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
  26 #include <trace/events/power.h>
  27 #define CREATE_TRACE_POINTS
  28 #include <trace/events/cpuhp.h>
  29
  30 #include "smpboot.h"
  31
  32 /**
  33  * cpuhp_cpu_state - Per cpu hotplug state storage
  34  * @state:      The current cpu state
  35  * @target:     The target state
  36  */
  37 struct cpuhp_cpu_state {
  38         enum cpuhp_state        state;
  39         enum cpuhp_state        target;
  40 };
  41
  42 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
  43
  44 /**
  45  * cpuhp_step - Hotplug state machine step
  46  * @name:       Name of the step
  47  * @startup:    Startup function of the step
  48  * @teardown:   Teardown function of the step
  49  * @skip_onerr: Do not invoke the functions on error rollback
  50  *              Will go away once the notifiers are gone
  51  */
  52 struct cpuhp_step {
  53         const char      *name;
  54         int             (*startup)(unsigned int cpu);
  55         int             (*teardown)(unsigned int cpu);
  56         bool            skip_onerr;
  57 };
  58
  59 static struct cpuhp_step cpuhp_bp_states[];
  60 static struct cpuhp_step cpuhp_ap_states[];
  61
  62 /**
  63  * cpuhp_invoke_callback _ Invoke the callbacks for a given state
  64  * @cpu:        The cpu for which the callback should be invoked
  65  * @step:       The step in the state machine
  66  * @cb:         The callback function to invoke
  67  *
  68  * Called from cpu hotplug and from the state register machinery
  69  */
  70 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state step,
  71                                  int (*cb)(unsigned int))
  72 {
  73         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
  74         int ret = 0;
  75
  76         if (cb) {
  77                 trace_cpuhp_enter(cpu, st->target, step, cb);
  78                 ret = cb(cpu);
  79                 trace_cpuhp_exit(cpu, st->state, step, ret);
  80         }
  81         return ret;
  82 }
  83
  84 #ifdef CONFIG_SMP
  85 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
  86 static DEFINE_MUTEX(cpu_add_remove_lock);
  87 bool cpuhp_tasks_frozen;
  88 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
  89
  90 /*
  91  * The following two APIs (cpu_maps_update_begin/done) must be used when
  92  * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
  93  * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
  94  * hotplug callback (un)registration performed using __register_cpu_notifier()
  95  * or __unregister_cpu_notifier().
  96  */
  97 void cpu_maps_update_begin(void)
  98 {
  99         mutex_lock(&cpu_add_remove_lock);
 100 }
 101 EXPORT_SYMBOL(cpu_notifier_register_begin);
 102
 103 void cpu_maps_update_done(void)
 104 {
 105         mutex_unlock(&cpu_add_remove_lock);
 106 }
 107 EXPORT_SYMBOL(cpu_notifier_register_done);
 108
 109 static RAW_NOTIFIER_HEAD(cpu_chain);
 110
 111 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
 112  * Should always be manipulated under cpu_add_remove_lock
 113  */
 114 static int cpu_hotplug_disabled;
 115
 116 #ifdef CONFIG_HOTPLUG_CPU
 117
 118 static struct {
 119         struct task_struct *active_writer;
 120         /* wait queue to wake up the active_writer */
 121         wait_queue_head_t wq;
 122         /* verifies that no writer will get active while readers are active */
 123         struct mutex lock;
 124         /*
 125          * Also blocks the new readers during
 126          * an ongoing cpu hotplug operation.
