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         bool                    executed;       /* actually executed? */
  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         spinlock_t              lock;
  39         bool                    enabled;        /* is this stopper enabled? */
  40         struct list_head        works;          /* list of pending works */
  41 };
  42
  43 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
  44 static DEFINE_PER_CPU(struct task_struct *, cpu_stopper_task);
  45 static bool stop_machine_initialized = false;
  46
  47 /*
  48  * Avoids a race between stop_two_cpus and global stop_cpus, where
  49  * the stoppers could get queued up in reverse order, leading to
  50  * system deadlock. Using an lglock means stop_two_cpus remains
  51  * relatively cheap.
  52  */
  53 DEFINE_STATIC_LGLOCK(stop_cpus_lock);
  54
  55 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
  56 {
  57         memset(done, 0, sizeof(*done));
  58         atomic_set(&done->nr_todo, nr_todo);
  59         init_completion(&done->completion);
  60 }
  61
  62 /* signal completion unless @done is NULL */
  63 static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
  64 {
  65         if (done) {
  66                 if (executed)
  67                         done->executed = true;
  68                 if (atomic_dec_and_test(&done->nr_todo))
  69                         complete(&done->completion);
  70         }
  71 }
  72
  73 /* queue @work to @stopper.  if offline, @work is completed immediately */
  74 static void cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
  75 {
  76         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
  77         struct task_struct *p = per_cpu(cpu_stopper_task, cpu);
  78
  79         unsigned long flags;
  80
  81         spin_lock_irqsave(&stopper->lock, flags);
  82
  83         if (stopper->enabled) {
  84                 list_add_tail(&work->list, &stopper->works);
  85                 wake_up_process(p);
  86         } else
  87                 cpu_stop_signal_done(work->done, false);
  88
  89         spin_unlock_irqrestore(&stopper->lock, flags);
  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         cpu_stop_queue_work(cpu, &work);
 123         wait_for_completion(&done.completion);
 124         return done.executed ? done.ret : -ENOENT;
 125 }
 126
 127 /* This controls the threads on each CPU. */
 128 enum multi_stop_state {
 129         /* Dummy starting state for thread. */
 130         MULTI_STOP_NONE,
 131         /* Awaiting everyone to be scheduled. */
 132         MULTI_STOP_PREPARE,
 133         /* Disable interrupts. */
 134         MULTI_STOP_DISABLE_IRQ,
 135         /* Run the function */
 136         MULTI_STOP_RUN,
 137         /* Exit */
 138         MULTI_STOP_EXIT,
 139 };
 140
 141 struct multi_stop_data {
 142         int                     (*fn)(void *);
 143         void                    *data;
 144         /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
 145         unsigned int            num_threads;
 146         const struct cpumask    *active_cpus;
 147
 148         enum multi_stop_state   state;
 149         atomic_t                thread_ack;
 150 };
 151
 152 static void set_state(struct multi_stop_data *msdata,
 153                       enum multi_stop_state newstate)
 154 {
 155         /* Reset ack counter. */
 156         atomic_set(&msdata->thread_ack, msdata->num_threads);
 157         smp_wmb();
 158         msdata->state = newstate;
 159 }
 160
 161 /* Last one to ack a state moves to the next state. */
 162 static void ack_state(struct multi_stop_data *msdata)
 163 {
 164         if (atomic_dec_and_test(&msdata->thread_ack))
 165                 set_state(msdata, msdata->state + 1);
 166 }
 167
 168 /* This is the cpu_stop function which stops the CPU. */
 169 static int multi_cpu_stop(void *data)
 170 {
 171         struct multi_stop_data *msdata = data;
 172         enum multi_stop_state curstate = MULTI_STOP_NONE;
 173         int cpu = smp_processor_id(), err = 0;
 174         unsigned long flags;
 175         bool is_active;
 176
 177         /*
 178          * When called from stop_machine_from_inactive_cpu(), irq might
 179          * already be disabled.  Save the state and restore it on exit.
