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 #include <linux/relay.h>
  27 #include <linux/slab.h>
  28
  29 #include <trace/events/power.h>
  30 #define CREATE_TRACE_POINTS
  31 #include <trace/events/cpuhp.h>
  32
  33 #include "smpboot.h"
  34
  35 /**
  36  * cpuhp_cpu_state - Per cpu hotplug state storage
  37  * @state:      The current cpu state
  38  * @target:     The target state
  39  * @thread:     Pointer to the hotplug thread
  40  * @should_run: Thread should execute
  41  * @rollback:   Perform a rollback
  42  * @single:     Single callback invocation
  43  * @bringup:    Single callback bringup or teardown selector
  44  * @cb_state:   The state for a single callback (install/uninstall)
  45  * @result:     Result of the operation
  46  * @done:       Signal completion to the issuer of the task
  47  */
  48 struct cpuhp_cpu_state {
  49         enum cpuhp_state        state;
  50         enum cpuhp_state        target;
  51 #ifdef CONFIG_SMP
  52         struct task_struct      *thread;
  53         bool                    should_run;
  54         bool                    rollback;
  55         bool                    single;
  56         bool                    bringup;
  57         struct hlist_node       *node;
  58         enum cpuhp_state        cb_state;
  59         int                     result;
  60         struct completion       done;
  61 #endif
  62 };
  63
  64 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
  65
  66 /**
  67  * cpuhp_step - Hotplug state machine step
  68  * @name:       Name of the step
  69  * @startup:    Startup function of the step
  70  * @teardown:   Teardown function of the step
  71  * @skip_onerr: Do not invoke the functions on error rollback
  72  *              Will go away once the notifiers are gone
  73  * @cant_stop:  Bringup/teardown can't be stopped at this step
  74  */
  75 struct cpuhp_step {
  76         const char              *name;
  77         union {
  78                 int             (*single)(unsigned int cpu);
  79                 int             (*multi)(unsigned int cpu,
  80                                          struct hlist_node *node);
  81         } startup;
  82         union {
  83                 int             (*single)(unsigned int cpu);
  84                 int             (*multi)(unsigned int cpu,
  85                                          struct hlist_node *node);
  86         } teardown;
  87         struct hlist_head       list;
  88         bool                    skip_onerr;
  89         bool                    cant_stop;
  90         bool                    multi_instance;
  91 };
  92
  93 static DEFINE_MUTEX(cpuhp_state_mutex);
  94 static struct cpuhp_step cpuhp_bp_states[];
  95 static struct cpuhp_step cpuhp_ap_states[];
  96
  97 static bool cpuhp_is_ap_state(enum cpuhp_state state)
  98 {
  99         /*
 100          * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
 101          * purposes as that state is handled explicitly in cpu_down.
 102          */
 103         return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
 104 }
 105
 106 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
 107 {
 108         struct cpuhp_step *sp;
 109
 110         sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
 111         return sp + state;
 112 }
 113
 114 /**
 115  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
 116  * @cpu:        The cpu for which the callback should be invoked
 117  * @step:       The step in the state machine
 118  * @bringup:    True if the bringup callback should be invoked
 119  *
 120  * Called from cpu hotplug and from the state register machinery.
 121  */
 122 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state state,
 123                                  bool bringup, struct hlist_node *node)
 124 {
 125         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 126         struct cpuhp_step *step = cpuhp_get_step(state);
 127         int (*cbm)(unsigned int cpu, struct hlist_node *node);
 128         int (*cb)(unsigned int cpu);
 129         int ret, cnt;
 130
 131         if (!step->multi_instance) {
 132                 cb = bringup ? step->startup.single : step->teardown.single;
 133                 if (!cb)
 134                         return 0;
 135                 trace_cpuhp_enter(cpu, st->target, state, cb);
 136                 ret = cb(cpu);
 137                 trace_cpuhp_exit(cpu, st->state, state, ret);
 138                 return ret;
 139         }
 140         cbm = bringup ? step->startup.multi : step->teardown.multi;
 141         if (!cbm)
 142                 return 0;
 143
 144         /* Single invocation for instance add/remove */
 145         if (node) {
 146                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 147                 ret = cbm(cpu, node);
 148                 trace_cpuhp_exit(cpu, st->state, state, ret);
 149                 return ret;
 150         }
 151
 152         /* State transition. Invoke on all instances */
 153         cnt = 0;
 154         hlist_for_each(node, &step->list) {
 155                 trace_cpuhp_multi_enter(cpu, st->target, state, cbm, node);
 156                 ret = cbm(cpu, node);
 157                 trace_cpuhp_exit(cpu, st->state, state, ret);
 158                 if (ret)
 159                         goto err;
 160                 cnt++;
 161         }
 162         return 0;
 163 err:
 164         /* Rollback the instances if one failed */
 165         cbm = !bringup ? step->startup.multi : step->teardown.multi;
 166         if (!cbm)
 167                 return ret;
 168
 169         hlist_for_each(node, &step->list) {
 170                 if (!cnt--)
 171                         break;
 172                 cbm(cpu, node);
 173         }
 174         return ret;
 175 }
 176
 177 #ifdef CONFIG_SMP
 178 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
 179 static DEFINE_MUTEX(cpu_add_remove_lock);
 180 bool cpuhp_tasks_frozen;
 181 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
 182
 183 /*
 184  * The following two APIs (cpu_maps_update_begin/done) must be used when
 185  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
 186  * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
 187  * hotplug callback (un)registration performed using __register_cpu_notifier()
 188  * or __unregister_cpu_notifier().
 189  */
 190 void cpu_maps_update_begin(void)
 191 {
 192         mutex_lock(&cpu_add_remove_lock);
 193 }
 194 EXPORT_SYMBOL(cpu_notifier_register_begin);
 195
 196 void cpu_maps_update_done(void)
 197 {
 198         mutex_unlock(&cpu_add_remove_lock);
 199 }
 200 EXPORT_SYMBOL(cpu_notifier_register_done);
 201
 202 static RAW_NOTIFIER_HEAD(cpu_chain);
 203
 204 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 205  * Should always be manipulated under cpu_add_remove_lock
 206  */
 207 static int cpu_hotplug_disabled;
 208
 209 #ifdef CONFIG_HOTPLUG_CPU
 210
 211 static struct {
 212         struct task_struct *active_writer;
 213         /* wait queue to wake up the active_writer */
 214         wait_queue_head_t wq;
 215         /* verifies that no writer will get active while readers are active */
 216         struct mutex lock;
 217         /*
 218          * Also blocks the new readers during
 219          * an ongoing cpu hotplug operation.
