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/lglock.h>
  24 #include <linux/nmi.h>
  25
  26 /*
  27  * Structure to determine completion condition and record errors.  May
  28  * be shared by works on different cpus.
  29  */
  30 struct cpu_stop_done {
  31         atomic_t                nr_todo;        /* nr left to execute */
  32         int                     ret;            /* collected return value */
  33         struct completion       completion;     /* fired if nr_todo reaches 0 */
  34 };
  35
  36 /* the actual stopper, one per every possible cpu, enabled on online cpus */
  37 struct cpu_stopper {
  38         struct task_struct      *thread;
  39
  40         spinlock_t              lock;
  41         bool                    enabled;        /* is this stopper enabled? */
  42         struct list_head        works;          /* list of pending works */
  43
  44         struct cpu_stop_work    stop_work;      /* for stop_cpus */
  45 };
  46
  47 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
  48 static bool stop_machine_initialized = false;
  49
  50 /*
  51  * Avoids a race between stop_two_cpus and global stop_cpus, where
  52  * the stoppers could get queued up in reverse order, leading to
  53  * system deadlock. Using an lglock means stop_two_cpus remains
  54  * relatively cheap.
  55  */
  56 DEFINE_STATIC_LGLOCK(stop_cpus_lock);
  57
  58 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
  59 {
  60         memset(done, 0, sizeof(*done));
  61         atomic_set(&done->nr_todo, nr_todo);
  62         init_completion(&done->completion);
  63 }
  64
  65 /* signal completion unless @done is NULL */
  66 static void cpu_stop_signal_done(struct cpu_stop_done *done)
  67 {
  68         if (atomic_dec_and_test(&done->nr_todo))
  69                 complete(&done->completion);
  70 }
  71
  72 static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
  73                                         struct cpu_stop_work *work)
  74 {
  75         list_add_tail(&work->list, &stopper->works);
  76         wake_up_process(stopper->thread);
  77 }
  78
  79 /* queue @work to @stopper.  if offline, @work is completed immediately */
  80 static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
  81 {
  82         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
  83         unsigned long flags;
  84         bool enabled;
  85
  86         spin_lock_irqsave(&stopper->lock, flags);
  87         enabled = stopper->enabled;
  88         if (enabled)
  89                 __cpu_stop_queue_work(stopper, work);
  90         else if (work->done)
  91                 cpu_stop_signal_done(work->done);
  92         spin_unlock_irqrestore(&stopper->lock, flags);
  93
  94         return enabled;
  95 }
  96
  97 /**
  98  * stop_one_cpu - stop a cpu
  99  * @cpu: cpu to stop
 100  * @fn: function to execute
 101  * @arg: argument to @fn
 102  *
 103  * Execute @fn(@arg) on @cpu.  @fn is run in a process context with
 104  * the highest priority preempting any task on the cpu and
 105  * monopolizing it.  This function returns after the execution is
 106  * complete.
 107  *
 108  * This function doesn't guarantee @cpu stays online till @fn
 109  * completes.  If @cpu goes down in the middle, execution may happen
 110  * partially or fully on different cpus.  @fn should either be ready
 111  * for that or the caller should ensure that @cpu stays online until
 112  * this function completes.
 113  *
 114  * CONTEXT:
 115  * Might sleep.
 116  *
 117  * RETURNS:
 118  * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
 119  * otherwise, the return value of @fn.
 120  */
 121 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
 122 {
 123         struct cpu_stop_done done;
 124         struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
 125
 126         cpu_stop_init_done(&done, 1);
 127         if (!cpu_stop_queue_work(cpu, &work))
 128                 return -ENOENT;
 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();
 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
 234         lg_double_lock(&stop_cpus_lock, cpu1, cpu2);
 235         spin_lock_irq(&stopper1->lock);
 236         spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
 237
 238         err = -ENOENT;
 239         if (!stopper1->enabled || !stopper2->enabled)
 240                 goto unlock;
 241
 242         err = 0;
 243         __cpu_stop_queue_work(stopper1, work1);
 244         __cpu_stop_queue_work(stopper2, work2);
 245 unlock:
 246         spin_unlock(&stopper2->lock);
 247         spin_unlock_irq(&stopper1->lock);
 248         lg_double_unlock(&stop_cpus_lock, cpu1, cpu2);
 249
 250         return err;
 251 }
 252 /**
 253  * stop_two_cpus - stops two cpus
 254  * @cpu1: the cpu to stop
 255  * @cpu2: the other cpu to stop
 256  * @fn: function to execute
 257  * @arg: argument to @fn
 258  *
 259  * Stops both the current and specified CPU and runs @fn on one of them.
