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 static void __cpu_hotplug_enable(void)
 335 {
 336         if (WARN_ONCE(!cpu_hotplug_disabled, "Unbalanced cpu hotplug enable\n"))
 337                 return;
 338         cpu_hotplug_disabled--;
 339 }
 340
 341 void cpu_hotplug_enable(void)
 342 {
 343         cpu_maps_update_begin();
 344         __cpu_hotplug_enable();
 345         cpu_maps_update_done();
 346 }
 347 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 348 #endif  /* CONFIG_HOTPLUG_CPU */
 349
 350 /* Need to know about CPUs going up/down? */
 351 int register_cpu_notifier(struct notifier_block *nb)
 352 {
 353         int ret;
 354         cpu_maps_update_begin();
 355         ret = raw_notifier_chain_register(&cpu_chain, nb);
 356         cpu_maps_update_done();
 357         return ret;
 358 }
 359
 360 int __register_cpu_notifier(struct notifier_block *nb)
 361 {
 362         return raw_notifier_chain_register(&cpu_chain, nb);
 363 }
 364
 365 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
 366                         int *nr_calls)
 367 {
 368         unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
 369         void *hcpu = (void *)(long)cpu;
 370
 371         int ret;
 372
 373         ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
 374                                         nr_calls);
 375
 376         return notifier_to_errno(ret);
 377 }
 378
 379 static int cpu_notify(unsigned long val, unsigned int cpu)
 380 {
 381         return __cpu_notify(val, cpu, -1, NULL);
 382 }
 383
 384 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
 385 {
 386         BUG_ON(cpu_notify(val, cpu));
 387 }
 388
 389 /* Notifier wrappers for transitioning to state machine */
 390 static int notify_prepare(unsigned int cpu)
 391 {
 392         int nr_calls = 0;
 393         int ret;
 394
 395         ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
 396         if (ret) {
 397                 nr_calls--;
 398                 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
 399                                 __func__, cpu);
 400                 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
 401         }
 402         return ret;
 403 }
 404
 405 static int notify_online(unsigned int cpu)
 406 {
 407         cpu_notify(CPU_ONLINE, cpu);
 408         return 0;
 409 }
 410
 411 static int notify_starting(unsigned int cpu)
 412 {
 413         cpu_notify(CPU_STARTING, cpu);
 414         return 0;
 415 }
 416
 417 static int bringup_wait_for_ap(unsigned int cpu)
 418 {
 419         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 420
 421         wait_for_completion(&st->done);
 422         return st->result;
 423 }
 424
 425 static int bringup_cpu(unsigned int cpu)
 426 {
 427         struct task_struct *idle = idle_thread_get(cpu);
 428         int ret;
 429
 430         /*
 431          * Some architectures have to walk the irq descriptors to
 432          * setup the vector space for the cpu which comes online.
 433          * Prevent irq alloc/free across the bringup.
 434          */
 435         irq_lock_sparse();
 436
 437         /* Arch-specific enabling code. */
 438         ret = __cpu_up(cpu, idle);
 439         irq_unlock_sparse();
 440         if (ret) {
 441                 cpu_notify(CPU_UP_CANCELED, cpu);
 442                 return ret;
 443         }
 444         ret = bringup_wait_for_ap(cpu);
 445         BUG_ON(!cpu_online(cpu));
 446         return ret;
 447 }
 448
 449 /*
 450  * Hotplug state machine related functions
 451  */
 452 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 453 {
 454         for (st->state++; st->state < st->target; st->state++) {
 455                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 456
 457                 if (!step->skip_onerr)
 458                         cpuhp_invoke_callback(cpu, st->state, true, NULL);
 459         }
 460 }
 461
 462 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 463                                 enum cpuhp_state target)
 464 {
 465         enum cpuhp_state prev_state = st->state;
 466         int ret = 0;
 467
 468         for (; st->state > target; st->state--) {
 469                 ret = cpuhp_invoke_callback(cpu, st->state, false, NULL);
 470                 if (ret) {
 471                         st->target = prev_state;
 472                         undo_cpu_down(cpu, st);
 473                         break;
 474                 }
 475         }
 476         return ret;
 477 }
 478
 479 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 480 {
 481         for (st->state--; st->state > st->target; st->state--) {
 482                 struct cpuhp_step *step = cpuhp_get_step(st->state);
 483
 484                 if (!step->skip_onerr)
 485                         cpuhp_invoke_callback(cpu, st->state, false, NULL);
 486         }
 487 }
 488
 489 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
 490                               enum cpuhp_state target)
 491 {
 492         enum cpuhp_state prev_state = st->state;
 493         int ret = 0;
 494
 495         while (st->state < target) {
 496                 st->state++;
 497                 ret = cpuhp_invoke_callback(cpu, st->state, true, NULL);
 498                 if (ret) {
 499                         st->target = prev_state;
 500                         undo_cpu_up(cpu, st);
 501                         break;
 502                 }
 503         }
 504         return ret;
 505 }
 506
 507 /*
 508  * The cpu hotplug threads manage the bringup and teardown of the cpus
 509  */
 510 static void cpuhp_create(unsigned int cpu)
 511 {
 512         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 513
 514         init_completion(&st->done);
 515 }
 516
 517 static int cpuhp_should_run(unsigned int cpu)
 518 {
 519         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 520
 521         return st->should_run;
 522 }
 523
 524 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
 525 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
 526 {
 527         enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
 528
 529         return cpuhp_down_callbacks(cpu, st, target);
 530 }
 531
 532 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
 533 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
 534 {
 535         return cpuhp_up_callbacks(cpu, st, st->target);
 536 }
 537
 538 /*
 539  * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
 540  * callbacks when a state gets [un]installed at runtime.
 541  */
 542 static void cpuhp_thread_fun(unsigned int cpu)
 543 {
 544         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 545         int ret = 0;
 546
 547         /*
 548          * Paired with the mb() in cpuhp_kick_ap_work and
 549          * cpuhp_invoke_ap_callback, so the work set is consistent visible.