 220          */
 221         atomic_t refcount;
 222
 223 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 224         struct lockdep_map dep_map;
 225 #endif
 226 } cpu_hotplug = {
 227         .active_writer = NULL,
 228         .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
 229         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
 230 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 231         .dep_map = STATIC_LOCKDEP_MAP_INIT("cpu_hotplug.dep_map", &cpu_hotplug.dep_map),
 232 #endif
 233 };
 234
 235 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
 236 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
 237 #define cpuhp_lock_acquire_tryread() \
 238                                   lock_map_acquire_tryread(&cpu_hotplug.dep_map)
 239 #define cpuhp_lock_acquire()      lock_map_acquire(&cpu_hotplug.dep_map)
 240 #define cpuhp_lock_release()      lock_map_release(&cpu_hotplug.dep_map)
 241
 242
 243 void get_online_cpus(void)
 244 {
 245         might_sleep();
 246         if (cpu_hotplug.active_writer == current)
 247                 return;
 248         cpuhp_lock_acquire_read();
 249         mutex_lock(&cpu_hotplug.lock);
 250         atomic_inc(&cpu_hotplug.refcount);
 251         mutex_unlock(&cpu_hotplug.lock);
 252 }
 253 EXPORT_SYMBOL_GPL(get_online_cpus);
 254
 255 void put_online_cpus(void)
 256 {
 257         int refcount;
 258
 259         if (cpu_hotplug.active_writer == current)
 260                 return;
 261
 262         refcount = atomic_dec_return(&cpu_hotplug.refcount);
 263         if (WARN_ON(refcount < 0)) /* try to fix things up */
 264                 atomic_inc(&cpu_hotplug.refcount);
 265
 266         if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
 267                 wake_up(&cpu_hotplug.wq);
 268
 269         cpuhp_lock_release();
 270
 271 }
 272 EXPORT_SYMBOL_GPL(put_online_cpus);
 273
 274 /*
 275  * This ensures that the hotplug operation can begin only when the
 276  * refcount goes to zero.
 277  *
 278  * Note that during a cpu-hotplug operation, the new readers, if any,
 279  * will be blocked by the cpu_hotplug.lock
 280  *
 281  * Since cpu_hotplug_begin() is always called after invoking
 282  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 283  *
 284  * Note that theoretically, there is a possibility of a livelock:
 285  * - Refcount goes to zero, last reader wakes up the sleeping
 286  *   writer.
 287  * - Last reader unlocks the cpu_hotplug.lock.
 288  * - A new reader arrives at this moment, bumps up the refcount.
 289  * - The writer acquires the cpu_hotplug.lock finds the refcount
 290  *   non zero and goes to sleep again.
 291  *
 292  * However, this is very difficult to achieve in practice since
 293  * get_online_cpus() not an api which is called all that often.
 294  *
 295  */
 296 void cpu_hotplug_begin(void)
 297 {
 298         DEFINE_WAIT(wait);
 299
 300         cpu_hotplug.active_writer = current;
 301         cpuhp_lock_acquire();
 302
 303         for (;;) {
 304                 mutex_lock(&cpu_hotplug.lock);
 305                 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
 306                 if (likely(!atomic_read(&cpu_hotplug.refcount)))
 307                                 break;
 308                 mutex_unlock(&cpu_hotplug.lock);
 309                 schedule();
 310         }
 311         finish_wait(&cpu_hotplug.wq, &wait);
 312 }
 313
 314 void cpu_hotplug_done(void)
 315 {
 316         cpu_hotplug.active_writer = NULL;
 317         mutex_unlock(&cpu_hotplug.lock);
 318         cpuhp_lock_release();
 319 }
 320
 321 /*
 322  * Wait for currently running CPU hotplug operations to complete (if any) and
 323  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 324  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 325  * hotplug path before performing hotplug operations. So acquiring that lock
 326  * guarantees mutual exclusion from any currently running hotplug operations.
 327  */
 328 void cpu_hotplug_disable(void)
 329 {
 330         cpu_maps_update_begin();
 331         cpu_hotplug_disabled++;
 332         cpu_maps_update_done();
 333 }
 334 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 335
 336 static void __cpu_hotplug_enable(void)
 337 {
 338         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 339                 return;
 340         cpu_hotplug_disabled--;
 341 }
 342
 343 void cpu_hotplug_enable(void)
 344 {
 345         cpu_maps_update_begin();
 346         __cpu_hotplug_enable();
 347         cpu_maps_update_done();
 348 }
 349 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 350 #endif  /* CONFIG_HOTPLUG_CPU */
 351
 352 /* Need to know about CPUs going up/down? */
 353 int register_cpu_notifier(struct notifier_block *nb)
 354 {
 355         int ret;
 356         cpu_maps_update_begin();
 357         ret = raw_notifier_chain_register(&cpu_chain, nb);
 358         cpu_maps_update_done();
 359         return ret;
 360 }
 361
 362 int __register_cpu_notifier(struct notifier_block *nb)
 363 {
 364         return raw_notifier_chain_register(&cpu_chain, nb);
 365 }
 366
 367 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
 368                         int *nr_calls)
 369 {
 370         unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
 371         void *hcpu = (void *)(long)cpu;
 372
 373         int ret;
 374
 375         ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
 376                                         nr_calls);
 377
 378         return notifier_to_errno(ret);
 379 }
 380
 381 static int cpu_notify(unsigned long val, unsigned int cpu)
 382 {
 383         return __cpu_notify(val, cpu, -1, NULL);
 384 }
 385
 386 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
 387 {
 388         BUG_ON(cpu_notify(val, cpu));
 389 }
 390
 391 /* Notifier wrappers for transitioning to state machine */
 392 static int notify_prepare(unsigned int cpu)
 393 {
 394         int nr_calls = 0;
 395         int ret;
 396
 397         ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
 398         if (ret) {
 399                 nr_calls--;
 400                 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
 401                                 __func__, cpu);
 402                 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
 403         }
 404         return ret;
 405 }
 406
 407 static int notify_online(unsigned int cpu)
 408 {
 409         cpu_notify(CPU_ONLINE, cpu);
 410         return 0;
 411 }
 412
 413 static int bringup_wait_for_ap(unsigned int cpu)
 414 {
 415         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 416
 417         wait_for_completion(&st->done);
 418         return st->result;
 419 }
 420
 421 static int bringup_cpu(unsigned int cpu)
 422 {
 423         struct task_struct *idle = idle_thread_get(cpu);
 424         int ret;
 425
 426         /*
 427          * Some architectures have to walk the irq descriptors to
 428          * setup the vector space for the cpu which comes online.
 429          * Prevent irq alloc/free across the bringup.
 430          */
 431         irq_lock_sparse();
 432
 433         /* Arch-specific enabling code. */
 434         ret = __cpu_up(cpu, idle);
 435         irq_unlock_sparse();
 436         if (ret) {
 437                 cpu_notify(CPU_UP_CANCELED, cpu);
 438                 return ret;
 439         }
 440         ret = bringup_wait_for_ap(cpu);
 441         BUG_ON(!cpu_online(cpu));
 442         return ret;
 443 }
 444
 445 /*
 446  * Hotplug state machine related functions
 447  */
 448 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 449 {
 450         for (st->state++; st->state < st->target; st->state++) {
 451                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 452
 453                 if (!step->skip_onerr)
 454                         cpuhp_invoke_callback(cpu, st->state, true, NULL);
 455         }
 456 }
 457
 458 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 459                                 enum cpuhp_state target)
 460 {
 461         enum cpuhp_state prev_state = st->state;
 462         int ret = 0;
 463
 464         for (; st->state > target; st->state--) {
 465                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL);
 466                 if (ret) {
 467                         st->target = prev_state;
 468                         undo_cpu_down(cpu, st);
 469                         break;
 470                 }
 471         }
 472         return ret;
 473 }
 474
 475 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 476 {
 477         for (st->state--; st->state > st->target; st->state--) {
 478                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 479
 480                 if (!step->skip_onerr)
 481                         cpuhp_invoke_callback(cpu, st->state, false, NULL);
 482         }
 483 }
 484
 485 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 486                               enum cpuhp_state target)
 487 {
 488         enum cpuhp_state prev_state = st->state;
 489         int ret = 0;
 490
 491         while (st->state < target) {
 492                 st->state++;
 493                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
 494                 if (ret) {
 495                         st->target = prev_state;
 496                         undo_cpu_up(cpu, st);
 497                         break;
 498                 }
 499         }
 500         return ret;
 501 }
 502
 503 /*
 504  * The cpu hotplug threads manage the bringup and teardown of the cpus
 505  */
 506 static void cpuhp_create(unsigned int cpu)
 507 {
 508         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 509
 510         init_completion(&st->done);
 511 }
 512
 513 static int cpuhp_should_run(unsigned int cpu)
 514 {
 515         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 516
 517         return st->should_run;
 518 }
 519
 520 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
 521 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
 522 {
 523         enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
 524
 525         return cpuhp_down_callbacks(cpu, st, target);
 526 }
 527
 528 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
 529 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
 530 {
 531         return cpuhp_up_callbacks(cpu, st, st->target);
 532 }
 533
 534 /*
 535  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 536  * callbacks when a state gets [un]installed at runtime.