 260  *
 261  * returns when both are completed.
 262  */
 263 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
 264 {
 265         struct cpu_stop_done done;
 266         struct cpu_stop_work work1, work2;
 267         struct multi_stop_data msdata;
 268
 269         msdata = (struct multi_stop_data){
 270                 .fn = fn,
 271                 .data = arg,
 272                 .num_threads = 2,
 273                 .active_cpus = cpumask_of(cpu1),
 274         };
 275
 276         work1 = work2 = (struct cpu_stop_work){
 277                 .fn = multi_cpu_stop,
 278                 .arg = &msdata,
 279                 .done = &done
 280         };
 281
 282         cpu_stop_init_done(&done, 2);
 283         set_state(&msdata, MULTI_STOP_PREPARE);
 284
 285         if (cpu1 > cpu2)
 286                 swap(cpu1, cpu2);
 287         if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
 288                 return -ENOENT;
 289
 290         wait_for_completion(&done.completion);
 291         return done.ret;
 292 }
 293
 294 /**
 295  * stop_one_cpu_nowait - stop a cpu but don't wait for completion
 296  * @cpu: cpu to stop
 297  * @fn: function to execute
 298  * @arg: argument to @fn
 299  * @work_buf: pointer to cpu_stop_work structure
 300  *
 301  * Similar to stop_one_cpu() but doesn't wait for completion.  The
 302  * caller is responsible for ensuring @work_buf is currently unused
 303  * and will remain untouched until stopper starts executing @fn.
 304  *
 305  * CONTEXT:
 306  * Don't care.
 307  *
 308  * RETURNS:
 309  * true if cpu_stop_work was queued successfully and @fn will be called,
 310  * false otherwise.
 311  */
 312 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
 313                         struct cpu_stop_work *work_buf)
 314 {
 315         *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
 316         return cpu_stop_queue_work(cpu, work_buf);
 317 }
 318
 319 /* static data for stop_cpus */
 320 static DEFINE_MUTEX(stop_cpus_mutex);
 321
 322 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
 323                                  cpu_stop_fn_t fn, void *arg,
 324                                  struct cpu_stop_done *done)
 325 {
 326         struct cpu_stop_work *work;
 327         unsigned int cpu;
 328         bool queued = false;
 329
 330         /*
 331          * Disable preemption while queueing to avoid getting
 332          * preempted by a stopper which might wait for other stoppers
 333          * to enter @fn which can lead to deadlock.
 334          */
 335         lg_global_lock(&stop_cpus_lock);
 336         for_each_cpu(cpu, cpumask) {
 337                 work = &per_cpu(cpu_stopper.stop_work, cpu);
 338                 work->fn = fn;
 339                 work->arg = arg;
 340                 work->done = done;
 341                 if (cpu_stop_queue_work(cpu, work))
 342                         queued = true;
 343         }
 344         lg_global_unlock(&stop_cpus_lock);
 345
 346         return queued;
 347 }
 348
 349 static int __stop_cpus(const struct cpumask *cpumask,
 350                        cpu_stop_fn_t fn, void *arg)
 351 {
 352         struct cpu_stop_done done;
 353
 354         cpu_stop_init_done(&done, cpumask_weight(cpumask));
 355         if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
 356                 return -ENOENT;
 357         wait_for_completion(&done.completion);
 358         return done.ret;
 359 }
 360
 361 /**
 362  * stop_cpus - stop multiple cpus
 363  * @cpumask: cpus to stop
 364  * @fn: function to execute
 365  * @arg: argument to @fn
 366  *
 367  * Execute @fn(@arg) on online cpus in @cpumask.  On each target cpu,
 368  * @fn is run in a process context with the highest priority
 369  * preempting any task on the cpu and monopolizing it.  This function
 370  * returns after all executions are complete.
 371  *
 372  * This function doesn't guarantee the cpus in @cpumask stay online
 373  * till @fn completes.  If some cpus go down in the middle, execution
 374  * on the cpu may happen partially or fully on different cpus.  @fn
 375  * should either be ready for that or the caller should ensure that
 376  * the cpus stay online until this function completes.
