kernel/stop_machine.c

   1 /*
   2  * kernel/stop_machine.c
   3  *
   4  * Copyright (C) 2008, 2005     IBM Corporation.
   5  * Copyright (C) 2008, 2005     Rusty Russell rusty@rustcorp.com.au
   6  * Copyright (C) 2010           SUSE Linux Products GmbH
   7  * Copyright (C) 2010           Tejun Heo <tj@kernel.org>
   8  *
   9  * This file is released under the GPLv2 and any later version.
  10  */
  11 #include <linux/completion.h>
  12 #include <linux/cpu.h>
  13 #include <linux/init.h>
  14 #include <linux/kthread.h>
  15 #include <linux/export.h>
  16 #include <linux/percpu.h>
  17 #include <linux/sched.h>
  18 #include <linux/stop_machine.h>
  19 #include <linux/interrupt.h>
  20 #include <linux/kallsyms.h>
  21 #include <linux/smpboot.h>
  22 #include <linux/atomic.h>
  23 #include <linux/nmi.h>
  24
  25 /*
  26  * Structure to determine completion condition and record errors.  May
  27  * be shared by works on different cpus.
  28  */
  29 struct cpu_stop_done {
  30         atomic_t                nr_todo;        /* nr left to execute */
  31         int                     ret;            /* collected return value */
  32         struct completion       completion;     /* fired if nr_todo reaches 0 */
  33 };
  34
  35 /* the actual stopper, one per every possible cpu, enabled on online cpus */
  36 struct cpu_stopper {
  37         struct task_struct      *thread;
  38
  39         spinlock_t              lock;
  40         bool                    enabled;        /* is this stopper enabled? */
  41         struct list_head        works;          /* list of pending works */
  42
  43         struct cpu_stop_work    stop_work;      /* for stop_cpus */
  44 };
  45
  46 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
  47 static bool stop_machine_initialized = false;
  48
  49 /* static data for stop_cpus */
  50 static DEFINE_MUTEX(stop_cpus_mutex);
  51 static bool stop_cpus_in_progress;
  52
  53 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
  54 {
  55         memset(done, 0, sizeof(*done));
  56         atomic_set(&done->nr_todo, nr_todo);
  57         init_completion(&done->completion);
  58 }
  59
  60 /* signal completion unless @done is NULL */
  61 static void cpu_stop_signal_done(struct cpu_stop_done *done)
  62 {
  63         if (atomic_dec_and_test(&done->nr_todo))
  64                 complete(&done->completion);
  65 }
  66
  67 static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
  68                                         struct cpu_stop_work *work)
  69 {
  70         list_add_tail(&work->list, &stopper->works);
  71         wake_up_process(stopper->thread);
  72 }
  73
  74 /* queue @work to @stopper.  if offline, @work is completed immediately */
  75 static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
  76 {
  77         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
  78         unsigned long flags;
  79         bool enabled;
  80
  81         spin_lock_irqsave(&stopper->lock, flags);
  82         enabled = stopper->enabled;
  83         if (enabled)
  84                 __cpu_stop_queue_work(stopper, work);
  85         else if (work->done)
  86                 cpu_stop_signal_done(work->done);
  87         spin_unlock_irqrestore(&stopper->lock, flags);
  88
  89         return enabled;
  90 }
  91
  92 /**
  93  * stop_one_cpu - stop a cpu
  94  * @cpu: cpu to stop
  95  * @fn: function to execute
  96  * @arg: argument to @fn
  97  *
  98  * Execute @fn(@arg) on @cpu.  @fn is run in a process context with
  99  * the highest priority preempting any task on the cpu and
 100  * monopolizing it.  This function returns after the execution is
 101  * complete.
 102  *
 103  * This function doesn't guarantee @cpu stays online till @fn
 104  * completes.  If @cpu goes down in the middle, execution may happen
 105  * partially or fully on different cpus.  @fn should either be ready
 106  * for that or the caller should ensure that @cpu stays online until
 107  * this function completes.
 108  *
 109  * CONTEXT:
 110  * Might sleep.
 111  *
 112  * RETURNS:
 113  * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
 114  * otherwise, the return value of @fn.