 537  */
 538 static void cpuhp_thread_fun(unsigned int cpu)
 539 {
 540         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 541         int ret = 0;
 542
 543         /*
 544          * Paired with the mb() in cpuhp_kick_ap_work and
 545          * cpuhp_invoke_ap_callback, so the work set is consistent visible.
 546          */
 547         smp_mb();
 548         if (!st->should_run)
 549                 return;
 550
 551         st->should_run = false;
 552
 553         /* Single callback invocation for [un]install ? */
 554         if (st->single) {
 555                 if (st->cb_state < CPUHP_AP_ONLINE) {
 556                         local_irq_disable();
 557                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 558                                                     st->bringup, st->node);
 559                         local_irq_enable();
 560                 } else {
 561                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 562                                                     st->bringup, st->node);
 563                 }
 564         } else if (st->rollback) {
 565                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 566
 567                 undo_cpu_down(cpu, st);
 568                 /*
 569                  * This is a momentary workaround to keep the notifier users
 570                  * happy. Will go away once we got rid of the notifiers.
 571                  */
 572                 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
 573                 st->rollback = false;
 574         } else {
 575                 /* Cannot happen .... */
 576                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 577
 578                 /* Regular hotplug work */
 579                 if (st->state < st->target)
 580                         ret = cpuhp_ap_online(cpu, st);
 581                 else if (st->state > st->target)
 582                         ret = cpuhp_ap_offline(cpu, st);
 583         }
 584         st->result = ret;
 585         complete(&st->done);
 586 }
 587
 588 /* Invoke a single callback on a remote cpu */
 589 static int
 590 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 591                          struct hlist_node *node)
 592 {
 593         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 594
 595         if (!cpu_online(cpu))
 596                 return 0;
 597
 598         /*
 599          * If we are up and running, use the hotplug thread. For early calls
 600          * we invoke the thread function directly.
 601          */
 602         if (!st->thread)
 603                 return cpuhp_invoke_callback(cpu, state, bringup, node);
 604
 605         st->cb_state = state;
 606         st->single = true;
 607         st->bringup = bringup;
 608         st->node = node;
 609
 610         /*
 611          * Make sure the above stores are visible before should_run becomes
 612          * true. Paired with the mb() above in cpuhp_thread_fun()
 613          */
 614         smp_mb();
 615         st->should_run = true;
 616         wake_up_process(st->thread);
 617         wait_for_completion(&st->done);
 618         return st->result;
 619 }
 620
 621 /* Regular hotplug invocation of the AP hotplug thread */
 622 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
 623 {
 624         st->result = 0;
 625         st->single = false;
 626         /*
 627          * Make sure the above stores are visible before should_run becomes
 628          * true. Paired with the mb() above in cpuhp_thread_fun()
 629          */
 630         smp_mb();
 631         st->should_run = true;
 632         wake_up_process(st->thread);
 633 }
 634
 635 static int cpuhp_kick_ap_work(unsigned int cpu)
 636 {
 637         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 638         enum cpuhp_state state = st->state;
 639
 640         trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
 641         __cpuhp_kick_ap_work(st);
 642         wait_for_completion(&st->done);
 643         trace_cpuhp_exit(cpu, st->state, state, st->result);
 644         return st->result;
 645 }
 646
 647 static struct smp_hotplug_thread cpuhp_threads = {
 648         .store                  = &cpuhp_state.thread,
 649         .create                 = &cpuhp_create,
 650         .thread_should_run      = cpuhp_should_run,
 651         .thread_fn              = cpuhp_thread_fun,
 652         .thread_comm            = "cpuhp/%u",
 653         .selfparking            = true,
 654 };
 655
 656 void __init cpuhp_threads_init(void)
 657 {
 658         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 659         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 660 }
 661
 662 #ifdef CONFIG_HOTPLUG_CPU
 663 EXPORT_SYMBOL(register_cpu_notifier);
 664 EXPORT_SYMBOL(__register_cpu_notifier);
 665 void unregister_cpu_notifier(struct notifier_block *nb)
 666 {
 667         cpu_maps_update_begin();
 668         raw_notifier_chain_unregister(&cpu_chain, nb);
 669         cpu_maps_update_done();
 670 }
 671 EXPORT_SYMBOL(unregister_cpu_notifier);
 672
 673 void __unregister_cpu_notifier(struct notifier_block *nb)
 674 {
 675         raw_notifier_chain_unregister(&cpu_chain, nb);
 676 }
 677 EXPORT_SYMBOL(__unregister_cpu_notifier);
 678
 679 /**
 680  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 681  * @cpu: a CPU id
 682  *
 683  * This function walks all processes, finds a valid mm struct for each one and
 684  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 685  * trivial, there are various non-obvious corner cases, which this function
 686  * tries to solve in a safe manner.
 687  *
 688  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 689  * be called only for an already offlined CPU.
 690  */
 691 void clear_tasks_mm_cpumask(int cpu)
 692 {
 693         struct task_struct *p;
 694
 695         /*
 696          * This function is called after the cpu is taken down and marked
 697          * offline, so its not like new tasks will ever get this cpu set in
 698          * their mm mask. -- Peter Zijlstra
 699          * Thus, we may use rcu_read_lock() here, instead of grabbing
 700          * full-fledged tasklist_lock.
 701          */
 702         WARN_ON(cpu_online(cpu));
 703         rcu_read_lock();
 704         for_each_process(p) {
 705                 struct task_struct *t;
 706
 707                 /*
 708                  * Main thread might exit, but other threads may still have
 709                  * a valid mm. Find one.
 710                  */
 711                 t = find_lock_task_mm(p);
 712                 if (!t)
 713                         continue;
 714                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 715                 task_unlock(t);
 716         }
 717         rcu_read_unlock();
 718 }
 719
 720 static inline void check_for_tasks(int dead_cpu)
 721 {
 722         struct task_struct *g, *p;
 723
 724         read_lock(&tasklist_lock);
 725         for_each_process_thread(g, p) {
 726                 if (!p->on_rq)
 727                         continue;
 728                 /*
 729                  * We do the check with unlocked task_rq(p)->lock.
 730                  * Order the reading to do not warn about a task,
 731                  * which was running on this cpu in the past, and
 732                  * it's just been woken on another cpu.
