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