule();
 308         }
 309         finish_wait(&cpu_hotplug.wq, &wait);
 310 }
 311
 312 void cpu_hotplug_done(void)
 313 {
 314         cpu_hotplug.active_writer = NULL;
 315         mutex_unlock(&cpu_hotplug.lock);
 316         cpuhp_lock_release();
 317 }
 318
 319 /*
 320  * Wait for currently running CPU hotplug operations to complete (if any) and
 321  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 322  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 323  * hotplug path before performing hotplug operations. So acquiring that lock
 324  * guarantees mutual exclusion from any currently running hotplug operations.
 325  */
 326 void cpu_hotplug_disable(void)
 327 {
 328         cpu_maps_update_begin();
 329         cpu_hotplug_disabled++;
 330         cpu_maps_update_done();
 331 }
 332 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 333
 334 void cpu_hotplug_enable(void)
 335 {
 336         cpu_maps_update_begin();
 337         WARN_ON(--cpu_hotplug_disabled < 0);
 338         cpu_maps_update_done();
 339 }
 340 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 341 #endif  /* CONFIG_HOTPLUG_CPU */
 342
 343 /* Need to know about CPUs going up/down? */
 344 int register_cpu_notifier(struct notifier_block *nb)
 345 {
 346         int ret;
 347         cpu_maps_update_begin();
 348         ret = raw_notifier_chain_register(&cpu_chain, nb);
 349         cpu_maps_update_done();
 350         return ret;
 351 }
 352
 353 int __register_cpu_notifier(struct notifier_block *nb)
 354 {
 355         return raw_notifier_chain_register(&cpu_chain, nb);
 356 }
 357
 358 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
 359                         int *nr_calls)
 360 {
 361         unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
 362         void *hcpu = (void *)(long)cpu;
 363
 364         int ret;
 365
 366         ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
 367                                         nr_calls);
 368
 369         return notifier_to_errno(ret);
 370 }
 371
 372 static int cpu_notify(unsigned long val, unsigned int cpu)
 373 {
 374         return __cpu_notify(val, cpu, -1, NULL);
 375 }
 376
 377 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
 378 {
 379         BUG_ON(cpu_notify(val, cpu));
 380 }
 381
 382 /* Notifier wrappers for transitioning to state machine */
 383 static int notify_prepare(unsigned int cpu)
 384 {
 385         int nr_calls = 0;
 386         int ret;
 387
 388         ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
 389         if (ret) {
 390                 nr_calls--;
 391                 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
 392                                 __func__, cpu);
 393                 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
 394         }
 395         return ret;
 396 }
 397
 398 static int notify_online(unsigned int cpu)
 399 {
 400         cpu_notify(CPU_ONLINE, cpu);
 401         return 0;
 402 }
 403
 404 static int notify_starting(unsigned int cpu)
 405 {
 406         cpu_notify(CPU_STARTING, cpu);
 407         return 0;
 408 }
 409
 410 static int bringup_wait_for_ap(unsigned int cpu)
 411 {
 412         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 413
 414         wait_for_completion(&st->done);
 415         return st->result;
 416 }
 417
 418 static int bringup_cpu(unsigned int cpu)
 419 {
 420         struct task_struct *idle = idle_thread_get(cpu);
 421         int ret;
 422
 423         /*
 424          * Some architectures have to walk the irq descriptors to
 425          * setup the vector space for the cpu which comes online.
 426          * Prevent irq alloc/free across the bringup.
 427          */
 428         irq_lock_sparse();
 429
 430         /* Arch-specific enabling code. */
 431         ret = __cpu_up(cpu, idle);
 432         irq_unlock_sparse();
 433         if (ret) {
 434                 cpu_notify(CPU_UP_CANCELED, cpu);
 435                 return ret;
 436         }
 437         ret = bringup_wait_for_ap(cpu);
 438         BUG_ON(!cpu_online(cpu));
 439         return ret;
 440 }
 441
 442 /*
 443  * Hotplug state machine related functions
 444  */
 445 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 446 {
 447         for (st->state++; st->state < st->target; st->state++) {
 448                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 449
 450                 if (!step->skip_onerr)
 451                         cpuhp_invoke_callback(cpu, st->state, true, NULL);
 452         }
 453 }
 454
 455 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 456                                 enum cpuhp_state target)
 457 {
 458         enum cpuhp_state prev_state = st->state;
 459         int ret = 0;
 460
 461         for (; st->state > target; st->state--) {
 462                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL);
 463                 if (ret) {
 464                         st->target = prev_state;
 465                         undo_cpu_down(cpu, st);
 466                         break;
 467                 }
 468         }
 469         return ret;
 470 }
 471
 472 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 473 {
 474         for (st->state--; st->state > st->target; st->state--) {
 475                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 476
 477                 if (!step->skip_onerr)
 478                         cpuhp_invoke_callback(cpu, st->state, false, NULL);
 479         }
 480 }
 481
 482 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 483                               enum cpuhp_state target)
 484 {
 485         enum cpuhp_state prev_state = st->state;
 486         int ret = 0;
 487
 488         while (st->state < target) {
 489                 st->state++;
 490                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
 491                 if (ret) {
 492                         st->target = prev_state;
 493                         undo_cpu_up(cpu, st);
 494                         break;
 495                 }
 496         }
 497         return ret;
 498 }
 499
 500 /*
 501  * The cpu hotplug threads manage the bringup and teardown of the cpus
 502  */
 503 static void cpuhp_create(unsigned int cpu)
 504 {
 505         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 506
 507         init_completion(&st->done);
 508 }
 509
 510 static int cpuhp_should_run(unsigned int cpu)
 511 {
 512         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 513
 514         return st->should_run;
 515 }
 516
 517 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
 518 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
 519 {
 520         enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
 521
 522         return cpuhp_down_callbacks(cpu, st, target);
 523 }
 524
 525 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
 526 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
 527 {
 528         return cpuhp_up_callbacks(cpu, st, st->target);
 529 }
 530
 531 /*
 532  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 533  * callbacks when a state gets [un]installed at runtime.