 127          */
 128         atomic_t refcount;
 129
 130 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 131         struct lockdep_map dep_map;
 132 #endif
 133 } cpu_hotplug = {
 134         .active_writer = NULL,
 135         .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
 136         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
 137 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 138         .dep_map = {.name = "cpu_hotplug.lock" },
 139 #endif
 140 };
 141
 142 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
 143 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
 144 #define cpuhp_lock_acquire_tryread() \
 145                                   lock_map_acquire_tryread(&cpu_hotplug.dep_map)
 146 #define cpuhp_lock_acquire()      lock_map_acquire(&cpu_hotplug.dep_map)
 147 #define cpuhp_lock_release()      lock_map_release(&cpu_hotplug.dep_map)
 148
 149
 150 void get_online_cpus(void)
 151 {
 152         might_sleep();
 153         if (cpu_hotplug.active_writer == current)
 154                 return;
 155         cpuhp_lock_acquire_read();
 156         mutex_lock(&cpu_hotplug.lock);
 157         atomic_inc(&cpu_hotplug.refcount);
 158         mutex_unlock(&cpu_hotplug.lock);
 159 }
 160 EXPORT_SYMBOL_GPL(get_online_cpus);
 161
 162 void put_online_cpus(void)
 163 {
 164         int refcount;
 165
 166         if (cpu_hotplug.active_writer == current)
 167                 return;
 168
 169         refcount = atomic_dec_return(&cpu_hotplug.refcount);
 170         if (WARN_ON(refcount < 0)) /* try to fix things up */
 171                 atomic_inc(&cpu_hotplug.refcount);
 172
 173         if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
 174                 wake_up(&cpu_hotplug.wq);
 175
 176         cpuhp_lock_release();
 177
 178 }
 179 EXPORT_SYMBOL_GPL(put_online_cpus);
 180
 181 /*
 182  * This ensures that the hotplug operation can begin only when the
 183  * refcount goes to zero.
 184  *
 185  * Note that during a cpu-hotplug operation, the new readers, if any,
 186  * will be blocked by the cpu_hotplug.lock
 187  *
 188  * Since cpu_hotplug_begin() is always called after invoking
 189  * cpu_maps_update_begin(), we can be sure that only one writer is active.
 190  *
 191  * Note that theoretically, there is a possibility of a livelock:
 192  * - Refcount goes to zero, last reader wakes up the sleeping
 193  *   writer.
 194  * - Last reader unlocks the cpu_hotplug.lock.
 195  * - A new reader arrives at this moment, bumps up the refcount.
 196  * - The writer acquires the cpu_hotplug.lock finds the refcount
 197  *   non zero and goes to sleep again.
 198  *
 199  * However, this is very difficult to achieve in practice since
 200  * get_online_cpus() not an api which is called all that often.
 201  *
 202  */
 203 void cpu_hotplug_begin(void)
 204 {
 205         DEFINE_WAIT(wait);
 206
 207         cpu_hotplug.active_writer = current;
 208         cpuhp_lock_acquire();
 209
 210         for (;;) {
 211                 mutex_lock(&cpu_hotplug.lock);
 212                 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
 213                 if (likely(!atomic_read(&cpu_hotplug.refcount)))
 214                                 break;
 215                 mutex_unlock(&cpu_hotplug.lock);
 216                 schedule();
 217         }
 218         finish_wait(&cpu_hotplug.wq, &wait);
 219 }
 220
 221 void cpu_hotplug_done(void)
 222 {
 223         cpu_hotplug.active_writer = NULL;
 224         mutex_unlock(&cpu_hotplug.lock);
 225         cpuhp_lock_release();
 226 }
 227
 228 /*
 229  * Wait for currently running CPU hotplug operations to complete (if any) and
 230  * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
 231  * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
 232  * hotplug path before performing hotplug operations. So acquiring that lock
 233  * guarantees mutual exclusion from any currently running hotplug operations.