 115  */
 116 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
 117 {
 118         struct cpu_stop_done done;
 119         struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
 120
 121         cpu_stop_init_done(&done, 1);
 122         if (!cpu_stop_queue_work(cpu, &work))
 123                 return -ENOENT;
 124         /*
 125          * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
 126          * cycle by doing a preemption:
 127          */
 128         cond_resched();
 129         wait_for_completion(&done.completion);
 130         return done.ret;
 131 }
 132
 133 /* This controls the threads on each CPU. */
 134 enum multi_stop_state {
 135         /* Dummy starting state for thread. */
 136         MULTI_STOP_NONE,
 137         /* Awaiting everyone to be scheduled. */
 138         MULTI_STOP_PREPARE,
 139         /* Disable interrupts. */
 140         MULTI_STOP_DISABLE_IRQ,
 141         /* Run the function */
 142         MULTI_STOP_RUN,
 143         /* Exit */
 144         MULTI_STOP_EXIT,
 145 };
 146
 147 struct multi_stop_data {
 148         cpu_stop_fn_t           fn;
 149         void                    *data;
 150         /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
 151         unsigned int            num_threads;
 152         const struct cpumask    *active_cpus;
 153
 154         enum multi_stop_state   state;
 155         atomic_t                thread_ack;
 156 };
 157
 158 static void set_state(struct multi_stop_data *msdata,
 159                       enum multi_stop_state newstate)
 160 {
 161         /* Reset ack counter. */
 162         atomic_set(&msdata->thread_ack, msdata->num_threads);
 163         smp_wmb();
 164         msdata->state = newstate;
 165 }
 166
 167 /* Last one to ack a state moves to the next state. */
 168 static void ack_state(struct multi_stop_data *msdata)
 169 {
 170         if (atomic_dec_and_test(&msdata->thread_ack))
 171                 set_state(msdata, msdata->state + 1);
 172 }
 173
 174 /* This is the cpu_stop function which stops the CPU. */
 175 static int multi_cpu_stop(void *data)
 176 {
 177         struct multi_stop_data *msdata = data;
 178         enum multi_stop_state curstate = MULTI_STOP_NONE;
 179         int cpu = smp_processor_id(), err = 0;
 180         unsigned long flags;
 181         bool is_active;
 182
 183         /*
 184          * When called from stop_machine_from_inactive_cpu(), irq might
 185          * already be disabled.  Save the state and restore it on exit.
 186          */
 187         local_save_flags(flags);
 188
 189         if (!msdata->active_cpus)
 190                 is_active = cpu == cpumask_first(cpu_online_mask);
 191         else
 192                 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
 193
 194         /* Simple state machine */
 195         do {
 196                 /* Chill out and ensure we re-read multi_stop_state. */
 197                 cpu_relax_yield();
 198                 if (msdata->state != curstate) {
 199                         curstate = msdata->state;
 200                         switch (curstate) {
 201                         case MULTI_STOP_DISABLE_IRQ:
 202                                 local_irq_disable();
 203                                 hard_irq_disable();
 204                                 break;
 205                         case MULTI_STOP_RUN:
 206                                 if (is_active)
 207                                         err = msdata->fn(msdata->data);
 208                                 break;
 209                         default:
 210                                 break;
 211                         }
 212                         ack_state(msdata);
 213                 } else if (curstate > MULTI_STOP_PREPARE) {
 214                         /*
 215                          * At this stage all other CPUs we depend on must spin
 216                          * in the same loop. Any reason for hard-lockup should
 217                          * be detected and reported on their side.
 218                          */
 219                         touch_nmi_watchdog();
 220                 }
 221         } while (curstate != MULTI_STOP_EXIT);
 222
 223         local_irq_restore(flags);
 224         return err;
 225 }
 226
 227 static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
 228                                     int cpu2, struct cpu_stop_work *work2)
 229 {
 230         struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
 231         struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
 232         int err;
 233 retry:
 234         spin_lock_irq(&stopper1->lock);
 235         spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
 236
 237         err = -ENOENT;
 238         if (!stopper1->enabled || !stopper2->enabled)
 239                 goto unlock;
 240         /*
 241          * Ensure that if we race with __stop_cpus() the stoppers won't get
 242          * queued up in reverse order leading to system deadlock.
 243          *
 244          * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
 245          * queued a work on cpu1 but not on cpu2, we hold both locks.
 246          *
 247          * It can be falsely true but it is safe to spin until it is cleared,
 248          * queue_stop_cpus_work() does everything under preempt_disable().