 550          */
 551         smp_mb();
 552         if (!st->should_run)
 553                 return;
 554
 555         st->should_run = false;
 556
 557         /* Single callback invocation for [un]install ? */
 558         if (st->single) {
 559                 if (st->cb_state < CPUHP_AP_ONLINE) {
 560                         local_irq_disable();
 561                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 562                                                     st->bringup, st->node);
 563                         local_irq_enable();
 564                 } else {
 565                         ret = cpuhp_invoke_callback(cpu, st->cb_state,
 566                                                     st->bringup, st->node);
 567                 }
 568         } else if (st->rollback) {
 569                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 570
 571                 undo_cpu_down(cpu, st);
 572                 /*
 573                  * This is a momentary workaround to keep the notifier users
 574                  * happy. Will go away once we got rid of the notifiers.
 575                  */
 576                 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
 577                 st->rollback = false;
 578         } else {
 579                 /* Cannot happen .... */
 580                 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
 581
 582                 /* Regular hotplug work */
 583                 if (st->state < st->target)
 584                         ret = cpuhp_ap_online(cpu, st);
 585                 else if (st->state > st->target)
 586                         ret = cpuhp_ap_offline(cpu, st);
 587         }
 588         st->result = ret;
 589         complete(&st->done);
 590 }
 591
 592 /* Invoke a single callback on a remote cpu */
 593 static int
 594 cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state, bool bringup,
 595                          struct hlist_node *node)
 596 {
 597         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 598
 599         if (!cpu_online(cpu))
 600                 return 0;
 601
 602         /*
 603          * If we are up and running, use the hotplug thread. For early calls
 604          * we invoke the thread function directly.
 605          */
 606         if (!st->thread)
 607                 return cpuhp_invoke_callback(cpu, state, bringup, node);
 608
 609         st->cb_state = state;
 610         st->single = true;
 611         st->bringup = bringup;
 612         st->node = node;
 613
 614         /*
 615          * Make sure the above stores are visible before should_run becomes
 616          * true. Paired with the mb() above in cpuhp_thread_fun()
 617          */
 618         smp_mb();
 619         st->should_run = true;
 620         wake_up_process(st->thread);
 621         wait_for_completion(&st->done);
 622         return st->result;
 623 }
 624
 625 /* Regular hotplug invocation of the AP hotplug thread */
 626 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
 627 {
 628         st->result = 0;
 629         st->single = false;
 630         /*
 631          * Make sure the above stores are visible before should_run becomes
 632          * true. Paired with the mb() above in cpuhp_thread_fun()
 633          */
 634         smp_mb();
 635         st->should_run = true;
 636         wake_up_process(st->thread);
 637 }
 638
 639 static int cpuhp_kick_ap_work(unsigned int cpu)
 640 {
 641         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 642         enum cpuhp_state state = st->state;
 643
 644         trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
 645         __cpuhp_kick_ap_work(st);
 646         wait_for_completion(&st->done);
 647         trace_cpuhp_exit(cpu, st->state, state, st->result);
 648         return st->result;
 649 }
 650
 651 static struct smp_hotplug_thread cpuhp_threads = {
 652         .store                  = &cpuhp_state.thread,
 653         .create                 = &cpuhp_create,
 654         .thread_should_run      = cpuhp_should_run,
 655         .thread_fn              = cpuhp_thread_fun,
 656         .thread_comm            = "cpuhp/%u",
 657         .selfparking            = true,
 658 };
 659
 660 void __init cpuhp_threads_init(void)
 661 {
 662         BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
 663         kthread_unpark(this_cpu_read(cpuhp_state.thread));
 664 }
 665
 666 #ifdef CONFIG_HOTPLUG_CPU
 667 EXPORT_SYMBOL(register_cpu_notifier);
 668 EXPORT_SYMBOL(__register_cpu_notifier);
 669 void unregister_cpu_notifier(struct notifier_block *nb)
 670 {
 671         cpu_maps_update_begin();
 672         raw_notifier_chain_unregister(&cpu_chain, nb);
 673         cpu_maps_update_done();
 674 }
 675 EXPORT_SYMBOL(unregister_cpu_notifier);
 676
 677 void __unregister_cpu_notifier(struct notifier_block *nb)
 678 {
 679         raw_notifier_chain_unregister(&cpu_chain, nb);
 680 }
 681 EXPORT_SYMBOL(__unregister_cpu_notifier);
 682
 683 /**
 684  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 685  * @cpu: a CPU id
 686  *
 687  * This function walks all processes, finds a valid mm struct for each one and
 688  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 689  * trivial, there are various non-obvious corner cases, which this function
 690  * tries to solve in a safe manner.
 691  *
 692  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 693  * be called only for an already offlined CPU.
 694  */
 695 void clear_tasks_mm_cpumask(int cpu)
 696 {
 697         struct task_struct *p;
 698
 699         /*
 700          * This function is called after the cpu is taken down and marked
 701          * offline, so its not like new tasks will ever get this cpu set in
 702          * their mm mask. -- Peter Zijlstra
 703          * Thus, we may use rcu_read_lock() here, instead of grabbing
 704          * full-fledged tasklist_lock.
 705          */
 706         WARN_ON(cpu_online(cpu));
 707         rcu_read_lock();
 708         for_each_process(p) {
 709                 struct task_struct *t;
 710
 711                 /*
 712                  * Main thread might exit, but other threads may still have
 713                  * a valid mm. Find one.
 714                  */
 715                 t = find_lock_task_mm(p);
 716                 if (!t)
 717                         continue;
 718                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 719                 task_unlock(t);
 720         }
 721         rcu_read_unlock();
 722 }
 723
 724 static inline void check_for_tasks(int dead_cpu)
 725 {
 726         struct task_struct *g, *p;
 727
 728         read_lock(&tasklist_lock);
 729         for_each_process_thread(g, p) {
 730                 if (!p->on_rq)
 731                         continue;
 732                 /*
 733                  * We do the check with unlocked task_rq(p)->lock.
 734                  * Order the reading to do not warn about a task,
 735                  * which was running on this cpu in the past, and
 736                  * it's just been woken on another cpu.