 234  */
 235 void cpu_hotplug_disable(void)
 236 {
 237         cpu_maps_update_begin();
 238         cpu_hotplug_disabled++;
 239         cpu_maps_update_done();
 240 }
 241 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
 242
 243 void cpu_hotplug_enable(void)
 244 {
 245         cpu_maps_update_begin();
 246         WARN_ON(--cpu_hotplug_disabled < 0);
 247         cpu_maps_update_done();
 248 }
 249 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
 250 #endif  /* CONFIG_HOTPLUG_CPU */
 251
 252 /* Need to know about CPUs going up/down? */
 253 int register_cpu_notifier(struct notifier_block *nb)
 254 {
 255         int ret;
 256         cpu_maps_update_begin();
 257         ret = raw_notifier_chain_register(&cpu_chain, nb);
 258         cpu_maps_update_done();
 259         return ret;
 260 }
 261
 262 int __register_cpu_notifier(struct notifier_block *nb)
 263 {
 264         return raw_notifier_chain_register(&cpu_chain, nb);
 265 }
 266
 267 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
 268                         int *nr_calls)
 269 {
 270         unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
 271         void *hcpu = (void *)(long)cpu;
 272
 273         int ret;
 274
 275         ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
 276                                         nr_calls);
 277
 278         return notifier_to_errno(ret);
 279 }
 280
 281 static int cpu_notify(unsigned long val, unsigned int cpu)
 282 {
 283         return __cpu_notify(val, cpu, -1, NULL);
 284 }
 285
 286 /* Notifier wrappers for transitioning to state machine */
 287 static int notify_prepare(unsigned int cpu)
 288 {
 289         int nr_calls = 0;
 290         int ret;
 291
 292         ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
 293         if (ret) {
 294                 nr_calls--;
 295                 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
 296                                 __func__, cpu);
 297                 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
 298         }
 299         return ret;
 300 }
 301
 302 static int notify_online(unsigned int cpu)
 303 {
 304         cpu_notify(CPU_ONLINE, cpu);
 305         return 0;
 306 }
 307
 308 static int notify_starting(unsigned int cpu)
 309 {
 310         cpu_notify(CPU_STARTING, cpu);
 311         return 0;
 312 }
 313
 314 static int bringup_cpu(unsigned int cpu)
 315 {
 316         struct task_struct *idle = idle_thread_get(cpu);
 317         int ret;
 318
 319         /* Arch-specific enabling code. */
 320         ret = __cpu_up(cpu, idle);
 321         if (ret) {
 322                 cpu_notify(CPU_UP_CANCELED, cpu);
 323                 return ret;
 324         }
 325         BUG_ON(!cpu_online(cpu));
 326         return 0;
 327 }
 328
 329 #ifdef CONFIG_HOTPLUG_CPU
 330 EXPORT_SYMBOL(register_cpu_notifier);
 331 EXPORT_SYMBOL(__register_cpu_notifier);
 332
 333 void unregister_cpu_notifier(struct notifier_block *nb)
 334 {
 335         cpu_maps_update_begin();
 336         raw_notifier_chain_unregister(&cpu_chain, nb);
 337         cpu_maps_update_done();
 338 }
 339 EXPORT_SYMBOL(unregister_cpu_notifier);
 340
 341 void __unregister_cpu_notifier(struct notifier_block *nb)
 342 {
 343         raw_notifier_chain_unregister(&cpu_chain, nb);
 344 }
 345 EXPORT_SYMBOL(__unregister_cpu_notifier);
 346
 347 /**
 348  * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
 349  * @cpu: a CPU id
 350  *
 351  * This function walks all processes, finds a valid mm struct for each one and
 352  * then clears a corresponding bit in mm's cpumask.  While this all sounds
 353  * trivial, there are various non-obvious corner cases, which this function
 354  * tries to solve in a safe manner.
 355  *
 356  * Also note that the function uses a somewhat relaxed locking scheme, so it may
 357  * be called only for an already offlined CPU.
 358  */
 359 void clear_tasks_mm_cpumask(int cpu)
 360 {
 361         struct task_struct *p;
 362
 363         /*
 364          * This function is called after the cpu is taken down and marked
 365          * offline, so its not like new tasks will ever get this cpu set in
 366          * their mm mask. -- Peter Zijlstra
 367          * Thus, we may use rcu_read_lock() here, instead of grabbing
 368          * full-fledged tasklist_lock.
 369          */
 370         WARN_ON(cpu_online(cpu));
 371         rcu_read_lock();
 372         for_each_process(p) {
 373                 struct task_struct *t;
 374
 375                 /*
 376                  * Main thread might exit, but other threads may still have
 377                  * a valid mm. Find one.
 378                  */
 379                 t = find_lock_task_mm(p);
 380                 if (!t)
 381                         continue;
 382                 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
 383                 task_unlock(t);
 384         }
 385         rcu_read_unlock();
 386 }
 387
 388 static inline void check_for_tasks(int dead_cpu)
 389 {
 390         struct task_struct *g, *p;
 391
 392         read_lock(&tasklist_lock);
 393         for_each_process_thread(g, p) {
 394                 if (!p->on_rq)
 395                         continue;
 396                 /*
 397                  * We do the check with unlocked task_rq(p)->lock.