 534  */
 535 static void cpuhp_thread_fun(unsigned int cpu)
 536 {
 537         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 538         int ret = 0;
 539
 540         /*
 541          * Paired with the mb() in cpuhp_kick_ap_work and
 542          * cpuhp_invoke_ap_callback, so the work set is consistent visible.
 543          */
 544         smp_mb();
 545         if (!st->should_run)
 546                 return;
 547
 548         st->should_run = false;
 549
 550         /* Single callback invocation for [un]install ? */
 551         if (st->single) {
 552                 if (st->cb_state < CPUHP_AP_ONLINE) {
 553                         local_irq_disable();
 554                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 555                                                     st->bringup, st->node);
 556                         local_irq_enable();
 557                 } else {
 558                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 559                                                     st->bringup, st->node);
 560                 }
 561         } else if (st->rollback) {
 562                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 563
 564                 undo_cpu_down(cpu, st);
 565                 /*
 566                  * This is a momentary workaround to keep the notifier users
 567                  * happy. Will go away once we got rid of the notifiers.
 568                  */
 569                 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
 570                 st->rollback = false;
 571         } else {
 572                 /* Cannot happen .... */
 573                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 574
 575                 /* Regular hotplug work */
 576                 if (st->state < st->target)
 577                         ret = cpuhp_ap_online(cpu, st);
 578                 else if (st->state > st->target)
 579                         ret = cpuhp_ap_offline(cpu, st);
 580         }
 581         st->result = ret;
 582         complete(&st->done);
 583 }
 584
 585 /* Invoke a single callback on a remote cpu */
 586 static int
 587 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 588                          struct hlist_node *node)
 589 {
 590         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 591
 592         if (!cpu_online(cpu))
 593                 return 0;
 594
 595         /*
 596          * If we are up and running, use the hotplug thread. For early calls
 597          * we invoke the thread function directly.
 598          */
 599         if (!st->thread)
 600                 return cpuhp_invoke_callback(cpu, state, bringup, node);
 601
 602         st->cb_state = state;
 603         st->single = true;
 604         st->bringup = bringup;
 605         st->node = node;
 606
 607         /*
 608          * Make sure the above stores are visible before should_run becomes
 609          * true. Paired with the mb() above in cpuhp_thread_fun()
 610          */
 611         smp_mb();
 612         st->should_run = true;
 613         wake_up_process(st->thread);
 614         wait_for_completion(&st->done);
 615         return st->result;
 616 }
 617
 618 /* Regular hotplug invocation of the AP hotplug thread */
 619 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
 620 {
 621         st->result = 0;
 622         st->single = false;
 623         /*
 624          * Make sure the above stores are visible before should_run becomes
 625          * true. Paired with the mb() above in cpuhp_thread_fun()
 626          */
 627         smp_mb();
 628         st->should_run = true;
 629         wake_up_process(st->thread);
 630 }
 631
 632 static int cpuhp_kick_ap_work(unsigned int cpu)
 633 {
 634         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 635         enum cpuhp_state state = st->state;
 636
 637         trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
 638         __cpuhp_kick_ap_work(st);
 639         wait_for_completion(&st->done);
 640         trace_cpuhp_exit(cpu, st->state, state, st->result);
 641         return st->result;
 642 }
 643
 644 static struct smp_hotplug_thread cpuhp_threads = {
 645         .store                  = &cpuhp_state.thread,
 646         .create                 = &cpuhp_create,
 647         .thread_should_run      = cpuhp_should_run,
 648         .thread_fn              = cpuhp_thread_fun,
 649         .thread_comm            = "cpuhp/%u",
 650         .selfparking            = true,
 651 };
 652
 653 void __init cpuhp_threads_init(void)
 654 {
 655         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 656         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 657 }
 658
 659 #ifdef CONFIG_HOTPLUG_CPU
 660 EXPORT_SYMBOL(register_cpu_notifier);
 661 EXPORT_SYMBOL(__register_cpu_notifier);
 662 void unregister_cpu_notifier(struct notifier_block *nb)
 663 {
 664         cpu_maps_update_begin();
 665         raw_notifier_chain_unregister(&cpu_chain, nb);
 666         cpu_maps_update_done();
 667 }
 668 EXPORT_SYMBOL(unregister_cpu_notifier);
 669
 670 void __unregister_cpu_notifier(struct notifier_block *nb)
 671 {
 672         raw_notifier_chain_unregister(&cpu_chain, nb);
 673 }
 674 EXPORT_SYMBOL(__unregister_cpu_notifier);
 675
 676 /**
 677  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 678  * @cpu: a CPU id
 679  *
 680  * This function walks all processes, finds a valid mm struct for each one and
 681  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 682  * trivial, there are various non-obvious corner cases, which this function
 683  * tries to solve in a safe manner.
 684  *
 685  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 686  * be called only for an already offlined CPU.
 687  */
 688 void clear_tasks_mm_cpumask(int cpu)
 689 {
 690         struct task_struct *p;
 691
 692         /*
 693          * This function is called after the cpu is taken down and marked
 694          * offline, so its not like new tasks will ever get this cpu set in
 695          * their mm mask. -- Peter Zijlstra
 696          * Thus, we may use rcu_read_lock() here, instead of grabbing
 697          * full-fledged tasklist_lock.
 698          */
 699         WARN_ON(cpu_online(cpu));
 700         rcu_read_lock();
 701         for_each_process(p) {
 702                 struct task_struct *t;
 703
 704                 /*
 705                  * Main thread might exit, but other threads may still have
 706                  * a valid mm. Find one.
 707                  */
 708                 t = find_lock_task_mm(p);
 709                 if (!t)
 710                         continue;
 711                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 712                 task_unlock(t);
 713         }
 714         rcu_read_unlock();
 715 }
 716
 717 static inline void check_for_tasks(int dead_cpu)
 718 {
 719         struct task_struct *g, *p;
 720
 721         read_lock(&tasklist_lock);
 722         for_each_process_thread(g, p) {
 723                 if (!p->on_rq)
 724                         continue;
 725                 /*
 726                  * We do the check with unlocked task_rq(p)->lock.
 727                  * Order the reading to do not warn about a task,
 728                  * which was running on this cpu in the past, and
 729                  * it's just been woken on another cpu.