 737                  */
 738                 rmb();
 739                 if (task_cpu(p) != dead_cpu)
 740                         continue;
 741
 742                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 743                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 744         }
 745         read_unlock(&tasklist_lock);
 746 }
 747
 748 static int notify_down_prepare(unsigned int cpu)
 749 {
 750         int err, nr_calls = 0;
 751
 752         err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
 753         if (err) {
 754                 nr_calls--;
 755                 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
 756                 pr_warn("%s: attempt to take down CPU %u failed\n",
 757                                 __func__, cpu);
 758         }
 759         return err;
 760 }
 761
 762 static int notify_dying(unsigned int cpu)
 763 {
 764         cpu_notify(CPU_DYING, cpu);
 765         return 0;
 766 }
 767
 768 /* Take this CPU down. */
 769 static int take_cpu_down(void *_param)
 770 {
 771         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 772         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 773         int err, cpu = smp_processor_id();
 774
 775         /* Ensure this CPU doesn't handle any more interrupts. */
 776         err = __cpu_disable();
 777         if (err < 0)
 778                 return err;
 779
 780         /*
 781          * We get here while we are in CPUHP_TEARDOWN_CPU state and we must not
 782          * do this step again.
 783          */
 784         WARN_ON(st->state != CPUHP_TEARDOWN_CPU);
 785         st->state--;
 786         /* Invoke the former CPU_DYING callbacks */
 787         for (; st->state > target; st->state--)
 788                 cpuhp_invoke_callback(cpu, st->state, false, NULL);
 789
 790         /* Give up timekeeping duties */
 791         tick_handover_do_timer();
 792         /* Park the stopper thread */
 793         stop_machine_park(cpu);
 794         return 0;
 795 }
 796
 797 static int takedown_cpu(unsigned int cpu)
 798 {
 799         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 800         int err;
 801
 802         /* Park the smpboot threads */
 803         kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 804         smpboot_park_threads(cpu);
 805
 806         /*
 807          * Prevent irq alloc/free while the dying cpu reorganizes the
 808          * interrupt affinities.
 809          */
 810         irq_lock_sparse();
 811
 812         /*
 813          * So now all preempt/rcu users must observe !cpu_active().
 814          */
 815         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 816         if (err) {
 817                 /* CPU refused to die */
 818                 irq_unlock_sparse();
 819                 /* Unpark the hotplug thread so we can rollback there */
 820                 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
 821                 return err;
 822         }
 823         BUG_ON(cpu_online(cpu));
 824
 825         /*
 826          * The migration_call() CPU_DYING callback will have removed all
 827          * runnable tasks from the cpu, there's only the idle task left now
 828          * that the migration thread is done doing the stop_machine thing.
 829          *
 830          * Wait for the stop thread to go away.
 831          */
 832         wait_for_completion(&st->done);
 833         BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
 834
 835         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 836         irq_unlock_sparse();
 837
 838         hotplug_cpu__broadcast_tick_pull(cpu);
 839         /* This actually kills the CPU. */
 840         __cpu_die(cpu);
 841
 842         tick_cleanup_dead_cpu(cpu);
 843         return 0;
 844 }
 845
 846 static int notify_dead(unsigned int cpu)
 847 {
 848         cpu_notify_nofail(CPU_DEAD, cpu);
 849         check_for_tasks(cpu);
 850         return 0;
 851 }
 852
 853 static void cpuhp_complete_idle_dead(void *arg)
 854 {
 855         struct cpuhp_cpu_state *st = arg;
 856
 857         complete(&st->done);
 858 }
 859
 860 void cpuhp_report_idle_dead(void)
 861 {
 862         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 863
 864         BUG_ON(st->state != CPUHP_AP_OFFLINE);
 865         rcu_report_dead(smp_processor_id());
 866         st->state = CPUHP_AP_IDLE_DEAD;
 867         /*
 868          * We cannot call complete after rcu_report_dead() so we delegate it
 869          * to an online cpu.
 870          */
 871         smp_call_function_single(cpumask_first(cpu_online_mask),
 872                                  cpuhp_complete_idle_dead, st, 0);
 873 }
 874
 875 #else
 876 #define notify_down_prepare     NULL
 877 #define takedown_cpu            NULL
 878 #define notify_dead             NULL
 879 #define notify_dying            NULL
 880 #endif
 881
 882 #ifdef CONFIG_HOTPLUG_CPU
 883
 884 /* Requires cpu_add_remove_lock to be held */
 885 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 886                            enum cpuhp_state target)
 887 {
 888         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 889         int prev_state, ret = 0;
 890         bool hasdied = false;
 891
 892         if (num_online_cpus() == 1)
 893                 return -EBUSY;
 894
 895         if (!cpu_present(cpu))
 896                 return -EINVAL;
 897
 898         cpu_hotplug_begin();
 899
 900         cpuhp_tasks_frozen = tasks_frozen;
 901
 902         prev_state = st->state;
 903         st->target = target;
 904         /*
 905          * If the current CPU state is in the range of the AP hotplug thread,
 906          * then we need to kick the thread.
 907          */
 908         if (st->state > CPUHP_TEARDOWN_CPU) {
 909                 ret = cpuhp_kick_ap_work(cpu);
 910                 /*
 911                  * The AP side has done the error rollback already. Just
 912                  * return the error code..
 913                  */
 914                 if (ret)
 915                         goto out;
 916
 917                 /*
 918                  * We might have stopped still in the range of the AP hotplug
 919                  * thread. Nothing to do anymore.
 920                  */
 921                 if (st->state > CPUHP_TEARDOWN_CPU)
 922                         goto out;
 923         }
 924         /*
 925          * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
 926          * to do the further cleanups.
 927          */
 928         ret = cpuhp_down_callbacks(cpu, st, target);
 929         if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
 930                 st->target = prev_state;
 931                 st->rollback = true;
 932                 cpuhp_kick_ap_work(cpu);
 933         }
 934
 935         hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
 936 out:
 937         cpu_hotplug_done();
 938         /* This post dead nonsense must die */
 939         if (!ret && hasdied)
 940                 cpu_notify_nofail(CPU_POST_DEAD, cpu);
 941         return ret;
 942 }
 943
 944 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 945 {
 946         int err;
 947
 948         cpu_maps_update_begin();
 949
 950         if (cpu_hotplug_disabled) {
 951                 err = -EBUSY;
 952                 goto out;
 953         }
 954
 955         err = _cpu_down(cpu, 0, target);
 956
 957 out:
 958         cpu_maps_update_done();
 959         return err;
 960 }
 961 int cpu_down(unsigned int cpu)
 962 {
 963         return do_cpu_down(cpu, CPUHP_OFFLINE);
 964 }
 965 EXPORT_SYMBOL(cpu_down);
 966 #endif /*CONFIG_HOTPLUG_CPU*/
 967
 968 /**
 969  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 970  * @cpu: cpu that just started
 971  *
 972  * This function calls the cpu_chain notifiers with CPU_STARTING.