 249          */
 250         err = -EDEADLK;
 251         if (unlikely(stop_cpus_in_progress))
 252                         goto unlock;
 253
 254         err = 0;
 255         __cpu_stop_queue_work(stopper1, work1);
 256         __cpu_stop_queue_work(stopper2, work2);
 257 unlock:
 258         spin_unlock(&stopper2->lock);
 259         spin_unlock_irq(&stopper1->lock);
 260
 261         if (unlikely(err == -EDEADLK)) {
 262                 while (stop_cpus_in_progress)
 263                         cpu_relax();
 264                 goto retry;
 265         }
 266         return err;
 267 }
 268 /**
 269  * stop_two_cpus - stops two cpus
 270  * @cpu1: the cpu to stop
 271  * @cpu2: the other cpu to stop
 272  * @fn: function to execute
 273  * @arg: argument to @fn
 274  *
 275  * Stops both the current and specified CPU and runs @fn on one of them.
 276  *
 277  * returns when both are completed.
 278  */
 279 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
 280 {
 281         struct cpu_stop_done done;
 282         struct cpu_stop_work work1, work2;
 283         struct multi_stop_data msdata;
 284
 285         msdata = (struct multi_stop_data){
 286                 .fn = fn,
 287                 .data = arg,
 288                 .num_threads = 2,
 289                 .active_cpus = cpumask_of(cpu1),
 290         };
 291
 292         work1 = work2 = (struct cpu_stop_work){
 293                 .fn = multi_cpu_stop,
 294                 .arg = &msdata,
 295                 .done = &done
 296         };
 297
 298         cpu_stop_init_done(&done, 2);
 299         set_state(&msdata, MULTI_STOP_PREPARE);
 300
 301         if (cpu1 > cpu2)
 302                 swap(cpu1, cpu2);
 303         if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
 304                 return -ENOENT;
 305
 306         wait_for_completion(&done.completion);
 307         return done.ret;
 308 }
 309
 310 /**
 311  * stop_one_cpu_nowait - stop a cpu but don't wait for completion
 312  * @cpu: cpu to stop
 313  * @fn: function to execute
 314  * @arg: argument to @fn
 315  * @work_buf: pointer to cpu_stop_work structure
 316  *
 317  * Similar to stop_one_cpu() but doesn't wait for completion.  The
 318  * caller is responsible for ensuring @work_buf is currently unused
 319  * and will remain untouched until stopper starts executing @fn.
 320  *
 321  * CONTEXT:
 322  * Don't care.
 323  *
 324  * RETURNS:
 325  * true if cpu_stop_work was queued successfully and @fn will be called,
 326  * false otherwise.
 327  */
 328 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
 329                         struct cpu_stop_work *work_buf)
 330 {
 331         *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
 332         return cpu_stop_queue_work(cpu, work_buf);
 333 }
 334
 335 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
 336                                  cpu_stop_fn_t fn, void *arg,
 337                                  struct cpu_stop_done *done)
 338 {
 339         struct cpu_stop_work *work;
 340         unsigned int cpu;
 341         bool queued = false;
 342
 343         /*
 344          * Disable preemption while queueing to avoid getting
 345          * preempted by a stopper which might wait for other stoppers
 346          * to enter @fn which can lead to deadlock.
 347          */
 348         preempt_disable();
 349         stop_cpus_in_progress = true;
 350         for_each_cpu(cpu, cpumask) {
 351                 work = &per_cpu(cpu_stopper.stop_work, cpu);
 352                 work->fn = fn;
 353                 work->arg = arg;
 354                 work->done = done;
 355                 if (cpu_stop_queue_work(cpu, work))
 356                         queued = true;
 357         }
 358         stop_cpus_in_progress = false;
 359         preempt_enable();
 360
 361         return queued;
 362 }
 363
 364 static int __stop_cpus(const struct cpumask *cpumask,
 365                        cpu_stop_fn_t fn, void *arg)
 366 {
 367         struct cpu_stop_done done;
 368
 369         cpu_stop_init_done(&done, cpumask_weight(cpumask));
 370         if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
 371                 return -ENOENT;
 372         wait_for_completion(&done.completion);
 373         return done.ret;
 374 }
 375
 376 /**
 377  * stop_cpus - stop multiple cpus
 378  * @cpumask: cpus to stop
 379  * @fn: function to execute
 380  * @arg: argument to @fn
 381  *
 382  * Execute @fn(@arg) on online cpus in @cpumask.  On each target cpu,
 383  * @fn is run in a process context with the highest priority
 384  * preempting any task on the cpu and monopolizing it.  This function
 385  * returns after all executions are complete.