 730                  */
 731                 rmb();
 732                 if (task_cpu(p) != dead_cpu)
 733                         continue;
 734
 735                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 736                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 737         }
 738         read_unlock(&tasklist_lock);
 739 }
 740
 741 static int notify_down_prepare(unsigned int cpu)
 742 {
 743         int err, nr_calls = 0;
 744
 745         err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
 746         if (err) {
 747                 nr_calls--;
 748                 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
 749                 pr_warn("%s: attempt to take down CPU %u failed\n",
 750                                 __func__, cpu);
 751         }
 752         return err;
 753 }
 754
 755 static int notify_dying(unsigned int cpu)
 756 {
 757         cpu_notify(CPU_DYING, cpu);
 758         return 0;
 759 }
 760
 761 /* Take this CPU down. */
 762 static int take_cpu_down(void *_param)
 763 {
 764         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 765         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 766         int err, cpu = smp_processor_id();
 767
 768         /* Ensure this CPU doesn't handle any more interrupts. */
 769         err = __cpu_disable();
 770         if (err < 0)
 771                 return err;
 772
 773         /*
 774          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 775          * do this step again.
 776          */
 777         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 778         st->state--;
 779         /* Invoke the former CPU_DYING callbacks */
 780         for (; st->state > target; st->state--)
 781                 cpuhp_invoke_callback(cpu, st->state, false, NULL);
 782
 783         /* Give up timekeeping duties */
 784         tick_handover_do_timer();
 785         /* Park the stopper thread */
 786         stop_machine_park(cpu);
 787         return 0;
 788 }
 789
 790 static int takedown_cpu(unsigned int cpu)
 791 {
 792         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 793         int err;
 794
 795         /* Park the smpboot threads */
 796         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 797         smpboot_park_threads(cpu);
 798
 799         /*
 800          * Prevent irq alloc/free while the dying cpu reorganizes the
 801          * interrupt affinities.
 802          */
 803         irq_lock_sparse();
 804
 805         /*
 806          * So now all preempt/rcu users must observe !cpu_active().
 807          */
 808         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 809         if (err) {
 810                 /* CPU refused to die */
 811                 irq_unlock_sparse();
 812                 /* Unpark the hotplug thread so we can rollback there */
 813                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 814                 return err;
 815         }
 816         BUG_ON(cpu_online(cpu));
 817
 818         /*
 819          * The migration_call() CPU_DYING callback will have removed all
 820          * runnable tasks from the cpu, there's only the idle task left now
 821          * that the migration thread is done doing the stop_machine thing.
 822          *
 823          * Wait for the stop thread to go away.
 824          */
 825         wait_for_completion(&st->done);
 826         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 827
 828         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 829         irq_unlock_sparse();
 830
 831         hotplug_cpu__broadcast_tick_pull(cpu);
 832         /* This actually kills the CPU. */
 833         __cpu_die(cpu);
 834
 835         tick_cleanup_dead_cpu(cpu);
 836         return 0;
 837 }
 838
 839 static int notify_dead(unsigned int cpu)
 840 {
 841         cpu_notify_nofail(CPU_DEAD, cpu);
 842         check_for_tasks(cpu);
 843         return 0;
 844 }
 845
 846 static void cpuhp_complete_idle_dead(void *arg)
 847 {
 848         struct cpuhp_cpu_state *st = arg;
 849
 850         complete(&st->done);
 851 }
 852
 853 void cpuhp_report_idle_dead(void)
 854 {
 855         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 856
 857         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 858         rcu_report_dead(smp_processor_id());
 859         st->state = CPUHP_AP_IDLE_DEAD;
 860         /*
 861          * We cannot call complete after rcu_report_dead() so we delegate it
 862          * to an online cpu.
 863          */
 864         smp_call_function_single(cpumask_first(cpu_online_mask),
 865                                  cpuhp_complete_idle_dead, st, 0);
 866 }
 867
 868 #else
 869 #define notify_down_prepare     NULL
 870 #define takedown_cpu            NULL
 871 #define notify_dead             NULL
 872 #define notify_dying            NULL
 873 #endif
 874
 875 #ifdef CONFIG_HOTPLUG_CPU
 876
 877 /* Requires cpu_add_remove_lock to be held */
 878 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 879                            enum cpuhp_state target)
 880 {
 881         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 882         int prev_state, ret = 0;
 883         bool hasdied = false;
 884
 885         if (num_online_cpus() == 1)
 886                 return -EBUSY;
 887
 888         if (!cpu_present(cpu))
 889                 return -EINVAL;
 890
 891         cpu_hotplug_begin();
 892
 893         cpuhp_tasks_frozen = tasks_frozen;
 894
 895         prev_state = st->state;
 896         st->target = target;
 897         /*
 898          * If the current CPU state is in the range of the AP hotplug thread,
 899          * then we need to kick the thread.
 900          */
 901         if (st->state > CPUHP_TEARDOWN_CPU) {
 902                 ret = cpuhp_kick_ap_work(cpu);
 903                 /*
 904                  * The AP side has done the error rollback already. Just
 905                  * return the error code..
 906                  */
 907                 if (ret)
 908                         goto out;
 909
 910                 /*
 911                  * We might have stopped still in the range of the AP hotplug
 912                  * thread. Nothing to do anymore.
 913                  */
 914                 if (st->state > CPUHP_TEARDOWN_CPU)
 915                         goto out;
 916         }
 917         /*
 918          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 919          * to do the further cleanups.
 920          */
 921         ret = cpuhp_down_callbacks(cpu, st, target);
 922         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 923                 st->target = prev_state;
 924                 st->rollback = true;
 925                 cpuhp_kick_ap_work(cpu);
 926         }
 927
 928         hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
 929 out:
 930         cpu_hotplug_done();
 931         /* This post dead nonsense must die */
 932         if (!ret && hasdied)
 933                 cpu_notify_nofail(CPU_POST_DEAD, cpu);
 934         return ret;
 935 }
 936
 937 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 938 {
 939         int err;
 940
 941         cpu_maps_update_begin();
 942
 943         if (cpu_hotplug_disabled) {
 944                 err = -EBUSY;
 945                 goto out;
 946         }
 947
 948         err = _cpu_down(cpu, 0, target);
 949
 950 out:
 951         cpu_maps_update_done();
 952         return err;
 953 }
 954 int cpu_down(unsigned int cpu)
 955 {
 956         return do_cpu_down(cpu, CPUHP_OFFLINE);
 957 }
 958 EXPORT_SYMBOL(cpu_down);
 959 #endif /*CONFIG_HOTPLUG_CPU*/
 960
 961 /**
 962  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 963  * @cpu: cpu that just started
 964  *
 965  * This function calls the cpu_chain notifiers with CPU_STARTING.