 973  * It must be called by the arch code on the new cpu, before the new cpu
 974  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 975  */
 976 void notify_cpu_starting(unsigned int cpu)
 977 {
 978         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 979         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 980
 981         while (st->state < target) {
 982                 st->state++;
 983                 cpuhp_invoke_callback(cpu, st->state, true, NULL);
 984         }
 985 }
 986
 987 /*
 988  * Called from the idle task. We need to set active here, so we can kick off
 989  * the stopper thread and unpark the smpboot threads. If the target state is
 990  * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
 991  * cpu further.
 992  */
 993 void cpuhp_online_idle(enum cpuhp_state state)
 994 {
 995         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 996         unsigned int cpu = smp_processor_id();
 997
 998         /* Happens for the boot cpu */
 999         if (state != CPUHP_AP_ONLINE_IDLE)
1000                 return;
1001
1002         st->state = CPUHP_AP_ONLINE_IDLE;
1003
1004         /* Unpark the stopper thread and the hotplug thread of this cpu */
1005         stop_machine_unpark(cpu);
1006         kthread_unpark(st->thread);
1007
1008         /* Should we go further up ? */
1009         if (st->target > CPUHP_AP_ONLINE_IDLE)
1010                 __cpuhp_kick_ap_work(st);
1011         else
1012                 complete(&st->done);
1013 }
1014
1015 /* Requires cpu_add_remove_lock to be held */
1016 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
1017 {
1018         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1019         struct task_struct *idle;
1020         int ret = 0;
1021
1022         cpu_hotplug_begin();
1023
1024         if (!cpu_present(cpu)) {
1025                 ret = -EINVAL;
1026                 goto out;
1027         }
1028
1029         /*
1030          * The caller of do_cpu_up might have raced with another
1031          * caller. Ignore it for now.
1032          */
1033         if (st->state >= target)
1034                 goto out;
1035
1036         if (st->state == CPUHP_OFFLINE) {
1037                 /* Let it fail before we try to bring the cpu up */
1038                 idle = idle_thread_get(cpu);
1039                 if (IS_ERR(idle)) {
1040                         ret = PTR_ERR(idle);
1041                         goto out;
1042                 }
1043         }
1044
1045         cpuhp_tasks_frozen = tasks_frozen;
1046
1047         st->target = target;
1048         /*
1049          * If the current CPU state is in the range of the AP hotplug thread,
1050          * then we need to kick the thread once more.
1051          */
1052         if (st->state > CPUHP_BRINGUP_CPU) {
1053                 ret = cpuhp_kick_ap_work(cpu);
1054                 /*
1055                  * The AP side has done the error rollback already. Just
1056                  * return the error code..
1057                  */
1058                 if (ret)
1059                         goto out;
1060         }
1061
1062         /*
1063          * Try to reach the target state. We max out on the BP at
1064          * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
1065          * responsible for bringing it up to the target state.
1066          */
1067         target = min((int)target, CPUHP_BRINGUP_CPU);
1068         ret = cpuhp_up_callbacks(cpu, st, target);
1069 out:
1070         cpu_hotplug_done();
1071         return ret;
1072 }
1073
1074 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1075 {
1076         int err = 0;
1077
1078         if (!cpu_possible(cpu)) {
1079                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1080                        cpu);
1081 #if defined(CONFIG_IA64)
1082                 pr_err("please check additional_cpus= boot parameter\n");
1083 #endif
1084                 return -EINVAL;
1085         }
1086
1087         err = try_online_node(cpu_to_node(cpu));
1088         if (err)
1089                 return err;
1090
1091         cpu_maps_update_begin();
1092
1093         if (cpu_hotplug_disabled) {
1094                 err = -EBUSY;
1095                 goto out;
1096         }
1097
1098         err = _cpu_up(cpu, 0, target);
1099 out:
1100         cpu_maps_update_done();
1101         return err;
1102 }
1103
1104 int cpu_up(unsigned int cpu)
1105 {
1106         return do_cpu_up(cpu, CPUHP_ONLINE);
1107 }
1108 EXPORT_SYMBOL_GPL(cpu_up);
1109
1110 #ifdef CONFIG_PM_SLEEP_SMP
1111 static cpumask_var_t frozen_cpus;
1112
1113 int disable_nonboot_cpus(void)
1114 {
1115         int cpu, first_cpu, error = 0;
1116
1117         cpu_maps_update_begin();
1118         first_cpu = cpumask_first(cpu_online_mask);
1119         /*
1120          * We take down all of the non-boot CPUs in one shot to avoid races
1121          * with the userspace trying to use the CPU hotplug at the same time
1122          */
1123         cpumask_clear(frozen_cpus);
1124
1125         pr_info("Disabling non-boot CPUs ...\n");
1126         for_each_online_cpu(cpu) {
1127                 if (cpu == first_cpu)
1128                         continue;
1129                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1130                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1131                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1132                 if (!error)
1133                         cpumask_set_cpu(cpu, frozen_cpus);
1134                 else {
1135                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
1136                         break;
1137                 }
1138         }
1139
1140         if (!error)
1141                 BUG_ON(num_online_cpus() > 1);
1142         else
1143                 pr_err("Non-boot CPUs are not disabled\n");
1144
1145         /*
1146          * Make sure the CPUs won't be enabled by someone else. We need to do
1147          * this even in case of failure as all disable_nonboot_cpus() users are
1148          * supposed to do enable_nonboot_cpus() on the failure path.