 386  *
 387  * This function doesn't guarantee the cpus in @cpumask stay online
 388  * till @fn completes.  If some cpus go down in the middle, execution
 389  * on the cpu may happen partially or fully on different cpus.  @fn
 390  * should either be ready for that or the caller should ensure that
 391  * the cpus stay online until this function completes.
 392  *
 393  * All stop_cpus() calls are serialized making it safe for @fn to wait
 394  * for all cpus to start executing it.
 395  *
 396  * CONTEXT:
 397  * Might sleep.
 398  *
 399  * RETURNS:
 400  * -ENOENT if @fn(@arg) was not executed at all because all cpus in
 401  * @cpumask were offline; otherwise, 0 if all executions of @fn
 402  * returned 0, any non zero return value if any returned non zero.
 403  */
 404 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 405 {
 406         int ret;
 407
 408         /* static works are used, process one request at a time */
 409         mutex_lock(&stop_cpus_mutex);
 410         ret = __stop_cpus(cpumask, fn, arg);
 411         mutex_unlock(&stop_cpus_mutex);
 412         return ret;
 413 }
 414
 415 /**
 416  * try_stop_cpus - try to stop multiple cpus
 417  * @cpumask: cpus to stop
 418  * @fn: function to execute
 419  * @arg: argument to @fn
 420  *
 421  * Identical to stop_cpus() except that it fails with -EAGAIN if
 422  * someone else is already using the facility.
 423  *
 424  * CONTEXT:
 425  * Might sleep.
 426  *
 427  * RETURNS:
 428  * -EAGAIN if someone else is already stopping cpus, -ENOENT if
 429  * @fn(@arg) was not executed at all because all cpus in @cpumask were
 430  * offline; otherwise, 0 if all executions of @fn returned 0, any non
 431  * zero return value if any returned non zero.
 432  */
 433 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 434 {
 435         int ret;
 436
 437         /* static works are used, process one request at a time */
 438         if (!mutex_trylock(&stop_cpus_mutex))
 439                 return -EAGAIN;
 440         ret = __stop_cpus(cpumask, fn, arg);
 441         mutex_unlock(&stop_cpus_mutex);
 442         return ret;
 443 }
 444
 445 static int cpu_stop_should_run(unsigned int cpu)
 446 {
 447         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 448         unsigned long flags;
 449         int run;
 450
 451         spin_lock_irqsave(&stopper->lock, flags);
 452         run = !list_empty(&stopper->works);
 453         spin_unlock_irqrestore(&stopper->lock, flags);
 454         return run;
 455 }
 456
 457 static void cpu_stopper_thread(unsigned int cpu)
 458 {
 459         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 460         struct cpu_stop_work *work;
 461
 462 repeat:
 463         work = NULL;
 464         spin_lock_irq(&stopper->lock);
 465         if (!list_empty(&stopper->works)) {
 466                 work = list_first_entry(&stopper->works,
 467                                         struct cpu_stop_work, list);
 468                 list_del_init(&work->list);
 469         }
 470         spin_unlock_irq(&stopper->lock);
 471
 472         if (work) {
 473                 cpu_stop_fn_t fn = work->fn;
 474                 void *arg = work->arg;
 475                 struct cpu_stop_done *done = work->done;
 476                 int ret;
 477
 478                 /* cpu stop callbacks must not sleep, make in_atomic() == T */
 479                 preempt_count_inc();
 480                 ret = fn(arg);
 481                 if (done) {
 482                         if (ret)
 483                                 done->ret = ret;
 484                         cpu_stop_signal_done(done);
 485                 }
 486                 preempt_count_dec();
 487                 WARN_ONCE(preempt_count(),
 488                           "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
 489                 goto repeat;
 490         }
 491 }
 492
 493 void stop_machine_park(int cpu)
 494 {
 495         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 496         /*
 497          * Lockless. cpu_stopper_thread() will take stopper->lock and flush
 498          * the pending works before it parks, until then it is fine to queue
 499          * the new works.