 966  * It must be called by the arch code on the new cpu, before the new cpu
 967  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 968  */
 969 void notify_cpu_starting(unsigned int cpu)
 970 {
 971         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 972         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 973
 974         while (st->state < target) {
 975                 st->state++;
 976                 cpuhp_invoke_callback(cpu, st->state, true, NULL);
 977         }
 978 }
 979
 980 /*
 981  * Called from the idle task. We need to set active here, so we can kick off
 982  * the stopper thread and unpark the smpboot threads. If the target state is
 983  * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
 984  * cpu further.
 985  */
 986 void cpuhp_online_idle(enum cpuhp_state state)
 987 {
 988         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 989         unsigned int cpu = smp_processor_id();
 990
 991         /* Happens for the boot cpu */
 992         if (state != CPUHP_AP_ONLINE_IDLE)
 993                 return;
 994
 995         st->state = CPUHP_AP_ONLINE_IDLE;
 996
 997         /* Unpark the stopper thread and the hotplug thread of this cpu */
 998         stop_machine_unpark(cpu);
 999         kthread_unpark(st->thread);
1000
1001         /* Should we go further up ? */
1002         if (st->target > CPUHP_AP_ONLINE_IDLE)
1003                 __cpuhp_kick_ap_work(st);
1004         else
1005                 complete(&st->done);
1006 }
1007
1008 /* Requires cpu_add_remove_lock to be held */
1009 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1010 {
1011         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1012         struct task_struct *idle;
1013         int ret = 0;
1014
1015         cpu_hotplug_begin();
1016
1017         if (!cpu_present(cpu)) {
1018                 ret = -EINVAL;
1019                 goto out;
1020         }
1021
1022         /*
1023          * The caller of do_cpu_up might have raced with another
1024          * caller. Ignore it for now.
1025          */
1026         if (st->state >= target)
1027                 goto out;
1028
1029         if (st->state == CPUHP_OFFLINE) {
1030                 /* Let it fail before we try to bring the cpu up */
1031                 idle = idle_thread_get(cpu);
1032                 if (IS_ERR(idle)) {
1033                         ret = PTR_ERR(idle);
1034                         goto out;
1035                 }
1036         }
1037
1038         cpuhp_tasks_frozen = tasks_frozen;
1039
1040         st->target = target;
1041         /*
1042          * If the current CPU state is in the range of the AP hotplug thread,
1043          * then we need to kick the thread once more.
1044          */
1045         if (st->state > CPUHP_BRINGUP_CPU) {
1046                 ret = cpuhp_kick_ap_work(cpu);
1047                 /*
1048                  * The AP side has done the error rollback already. Just
1049                  * return the error code..
1050                  */
1051                 if (ret)
1052                         goto out;
1053         }
1054
1055         /*
1056          * Try to reach the target state. We max out on the BP at
1057          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1058          * responsible for bringing it up to the target state.
1059          */
1060         target = min((int)target, CPUHP_BRINGUP_CPU);
1061         ret = cpuhp_up_callbacks(cpu, st, target);
1062 out:
1063         cpu_hotplug_done();
1064         return ret;
1065 }
1066
1067 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1068 {
1069         int err = 0;
1070
1071         if (!cpu_possible(cpu)) {
1072                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1073                        cpu);
1074 #if defined(CONFIG_IA64)
1075                 pr_err("please check additional_cpus= boot parameter\n");
1076 #endif
1077                 return -EINVAL;
1078         }
1079
1080         err = try_online_node(cpu_to_node(cpu));
1081         if (err)
1082                 return err;
1083
1084         cpu_maps_update_begin();
1085
1086         if (cpu_hotplug_disabled) {
1087                 err = -EBUSY;
1088                 goto out;
1089         }
1090
1091         err = _cpu_up(cpu, 0, target);
1092 out:
1093         cpu_maps_update_done();
1094         return err;
1095 }
1096
1097 int cpu_up(unsigned int cpu)
1098 {
1099         return do_cpu_up(cpu, CPUHP_ONLINE);
1100 }
1101 EXPORT_SYMBOL_GPL(cpu_up);
1102
1103 #ifdef CONFIG_PM_SLEEP_SMP
1104 static cpumask_var_t frozen_cpus;
1105
1106 int disable_nonboot_cpus(void)
1107 {
1108         int cpu, first_cpu, error = 0;
1109
1110         cpu_maps_update_begin();
1111         first_cpu = cpumask_first(cpu_online_mask);
1112         /*
1113          * We take down all of the non-boot CPUs in one shot to avoid races
1114          * with the userspace trying to use the CPU hotplug at the same time
1115          */
1116         cpumask_clear(frozen_cpus);
1117
1118         pr_info("Disabling non-boot CPUs ...\n");
1119         for_each_online_cpu(cpu) {
1120                 if (cpu == first_cpu)
1121                         continue;
1122                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1123                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1124                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1125                 if (!error)
1126                         cpumask_set_cpu(cpu, frozen_cpus);
1127                 else {
1128                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1129                         break;
1130                 }
1131         }
1132
1133         if (!error)
1134                 BUG_ON(num_online_cpus() > 1);
1135         else
1136                 pr_err("Non-boot CPUs are not disabled\n");
1137
1138         /*
1139          * Make sure the CPUs won't be enabled by someone else. We need to do
1140          * this even in case of failure as all disable_nonboot_cpus() users are
1141          * supposed to do enable_nonboot_cpus() on the failure path.