 398                  * Order the reading to do not warn about a task,
 399                  * which was running on this cpu in the past, and
 400                  * it's just been woken on another cpu.
 401                  */
 402                 rmb();
 403                 if (task_cpu(p) != dead_cpu)
 404                         continue;
 405
 406                 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
 407                         p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
 408         }
 409         read_unlock(&tasklist_lock);
 410 }
 411
 412 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
 413 {
 414         BUG_ON(cpu_notify(val, cpu));
 415 }
 416
 417 static int notify_down_prepare(unsigned int cpu)
 418 {
 419         int err, nr_calls = 0;
 420
 421         err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
 422         if (err) {
 423                 nr_calls--;
 424                 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
 425                 pr_warn("%s: attempt to take down CPU %u failed\n",
 426                                 __func__, cpu);
 427         }
 428         return err;
 429 }
 430
 431 static int notify_dying(unsigned int cpu)
 432 {
 433         cpu_notify(CPU_DYING, cpu);
 434         return 0;
 435 }
 436
 437 /* Take this CPU down. */
 438 static int take_cpu_down(void *_param)
 439 {
 440         struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
 441         enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
 442         int err, cpu = smp_processor_id();
 443
 444         /* Ensure this CPU doesn't handle any more interrupts. */
 445         err = __cpu_disable();
 446         if (err < 0)
 447                 return err;
 448
 449         /* Invoke the former CPU_DYING callbacks */
 450         for (; st->state > target; st->state--) {
 451                 struct cpuhp_step *step = cpuhp_ap_states + st->state;
 452
 453                 cpuhp_invoke_callback(cpu, st->state, step->teardown);
 454         }
 455         /* Give up timekeeping duties */
 456         tick_handover_do_timer();
 457         /* Park the stopper thread */
 458         stop_machine_park(cpu);
 459         return 0;
 460 }
 461
 462 static int takedown_cpu(unsigned int cpu)
 463 {
 464         int err;
 465
 466         /*
 467          * By now we've cleared cpu_active_mask, wait for all preempt-disabled
 468          * and RCU users of this state to go away such that all new such users
 469          * will observe it.
 470          *
 471          * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
 472          * not imply sync_sched(), so wait for both.
 473          *
 474          * Do sync before park smpboot threads to take care the rcu boost case.
 475          */
 476         if (IS_ENABLED(CONFIG_PREEMPT))
 477                 synchronize_rcu_mult(call_rcu, call_rcu_sched);
 478         else
 479                 synchronize_rcu();
 480
 481         smpboot_park_threads(cpu);
 482
 483         /*
 484          * Prevent irq alloc/free while the dying cpu reorganizes the
 485          * interrupt affinities.
 486          */
 487         irq_lock_sparse();
 488
 489         /*
 490          * So now all preempt/rcu users must observe !cpu_active().
 491          */
 492         err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
 493         if (err) {
 494                 /* CPU didn't die: tell everyone.  Can't complain. */
 495                 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
 496                 irq_unlock_sparse();
 497                 return err;
 498         }
 499         BUG_ON(cpu_online(cpu));
 500
 501         /*
 502          * The migration_call() CPU_DYING callback will have removed all
 503          * runnable tasks from the cpu, there's only the idle task left now
 504          * that the migration thread is done doing the stop_machine thing.
 505          *
 506          * Wait for the stop thread to go away.
