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