1149          */
1150         cpu_hotplug_disabled++;
1151
1152         cpu_maps_update_done();
1153         return error;
1154 }
1155
1156 void __weak arch_enable_nonboot_cpus_begin(void)
1157 {
1158 }
1159
1160 void __weak arch_enable_nonboot_cpus_end(void)
1161 {
1162 }
1163
1164 void enable_nonboot_cpus(void)
1165 {
1166         int cpu, error;
1167
1168         /* Allow everyone to use the CPU hotplug again */
1169         cpu_maps_update_begin();
1170         __cpu_hotplug_enable();
1171         if (cpumask_empty(frozen_cpus))
1172                 goto out;
1173
1174         pr_info("Enabling non-boot CPUs ...\n");
1175
1176         arch_enable_nonboot_cpus_begin();
1177
1178         for_each_cpu(cpu, frozen_cpus) {
1179                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1180                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1181                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1182                 if (!error) {
1183                         pr_info("CPU%d is up\n", cpu);
1184                         continue;
1185                 }
1186                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1187         }
1188
1189         arch_enable_nonboot_cpus_end();
1190
1191         cpumask_clear(frozen_cpus);
1192 out:
1193         cpu_maps_update_done();
1194 }
1195
1196 static int __init alloc_frozen_cpus(void)
1197 {
1198         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1199                 return -ENOMEM;
1200         return 0;
1201 }
1202 core_initcall(alloc_frozen_cpus);
1203
1204 /*
1205  * When callbacks for CPU hotplug notifications are being executed, we must
1206  * ensure that the state of the system with respect to the tasks being frozen
1207  * or not, as reported by the notification, remains unchanged *throughout the
1208  * duration* of the execution of the callbacks.
1209  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1210  *
1211  * This synchronization is implemented by mutually excluding regular CPU
1212  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1213  * Hibernate notifications.
1214  */
1215 static int
1216 cpu_hotplug_pm_callback(struct notifier_block *nb,
1217                         unsigned long action, void *ptr)
1218 {
1219         switch (action) {
1220
1221         case PM_SUSPEND_PREPARE:
1222         case PM_HIBERNATION_PREPARE:
1223                 cpu_hotplug_disable();
1224                 break;
1225
1226         case PM_POST_SUSPEND:
1227         case PM_POST_HIBERNATION:
1228                 cpu_hotplug_enable();
1229                 break;
1230
1231         default:
1232                 return NOTIFY_DONE;
1233         }
1234
1235         return NOTIFY_OK;
1236 }
1237
1238
1239 static int __init cpu_hotplug_pm_sync_init(void)
1240 {
1241         /*
1242          * cpu_hotplug_pm_callback has higher priority than x86
1243          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1244          * to disable cpu hotplug to avoid cpu hotplug race.
1245          */
1246         pm_notifier(cpu_hotplug_pm_callback, 0);
1247         return 0;
1248 }
1249 core_initcall(cpu_hotplug_pm_sync_init);
1250
1251 #endif /* CONFIG_PM_SLEEP_SMP */
1252
1253 #endif /* CONFIG_SMP */
1254
1255 /* Boot processor state steps */
1256 static struct cpuhp_step cpuhp_bp_states[] = {
1257         [CPUHP_OFFLINE] = {
1258                 .name                   = "offline",
1259                 .startup                = NULL,
1260                 .teardown               = NULL,
1261         },
1262 #ifdef CONFIG_SMP
1263         [CPUHP_CREATE_THREADS]= {
1264                 .name                   = "threads:create",
1265                 .startup                = smpboot_create_threads,
1266                 .teardown               = NULL,
1267                 .cant_stop              = true,
1268         },
1269         [CPUHP_PERF_PREPARE] = {
1270                 .name = "perf prepare",
1271                 .startup = perf_event_init_cpu,
1272                 .teardown = perf_event_exit_cpu,
1273         },
1274         [CPUHP_WORKQUEUE_PREP] = {
1275                 .name = "workqueue prepare",
1276                 .startup = workqueue_prepare_cpu,
1277                 .teardown = NULL,
1278         },
1279         [CPUHP_HRTIMERS_PREPARE] = {
1280                 .name = "hrtimers prepare",
1281                 .startup = hrtimers_prepare_cpu,
1282                 .teardown = hrtimers_dead_cpu,
1283         },
1284         [CPUHP_SMPCFD_PREPARE] = {
1285                 .name = "SMPCFD prepare",
1286                 .startup = smpcfd_prepare_cpu,
1287                 .teardown = smpcfd_dead_cpu,
1288         },
1289         [CPUHP_RCUTREE_PREP] = {
1290                 .name = "RCU-tree prepare",
1291                 .startup = rcutree_prepare_cpu,
1292                 .teardown = 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                = notify_prepare,
1301                 .teardown               = 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 = NULL,
1313                 .teardown = timers_dead_cpu,
1314         },
1315         /* Kicks the plugged cpu into life */
1316         [CPUHP_BRINGUP_CPU] = {
1317                 .name                   = "cpu:bringup",
1318                 .startup                = bringup_cpu,
1319                 .teardown               = NULL,
1320                 .cant_stop              = true,
1321         },
1322         [CPUHP_AP_SMPCFD_DYING] = {
1323                 .startup = NULL,
1324                 .teardown = smpcfd_dying_cpu,
1325         },
1326         /*
1327          * Handled on controll processor until the plugged processor manages
1328          * this itself.
1329          */
1330         [CPUHP_TEARDOWN_CPU] = {
1331                 .name                   = "cpu:teardown",
1332                 .startup                = NULL,
1333                 .teardown               = takedown_cpu,
1334                 .cant_stop              = true,
1335         },
1336 #else
1337         [CPUHP_BRINGUP_CPU] = { },
1338 #endif
1339 };
1340
1341 /* Application processor state steps */
1342 static struct cpuhp_step cpuhp_ap_states[] = {
1343 #ifdef CONFIG_SMP
1344         /* Final state before CPU kills itself */
1345         [CPUHP_AP_IDLE_DEAD] = {
1346                 .name                   = "idle:dead",
1347         },
1348         /*
1349          * Last state before CPU enters the idle loop to die. Transient state
1350          * for synchronization.