 507          */
 508         while (!per_cpu(cpu_dead_idle, cpu))
 509                 cpu_relax();
 510         smp_mb(); /* Read from cpu_dead_idle before __cpu_die(). */
 511         per_cpu(cpu_dead_idle, cpu) = false;
 512
 513         /* Interrupts are moved away from the dying cpu, reenable alloc/free */
 514         irq_unlock_sparse();
 515
 516         hotplug_cpu__broadcast_tick_pull(cpu);
 517         /* This actually kills the CPU. */
 518         __cpu_die(cpu);
 519
 520         tick_cleanup_dead_cpu(cpu);
 521         return 0;
 522 }
 523
 524 static int notify_dead(unsigned int cpu)
 525 {
 526         cpu_notify_nofail(CPU_DEAD, cpu);
 527         check_for_tasks(cpu);
 528         return 0;
 529 }
 530
 531 #else
 532 #define notify_down_prepare     NULL
 533 #define takedown_cpu            NULL
 534 #define notify_dead             NULL
 535 #define notify_dying            NULL
 536 #endif
 537
 538 #ifdef CONFIG_HOTPLUG_CPU
 539 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st)
 540 {
 541         for (st->state++; st->state < st->target; st->state++) {
 542                 struct cpuhp_step *step = cpuhp_bp_states + st->state;
 543
 544                 if (!step->skip_onerr)
 545                         cpuhp_invoke_callback(cpu, st->state, step->startup);
 546         }
 547 }
 548
 549 /* Requires cpu_add_remove_lock to be held */
 550 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
 551                            enum cpuhp_state target)
 552 {
 553         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 554         int prev_state, ret = 0;
 555         bool hasdied = false;
 556
 557         if (num_online_cpus() == 1)
 558                 return -EBUSY;
 559
 560         if (!cpu_online(cpu))
 561                 return -EINVAL;
 562
 563         cpu_hotplug_begin();
 564
 565         cpuhp_tasks_frozen = tasks_frozen;
 566
 567         prev_state = st->state;
 568         st->target = target;
 569         for (; st->state > st->target; st->state--) {
 570                 struct cpuhp_step *step = cpuhp_bp_states + st->state;
 571
 572                 ret = cpuhp_invoke_callback(cpu, st->state, step->teardown);
 573                 if (ret) {
 574                         st->target = prev_state;
 575                         undo_cpu_down(cpu, st);
 576                         break;
 577                 }
 578         }
 579         hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
 580
 581         cpu_hotplug_done();
 582         /* This post dead nonsense must die */
 583         if (!ret && hasdied)
 584                 cpu_notify_nofail(CPU_POST_DEAD, cpu);
 585         return ret;
 586 }
 587
 588 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
 589 {
 590         int err;
 591
 592         cpu_maps_update_begin();
 593
 594         if (cpu_hotplug_disabled) {
 595                 err = -EBUSY;
 596                 goto out;
 597         }
 598
 599         err = _cpu_down(cpu, 0, target);
 600
 601 out:
 602         cpu_maps_update_done();
 603         return err;
 604 }
 605 int cpu_down(unsigned int cpu)
 606 {
 607         return do_cpu_down(cpu, CPUHP_OFFLINE);
 608 }
 609 EXPORT_SYMBOL(cpu_down);
 610 #endif /*CONFIG_HOTPLUG_CPU*/
 611
 612 /*
 613  * Unpark per-CPU smpboot kthreads at CPU-online time.
 614  */
 615 static int smpboot_thread_call(struct notifier_block *nfb,
 616                                unsigned long action, void *hcpu)
 617 {
 618         int cpu = (long)hcpu;
 619
 620         switch (action & ~CPU_TASKS_FROZEN) {
 621
 622         case CPU_DOWN_FAILED:
 623         case CPU_ONLINE:
 624                 smpboot_unpark_threads(cpu);
 625                 break;
 626
 627         default:
 628                 break;
 629         }
 630
 631         return NOTIFY_OK;
 632 }
 633
 634 static struct notifier_block smpboot_thread_notifier = {
 635         .notifier_call = smpboot_thread_call,
 636         .priority = CPU_PRI_SMPBOOT,
 637 };
 638
 639 void smpboot_thread_init(void)
 640 {
 641         register_cpu_notifier(&smpboot_thread_notifier);
 642 }
 643
 644 /**
 645  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
 646  * @cpu: cpu that just started
 647  *
 648  * This function calls the cpu_chain notifiers with CPU_STARTING.
 649  * It must be called by the arch code on the new cpu, before the new cpu
 650  * enables interrupts and before the "boot" cpu returns from __cpu_up().