 180          */
 181         local_save_flags(flags);
 182
 183         if (!msdata->active_cpus)
 184                 is_active = cpu == cpumask_first(cpu_online_mask);
 185         else
 186                 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
 187
 188         /* Simple state machine */
 189         do {
 190                 /* Chill out and ensure we re-read multi_stop_state. */
 191                 cpu_relax();
 192                 if (msdata->state != curstate) {
 193                         curstate = msdata->state;
 194                         switch (curstate) {
 195                         case MULTI_STOP_DISABLE_IRQ:
 196                                 local_irq_disable();
 197                                 hard_irq_disable();
 198                                 break;
 199                         case MULTI_STOP_RUN:
 200                                 if (is_active)
 201                                         err = msdata->fn(msdata->data);
 202                                 break;
 203                         default:
 204                                 break;
 205                         }
 206                         ack_state(msdata);
 207                 }
 208         } while (curstate != MULTI_STOP_EXIT);
 209
 210         local_irq_restore(flags);
 211         return err;
 212 }
 213
 214 struct irq_cpu_stop_queue_work_info {
 215         int cpu1;
 216         int cpu2;
 217         struct cpu_stop_work *work1;
 218         struct cpu_stop_work *work2;
 219 };
 220
 221 /*
 222  * This function is always run with irqs and preemption disabled.
 223  * This guarantees that both work1 and work2 get queued, before
 224  * our local migrate thread gets the chance to preempt us.
 225  */
 226 static void irq_cpu_stop_queue_work(void *arg)
 227 {
 228         struct irq_cpu_stop_queue_work_info *info = arg;
 229         cpu_stop_queue_work(info->cpu1, info->work1);
 230         cpu_stop_queue_work(info->cpu2, info->work2);
 231 }
 232
 233 /**
 234  * stop_two_cpus - stops two cpus
 235  * @cpu1: the cpu to stop
 236  * @cpu2: the other cpu to stop
 237  * @fn: function to execute
 238  * @arg: argument to @fn
 239  *
 240  * Stops both the current and specified CPU and runs @fn on one of them.
 241  *
 242  * returns when both are completed.
 243  */
 244 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
 245 {
 246         struct cpu_stop_done done;
 247         struct cpu_stop_work work1, work2;
 248         struct irq_cpu_stop_queue_work_info call_args;
 249         struct multi_stop_data msdata;
 250
 251         preempt_disable();
 252         msdata = (struct multi_stop_data){
 253                 .fn = fn,
 254                 .data = arg,
 255                 .num_threads = 2,
 256                 .active_cpus = cpumask_of(cpu1),
 257         };
 258
 259         work1 = work2 = (struct cpu_stop_work){
 260                 .fn = multi_cpu_stop,
 261                 .arg = &msdata,
 262                 .done = &done
 263         };
 264
 265         call_args = (struct irq_cpu_stop_queue_work_info){
 266                 .cpu1 = cpu1,
 267                 .cpu2 = cpu2,
 268                 .work1 = &work1,
 269                 .work2 = &work2,
 270         };
 271
 272         cpu_stop_init_done(&done, 2);
 273         set_state(&msdata, MULTI_STOP_PREPARE);
 274
 275         /*
 276          * If we observe both CPUs active we know _cpu_down() cannot yet have
 277          * queued its stop_machine works and therefore ours will get executed
 278          * first. Or its not either one of our CPUs that's getting unplugged,
 279          * in which case we don't care.
 280          *
 281          * This relies on the stopper workqueues to be FIFO.
 282          */
 283         if (!cpu_active(cpu1) || !cpu_active(cpu2)) {
 284                 preempt_enable();
 285                 return -ENOENT;
 286         }
 287
 288         lg_local_lock(&stop_cpus_lock);
 289         /*
 290          * Queuing needs to be done by the lowest numbered CPU, to ensure
 291          * that works are always queued in the same order on every CPU.