1351          */
1352         [CPUHP_AP_OFFLINE] = {
1353                 .name                   = "ap:offline",
1354                 .cant_stop              = true,
1355         },
1356         /* First state is scheduler control. Interrupts are disabled */
1357         [CPUHP_AP_SCHED_STARTING] = {
1358                 .name                   = "sched:starting",
1359                 .startup                = sched_cpu_starting,
1360                 .teardown               = sched_cpu_dying,
1361         },
1362         [CPUHP_AP_RCUTREE_DYING] = {
1363                 .startup = NULL,
1364                 .teardown = rcutree_dying_cpu,
1365         },
1366         /*
1367          * Low level startup/teardown notifiers. Run with interrupts
1368          * disabled. Will be removed once the notifiers are converted to
1369          * states.
1370          */
1371         [CPUHP_AP_NOTIFY_STARTING] = {
1372                 .name                   = "notify:starting",
1373                 .startup                = notify_starting,
1374                 .teardown               = notify_dying,
1375                 .skip_onerr             = true,
1376                 .cant_stop              = true,
1377         },
1378         /* Entry state on starting. Interrupts enabled from here on. Transient
1379          * state for synchronsization */
1380         [CPUHP_AP_ONLINE] = {
1381                 .name                   = "ap:online",
1382         },
1383         /* Handle smpboot threads park/unpark */
1384         [CPUHP_AP_SMPBOOT_THREADS] = {
1385                 .name                   = "smpboot:threads",
1386                 .startup                = smpboot_unpark_threads,
1387                 .teardown               = NULL,
1388         },
1389         [CPUHP_AP_PERF_ONLINE] = {
1390                 .name = "perf online",
1391                 .startup = perf_event_init_cpu,
1392                 .teardown = perf_event_exit_cpu,
1393         },
1394         [CPUHP_AP_WORKQUEUE_ONLINE] = {
1395                 .name = "workqueue online",
1396                 .startup = workqueue_online_cpu,
1397                 .teardown = workqueue_offline_cpu,
1398         },
1399         [CPUHP_AP_RCUTREE_ONLINE] = {
1400                 .name = "RCU-tree online",
1401                 .startup = rcutree_online_cpu,
1402                 .teardown = rcutree_offline_cpu,
1403         },
1404
1405         /*
1406          * Online/down_prepare notifiers. Will be removed once the notifiers
1407          * are converted to states.
1408          */
1409         [CPUHP_AP_NOTIFY_ONLINE] = {
1410                 .name                   = "notify:online",
1411                 .startup                = notify_online,
1412                 .teardown               = notify_down_prepare,
1413                 .skip_onerr             = true,
1414         },
1415 #endif
1416         /*
1417          * The dynamically registered state space is here
1418          */
1419
1420 #ifdef CONFIG_SMP
1421         /* Last state is scheduler control setting the cpu active */
1422         [CPUHP_AP_ACTIVE] = {
1423                 .name                   = "sched:active",
1424                 .startup                = sched_cpu_activate,
1425                 .teardown               = sched_cpu_deactivate,
1426         },
1427 #endif
1428
1429         /* CPU is fully up and running. */
1430         [CPUHP_ONLINE] = {
1431                 .name                   = "online",
1432                 .startup                = NULL,
1433                 .teardown               = NULL,
1434         },
1435 };
1436
1437 /* Sanity check for callbacks */
1438 static int cpuhp_cb_check(enum cpuhp_state state)
1439 {
1440         if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1441                 return -EINVAL;
1442         return 0;
1443 }
1444
1445 static void cpuhp_store_callbacks(enum cpuhp_state state,
1446                                   const char *name,
1447                                   int (*startup)(unsigned int cpu),
1448                                   int (*teardown)(unsigned int cpu),
1449                                   bool multi_instance)
1450 {
1451         /* (Un)Install the callbacks for further cpu hotplug operations */
1452         struct cpuhp_step *sp;
1453
1454         mutex_lock(&cpuhp_state_mutex);
1455         sp = cpuhp_get_step(state);
1456         sp->startup = startup;
1457         sp->teardown = teardown;
1458         sp->name = name;
1459         sp->multi_instance = multi_instance;
1460         INIT_HLIST_HEAD(&sp->list);
1461         mutex_unlock(&cpuhp_state_mutex);
1462 }
1463
1464 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1465 {
1466         return cpuhp_get_step(state)->teardown;
1467 }
1468
1469 /*
1470  * Call the startup/teardown function for a step either on the AP or
1471  * on the current CPU.
1472  */
1473 static int cpuhp_issue_call(int cpu, enum cpuhp_state state, bool bringup,
1474                             struct hlist_node *node)
1475 {
1476         struct cpuhp_step *sp = cpuhp_get_step(state);
1477         int ret;
1478
1479         if ((bringup && !sp->startup) || (!bringup && !sp->teardown))
1480                 return 0;
1481         /*
1482          * The non AP bound callbacks can fail on bringup. On teardown
1483          * e.g. module removal we crash for now.
1484          */
1485 #ifdef CONFIG_SMP
1486         if (cpuhp_is_ap_state(state))
1487                 ret = cpuhp_invoke_ap_callback(cpu, state, bringup, node);
1488         else
1489                 ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1490 #else
1491         ret = cpuhp_invoke_callback(cpu, state, bringup, node);
1492 #endif
1493         BUG_ON(ret && !bringup);
1494         return ret;
1495 }
1496
1497 /*
1498  * Called from __cpuhp_setup_state on a recoverable failure.
1499  *
1500  * Note: The teardown callbacks for rollback are not allowed to fail!
1501  */
1502 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1503                                    struct hlist_node *node)
1504 {
1505         int cpu;
1506
1507         /* Roll back the already executed steps on the other cpus */
1508         for_each_present_cpu(cpu) {
1509                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1510                 int cpustate = st->state;
1511
1512                 if (cpu >= failedcpu)
1513                         break;
1514
1515                 /* Did we invoke the startup call on that cpu ? */
1516                 if (cpustate >= state)
1517                         cpuhp_issue_call(cpu, state, false, node);
1518         }
1519 }
1520
1521 /*
1522  * Returns a free for dynamic slot assignment of the Online state. The states
1523  * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1524  * by having no name assigned.