 500          */
 501         stopper->enabled = false;
 502         kthread_park(stopper->thread);
 503 }
 504
 505 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
 506
 507 static void cpu_stop_create(unsigned int cpu)
 508 {
 509         sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
 510 }
 511
 512 static void cpu_stop_park(unsigned int cpu)
 513 {
 514         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 515
 516         WARN_ON(!list_empty(&stopper->works));
 517 }
 518
 519 void stop_machine_unpark(int cpu)
 520 {
 521         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 522
 523         stopper->enabled = true;
 524         kthread_unpark(stopper->thread);
 525 }
 526
 527 static struct smp_hotplug_thread cpu_stop_threads = {
 528         .store                  = &cpu_stopper.thread,
 529         .thread_should_run      = cpu_stop_should_run,
 530         .thread_fn              = cpu_stopper_thread,
 531         .thread_comm            = "migration/%u",
 532         .create                 = cpu_stop_create,
 533         .park                   = cpu_stop_park,
 534         .selfparking            = true,
 535 };
 536
 537 static int __init cpu_stop_init(void)
 538 {
 539         unsigned int cpu;
 540
 541         for_each_possible_cpu(cpu) {
 542                 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 543
 544                 spin_lock_init(&stopper->lock);
 545                 INIT_LIST_HEAD(&stopper->works);
 546         }
 547
 548         BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
 549         stop_machine_unpark(raw_smp_processor_id());
 550         stop_machine_initialized = true;
 551         return 0;
 552 }
 553 early_initcall(cpu_stop_init);
 554
 555 int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
 556                             const struct cpumask *cpus)
 557 {
 558         struct multi_stop_data msdata = {
 559                 .fn = fn,
 560                 .data = data,
 561                 .num_threads = num_online_cpus(),
 562                 .active_cpus = cpus,
 563         };
 564
 565         lockdep_assert_cpus_held();
 566
 567         if (!stop_machine_initialized) {
 568                 /*
 569                  * Handle the case where stop_machine() is called
 570                  * early in boot before stop_machine() has been
 571                  * initialized.
 572                  */
 573                 unsigned long flags;
 574                 int ret;
 575
 576                 WARN_ON_ONCE(msdata.num_threads != 1);
 577
 578                 local_irq_save(flags);
 579                 hard_irq_disable();
 580                 ret = (*fn)(data);
 581                 local_irq_restore(flags);
 582
 583                 return ret;
 584         }
 585
 586         /* Set the initial state and stop all online cpus. */
 587         set_state(&msdata, MULTI_STOP_PREPARE);
 588         return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
 589 }
 590
 591 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
 592 {
 593         int ret;
 594
 595         /* No CPUs can come up or down during this. */
 596         cpus_read_lock();
 597         ret = stop_machine_cpuslocked(fn, data, cpus);
 598         cpus_read_unlock();
 599         return ret;
 600 }
 601 EXPORT_SYMBOL_GPL(stop_machine);
 602
 603 /**
 604  * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
 605  * @fn: the function to run
 606  * @data: the data ptr for the @fn()
 607  * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
 608  *
 609  * This is identical to stop_machine() but can be called from a CPU which
 610  * is not active.  The local CPU is in the process of hotplug (so no other
 611  * CPU hotplug can start) and not marked active and doesn't have enough
 612  * context to sleep.
 613  *
 614  * This function provides stop_machine() functionality for such state by
 615  * using busy-wait for synchronization and executing @fn directly for local
 616  * CPU.
 617  *
 618  * CONTEXT:
 619  * Local CPU is inactive.  Temporarily stops all active CPUs.
 620  *
 621  * RETURNS:
 622  * 0 if all executions of @fn returned 0, any non zero return value if any
 623  * returned non zero.
 624  */
 625 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
 626                                   const struct cpumask *cpus)
 627 {
 628         struct multi_stop_data msdata = { .fn = fn, .data = data,
 629                                             .active_cpus = cpus };
 630         struct cpu_stop_done done;
 631         int ret;
 632
 633         /* Local CPU must be inactive and CPU hotplug in progress. */
 634         BUG_ON(cpu_active(raw_smp_processor_id()));
 635         msdata.num_threads = num_active_cpus() + 1;     /* +1 for local */
 636
 637         /* No proper task established and can't sleep - busy wait for lock. */
 638         while (!mutex_trylock(&stop_cpus_mutex))
 639                 cpu_relax();
 640
 641         /* Schedule work on other CPUs and execute directly for local CPU */
 642         set_state(&msdata, MULTI_STOP_PREPARE);
 643         cpu_stop_init_done(&done, num_active_cpus());
 644         queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
 645                              &done);
 646         ret = multi_cpu_stop(&msdata);
 647
 648         /* Busy wait for completion. */
 649         while (!completion_done(&done.completion))
 650                 cpu_relax();
 651
 652         mutex_unlock(&stop_cpus_mutex);
 653         return ret ?: done.ret;
 654 }