 377  *
 378  * All stop_cpus() calls are serialized making it safe for @fn to wait
 379  * for all cpus to start executing it.
 380  *
 381  * CONTEXT:
 382  * Might sleep.
 383  *
 384  * RETURNS:
 385  * -ENOENT if @fn(@arg) was not executed at all because all cpus in
 386  * @cpumask were offline; otherwise, 0 if all executions of @fn
 387  * returned 0, any non zero return value if any returned non zero.
 388  */
 389 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 390 {
 391         int ret;
 392
 393         /* static works are used, process one request at a time */
 394         mutex_lock(&stop_cpus_mutex);
 395         ret = __stop_cpus(cpumask, fn, arg);
 396         mutex_unlock(&stop_cpus_mutex);
 397         return ret;
 398 }
 399
 400 /**
 401  * try_stop_cpus - try to stop multiple cpus
 402  * @cpumask: cpus to stop
 403  * @fn: function to execute
 404  * @arg: argument to @fn
 405  *
 406  * Identical to stop_cpus() except that it fails with -EAGAIN if
 407  * someone else is already using the facility.
 408  *
 409  * CONTEXT:
 410  * Might sleep.
 411  *
 412  * RETURNS:
 413  * -EAGAIN if someone else is already stopping cpus, -ENOENT if
 414  * @fn(@arg) was not executed at all because all cpus in @cpumask were
 415  * offline; otherwise, 0 if all executions of @fn returned 0, any non
 416  * zero return value if any returned non zero.
 417  */
 418 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 419 {
 420         int ret;
 421
 422         /* static works are used, process one request at a time */
 423         if (!mutex_trylock(&stop_cpus_mutex))
 424                 return -EAGAIN;
 425         ret = __stop_cpus(cpumask, fn, arg);
 426         mutex_unlock(&stop_cpus_mutex);
 427         return ret;
 428 }
 429
 430 static int cpu_stop_should_run(unsigned int cpu)
 431 {
 432         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 433         unsigned long flags;
 434         int run;
 435
 436         spin_lock_irqsave(&stopper->lock, flags);
 437         run = !list_empty(&stopper->works);
 438         spin_unlock_irqrestore(&stopper->lock, flags);
 439         return run;
 440 }
 441
 442 static void cpu_stopper_thread(unsigned int cpu)
 443 {
 444         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 445         struct cpu_stop_work *work;
 446
 447 repeat:
 448         work = NULL;
 449         spin_lock_irq(&stopper->lock);
 450         if (!list_empty(&stopper->works)) {
 451                 work = list_first_entry(&stopper->works,
 452                                         struct cpu_stop_work, list);
 453                 list_del_init(&work->list);
 454         }
 455         spin_unlock_irq(&stopper->lock);
 456
 457         if (work) {
 458                 cpu_stop_fn_t fn = work->fn;
 459                 void *arg = work->arg;
 460                 struct cpu_stop_done *done = work->done;
 461                 int ret;
 462
 463                 /* cpu stop callbacks must not sleep, make in_atomic() == T */
 464                 preempt_count_inc();
 465                 ret = fn(arg);
 466                 if (done) {
 467                         if (ret)
 468                                 done->ret = ret;
 469                         cpu_stop_signal_done(done);
 470                 }
 471                 preempt_count_dec();
 472                 WARN_ONCE(preempt_count(),
 473                           "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
 474                 goto repeat;
 475         }
 476 }
 477
 478 void stop_machine_park(int cpu)
 479 {
 480         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 481         /*
 482          * Lockless. cpu_stopper_thread() will take stopper->lock and flush
 483          * the pending works before it parks, until then it is fine to queue
 484          * the new works.