 651  */
 652 void notify_cpu_starting(unsigned int cpu)
 653 {
 654         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 655         enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
 656
 657         while (st->state < target) {
 658                 struct cpuhp_step *step;
 659
 660                 st->state++;
 661                 step = cpuhp_ap_states + st->state;
 662                 cpuhp_invoke_callback(cpu, st->state, step->startup);
 663         }
 664 }
 665
 666 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st)
 667 {
 668         for (st->state--; st->state > st->target; st->state--) {
 669                 struct cpuhp_step *step = cpuhp_bp_states + st->state;
 670
 671                 if (!step->skip_onerr)
 672                         cpuhp_invoke_callback(cpu, st->state, step->teardown);
 673         }
 674 }
 675
 676 /* Requires cpu_add_remove_lock to be held */
 677 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
 678 {
 679         struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
 680         struct task_struct *idle;
 681         int prev_state, ret = 0;
 682
 683         cpu_hotplug_begin();
 684
 685         if (cpu_online(cpu) || !cpu_present(cpu)) {
 686                 ret = -EINVAL;
 687                 goto out;
 688         }
 689
 690         /* Let it fail before we try to bring the cpu up */
 691         idle = idle_thread_get(cpu);
 692         if (IS_ERR(idle)) {
 693                 ret = PTR_ERR(idle);
 694                 goto out;
 695         }
 696
 697         cpuhp_tasks_frozen = tasks_frozen;
 698
 699         prev_state = st->state;
 700         st->target = target;
 701         while (st->state < st->target) {
 702                 struct cpuhp_step *step;
 703
 704                 st->state++;
 705                 step = cpuhp_bp_states + st->state;
 706                 ret = cpuhp_invoke_callback(cpu, st->state, step->startup);
 707                 if (ret) {
 708                         st->target = prev_state;
 709                         undo_cpu_up(cpu, st);
 710                         break;
 711                 }
 712         }
 713 out:
 714         cpu_hotplug_done();
 715         return ret;
 716 }
 717
 718 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
 719 {
 720         int err = 0;
 721
 722         if (!cpu_possible(cpu)) {
 723                 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
 724                        cpu);
 725 #if defined(CONFIG_IA64)
 726                 pr_err("please check additional_cpus= boot parameter\n");
 727 #endif
 728                 return -EINVAL;
 729         }
 730
 731         err = try_online_node(cpu_to_node(cpu));
 732         if (err)
 733                 return err;
 734
 735         cpu_maps_update_begin();
 736
 737         if (cpu_hotplug_disabled) {
 738                 err = -EBUSY;
 739                 goto out;
 740         }
 741
 742         err = _cpu_up(cpu, 0, target);
 743 out:
 744         cpu_maps_update_done();
 745         return err;
 746 }
 747
 748 int cpu_up(unsigned int cpu)
 749 {
 750         return do_cpu_up(cpu, CPUHP_ONLINE);
 751 }
 752 EXPORT_SYMBOL_GPL(cpu_up);
 753
 754 #ifdef CONFIG_PM_SLEEP_SMP
 755 static cpumask_var_t frozen_cpus;
 756
 757 int disable_nonboot_cpus(void)
 758 {
 759         int cpu, first_cpu, error = 0;
 760
 761         cpu_maps_update_begin();
 762         first_cpu = cpumask_first(cpu_online_mask);
 763         /*
 764          * We take down all of the non-boot CPUs in one shot to avoid races
 765          * with the userspace trying to use the CPU hotplug at the same time
 766          */
 767         cpumask_clear(frozen_cpus);
 768
 769         pr_info("Disabling non-boot CPUs ...\n");
 770         for_each_online_cpu(cpu) {
 771                 if (cpu == first_cpu)
 772                         continue;
 773                 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
 774                 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
 775                 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
 776                 if (!error)
 777                         cpumask_set_cpu(cpu, frozen_cpus);
 778                 else {
 779                         pr_err("Error taking CPU%d down: %d\n", cpu, error);
 780                         break;
 781                 }
 782         }
 783
 784         if (!error)
 785                 BUG_ON(num_online_cpus() > 1);
 786         else
 787                 pr_err("Non-boot CPUs are not disabled\n");
 788
 789         /*
 790          * Make sure the CPUs won't be enabled by someone else. We need to do
 791          * this even in case of failure as all disable_nonboot_cpus() users are
 792          * supposed to do enable_nonboot_cpus() on the failure path.