1525  */
1526 static int cpuhp_reserve_state(enum cpuhp_state state)
1527 {
1528         enum cpuhp_state i;
1529
1530         mutex_lock(&cpuhp_state_mutex);
1531         for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1532                 if (cpuhp_ap_states[i].name)
1533                         continue;
1534
1535                 cpuhp_ap_states[i].name = "Reserved";
1536                 mutex_unlock(&cpuhp_state_mutex);
1537                 return i;
1538         }
1539         mutex_unlock(&cpuhp_state_mutex);
1540         WARN(1, "No more dynamic states available for CPU hotplug\n");
1541         return -ENOSPC;
1542 }
1543
1544 int __cpuhp_state_add_instance(enum cpuhp_state state, struct hlist_node *node,
1545                                bool invoke)
1546 {
1547         struct cpuhp_step *sp;
1548         int cpu;
1549         int ret;
1550
1551         sp = cpuhp_get_step(state);
1552         if (sp->multi_instance == false)
1553                 return -EINVAL;
1554
1555         get_online_cpus();
1556
1557         if (!invoke || !sp->startup_multi)
1558                 goto add_node;
1559
1560         /*
1561          * Try to call the startup callback for each present cpu
1562          * depending on the hotplug state of the cpu.
1563          */
1564         for_each_present_cpu(cpu) {
1565                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1566                 int cpustate = st->state;
1567
1568                 if (cpustate < state)
1569                         continue;
1570
1571                 ret = cpuhp_issue_call(cpu, state, true, node);
1572                 if (ret) {
1573                         if (sp->teardown_multi)
1574                                 cpuhp_rollback_install(cpu, state, node);
1575                         goto err;
1576                 }
1577         }
1578 add_node:
1579         ret = 0;
1580         mutex_lock(&cpuhp_state_mutex);
1581         hlist_add_head(node, &sp->list);
1582         mutex_unlock(&cpuhp_state_mutex);
1583
1584 err:
1585         put_online_cpus();
1586         return ret;
1587 }
1588 EXPORT_SYMBOL_GPL(__cpuhp_state_add_instance);
1589
1590 /**
1591  * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1592  * @state:      The state to setup
1593  * @invoke:     If true, the startup function is invoked for cpus where
1594  *              cpu state >= @state
1595  * @startup:    startup callback function
1596  * @teardown:   teardown callback function
1597  *
1598  * Returns 0 if successful, otherwise a proper error code
1599  */
1600 int __cpuhp_setup_state(enum cpuhp_state state,
1601                         const char *name, bool invoke,
1602                         int (*startup)(unsigned int cpu),
1603                         int (*teardown)(unsigned int cpu),
1604                         bool multi_instance)
1605 {
1606         int cpu, ret = 0;
1607         int dyn_state = 0;
1608
1609         if (cpuhp_cb_check(state) || !name)
1610                 return -EINVAL;
1611
1612         get_online_cpus();
1613
1614         /* currently assignments for the ONLINE state are possible */
1615         if (state == CPUHP_AP_ONLINE_DYN) {
1616                 dyn_state = 1;
1617                 ret = cpuhp_reserve_state(state);
1618                 if (ret < 0)
1619                         goto out;
1620                 state = ret;
1621         }
1622
1623         cpuhp_store_callbacks(state, name, startup, teardown, multi_instance);
1624
1625         if (!invoke || !startup)
1626                 goto out;
1627
1628         /*
1629          * Try to call the startup callback for each present cpu
1630          * depending on the hotplug state of the cpu.
1631          */
1632         for_each_present_cpu(cpu) {
1633                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1634                 int cpustate = st->state;
1635
1636                 if (cpustate < state)
1637                         continue;
1638
1639                 ret = cpuhp_issue_call(cpu, state, true, NULL);
1640                 if (ret) {
1641                         if (teardown)
1642                                 cpuhp_rollback_install(cpu, state, NULL);
1643                         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1644                         goto out;
1645                 }
1646         }
1647 out:
1648         put_online_cpus();
1649         if (!ret && dyn_state)
1650                 return state;
1651         return ret;
1652 }
1653 EXPORT_SYMBOL(__cpuhp_setup_state);
1654
1655 int __cpuhp_state_remove_instance(enum cpuhp_state state,
1656                                   struct hlist_node *node, bool invoke)
1657 {
1658         struct cpuhp_step *sp = cpuhp_get_step(state);
1659         int cpu;
1660
1661         BUG_ON(cpuhp_cb_check(state));
1662
1663         if (!sp->multi_instance)
1664                 return -EINVAL;
1665
1666         get_online_cpus();
1667         if (!invoke || !cpuhp_get_teardown_cb(state))
1668                 goto remove;
1669         /*
1670          * Call the teardown callback for each present cpu depending
1671          * on the hotplug state of the cpu. This function is not
1672          * allowed to fail currently!
1673          */
1674         for_each_present_cpu(cpu) {
1675                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1676                 int cpustate = st->state;
1677
1678                 if (cpustate >= state)
1679                         cpuhp_issue_call(cpu, state, false, node);
1680         }
1681
1682 remove:
1683         mutex_lock(&cpuhp_state_mutex);
1684         hlist_del(node);
1685         mutex_unlock(&cpuhp_state_mutex);
1686         put_online_cpus();
1687
1688         return 0;
1689 }
1690 EXPORT_SYMBOL_GPL(__cpuhp_state_remove_instance);
1691 /**
1692  * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1693  * @state:      The state to remove
1694  * @invoke:     If true, the teardown function is invoked for cpus where
1695  *              cpu state >= @state
1696  *
1697  * The teardown callback is currently not allowed to fail. Think
1698  * about module removal!
1699  */
1700 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1701 {
1702         struct cpuhp_step *sp = cpuhp_get_step(state);
1703         int cpu;
1704
1705         BUG_ON(cpuhp_cb_check(state));
1706
1707         get_online_cpus();
1708
1709         if (sp->multi_instance) {
1710                 WARN(!hlist_empty(&sp->list),
1711                      "Error: Removing state %d which has instances left.\n",
1712                      state);
1713                 goto remove;
1714         }
1715
1716         if (!invoke || !cpuhp_get_teardown_cb(state))
1717                 goto remove;
1718
1719         /*
1720          * Call the teardown callback for each present cpu depending
1721          * on the hotplug state of the cpu. This function is not
1722          * allowed to fail currently!