 733                  */
 734                 rmb();
 735                 if (task_cpu(p) != dead_cpu)
 736                         continue;
 737
 738                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 739                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 740         }
 741         read_unlock(&tasklist_lock);
 742 }
 743
 744 static int notify_down_prepare(unsigned int cpu)
 745 {
 746         int err, nr_calls = 0;
 747
 748         err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
 749         if (err) {
 750                 nr_calls--;
 751                 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
 752                 pr_warn("%s: attempt to take down CPU %u failed\n",
 753                                 __func__, cpu);
 754         }
 755         return err;
 756 }
 757
 758 /* Take this CPU down. */
 759 static int take_cpu_down(void *_param)
 760 {
 761         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 762         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 763         int err, cpu = smp_processor_id();
 764
 765         /* Ensure this CPU doesn't handle any more interrupts. */
 766         err = __cpu_disable();
 767         if (err < 0)
 768                 return err;
 769
 770         /*
 771          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 772          * do this step again.
 773          */
 774         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 775         st->state--;
 776         /* Invoke the former CPU_DYING callbacks */
 777         for (; st->state > target; st->state--)
 778                 cpuhp_invoke_callback(cpu, st->state, false, NULL);
 779
 780         /* Give up timekeeping duties */
 781         tick_handover_do_timer();
 782         /* Park the stopper thread */
 783         stop_machine_park(cpu);
 784         return 0;
 785 }
 786
 787 static int takedown_cpu(unsigned int cpu)
 788 {
 789         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 790         int err;
 791
 792         /* Park the smpboot threads */
 793         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 794         smpboot_park_threads(cpu);
 795
 796         /*
 797          * Prevent irq alloc/free while the dying cpu reorganizes the
 798          * interrupt affinities.
 799          */
 800         irq_lock_sparse();
 801
 802         /*
 803          * So now all preempt/rcu users must observe !cpu_active().
 804          */
 805         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 806         if (err) {
 807                 /* CPU refused to die */
 808                 irq_unlock_sparse();
 809                 /* Unpark the hotplug thread so we can rollback there */
 810                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 811                 return err;
 812         }
 813         BUG_ON(cpu_online(cpu));
 814
 815         /*
 816          * The CPUHP_AP_SCHED_MIGRATE_DYING callback will have removed all
 817          * runnable tasks from the cpu, there's only the idle task left now
 818          * that the migration thread is done doing the stop_machine thing.
 819          *
 820          * Wait for the stop thread to go away.
 821          */
 822         wait_for_completion(&st->done);
 823         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 824
 825         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 826         irq_unlock_sparse();
 827
 828         hotplug_cpu__broadcast_tick_pull(cpu);
 829         /* This actually kills the CPU. */
 830         __cpu_die(cpu);
 831
 832         tick_cleanup_dead_cpu(cpu);
 833         return 0;
 834 }
 835
 836 static int notify_dead(unsigned int cpu)
 837 {
 838         cpu_notify_nofail(CPU_DEAD, cpu);
 839         check_for_tasks(cpu);
 840         return 0;
 841 }
 842
 843 static void cpuhp_complete_idle_dead(void *arg)
 844 {
 845         struct cpuhp_cpu_state *st = arg;
 846
 847         complete(&st->done);
 848 }
 849
 850 void cpuhp_report_idle_dead(void)
 851 {
 852         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 853
 854         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 855         rcu_report_dead(smp_processor_id());
 856         st->state = CPUHP_AP_IDLE_DEAD;
 857         /*
 858          * We cannot call complete after rcu_report_dead() so we delegate it
 859          * to an online cpu.
 860          */
 861         smp_call_function_single(cpumask_first(cpu_online_mask),
 862                                  cpuhp_complete_idle_dead, st, 0);
 863 }
 864
 865 #else
 866 #define notify_down_prepare     NULL
 867 #define takedown_cpu            NULL
 868 #define notify_dead             NULL
 869 #endif
 870
 871 #ifdef CONFIG_HOTPLUG_CPU
 872
 873 /* Requires cpu_add_remove_lock to be held */
 874 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 875                            enum cpuhp_state target)
 876 {
 877         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 878         int prev_state, ret = 0;
 879         bool hasdied = false;
 880
 881         if (num_online_cpus() == 1)
 882                 return -EBUSY;
 883
 884         if (!cpu_present(cpu))
 885                 return -EINVAL;
 886
 887         cpu_hotplug_begin();
 888
 889         cpuhp_tasks_frozen = tasks_frozen;
 890
 891         prev_state = st->state;
 892         st->target = target;
 893         /*
 894          * If the current CPU state is in the range of the AP hotplug thread,
 895          * then we need to kick the thread.
 896          */
 897         if (st->state > CPUHP_TEARDOWN_CPU) {
 898                 ret = cpuhp_kick_ap_work(cpu);
 899                 /*
 900                  * The AP side has done the error rollback already. Just
 901                  * return the error code..
 902                  */
 903                 if (ret)
 904                         goto out;
 905
 906                 /*
 907                  * We might have stopped still in the range of the AP hotplug
 908                  * thread. Nothing to do anymore.
 909                  */
 910                 if (st->state > CPUHP_TEARDOWN_CPU)
 911                         goto out;
 912         }
 913         /*
 914          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 915          * to do the further cleanups.
 916          */
 917         ret = cpuhp_down_callbacks(cpu, st, target);
 918         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 919                 st->target = prev_state;
 920                 st->rollback = true;
 921                 cpuhp_kick_ap_work(cpu);
 922         }
 923
 924         hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
 925 out:
 926         cpu_hotplug_done();
 927         /* This post dead nonsense must die */
 928         if (!ret && hasdied)
 929                 cpu_notify_nofail(CPU_POST_DEAD, cpu);
 930         return ret;
 931 }
 932
 933 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 934 {
 935         int err;
 936
 937         cpu_maps_update_begin();
 938
 939         if (cpu_hotplug_disabled) {
 940                 err = -EBUSY;
 941                 goto out;
 942         }
 943
 944         err = _cpu_down(cpu, 0, target);
 945
 946 out:
 947         cpu_maps_update_done();
 948         return err;
 949 }
 950 int cpu_down(unsigned int cpu)
 951 {
 952         return do_cpu_down(cpu, CPUHP_OFFLINE);
 953 }
 954 EXPORT_SYMBOL(cpu_down);
 955 #endif /*CONFIG_HOTPLUG_CPU*/
 956
 957 /**
 958  * notify_cpu_starting(cpu) - Invoke the callbacks on the starting CPU
 959  * @cpu: cpu that just started
 960  *
 961  * It must be called by the arch code on the new cpu, before the new cpu
 962  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 963  */
 964 void notify_cpu_starting(unsigned int cpu)
 965 {
 966         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 967         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 968
 969         rcu_cpu_starting(cpu);  /* Enables RCU usage on this CPU. */
 970         while (st->state < target) {
 971                 st->state++;
 972                 cpuhp_invoke_callback(cpu, st->state, true, NULL);
 973         }
 974 }
 975
 976 /*
 977  * Called from the idle task. We need to set active here, so we can kick off
 978  * the stopper thread and unpark the smpboot threads. If the target state is
 979  * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
 980  * cpu further.