1142          */
1143         cpu_hotplug_disabled++;
1144
1145         cpu_maps_update_done();
1146         return error;
1147 }
1148
1149 void __weak arch_enable_nonboot_cpus_begin(void)
1150 {
1151 }
1152
1153 void __weak arch_enable_nonboot_cpus_end(void)
1154 {
1155 }
1156
1157 void enable_nonboot_cpus(void)
1158 {
1159         int cpu, error;
1160
1161         /* Allow everyone to use the CPU hotplug again */
1162         cpu_maps_update_begin();
1163         WARN_ON(--cpu_hotplug_disabled < 0);
1164         if (cpumask_empty(frozen_cpus))
1165                 goto out;
1166
1167         pr_info("Enabling non-boot CPUs ...\n");
1168
1169         arch_enable_nonboot_cpus_begin();
1170
1171         for_each_cpu(cpu, frozen_cpus) {
1172                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1173                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1174                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1175                 if (!error) {
1176                         pr_info("CPU%d is up\n", cpu);
1177                         continue;
1178                 }
1179                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1180         }
1181
1182         arch_enable_nonboot_cpus_end();
1183
1184         cpumask_clear(frozen_cpus);
1185 out:
1186         cpu_maps_update_done();
1187 }
1188
1189 static int __init alloc_frozen_cpus(void)
1190 {
1191         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1192                 return -ENOMEM;
1193         return 0;
1194 }
1195 core_initcall(alloc_frozen_cpus);
1196
1197 /*
1198  * When callbacks for CPU hotplug notifications are being executed, we must
1199  * ensure that the state of the system with respect to the tasks being frozen
1200  * or not, as reported by the notification, remains unchanged *throughout the
1201  * duration* of the execution of the callbacks.
1202  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1203  *
1204  * This synchronization is implemented by mutually excluding regular CPU
1205  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1206  * Hibernate notifications.
1207  */
1208 static int
1209 cpu_hotplug_pm_callback(struct notifier_block *nb,
1210                         unsigned long action, void *ptr)
1211 {
1212         switch (action) {
1213
1214         case PM_SUSPEND_PREPARE:
1215         case PM_HIBERNATION_PREPARE:
1216                 cpu_hotplug_disable();
1217                 break;
1218
1219         case PM_POST_SUSPEND:
1220         case PM_POST_HIBERNATION:
1221                 cpu_hotplug_enable();
1222                 break;
1223
1224         default:
1225                 return NOTIFY_DONE;
1226         }
1227
1228         return NOTIFY_OK;
1229 }
1230
1231
1232 static int __init cpu_hotplug_pm_sync_init(void)
1233 {
1234         /*
1235          * cpu_hotplug_pm_callback has higher priority than x86
1236          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1237          * to disable cpu hotplug to avoid cpu hotplug race.
1238          */
1239         pm_notifier(cpu_hotplug_pm_callback, 0);
1240         return 0;
1241 }
1242 core_initcall(cpu_hotplug_pm_sync_init);
1243
1244 #endif /* CONFIG_PM_SLEEP_SMP */
1245
1246 #endif /* CONFIG_SMP */
1247
1248 /* Boot processor state steps */
1249 static struct cpuhp_step cpuhp_bp_states[] = {
1250         [CPUHP_OFFLINE] = {
1251                 .name                   = "offline",
1252                 .startup                = NULL,
1253                 .teardown               = NULL,
1254         },
1255 #ifdef CONFIG_SMP
1256         [CPUHP_CREATE_THREADS]= {
1257                 .name                   = "threads:create",
1258                 .startup                = smpboot_create_threads,
1259                 .teardown               = NULL,
1260                 .cant_stop              = true,
1261         },
1262         [CPUHP_PERF_PREPARE] = {
1263                 .name = "perf prepare",
1264                 .startup = perf_event_init_cpu,
1265                 .teardown = perf_event_exit_cpu,
1266         },
1267         [CPUHP_WORKQUEUE_PREP] = {
1268                 .name = "workqueue prepare",
1269                 .startup = workqueue_prepare_cpu,
1270                 .teardown = NULL,
1271         },
1272         [CPUHP_HRTIMERS_PREPARE] = {
1273                 .name = "hrtimers prepare",
1274                 .startup = hrtimers_prepare_cpu,
1275                 .teardown = hrtimers_dead_cpu,
1276         },
1277         [CPUHP_SMPCFD_PREPARE] = {
1278                 .name = "SMPCFD prepare",
1279                 .startup = smpcfd_prepare_cpu,
1280                 .teardown = smpcfd_dead_cpu,
1281         },
1282         [CPUHP_RCUTREE_PREP] = {
1283                 .name = "RCU-tree prepare",
1284                 .startup = rcutree_prepare_cpu,
1285                 .teardown = rcutree_dead_cpu,
1286         },
1287         /*
1288          * Preparatory and dead notifiers. Will be replaced once the notifiers
1289          * are converted to states.
1290          */
1291         [CPUHP_NOTIFY_PREPARE] = {
1292                 .name                   = "notify:prepare",
1293                 .startup                = notify_prepare,
1294                 .teardown               = notify_dead,
1295                 .skip_onerr             = true,
1296                 .cant_stop              = true,
1297         },
1298         /*
1299          * On the tear-down path, timers_dead_cpu() must be invoked
1300          * before blk_mq_queue_reinit_notify() from notify_dead(),
1301          * otherwise a RCU stall occurs.
1302          */
1303         [CPUHP_TIMERS_DEAD] = {
1304                 .name = "timers dead",
1305                 .startup = NULL,
1306                 .teardown = timers_dead_cpu,
1307         },
1308         /* Kicks the plugged cpu into life */
1309         [CPUHP_BRINGUP_CPU] = {
1310                 .name                   = "cpu:bringup",
1311                 .startup                = bringup_cpu,
1312                 .teardown               = NULL,
1313                 .cant_stop              = true,
1314         },
1315         [CPUHP_AP_SMPCFD_DYING] = {
1316                 .startup = NULL,
1317                 .teardown = smpcfd_dying_cpu,
1318         },
1319         /*
1320          * Handled on controll processor until the plugged processor manages
1321          * this itself.