 292          * This prevents deadlocks.
 293          */
 294         smp_call_function_single(min(cpu1, cpu2),
 295                                  &irq_cpu_stop_queue_work,
 296                                  &call_args, 1);
 297         lg_local_unlock(&stop_cpus_lock);
 298         preempt_enable();
 299
 300         wait_for_completion(&done.completion);
 301
 302         return done.executed ? done.ret : -ENOENT;
 303 }
 304
 305 /**
 306  * stop_one_cpu_nowait - stop a cpu but don't wait for completion
 307  * @cpu: cpu to stop
 308  * @fn: function to execute
 309  * @arg: argument to @fn
 310  * @work_buf: pointer to cpu_stop_work structure
 311  *
 312  * Similar to stop_one_cpu() but doesn't wait for completion.  The
 313  * caller is responsible for ensuring @work_buf is currently unused
 314  * and will remain untouched until stopper starts executing @fn.
 315  *
 316  * CONTEXT:
 317  * Don't care.
 318  */
 319 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
 320                         struct cpu_stop_work *work_buf)
 321 {
 322         *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
 323         cpu_stop_queue_work(cpu, work_buf);
 324 }
 325
 326 /* static data for stop_cpus */
 327 static DEFINE_MUTEX(stop_cpus_mutex);
 328 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
 329
 330 static void queue_stop_cpus_work(const struct cpumask *cpumask,
 331                                  cpu_stop_fn_t fn, void *arg,
 332                                  struct cpu_stop_done *done)
 333 {
 334         struct cpu_stop_work *work;
 335         unsigned int cpu;
 336
 337         /* initialize works and done */
 338         for_each_cpu(cpu, cpumask) {
 339                 work = &per_cpu(stop_cpus_work, cpu);
 340                 work->fn = fn;
 341                 work->arg = arg;
 342                 work->done = done;
 343         }
 344
 345         /*
 346          * Disable preemption while queueing to avoid getting
 347          * preempted by a stopper which might wait for other stoppers
 348          * to enter @fn which can lead to deadlock.
 349          */
 350         lg_global_lock(&stop_cpus_lock);
 351         for_each_cpu(cpu, cpumask)
 352                 cpu_stop_queue_work(cpu, &per_cpu(stop_cpus_work, cpu));
 353         lg_global_unlock(&stop_cpus_lock);
 354 }
 355
 356 static int __stop_cpus(const struct cpumask *cpumask,
 357                        cpu_stop_fn_t fn, void *arg)
 358 {
 359         struct cpu_stop_done done;
 360
 361         cpu_stop_init_done(&done, cpumask_weight(cpumask));
 362         queue_stop_cpus_work(cpumask, fn, arg, &done);
 363         wait_for_completion(&done.completion);
 364         return done.executed ? done.ret : -ENOENT;
 365 }
 366
 367 /**
 368  * stop_cpus - stop multiple cpus
 369  * @cpumask: cpus to stop
 370  * @fn: function to execute
 371  * @arg: argument to @fn
 372  *
 373  * Execute @fn(@arg) on online cpus in @cpumask.  On each target cpu,
 374  * @fn is run in a process context with the highest priority
 375  * preempting any task on the cpu and monopolizing it.  This function
 376  * returns after all executions are complete.
 377  *
 378  * This function doesn't guarantee the cpus in @cpumask stay online
 379  * till @fn completes.  If some cpus go down in the middle, execution
 380  * on the cpu may happen partially or fully on different cpus.  @fn
 381  * should either be ready for that or the caller should ensure that
 382  * the cpus stay online until this function completes.
 383  *
 384  * All stop_cpus() calls are serialized making it safe for @fn to wait
 385  * for all cpus to start executing it.
 386  *
 387  * CONTEXT:
 388  * Might sleep.
 389  *
 390  * RETURNS:
 391  * -ENOENT if @fn(@arg) was not executed at all because all cpus in
 392  * @cpumask were offline; otherwise, 0 if all executions of @fn
 393  * returned 0, any non zero return value if any returned non zero.