 793          */
 794         cpu_hotplug_disabled++;
 795
 796         cpu_maps_update_done();
 797         return error;
 798 }
 799
 800 void __weak arch_enable_nonboot_cpus_begin(void)
 801 {
 802 }
 803
 804 void __weak arch_enable_nonboot_cpus_end(void)
 805 {
 806 }
 807
 808 void enable_nonboot_cpus(void)
 809 {
 810         int cpu, error;
 811
 812         /* Allow everyone to use the CPU hotplug again */
 813         cpu_maps_update_begin();
 814         WARN_ON(--cpu_hotplug_disabled < 0);
 815         if (cpumask_empty(frozen_cpus))
 816                 goto out;
 817
 818         pr_info("Enabling non-boot CPUs ...\n");
 819
 820         arch_enable_nonboot_cpus_begin();
 821
 822         for_each_cpu(cpu, frozen_cpus) {
 823                 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
 824                 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
 825                 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
 826                 if (!error) {
 827                         pr_info("CPU%d is up\n", cpu);
 828                         continue;
 829                 }
 830                 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
 831         }
 832
 833         arch_enable_nonboot_cpus_end();
 834
 835         cpumask_clear(frozen_cpus);
 836 out:
 837         cpu_maps_update_done();
 838 }
 839
 840 static int __init alloc_frozen_cpus(void)
 841 {
 842         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
 843                 return -ENOMEM;
 844         return 0;
 845 }
 846 core_initcall(alloc_frozen_cpus);
 847
 848 /*
 849  * When callbacks for CPU hotplug notifications are being executed, we must
 850  * ensure that the state of the system with respect to the tasks being frozen
 851  * or not, as reported by the notification, remains unchanged *throughout the
 852  * duration* of the execution of the callbacks.
 853  * Hence we need to prevent the freezer from racing with regular CPU hotplug.
 854  *
 855  * This synchronization is implemented by mutually excluding regular CPU
 856  * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
 857  * Hibernate notifications.
 858  */
 859 static int
 860 cpu_hotplug_pm_callback(struct notifier_block *nb,
 861                         unsigned long action, void *ptr)
 862 {
 863         switch (action) {
 864
 865         case PM_SUSPEND_PREPARE:
 866         case PM_HIBERNATION_PREPARE:
 867                 cpu_hotplug_disable();
 868                 break;
 869
 870         case PM_POST_SUSPEND:
 871         case PM_POST_HIBERNATION:
 872                 cpu_hotplug_enable();
 873                 break;
 874
 875         default:
 876                 return NOTIFY_DONE;
 877         }
 878
 879         return NOTIFY_OK;
 880 }
 881
 882
 883 static int __init cpu_hotplug_pm_sync_init(void)
 884 {
 885         /*
 886          * cpu_hotplug_pm_callback has higher priority than x86
 887          * bsp_pm_callback which depends on cpu_hotplug_pm_callback
 888          * to disable cpu hotplug to avoid cpu hotplug race.