1322          */
1323         [CPUHP_TEARDOWN_CPU] = {
1324                 .name                   = "cpu:teardown",
1325                 .startup                = NULL,
1326                 .teardown               = takedown_cpu,
1327                 .cant_stop              = true,
1328         },
1329 #else
1330         [CPUHP_BRINGUP_CPU] = { },
1331 #endif
1332 };
1333
1334 /* Application processor state steps */
1335 static struct cpuhp_step cpuhp_ap_states[] = {
1336 #ifdef CONFIG_SMP
1337         /* Final state before CPU kills itself */
1338         [CPUHP_AP_IDLE_DEAD] = {
1339                 .name                   = "idle:dead",
1340         },
1341         /*
1342          * Last state before CPU enters the idle loop to die. Transient state
1343          * for synchronization.
1344          */
1345         [CPUHP_AP_OFFLINE] = {
1346                 .name                   = "ap:offline",
1347                 .cant_stop              = true,
1348         },
1349         /* First state is scheduler control. Interrupts are disabled */
1350         [CPUHP_AP_SCHED_STARTING] = {
1351                 .name                   = "sched:starting",
1352                 .startup                = sched_cpu_starting,
1353                 .teardown               = sched_cpu_dying,
1354         },
1355         [CPUHP_AP_RCUTREE_DYING] = {
1356                 .startup = NULL,
1357                 .teardown = rcutree_dying_cpu,
1358         },
1359         /*
1360          * Low level startup/teardown notifiers. Run with interrupts
1361          * disabled. Will be removed once the notifiers are converted to
1362          * states.
1363          */
1364         [CPUHP_AP_NOTIFY_STARTING] = {
1365                 .name                   = "notify:starting",
1366                 .startup                = notify_starting,
1367                 .teardown               = notify_dying,
1368                 .skip_onerr             = true,
1369                 .cant_stop              = true,
1370         },
1371         /* Entry state on starting. Interrupts enabled from here on. Transient
1372          * state for synchronsization */
1373         [CPUHP_AP_ONLINE] = {
1374                 .name                   = "ap:online",
1375         },
1376         /* Handle smpboot threads park/unpark */
1377         [CPUHP_AP_SMPBOOT_THREADS] = {
1378                 .name                   = "smpboot:threads",
1379                 .startup                = smpboot_unpark_threads,
1380                 .teardown               = NULL,
1381         },
1382         [CPUHP_AP_PERF_ONLINE] = {
1383                 .name = "perf online",
1384                 .startup = perf_event_init_cpu,
1385                 .teardown = perf_event_exit_cpu,
1386         },
1387         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1388                 .name = "workqueue online",
1389                 .startup = workqueue_online_cpu,
1390                 .teardown = workqueue_offline_cpu,
1391         },
1392         [CPUHP_AP_RCUTREE_ONLINE] = {
1393                 .name = "RCU-tree online",
1394                 .startup = rcutree_online_cpu,
1395                 .teardown = rcutree_offline_cpu,
1396         },
1397
1398         /*
1399          * Online/down_prepare notifiers. Will be removed once the notifiers
1400          * are converted to states.
1401          */
1402         [CPUHP_AP_NOTIFY_ONLINE] = {
1403                 .name                   = "notify:online",
1404                 .startup                = notify_online,
1405                 .teardown               = notify_down_prepare,
1406                 .skip_onerr             = true,
1407         },
1408 #endif
1409         /*
1410          * The dynamically registered state space is here
1411          */
1412
1413 #ifdef CONFIG_SMP
1414         /* Last state is scheduler control setting the cpu active */
1415         [CPUHP_AP_ACTIVE] = {
1416                 .name                   = "sched:active",
1417                 .startup                = sched_cpu_activate,
1418                 .teardown               = sched_cpu_deactivate,
1419         },
1420 #endif
1421
1422         /* CPU is fully up and running. */
1423         [CPUHP_ONLINE] = {
1424                 .name                   = "online",
1425                 .startup                = NULL,
1426                 .teardown               = NULL,
1427         },
1428 };
1429
1430 /* Sanity check for callbacks */
1431 static int cpuhp_cb_check(enum cpuhp_state state)
1432 {
1433         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1434                 return -EINVAL;
1435         return 0;
1436 }
1437
1438 static void cpuhp_store_callbacks(enum cpuhp_state state,
1439                                   const char *name,
1440                                   int (*startup)(unsigned int cpu),
1441                                   int (*teardown)(unsigned int cpu),
1442                                   bool multi_instance)
1443 {
1444         /* (Un)Install the callbacks for further cpu hotplug operations */
1445         struct cpuhp_step *sp;
1446
1447         mutex_lock(&cpuhp_state_mutex);
1448         sp = cpuhp_get_step(state);
1449         sp->startup = startup;
1450         sp->teardown = teardown;
1451         sp->name = name;
1452         sp->multi_instance = multi_instance;
1453         INIT_HLIST_HEAD(&sp->list);
1454         mutex_unlock(&cpuhp_state_mutex);
1455 }
1456
1457 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1458 {
1459         return cpuhp_get_step(state)->teardown;
1460 }
1461
1462 /*
1463  * Call the startup/teardown function for a step either on the AP or
1464  * on the current CPU.
1465  */
1466 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1467                             struct hlist_node *node)
1468 {
1469         struct cpuhp_step *sp = cpuhp_get_step(state);
1470         int ret;
1471
1472         if ((bringup && !sp->startup) || (!bringup && !sp->teardown))
1473                 return 0;
1474         /*
1475          * The non AP bound callbacks can fail on bringup. On teardown
1476          * e.g. module removal we crash for now.
1477          */
1478 #ifdef CONFIG_SMP
1479         if (cpuhp_is_ap_state(state))
1480                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1481         else
1482                 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1483 #else
1484         ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1485 #endif
1486         BUG_ON(ret && !bringup);
1487         return ret;
1488 }
1489
1490 /*
1491  * Called from __cpuhp_setup_state on a recoverable failure.
1492  *
1493  * Note: The teardown callbacks for rollback are not allowed to fail!
1494  */
1495 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1496                                    struct hlist_node *node)
1497 {
1498         int cpu;
1499
1500         /* Roll back the already executed steps on the other cpus */
1501         for_each_present_cpu(cpu) {
1502                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1503                 int cpustate = st->state;
1504
1505                 if (cpu >= failedcpu)
1506                         break;
1507
1508                 /* Did we invoke the startup call on that cpu ? */
1509                 if (cpustate >= state)
1510                         cpuhp_issue_call(cpu, state, false, node);
1511         }
1512 }
1513
1514 /*
1515  * Returns a free for dynamic slot assignment of the Online state. The states
1516  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1517  * by having no name assigned.