1723          */
1724         for_each_present_cpu(cpu) {
1725                 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1726                 int cpustate = st->state;
1727
1728                 if (cpustate >= state)
1729                         cpuhp_issue_call(cpu, state, false, NULL);
1730         }
1731 remove:
1732         cpuhp_store_callbacks(state, NULL, NULL, NULL, false);
1733         put_online_cpus();
1734 }
1735 EXPORT_SYMBOL(__cpuhp_remove_state);
1736
1737 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1738 static ssize_t show_cpuhp_state(struct device *dev,
1739                                 struct device_attribute *attr, char *buf)
1740 {
1741         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1742
1743         return sprintf(buf, "%d\n", st->state);
1744 }
1745 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1746
1747 static ssize_t write_cpuhp_target(struct device *dev,
1748                                   struct device_attribute *attr,
1749                                   const char *buf, size_t count)
1750 {
1751         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1752         struct cpuhp_step *sp;
1753         int target, ret;
1754
1755         ret = kstrtoint(buf, 10, &target);
1756         if (ret)
1757                 return ret;
1758
1759 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1760         if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1761                 return -EINVAL;
1762 #else
1763         if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1764                 return -EINVAL;
1765 #endif
1766
1767         ret = lock_device_hotplug_sysfs();
1768         if (ret)
1769                 return ret;
1770
1771         mutex_lock(&cpuhp_state_mutex);
1772         sp = cpuhp_get_step(target);
1773         ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1774         mutex_unlock(&cpuhp_state_mutex);
1775         if (ret)
1776                 return ret;
1777
1778         if (st->state < target)
1779                 ret = do_cpu_up(dev->id, target);
1780         else
1781                 ret = do_cpu_down(dev->id, target);
1782
1783         unlock_device_hotplug();
1784         return ret ? ret : count;
1785 }
1786
1787 static ssize_t show_cpuhp_target(struct device *dev,
1788                                  struct device_attribute *attr, char *buf)
1789 {
1790         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1791
1792         return sprintf(buf, "%d\n", st->target);
1793 }
1794 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1795
1796 static struct attribute *cpuhp_cpu_attrs[] = {
1797         &dev_attr_state.attr,
1798         &dev_attr_target.attr,
1799         NULL
1800 };
1801
1802 static struct attribute_group cpuhp_cpu_attr_group = {
1803         .attrs = cpuhp_cpu_attrs,
1804         .name = "hotplug",
1805         NULL
1806 };
1807
1808 static ssize_t show_cpuhp_states(struct device *dev,
1809                                  struct device_attribute *attr, char *buf)
1810 {
1811         ssize_t cur, res = 0;
1812         int i;
1813
1814         mutex_lock(&cpuhp_state_mutex);
1815         for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1816                 struct cpuhp_step *sp = cpuhp_get_step(i);
1817
1818                 if (sp->name) {
1819                         cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1820                         buf += cur;
1821                         res += cur;
1822                 }
1823         }
1824         mutex_unlock(&cpuhp_state_mutex);
1825         return res;
1826 }
1827 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1828
1829 static struct attribute *cpuhp_cpu_root_attrs[] = {
1830         &dev_attr_states.attr,
1831         NULL
1832 };
1833
1834 static struct attribute_group cpuhp_cpu_root_attr_group = {
1835         .attrs = cpuhp_cpu_root_attrs,
1836         .name = "hotplug",
1837         NULL
1838 };
1839
1840 static int __init cpuhp_sysfs_init(void)
1841 {
1842         int cpu, ret;
1843
1844         ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1845                                  &cpuhp_cpu_root_attr_group);
1846         if (ret)
1847                 return ret;
1848
1849         for_each_possible_cpu(cpu) {
1850                 struct device *dev = get_cpu_device(cpu);
1851
1852                 if (!dev)
1853                         continue;
1854                 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1855                 if (ret)
1856                         return ret;
1857         }
1858         return 0;
1859 }
1860 device_initcall(cpuhp_sysfs_init);
1861 #endif
1862
1863 /*
1864  * cpu_bit_bitmap[] is a special, "compressed" data structure that
1865  * represents all NR_CPUS bits binary values of 1<<nr.
1866  *
1867  * It is used by cpumask_of() to get a constant address to a CPU
1868  * mask value that has a single bit set only.
1869  */
1870
1871 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1872 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
1873 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1874 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1875 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1876
1877 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1878
1879         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
1880         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
1881 #if BITS_PER_LONG > 32
1882         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
1883         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
1884 #endif
1885 };
1886 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1887
1888 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1889 EXPORT_SYMBOL(cpu_all_bits);
1890
1891 #ifdef CONFIG_INIT_ALL_POSSIBLE
1892 struct cpumask __cpu_possible_mask __read_mostly
1893         = {CPU_BITS_ALL};
1894 #else
1895 struct cpumask __cpu_possible_mask __read_mostly;
1896 #endif
1897 EXPORT_SYMBOL(__cpu_possible_mask);
1898
1899 struct cpumask __cpu_online_mask __read_mostly;
1900 EXPORT_SYMBOL(__cpu_online_mask);
1901
1902 struct cpumask __cpu_present_mask __read_mostly;
1903 EXPORT_SYMBOL(__cpu_present_mask);
1904
1905 struct cpumask __cpu_active_mask __read_mostly;
1906 EXPORT_SYMBOL(__cpu_active_mask);
1907
1908 void init_cpu_present(const struct cpumask *src)
1909 {
1910         cpumask_copy(&__cpu_present_mask, src);
1911 }
1912
1913 void init_cpu_possible(const struct cpumask *src)
1914 {
1915         cpumask_copy(&__cpu_possible_mask, src);
1916 }
1917
1918 void init_cpu_online(const struct cpumask *src)
1919 {
1920         cpumask_copy(&__cpu_online_mask, src);
1921 }
1922
1923 /*
1924  * Activate the first processor.
1925  */
1926 void __init boot_cpu_init(void)
1927 {
1928         int cpu = smp_processor_id();
1929
1930         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1931         set_cpu_online(cpu, true);
1932         set_cpu_active(cpu, true);
1933         set_cpu_present(cpu, true);
1934         set_cpu_possible(cpu, true);
1935 }
1936
1937 /*
1938  * Must be called _AFTER_ setting up the per_cpu areas
1939  */
1940 void __init boot_cpu_state_init(void)
1941 {
1942         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1943 }