 981  */
 982 void cpuhp_online_idle(enum cpuhp_state state)
 983 {
 984         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 985         unsigned int cpu = smp_processor_id();
 986
 987         /* Happens for the boot cpu */
 988         if (state != CPUHP_AP_ONLINE_IDLE)
 989                 return;
 990
 991         st->state = CPUHP_AP_ONLINE_IDLE;
 992
 993         /* Unpark the stopper thread and the hotplug thread of this cpu */
 994         stop_machine_unpark(cpu);
 995         kthread_unpark(st->thread);
 996
 997         /* Should we go further up ? */
 998         if (st->target > CPUHP_AP_ONLINE_IDLE)
 999                 __cpuhp_kick_ap_work(st);
1000         else
1001                 complete(&st->done);
1002 }
1003
1004 /* Requires cpu_add_remove_lock to be held */
1005 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1006 {
1007         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1008         struct task_struct *idle;
1009         int ret = 0;
1010
1011         cpu_hotplug_begin();
1012
1013         if (!cpu_present(cpu)) {
1014                 ret = -EINVAL;
1015                 goto out;
1016         }
1017
1018         /*
1019          * The caller of do_cpu_up might have raced with another
1020          * caller. Ignore it for now.
1021          */
1022         if (st->state >= target)
1023                 goto out;
1024
1025         if (st->state == CPUHP_OFFLINE) {
1026                 /* Let it fail before we try to bring the cpu up */
1027                 idle = idle_thread_get(cpu);
1028                 if (IS_ERR(idle)) {
1029                         ret = PTR_ERR(idle);
1030                         goto out;
1031                 }
1032         }
1033
1034         cpuhp_tasks_frozen = tasks_frozen;
1035
1036         st->target = target;
1037         /*
1038          * If the current CPU state is in the range of the AP hotplug thread,
1039          * then we need to kick the thread once more.
1040          */
1041         if (st->state > CPUHP_BRINGUP_CPU) {
1042                 ret = cpuhp_kick_ap_work(cpu);
1043                 /*
1044                  * The AP side has done the error rollback already. Just
1045                  * return the error code..
1046                  */
1047                 if (ret)
1048                         goto out;
1049         }
1050
1051         /*
1052          * Try to reach the target state. We max out on the BP at
1053          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1054          * responsible for bringing it up to the target state.
1055          */
1056         target = min((int)target, CPUHP_BRINGUP_CPU);
1057         ret = cpuhp_up_callbacks(cpu, st, target);
1058 out:
1059         cpu_hotplug_done();
1060         return ret;
1061 }
1062
1063 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1064 {
1065         int err = 0;
1066
1067         if (!cpu_possible(cpu)) {
1068                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1069                        cpu);
1070 #if defined(CONFIG_IA64)
1071                 pr_err("please check additional_cpus= boot parameter\n");
1072 #endif
1073                 return -EINVAL;
1074         }
1075
1076         err = try_online_node(cpu_to_node(cpu));
1077         if (err)
1078                 return err;
1079
1080         cpu_maps_update_begin();
1081
1082         if (cpu_hotplug_disabled) {
1083                 err = -EBUSY;
1084                 goto out;
1085         }
1086
1087         err = _cpu_up(cpu, 0, target);
1088 out:
1089         cpu_maps_update_done();
1090         return err;
1091 }
1092
1093 int cpu_up(unsigned int cpu)
1094 {
1095         return do_cpu_up(cpu, CPUHP_ONLINE);
1096 }
1097 EXPORT_SYMBOL_GPL(cpu_up);
1098
1099 #ifdef CONFIG_PM_SLEEP_SMP
1100 static cpumask_var_t frozen_cpus;
1101
1102 int freeze_secondary_cpus(int primary)
1103 {
1104         int cpu, error = 0;
1105
1106         cpu_maps_update_begin();
1107         if (!cpu_online(primary))
1108                 primary = cpumask_first(cpu_online_mask);
1109         /*
1110          * We take down all of the non-boot CPUs in one shot to avoid races
1111          * with the userspace trying to use the CPU hotplug at the same time
1112          */
1113         cpumask_clear(frozen_cpus);
1114
1115         pr_info("Disabling non-boot CPUs ...\n");
1116         for_each_online_cpu(cpu) {
1117                 if (cpu == primary)
1118                         continue;
1119                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1120                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1121                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1122                 if (!error)
1123                         cpumask_set_cpu(cpu, frozen_cpus);
1124                 else {
1125                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1126                         break;
1127                 }
1128         }
1129
1130         if (!error)
1131                 BUG_ON(num_online_cpus() > 1);
1132         else
1133                 pr_err("Non-boot CPUs are not disabled\n");
1134
1135         /*
1136          * Make sure the CPUs won't be enabled by someone else. We need to do
1137          * this even in case of failure as all disable_nonboot_cpus() users are
1138          * supposed to do enable_nonboot_cpus() on the failure path.
1139          */
1140         cpu_hotplug_disabled++;
1141
1142         cpu_maps_update_done();
1143         return error;
1144 }
1145
1146 void __weak arch_enable_nonboot_cpus_begin(void)
1147 {
1148 }
1149
1150 void __weak arch_enable_nonboot_cpus_end(void)
1151 {
1152 }
1153
1154 void enable_nonboot_cpus(void)
1155 {
1156         int cpu, error;
1157
1158         /* Allow everyone to use the CPU hotplug again */
1159         cpu_maps_update_begin();
1160         __cpu_hotplug_enable();
1161         if (cpumask_empty(frozen_cpus))
1162                 goto out;
1163
1164         pr_info("Enabling non-boot CPUs ...\n");
1165
1166         arch_enable_nonboot_cpus_begin();
1167
1168         for_each_cpu(cpu, frozen_cpus) {
1169                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1170                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1171                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1172                 if (!error) {
1173                         pr_info("CPU%d is up\n", cpu);
1174                         continue;
1175                 }
1176                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1177         }
1178
1179         arch_enable_nonboot_cpus_end();
1180
1181         cpumask_clear(frozen_cpus);
1182 out:
1183         cpu_maps_update_done();
1184 }
1185
1186 static int __init alloc_frozen_cpus(void)
1187 {
1188         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1189                 return -ENOMEM;
1190         return 0;
1191 }
1192 core_initcall(alloc_frozen_cpus);
1193
1194 /*
1195  * When callbacks for CPU hotplug notifications are being executed, we must
1196  * ensure that the state of the system with respect to the tasks being frozen
1197  * or not, as reported by the notification, remains unchanged *throughout the
1198  * duration* of the execution of the callbacks.
1199  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1200  *
1201  * This synchronization is implemented by mutually excluding regular CPU
1202  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1203  * Hibernate notifications.
1204  */
1205 static int
1206 cpu_hotplug_pm_callback(struct notifier_block *nb,
1207                         unsigned long action, void *ptr)
1208 {
1209         switch (action) {
1210
1211         case PM_SUSPEND_PREPARE:
1212         case PM_HIBERNATION_PREPARE:
1213                 cpu_hotplug_disable();
1214                 break;
1215
1216         case PM_POST_SUSPEND:
1217         case PM_POST_HIBERNATION:
1218                 cpu_hotplug_enable();
1219                 break;
1220
1221         default:
1222                 return NOTIFY_DONE;
1223         }
1224
1225         return NOTIFY_OK;
1226 }
1227
1228
1229 static int __init cpu_hotplug_pm_sync_init(void)
1230 {
1231         /*
1232          * cpu_hotplug_pm_callback has higher priority than x86
1233          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1234          * to disable cpu hotplug to avoid cpu hotplug race.