 485          */
 486         stopper->enabled = false;
 487         kthread_park(stopper->thread);
 488 }
 489
 490 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
 491
 492 static void cpu_stop_create(unsigned int cpu)
 493 {
 494         sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
 495 }
 496
 497 static void cpu_stop_park(unsigned int cpu)
 498 {
 499         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 500
 501         WARN_ON(!list_empty(&stopper->works));
 502 }
 503
 504 void stop_machine_unpark(int cpu)
 505 {
 506         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 507
 508         stopper->enabled = true;
 509         kthread_unpark(stopper->thread);
 510 }
 511
 512 static struct smp_hotplug_thread cpu_stop_threads = {
 513         .store                  = &cpu_stopper.thread,
 514         .thread_should_run      = cpu_stop_should_run,
 515         .thread_fn              = cpu_stopper_thread,
 516         .thread_comm            = "migration/%u",
 517         .create                 = cpu_stop_create,
 518         .park                   = cpu_stop_park,
 519         .selfparking            = true,
 520 };
 521
 522 static int __init cpu_stop_init(void)
 523 {
 524         unsigned int cpu;
 525
 526         for_each_possible_cpu(cpu) {
 527                 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 528
 529                 spin_lock_init(&stopper->lock);
 530                 INIT_LIST_HEAD(&stopper->works);
 531         }
 532
 533         BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
 534         stop_machine_unpark(raw_smp_processor_id());
 535         stop_machine_initialized = true;
 536         return 0;
 537 }
 538 early_initcall(cpu_stop_init);
 539
 540 static int __stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
 541 {
 542         struct multi_stop_data msdata = {
 543                 .fn = fn,
 544                 .data = data,
 545                 .num_threads = num_online_cpus(),
 546                 .active_cpus = cpus,
 547         };
 548
 549         if (!stop_machine_initialized) {
 550                 /*
 551                  * Handle the case where stop_machine() is called
 552                  * early in boot before stop_machine() has been
 553                  * initialized.
 554                  */
 555                 unsigned long flags;
 556                 int ret;
 557
 558                 WARN_ON_ONCE(msdata.num_threads != 1);
 559
 560                 local_irq_save(flags);
 561                 hard_irq_disable();
 562                 ret = (*fn)(data);
 563                 local_irq_restore(flags);
 564
 565                 return ret;
 566         }
 567
 568         /* Set the initial state and stop all online cpus. */
 569         set_state(&msdata, MULTI_STOP_PREPARE);
 570         return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
 571 }
 572
 573 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
 574 {
 575         int ret;
 576
 577         /* No CPUs can come up or down during this. */
 578         get_online_cpus();
 579         ret = __stop_machine(fn, data, cpus);
 580         put_online_cpus();
 581         return ret;
 582 }
 583 EXPORT_SYMBOL_GPL(stop_machine);
 584
 585 /**
 586  * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
 587  * @fn: the function to run
 588  * @data: the data ptr for the @fn()
 589  * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
 590  *
 591  * This is identical to stop_machine() but can be called from a CPU which
 592  * is not active.  The local CPU is in the process of hotplug (so no other
 593  * CPU hotplug can start) and not marked active and doesn't have enough
 594  * context to sleep.
 595  *
 596  * This function provides stop_machine() functionality for such state by
 597  * using busy-wait for synchronization and executing @fn directly for local
 598  * CPU.
 599  *
 600  * CONTEXT:
 601  * Local CPU is inactive.  Temporarily stops all active CPUs.
 602  *
 603  * RETURNS:
 604  * 0 if all executions of @fn returned 0, any non zero return value if any
 605  * returned non zero.
 606  */
 607 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
 608                                   const struct cpumask *cpus)
 609 {
 610         struct multi_stop_data msdata = { .fn = fn, .data = data,
 611                                             .active_cpus = cpus };
 612         struct cpu_stop_done done;
 613         int ret;
 614
 615         /* Local CPU must be inactive and CPU hotplug in progress. */
 616         BUG_ON(cpu_active(raw_smp_processor_id()));
 617         msdata.num_threads = num_active_cpus() + 1;     /* +1 for local */
 618
 619         /* No proper task established and can't sleep - busy wait for lock. */
 620         while (!mutex_trylock(&stop_cpus_mutex))
 621                 cpu_relax();
 622
 623         /* Schedule work on other CPUs and execute directly for local CPU */
 624         set_state(&msdata, MULTI_STOP_PREPARE);
 625         cpu_stop_init_done(&done, num_active_cpus());
 626         queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
 627                              &done);
 628         ret = multi_cpu_stop(&msdata);
 629
 630         /* Busy wait for completion. */
 631         while (!completion_done(&done.completion))
 632                 cpu_relax();
 633
 634         mutex_unlock(&stop_cpus_mutex);
 635         return ret ?: done.ret;
 636 }