 889          */
 890         pm_notifier(cpu_hotplug_pm_callback, 0);
 891         return 0;
 892 }
 893 core_initcall(cpu_hotplug_pm_sync_init);
 894
 895 #endif /* CONFIG_PM_SLEEP_SMP */
 896
 897 #endif /* CONFIG_SMP */
 898
 899 /* Boot processor state steps */
 900 static struct cpuhp_step cpuhp_bp_states[] = {
 901         [CPUHP_OFFLINE] = {
 902                 .name                   = "offline",
 903                 .startup                = NULL,
 904                 .teardown               = NULL,
 905         },
 906 #ifdef CONFIG_SMP
 907         [CPUHP_CREATE_THREADS]= {
 908                 .name                   = "threads:create",
 909                 .startup                = smpboot_create_threads,
 910                 .teardown               = NULL,
 911         },
 912         [CPUHP_NOTIFY_PREPARE] = {
 913                 .name                   = "notify:prepare",
 914                 .startup                = notify_prepare,
 915                 .teardown               = notify_dead,
 916                 .skip_onerr             = true,
 917         },
 918         [CPUHP_BRINGUP_CPU] = {
 919                 .name                   = "cpu:bringup",
 920                 .startup                = bringup_cpu,
 921                 .teardown               = NULL,
 922         },
 923         [CPUHP_TEARDOWN_CPU] = {
 924                 .name                   = "cpu:teardown",
 925                 .startup                = NULL,
 926                 .teardown               = takedown_cpu,
 927         },
 928         [CPUHP_NOTIFY_ONLINE] = {
 929                 .name                   = "notify:online",
 930                 .startup                = notify_online,
 931                 .teardown               = notify_down_prepare,
 932         },
 933 #endif
 934         [CPUHP_ONLINE] = {
 935                 .name                   = "online",
 936                 .startup                = NULL,
 937                 .teardown               = NULL,
 938         },
 939 };
 940
 941 /* Application processor state steps */
 942 static struct cpuhp_step cpuhp_ap_states[] = {
 943 #ifdef CONFIG_SMP
 944         [CPUHP_AP_NOTIFY_STARTING] = {
 945                 .name                   = "notify:starting",
 946                 .startup                = notify_starting,
 947                 .teardown               = notify_dying,
 948                 .skip_onerr             = true,
 949         },
 950 #endif
 951         [CPUHP_ONLINE] = {
 952                 .name                   = "online",
 953                 .startup                = NULL,
 954                 .teardown               = NULL,
 955         },
 956 };
 957
 958 /*
 959  * cpu_bit_bitmap[] is a special, "compressed" data structure that
 960  * represents all NR_CPUS bits binary values of 1<<nr.
 961  *
 962  * It is used by cpumask_of() to get a constant address to a CPU
 963  * mask value that has a single bit set only.
 964  */
 965
 966 /* cpu_bit_bitmap[0] is empty - so we can back into it */
 967 #define MASK_DECLARE_1(x)       [x+1][0] = (1UL << (x))
 968 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
 969 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
 970 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
 971
 972 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
 973
 974         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
 975         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
 976 #if BITS_PER_LONG > 32
 977         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
 978         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
 979 #endif
 980 };
 981 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
 982
 983 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
 984 EXPORT_SYMBOL(cpu_all_bits);
 985
 986 #ifdef CONFIG_INIT_ALL_POSSIBLE
 987 struct cpumask __cpu_possible_mask __read_mostly
 988         = {CPU_BITS_ALL};
 989 #else
 990 struct cpumask __cpu_possible_mask __read_mostly;
 991 #endif
 992 EXPORT_SYMBOL(__cpu_possible_mask);
 993
 994 struct cpumask __cpu_online_mask __read_mostly;
 995 EXPORT_SYMBOL(__cpu_online_mask);
 996
 997 struct cpumask __cpu_present_mask __read_mostly;
 998 EXPORT_SYMBOL(__cpu_present_mask);
 999
1000 struct cpumask __cpu_active_mask __read_mostly;
1001 EXPORT_SYMBOL(__cpu_active_mask);
1002
1003 void init_cpu_present(const struct cpumask *src)
1004 {
1005         cpumask_copy(&__cpu_present_mask, src);
1006 }
1007
1008 void init_cpu_possible(const struct cpumask *src)
1009 {
1010         cpumask_copy(&__cpu_possible_mask, src);
1011 }
1012
1013 void init_cpu_online(const struct cpumask *src)
1014 {
1015         cpumask_copy(&__cpu_online_mask, src);
1016 }
1017
1018 /*
1019  * Activate the first processor.
1020  */
1021 void __init boot_cpu_init(void)
1022 {
1023         int cpu = smp_processor_id();
1024
1025         /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1026         set_cpu_online(cpu, true);
1027         set_cpu_active(cpu, true);
1028         set_cpu_present(cpu, true);
1029         set_cpu_possible(cpu, true);
1030 }
1031
1032 /*
1033  * Must be called _AFTER_ setting up the per_cpu areas
1034  */
1035 void __init boot_cpu_state_init(void)
1036 {
1037         per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;
1038 }