1235          */
1236         pm_notifier(cpu_hotplug_pm_callback, 0);
1237         return 0;
1238 }
1239 core_initcall(cpu_hotplug_pm_sync_init);
1240
1241 #endif /* CONFIG_PM_SLEEP_SMP */
1242
1243 #endif /* CONFIG_SMP */
1244
1245 /* Boot processor state steps */
1246 static struct cpuhp_step cpuhp_bp_states[] = {
1247         [CPUHP_OFFLINE] = {
1248                 .name                   = "offline",
1249                 .startup.single         = NULL,
1250                 .teardown.single        = NULL,
1251         },
1252 #ifdef CONFIG_SMP
1253         [CPUHP_CREATE_THREADS]= {
1254                 .name                   = "threads:prepare",
1255                 .startup.single         = smpboot_create_threads,
1256                 .teardown.single        = NULL,
1257                 .cant_stop              = true,
1258         },
1259         [CPUHP_PERF_PREPARE] = {
1260                 .name                   = "perf:prepare",
1261                 .startup.single         = perf_event_init_cpu,
1262                 .teardown.single        = perf_event_exit_cpu,
1263         },
1264         [CPUHP_WORKQUEUE_PREP] = {
1265                 .name                   = "workqueue:prepare",
1266                 .startup.single         = workqueue_prepare_cpu,
1267                 .teardown.single        = NULL,
1268         },
1269         [CPUHP_HRTIMERS_PREPARE] = {
1270                 .name                   = "hrtimers:prepare",
1271                 .startup.single         = hrtimers_prepare_cpu,
1272                 .teardown.single        = hrtimers_dead_cpu,
1273         },
1274         [CPUHP_SMPCFD_PREPARE] = {
1275                 .name                   = "smpcfd:prepare",
1276                 .startup.single         = smpcfd_prepare_cpu,
1277                 .teardown.single        = smpcfd_dead_cpu,
1278         },
1279         [CPUHP_RELAY_PREPARE] = {
1280                 .name                   = "relay:prepare",
1281                 .startup.single         = relay_prepare_cpu,
1282                 .teardown.single        = NULL,
1283         },
1284         [CPUHP_SLAB_PREPARE] = {
1285                 .name                   = "slab:prepare",
1286                 .startup.single         = slab_prepare_cpu,
1287                 .teardown.single        = slab_dead_cpu,
1288         },
1289         [CPUHP_RCUTREE_PREP] = {
1290                 .name                   = "RCU/tree:prepare",
1291                 .startup.single         = rcutree_prepare_cpu,
1292                 .teardown.single        = rcutree_dead_cpu,
1293         },
1294         /*
1295          * Preparatory and dead notifiers. Will be replaced once the notifiers
1296          * are converted to states.
1297          */
1298         [CPUHP_NOTIFY_PREPARE] = {
1299                 .name                   = "notify:prepare",
1300                 .startup.single         = notify_prepare,
1301                 .teardown.single        = notify_dead,
1302                 .skip_onerr             = true,
1303                 .cant_stop              = true,
1304         },
1305         /*
1306          * On the tear-down path, timers_dead_cpu() must be invoked
1307          * before blk_mq_queue_reinit_notify() from notify_dead(),
1308          * otherwise a RCU stall occurs.
1309          */
1310         [CPUHP_TIMERS_DEAD] = {
1311                 .name                   = "timers:dead",
1312                 .startup.single         = NULL,
1313                 .teardown.single        = timers_dead_cpu,
1314         },
1315         /* Kicks the plugged cpu into life */
1316         [CPUHP_BRINGUP_CPU] = {
1317                 .name                   = "cpu:bringup",
1318                 .startup.single         = bringup_cpu,
1319                 .teardown.single        = NULL,
1320                 .cant_stop              = true,
1321         },
1322         [CPUHP_AP_SMPCFD_DYING] = {
1323                 .name                   = "smpcfd:dying",
1324                 .startup.single         = NULL,
1325                 .teardown.single        = smpcfd_dying_cpu,
1326         },
1327         /*
1328          * Handled on controll processor until the plugged processor manages
1329          * this itself.
1330          */
1331         [CPUHP_TEARDOWN_CPU] = {
1332                 .name                   = "cpu:teardown",
1333                 .startup.single         = NULL,
1334                 .teardown.single        = takedown_cpu,
1335                 .cant_stop              = true,
1336         },
1337 #else
1338         [CPUHP_BRINGUP_CPU] = { },
1339 #endif
1340 };
1341
1342 /* Application processor state steps */
1343 static struct cpuhp_step cpuhp_ap_states[] = {
1344 #ifdef CONFIG_SMP
1345         /* Final state before CPU kills itself */
1346         [CPUHP_AP_IDLE_DEAD] = {
1347                 .name                   = "idle:dead",
1348         },
1349         /*
1350          * Last state before CPU enters the idle loop to die. Transient state
1351          * for synchronization.
1352          */
1353         [CPUHP_AP_OFFLINE] = {
1354                 .name                   = "ap:offline",
1355                 .cant_stop              = true,
1356         },
1357         /* First state is scheduler control. Interrupts are disabled */
1358         [CPUHP_AP_SCHED_STARTING] = {
1359                 .name                   = "sched:starting",
1360                 .startup.single         = sched_cpu_starting,
1361                 .teardown.single        = sched_cpu_dying,
1362         },
1363         [CPUHP_AP_RCUTREE_DYING] = {
1364                 .name                   = "RCU/tree:dying",
1365                 .startup.single         = NULL,
1366                 .teardown.single        = rcutree_dying_cpu,
1367         },
1368         /* Entry state on starting. Interrupts enabled from here on. Transient
1369          * state for synchronsization */
1370         [CPUHP_AP_ONLINE] = {
1371                 .name                   = "ap:online",
1372         },
1373         /* Handle smpboot threads park/unpark */
1374         [CPUHP_AP_SMPBOOT_THREADS] = {
1375                 .name                   = "smpboot/threads:online",
1376                 .startup.single         = smpboot_unpark_threads,
1377                 .teardown.single        = NULL,
1378         },
1379         [CPUHP_AP_PERF_ONLINE] = {
1380                 .name                   = "perf:online",
1381                 .startup.single         = perf_event_init_cpu,
1382                 .teardown.single        = perf_event_exit_cpu,
1383         },
1384         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1385                 .name                   = "workqueue:online",
1386                 .startup.single         = workqueue_online_cpu,
1387                 .teardown.single        = workqueue_offline_cpu,
1388         },
1389         [CPUHP_AP_RCUTREE_ONLINE] = {
1390                 .name                   = "RCU/tree:online",
1391                 .startup.single         = rcutree_online_cpu,
1392                 .teardown.single        = rcutree_offline_cpu,
1393         },
1394
1395         /*
1396          * Online/down_prepare notifiers. Will be removed once the notifiers
1397          * are converted to states.
