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