1518  */
1519 static int cpuhp_reserve_state(enum cpuhp_state state)
1520 {
1521         enum cpuhp_state i;
1522
1523         mutex_lock(&cpuhp_state_mutex);
1524         for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1525                 if (cpuhp_ap_states[i].name)
1526                         continue;
1527
1528                 cpuhp_ap_states[i].name = "Reserved";
1529                 mutex_unlock(&cpuhp_state_mutex);
1530                 return i;
1531         }
1532         mutex_unlock(&cpuhp_state_mutex);
1533         WARN(1, "No more dynamic states available for CPU hotplug\n");
1534         return -ENOSPC;
1535 }
1536
1537 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1538                                bool invoke)
1539 {
1540         struct cpuhp_step *sp;
1541         int cpu;
1542         int ret;
1543
1544         sp = cpuhp_get_step(state);
1545         if (sp->multi_instance == false)
1546                 return -EINVAL;
1547
1548         get_online_cpus();
1549
1550         if (!invoke || !sp->startup_multi)
1551                 goto add_node;
1552
1553         /*
1554          * Try to call the startup callback for each present cpu
1555          * depending on the hotplug state of the cpu.
1556          */
1557         for_each_present_cpu(cpu) {
1558                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1559                 int cpustate = st->state;
1560
1561                 if (cpustate < state)
1562                         continue;
1563
1564                 ret = cpuhp_issue_call(cpu, state, true, node);
1565                 if (ret) {
1566                         if (sp->teardown_multi)
1567                                 cpuhp_rollback_install(cpu, state, node);
1568                         goto err;
1569                 }
1570         }
1571 add_node:
1572         ret = 0;
1573         mutex_lock(&cpuhp_state_mutex);
1574         hlist_add_head(node, &sp->list);
1575         mutex_unlock(&cpuhp_state_mutex);
1576
1577 err:
1578         put_online_cpus();
1579         return ret;
1580 }
1581 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1582
1583 /**
1584  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1585  * @state:      The state to setup
1586  * @invoke:     If true, the startup function is invoked for cpus where
1587  *              cpu state >= @state
1588  * @startup:    startup callback function
1589  * @teardown:   teardown callback function
1590  *
1591  * Returns 0 if successful, otherwise a proper error code
1592  */
1593 int __cpuhp_setup_state(enum cpuhp_state state,
1594                         const char *name, bool invoke,
1595                         int (*startup)(unsigned int cpu),
1596                         int (*teardown)(unsigned int cpu),
1597                         bool multi_instance)
1598 {
1599         int cpu, ret = 0;
1600         int dyn_state = 0;
1601
1602         if (cpuhp_cb_check(state) || !name)
1603                 return -EINVAL;
1604
1605         get_online_cpus();
1606
1607         /* currently assignments for the ONLINE state are possible */
1608         if (state == CPUHP_AP_ONLINE_DYN) {
1609                 dyn_state = 1;
1610                 ret = cpuhp_reserve_state(state);
1611                 if (ret < 0)
1612                         goto out;
1613                 state = ret;
1614         }
1615
1616         cpuhp_store_callbacks(state, name, startup, teardown, multi_instance);
1617
1618         if (!invoke || !startup)
1619                 goto out;
1620
1621         /*
1622          * Try to call the startup callback for each present cpu
1623          * depending on the hotplug state of the cpu.
1624          */
1625         for_each_present_cpu(cpu) {
1626                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1627                 int cpustate = st->state;
1628
1629                 if (cpustate < state)
1630                         continue;
1631
1632                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1633                 if (ret) {
1634                         if (teardown)
1635                                 cpuhp_rollback_install(cpu, state, NULL);
1636                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1637                         goto out;
1638                 }
1639         }
1640 out:
1641         put_online_cpus();
1642         if (!ret && dyn_state)
1643                 return state;
1644         return ret;
1645 }
1646 EXPORT_SYMBOL(__cpuhp_setup_state);
1647
1648 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1649                                   struct hlist_node *node, bool invoke)
1650 {
1651         struct cpuhp_step *sp = cpuhp_get_step(state);
1652         int cpu;
1653
1654         BUG_ON(cpuhp_cb_check(state));
1655
1656         if (!sp->multi_instance)
1657                 return -EINVAL;
1658
1659         get_online_cpus();
1660         if (!invoke || !cpuhp_get_teardown_cb(state))
1661                 goto remove;
1662         /*
1663          * Call the teardown callback for each present cpu depending
1664          * on the hotplug state of the cpu. This function is not
1665          * allowed to fail currently!
1666          */
1667         for_each_present_cpu(cpu) {
1668                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1669                 int cpustate = st->state;
1670
1671                 if (cpustate >= state)
1672                         cpuhp_issue_call(cpu, state, false, node);
1673         }
1674
1675 remove:
1676         mutex_lock(&cpuhp_state_mutex);
1677         hlist_del(node);
1678         mutex_unlock(&cpuhp_state_mutex);
1679         put_online_cpus();
1680
1681         return 0;
1682 }
1683 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1684 /**
1685  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1686  * @state:      The state to remove
1687  * @invoke:     If true, the teardown function is invoked for cpus where
1688  *              cpu state >= @state
1689  *
1690  * The teardown callback is currently not allowed to fail. Think
1691  * about module removal!
1692  */
1693 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1694 {
1695         struct cpuhp_step *sp = cpuhp_get_step(state);
1696         int cpu;
1697
1698         BUG_ON(cpuhp_cb_check(state));
1699
1700         get_online_cpus();
1701
1702         if (sp->multi_instance) {
1703                 WARN(!hlist_empty(&sp->list),
1704                      "Error: Removing state %d which has instances left.\n",
1705                      state);
1706                 goto remove;
1707         }
1708
1709         if (!invoke || !cpuhp_get_teardown_cb(state))
1710                 goto remove;
1711
1712         /*
1713          * Call the teardown callback for each present cpu depending
1714          * on the hotplug state of the cpu. This function is not
1715          * allowed to fail currently!