1398          */
1399         [CPUHP_AP_NOTIFY_ONLINE] = {
1400                 .name                   = "notify:online",
1401                 .startup.single         = notify_online,
1402                 .teardown.single        = notify_down_prepare,
1403                 .skip_onerr             = true,
1404         },
1405 #endif
1406         /*
1407          * The dynamically registered state space is here
1408          */
1409
1410 #ifdef CONFIG_SMP
1411         /* Last state is scheduler control setting the cpu active */
1412         [CPUHP_AP_ACTIVE] = {
1413                 .name                   = "sched:active",
1414                 .startup.single         = sched_cpu_activate,
1415                 .teardown.single        = sched_cpu_deactivate,
1416         },
1417 #endif
1418
1419         /* CPU is fully up and running. */
1420         [CPUHP_ONLINE] = {
1421                 .name                   = "online",
1422                 .startup.single         = NULL,
1423                 .teardown.single        = NULL,
1424         },
1425 };
1426
1427 /* Sanity check for callbacks */
1428 static int cpuhp_cb_check(enum cpuhp_state state)
1429 {
1430         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1431                 return -EINVAL;
1432         return 0;
1433 }
1434
1435 static void cpuhp_store_callbacks(enum cpuhp_state state,
1436                                   const char *name,
1437                                   int (*startup)(unsigned int cpu),
1438                                   int (*teardown)(unsigned int cpu),
1439                                   bool multi_instance)
1440 {
1441         /* (Un)Install the callbacks for further cpu hotplug operations */
1442         struct cpuhp_step *sp;
1443
1444         mutex_lock(&cpuhp_state_mutex);
1445         sp = cpuhp_get_step(state);
1446         sp->startup.single = startup;
1447         sp->teardown.single = teardown;
1448         sp->name = name;
1449         sp->multi_instance = multi_instance;
1450         INIT_HLIST_HEAD(&sp->list);
1451         mutex_unlock(&cpuhp_state_mutex);
1452 }
1453
1454 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1455 {
1456         return cpuhp_get_step(state)->teardown.single;
1457 }
1458
1459 /*
1460  * Call the startup/teardown function for a step either on the AP or
1461  * on the current CPU.
1462  */
1463 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1464                             struct hlist_node *node)
1465 {
1466         struct cpuhp_step *sp = cpuhp_get_step(state);
1467         int ret;
1468
1469         if ((bringup && !sp->startup.single) ||
1470             (!bringup && !sp->teardown.single))
1471                 return 0;
1472         /*
1473          * The non AP bound callbacks can fail on bringup. On teardown
1474          * e.g. module removal we crash for now.
1475          */
1476 #ifdef CONFIG_SMP
1477         if (cpuhp_is_ap_state(state))
1478                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1479         else
1480                 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1481 #else
1482         ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1483 #endif
1484         BUG_ON(ret && !bringup);
1485         return ret;
1486 }
1487
1488 /*
1489  * Called from __cpuhp_setup_state on a recoverable failure.
1490  *
1491  * Note: The teardown callbacks for rollback are not allowed to fail!
1492  */
1493 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1494                                    struct hlist_node *node)
1495 {
1496         int cpu;
1497
1498         /* Roll back the already executed steps on the other cpus */
1499         for_each_present_cpu(cpu) {
1500                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1501                 int cpustate = st->state;
1502
1503                 if (cpu >= failedcpu)
1504                         break;
1505
1506                 /* Did we invoke the startup call on that cpu ? */
1507                 if (cpustate >= state)
1508                         cpuhp_issue_call(cpu, state, false, node);
1509         }
1510 }
1511
1512 /*
1513  * Returns a free for dynamic slot assignment of the Online state. The states
1514  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1515  * by having no name assigned.
1516  */
1517 static int cpuhp_reserve_state(enum cpuhp_state state)
1518 {
1519         enum cpuhp_state i;
1520
1521         mutex_lock(&cpuhp_state_mutex);
1522         for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1523                 if (cpuhp_ap_states[i].name)
1524                         continue;
1525
1526                 cpuhp_ap_states[i].name = "Reserved";
1527                 mutex_unlock(&cpuhp_state_mutex);
1528                 return i;
1529         }
1530         mutex_unlock(&cpuhp_state_mutex);
1531         WARN(1, "No more dynamic states available for CPU hotplug\n");
1532         return -ENOSPC;
1533 }
1534
1535 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1536                                bool invoke)
1537 {
1538         struct cpuhp_step *sp;
1539         int cpu;
1540         int ret;
1541
1542         sp = cpuhp_get_step(state);
1543         if (sp->multi_instance == false)
1544                 return -EINVAL;
1545
1546         get_online_cpus();
1547
1548         if (!invoke || !sp->startup.multi)
1549                 goto add_node;
1550
1551         /*
1552          * Try to call the startup callback for each present cpu
1553          * depending on the hotplug state of the cpu.
1554          */
1555         for_each_present_cpu(cpu) {
1556                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1557                 int cpustate = st->state;
1558
1559                 if (cpustate < state)
1560                         continue;
1561
1562                 ret = cpuhp_issue_call(cpu, state, true, node);
1563                 if (ret) {
1564                         if (sp->teardown.multi)
1565                                 cpuhp_rollback_install(cpu, state, node);
1566                         goto err;
1567                 }
1568         }
1569 add_node:
1570         ret = 0;
1571         mutex_lock(&cpuhp_state_mutex);
1572         hlist_add_head(node, &sp->list);
1573         mutex_unlock(&cpuhp_state_mutex);
1574
1575 err:
1576         put_online_cpus();
1577         return ret;
1578 }
1579 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1580
1581 /**
1582  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1583  * @state:      The state to setup
1584  * @invoke:     If true, the startup function is invoked for cpus where
1585  *              cpu state >= @state
1586  * @startup:    startup callback function
1587  * @teardown:   teardown callback function
1588  *
1589  * Returns 0 if successful, otherwise a proper error code
1590  */
1591 int __cpuhp_setup_state(enum cpuhp_state state,
1592                         const char *name, bool invoke,
1593                         int (*startup)(unsigned int cpu),
1594                         int (*teardown)(unsigned int cpu),
1595                         bool multi_instance)
1596 {
1597         int cpu, ret = 0;
1598         int dyn_state = 0;
1599
1600         if (cpuhp_cb_check(state) || !name)
1601                 return -EINVAL;
1602
1603         get_online_cpus();
1604
1605         /* currently assignments for the ONLINE state are possible */
1606         if (state == CPUHP_AP_ONLINE_DYN) {
1607                 dyn_state = 1;
1608                 ret = cpuhp_reserve_state(state);
1609                 if (ret < 0)
1610                         goto out;
1611                 state = ret;
1612         }
1613
1614         cpuhp_store_callbacks(state, name, startup, teardown, multi_instance);
1615
1616         if (!invoke || !startup)
1617                 goto out;
1618
1619         /*
1620          * Try to call the startup callback for each present cpu
1621          * depending on the hotplug state of the cpu.
1622          */
1623         for_each_present_cpu(cpu) {
1624                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1625                 int cpustate = st->state;
1626
1627                 if (cpustate < state)
1628                         continue;
1629
1630                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1631                 if (ret) {
1632                         if (teardown)
1633                                 cpuhp_rollback_install(cpu, state, NULL);
1634                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1635                         goto out;
1636                 }
1637         }
1638 out:
1639         put_online_cpus();
1640         if (!ret && dyn_state)
1641                 return state;
1642         return ret;
1643 }
1644 EXPORT_SYMBOL(__cpuhp_setup_state);
1645
1646 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1647                                   struct hlist_node *node, bool invoke)
1648 {
1649         struct cpuhp_step *sp = cpuhp_get_step(state);
1650         int cpu;
1651
1652         BUG_ON(cpuhp_cb_check(state));
1653
1654         if (!sp->multi_instance)
1655                 return -EINVAL;
1656
1657         get_online_cpus();
1658         if (!invoke || !cpuhp_get_teardown_cb(state))
1659                 goto remove;
1660         /*
1661          * Call the teardown callback for each present cpu depending
1662          * on the hotplug state of the cpu. This function is not
1663          * allowed to fail currently!
1664          */
1665         for_each_present_cpu(cpu) {
1666                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1667                 int cpustate = st->state;
1668
1669                 if (cpustate >= state)
1670                         cpuhp_issue_call(cpu, state, false, node);
1671         }
1672
1673 remove:
1674         mutex_lock(&cpuhp_state_mutex);
1675         hlist_del(node);
1676         mutex_unlock(&cpuhp_state_mutex);
1677         put_online_cpus();
1678
1679         return 0;
1680 }
1681 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1682 /**
1683  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1684  * @state:      The state to remove
1685  * @invoke:     If true, the teardown function is invoked for cpus where
1686  *              cpu state >= @state
1687  *
1688  * The teardown callback is currently not allowed to fail. Think
1689  * about module removal!
1690  */
1691 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1692 {
1693         struct cpuhp_step *sp = cpuhp_get_step(state);
1694         int cpu;
1695
1696         BUG_ON(cpuhp_cb_check(state));
1697
1698         get_online_cpus();
1699
1700         if (sp->multi_instance) {
1701                 WARN(!hlist_empty(&sp->list),
1702                      "Error: Removing state %d which has instances left.\n",
1703                      state);
1704                 goto remove;
1705         }
1706
1707         if (!invoke || !cpuhp_get_teardown_cb(state))
1708                 goto remove;
1709
1710         /*
1711          * Call the teardown callback for each present cpu depending
1712          * on the hotplug state of the cpu. This function is not
1713          * allowed to fail currently!
1714          */
1715         for_each_present_cpu(cpu) {
1716                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1717                 int cpustate = st->state;
1718
1719                 if (cpustate >= state)
1720                         cpuhp_issue_call(cpu, state, false, NULL);
1721         }
1722 remove:
1723         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1724         put_online_cpus();
1725 }
1726 EXPORT_SYMBOL(__cpuhp_remove_state);
1727
1728 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1729 static ssize_t show_cpuhp_state(struct device *dev,
1730                                 struct device_attribute *attr, char *buf)
1731 {
1732         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1733
1734         return sprintf(buf, "%d\n", st->state);
1735 }
1736 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1737
1738 static ssize_t write_cpuhp_target(struct device *dev,
1739                                   struct device_attribute *attr,
1740                                   const char *buf, size_t count)
1741 {
1742         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1743         struct cpuhp_step *sp;
1744         int target, ret;
1745
1746         ret = kstrtoint(buf, 10, &target);
1747         if (ret)
1748                 return ret;
1749
1750 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1751         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1752                 return -EINVAL;
1753 #else
1754         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1755                 return -EINVAL;
1756 #endif
1757
1758         ret = lock_device_hotplug_sysfs();
1759         if (ret)
1760                 return ret;
1761
1762         mutex_lock(&cpuhp_state_mutex);
1763         sp = cpuhp_get_step(target);
1764         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1765         mutex_unlock(&cpuhp_state_mutex);
1766         if (ret)
1767                 return ret;
1768
1769         if (st->state < target)
1770                 ret = do_cpu_up(dev->id, target);
1771         else
1772                 ret = do_cpu_down(dev->id, target);
1773
1774         unlock_device_hotplug();
1775         return ret ? ret : count;
1776 }
1777
1778 static ssize_t show_cpuhp_target(struct device *dev,
1779                                  struct device_attribute *attr, char *buf)
1780 {
1781         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1782
1783         return sprintf(buf, "%d\n", st->target);
1784 }
1785 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1786
1787 static struct attribute *cpuhp_cpu_attrs[] = {
1788         &dev_attr_state.attr,
1789         &dev_attr_target.attr,
1790         NULL
1791 };
1792
1793 static struct attribute_group cpuhp_cpu_attr_group = {
1794         .attrs = cpuhp_cpu_attrs,
1795         .name = "hotplug",
1796         NULL
1797 };
1798
1799 static ssize_t show_cpuhp_states(struct device *dev,
1800                                  struct device_attribute *attr, char *buf)
1801 {
1802         ssize_t cur, res = 0;
1803         int i;
1804
1805         mutex_lock(&cpuhp_state_mutex);
1806         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1807                 struct cpuhp_step *sp = cpuhp_get_step(i);
1808
1809                 if (sp->name) {
1810                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1811                         buf += cur;
1812                         res += cur;
1813                 }
1814         }
1815         mutex_unlock(&cpuhp_state_mutex);
1816         return res;
1817 }
1818 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1819
1820 static struct attribute *cpuhp_cpu_root_attrs[] = {
1821         &dev_attr_states.attr,
1822         NULL
1823 };
1824
1825 static struct attribute_group cpuhp_cpu_root_attr_group = {
1826         .attrs = cpuhp_cpu_root_attrs,
1827         .name = "hotplug",
1828         NULL
1829 };
1830
1831 static int __init cpuhp_sysfs_init(void)
1832 {
1833         int cpu, ret;
1834
1835         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1836                                  &cpuhp_cpu_root_attr_group);
1837         if (ret)
1838                 return ret;
1839
1840         for_each_possible_cpu(cpu) {
1841                 struct device *dev = get_cpu_device(cpu);
1842
1843                 if (!dev)
1844                         continue;
1845                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1846                 if (ret)
1847                         return ret;
1848         }
1849         return 0;
1850 }
1851 device_initcall(cpuhp_sysfs_init);
1852 #endif
1853
1854 /*
1855  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1856  * represents all NR_CPUS bits binary values of 1<<nr.
1857  *
1858  * It is used by cpumask_of() to get a constant address to a CPU
1859  * mask value that has a single bit set only.
1860  */
1861
1862 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1863 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1864 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1865 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1866 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1867
1868 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1869
1870         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1871         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1872 #if BITS_PER_LONG > 32
1873         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1874         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1875 #endif
1876 };
1877 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1878
1879 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1880 EXPORT_SYMBOL(cpu_all_bits);
1881
1882 #ifdef CONFIG_INIT_ALL_POSSIBLE
1883 struct cpumask __cpu_possible_mask __read_mostly
1884         = {CPU_BITS_ALL};
1885 #else
1886 struct cpumask __cpu_possible_mask __read_mostly;
1887 #endif
1888 EXPORT_SYMBOL(__cpu_possible_mask);
1889
1890 struct cpumask __cpu_online_mask __read_mostly;
1891 EXPORT_SYMBOL(__cpu_online_mask);
1892
1893 struct cpumask __cpu_present_mask __read_mostly;
1894 EXPORT_SYMBOL(__cpu_present_mask);
1895
1896 struct cpumask __cpu_active_mask __read_mostly;
1897 EXPORT_SYMBOL(__cpu_active_mask);
1898
1899 void init_cpu_present(const struct cpumask *src)
1900 {
1901         cpumask_copy(&__cpu_present_mask, src);
1902 }
1903
1904 void init_cpu_possible(const struct cpumask *src)
1905 {
1906         cpumask_copy(&__cpu_possible_mask, src);
1907 }
1908
1909 void init_cpu_online(const struct cpumask *src)
1910 {
1911         cpumask_copy(&__cpu_online_mask, src);
1912 }
1913
1914 /*
1915  * Activate the first processor.
1916  */
1917 void __init boot_cpu_init(void)
1918 {
1919         int cpu = smp_processor_id();
1920
1921         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1922         set_cpu_online(cpu, true);
1923         set_cpu_active(cpu, true);
1924         set_cpu_present(cpu, true);
1925         set_cpu_possible(cpu, true);
1926 }
1927
1928 /*
1929  * Must be called _AFTER_ setting up the per_cpu areas
1930  */
1931 void __init boot_cpu_state_init(void)
1932 {
1933         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1934 }