1716          */
1717         for_each_present_cpu(cpu) {
1718                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1719                 int cpustate = st->state;
1720
1721                 if (cpustate >= state)
1722                         cpuhp_issue_call(cpu, state, false, NULL);
1723         }
1724 remove:
1725         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1726         put_online_cpus();
1727 }
1728 EXPORT_SYMBOL(__cpuhp_remove_state);
1729
1730 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1731 static ssize_t show_cpuhp_state(struct device *dev,
1732                                 struct device_attribute *attr, char *buf)
1733 {
1734         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1735
1736         return sprintf(buf, "%d\n", st->state);
1737 }
1738 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1739
1740 static ssize_t write_cpuhp_target(struct device *dev,
1741                                   struct device_attribute *attr,
1742                                   const char *buf, size_t count)
1743 {
1744         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1745         struct cpuhp_step *sp;
1746         int target, ret;
1747
1748         ret = kstrtoint(buf, 10, &target);
1749         if (ret)
1750                 return ret;
1751
1752 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1753         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1754                 return -EINVAL;
1755 #else
1756         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1757                 return -EINVAL;
1758 #endif
1759
1760         ret = lock_device_hotplug_sysfs();
1761         if (ret)
1762                 return ret;
1763
1764         mutex_lock(&cpuhp_state_mutex);
1765         sp = cpuhp_get_step(target);
1766         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1767         mutex_unlock(&cpuhp_state_mutex);
1768         if (ret)
1769                 return ret;
1770
1771         if (st->state < target)
1772                 ret = do_cpu_up(dev->id, target);
1773         else
1774                 ret = do_cpu_down(dev->id, target);
1775
1776         unlock_device_hotplug();
1777         return ret ? ret : count;
1778 }
1779
1780 static ssize_t show_cpuhp_target(struct device *dev,
1781                                  struct device_attribute *attr, char *buf)
1782 {
1783         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1784
1785         return sprintf(buf, "%d\n", st->target);
1786 }
1787 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1788
1789 static struct attribute *cpuhp_cpu_attrs[] = {
1790         &dev_attr_state.attr,
1791         &dev_attr_target.attr,
1792         NULL
1793 };
1794
1795 static struct attribute_group cpuhp_cpu_attr_group = {
1796         .attrs = cpuhp_cpu_attrs,
1797         .name = "hotplug",
1798         NULL
1799 };
1800
1801 static ssize_t show_cpuhp_states(struct device *dev,
1802                                  struct device_attribute *attr, char *buf)
1803 {
1804         ssize_t cur, res = 0;
1805         int i;
1806
1807         mutex_lock(&cpuhp_state_mutex);
1808         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1809                 struct cpuhp_step *sp = cpuhp_get_step(i);
1810
1811                 if (sp->name) {
1812                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1813                         buf += cur;
1814                         res += cur;
1815                 }
1816         }
1817         mutex_unlock(&cpuhp_state_mutex);
1818         return res;
1819 }
1820 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1821
1822 static struct attribute *cpuhp_cpu_root_attrs[] = {
1823         &dev_attr_states.attr,
1824         NULL
1825 };
1826
1827 static struct attribute_group cpuhp_cpu_root_attr_group = {
1828         .attrs = cpuhp_cpu_root_attrs,
1829         .name = "hotplug",
1830         NULL
1831 };
1832
1833 static int __init cpuhp_sysfs_init(void)
1834 {
1835         int cpu, ret;
1836
1837         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1838                                  &cpuhp_cpu_root_attr_group);
1839         if (ret)
1840                 return ret;
1841
1842         for_each_possible_cpu(cpu) {
1843                 struct device *dev = get_cpu_device(cpu);
1844
1845                 if (!dev)
1846                         continue;
1847                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1848                 if (ret)
1849                         return ret;
1850         }
1851         return 0;
1852 }
1853 device_initcall(cpuhp_sysfs_init);
1854 #endif
1855
1856 /*
1857  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1858  * represents all NR_CPUS bits binary values of 1<<nr.
1859  *
1860  * It is used by cpumask_of() to get a constant address to a CPU
1861  * mask value that has a single bit set only.
1862  */
1863
1864 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1865 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1866 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1867 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1868 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1869
1870 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1871
1872         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1873         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1874 #if BITS_PER_LONG > 32
1875         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1876         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1877 #endif
1878 };
1879 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1880
1881 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1882 EXPORT_SYMBOL(cpu_all_bits);
1883
1884 #ifdef CONFIG_INIT_ALL_POSSIBLE
1885 struct cpumask __cpu_possible_mask __read_mostly
1886         = {CPU_BITS_ALL};
1887 #else
1888 struct cpumask __cpu_possible_mask __read_mostly;
1889 #endif
1890 EXPORT_SYMBOL(__cpu_possible_mask);
1891
1892 struct cpumask __cpu_online_mask __read_mostly;
1893 EXPORT_SYMBOL(__cpu_online_mask);
1894
1895 struct cpumask __cpu_present_mask __read_mostly;
1896 EXPORT_SYMBOL(__cpu_present_mask);
1897
1898 struct cpumask __cpu_active_mask __read_mostly;
1899 EXPORT_SYMBOL(__cpu_active_mask);
1900
1901 void init_cpu_present(const struct cpumask *src)
1902 {
1903         cpumask_copy(&__cpu_present_mask, src);
1904 }
1905
1906 void init_cpu_possible(const struct cpumask *src)
1907 {
1908         cpumask_copy(&__cpu_possible_mask, src);
1909 }
1910
1911 void init_cpu_online(const struct cpumask *src)
1912 {
1913         cpumask_copy(&__cpu_online_mask, src);
1914 }
1915
1916 /*
1917  * Activate the first processor.
1918  */
1919 void __init boot_cpu_init(void)
1920 {
1921         int cpu = smp_processor_id();
1922
1923         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1924         set_cpu_online(cpu, true);
1925         set_cpu_active(cpu, true);
1926         set_cpu_present(cpu, true);
1927         set_cpu_possible(cpu, true);
1928 }
1929
1930 /*
1931  * Must be called _AFTER_ setting up the per_cpu areas
1932  */
1933 void __init boot_cpu_state_init(void)
1934 {
1935         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1936 }