 394  */
 395 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 396 {
 397         int ret;
 398
 399         /* static works are used, process one request at a time */
 400         mutex_lock(&stop_cpus_mutex);
 401         ret = __stop_cpus(cpumask, fn, arg);
 402         mutex_unlock(&stop_cpus_mutex);
 403         return ret;
 404 }
 405
 406 /**
 407  * try_stop_cpus - try to stop multiple cpus
 408  * @cpumask: cpus to stop
 409  * @fn: function to execute
 410  * @arg: argument to @fn
 411  *
 412  * Identical to stop_cpus() except that it fails with -EAGAIN if
 413  * someone else is already using the facility.
 414  *
 415  * CONTEXT:
 416  * Might sleep.
 417  *
 418  * RETURNS:
 419  * -EAGAIN if someone else is already stopping cpus, -ENOENT if
 420  * @fn(@arg) was not executed at all because all cpus in @cpumask were
 421  * offline; otherwise, 0 if all executions of @fn returned 0, any non
 422  * zero return value if any returned non zero.
 423  */
 424 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 425 {
 426         int ret;
 427
 428         /* static works are used, process one request at a time */
 429         if (!mutex_trylock(&stop_cpus_mutex))
 430                 return -EAGAIN;
 431         ret = __stop_cpus(cpumask, fn, arg);
 432         mutex_unlock(&stop_cpus_mutex);
 433         return ret;
 434 }
 435
 436 static int cpu_stop_should_run(unsigned int cpu)
 437 {
 438         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 439         unsigned long flags;
 440         int run;
 441
 442         spin_lock_irqsave(&stopper->lock, flags);
 443         run = !list_empty(&stopper->works);
 444         spin_unlock_irqrestore(&stopper->lock, flags);
 445         return run;
 446 }
 447
 448 static void cpu_stopper_thread(unsigned int cpu)
 449 {
 450         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 451         struct cpu_stop_work *work;
 452         int ret;
 453
 454 repeat:
 455         work = NULL;
 456         spin_lock_irq(&stopper->lock);
 457         if (!list_empty(&stopper->works)) {
 458                 work = list_first_entry(&stopper->works,
 459                                         struct cpu_stop_work, list);
 460                 list_del_init(&work->list);
 461         }
 462         spin_unlock_irq(&stopper->lock);
 463
 464         if (work) {
 465                 cpu_stop_fn_t fn = work->fn;
 466                 void *arg = work->arg;
 467                 struct cpu_stop_done *done = work->done;
 468                 char ksym_buf[KSYM_NAME_LEN] __maybe_unused;
 469
 470                 /* cpu stop callbacks are not allowed to sleep */
 471                 preempt_disable();
 472
 473                 ret = fn(arg);
 474                 if (ret)
 475                         done->ret = ret;
 476
 477                 /* restore preemption and check it's still balanced */
 478                 preempt_enable();
 479                 WARN_ONCE(preempt_count(),
 480                           "cpu_stop: %s(%p) leaked preempt count\n",
 481                           kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
 482                                           ksym_buf), arg);
 483
 484                 cpu_stop_signal_done(done, true);
 485                 goto repeat;
 486         }
 487 }
 488
 489 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
 490
 491 static void cpu_stop_create(unsigned int cpu)
 492 {
 493         sched_set_stop_task(cpu, per_cpu(cpu_stopper_task, cpu));
 494 }
 495
 496 static void cpu_stop_park(unsigned int cpu)
 497 {
 498         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 499         struct cpu_stop_work *work;
 500         unsigned long flags;
 501
 502         /* drain remaining works */
 503         spin_lock_irqsave(&stopper->lock, flags);
 504         list_for_each_entry(work, &stopper->works, list)
 505                 cpu_stop_signal_done(work->done, false);
 506         stopper->enabled = false;
 507         spin_unlock_irqrestore(&stopper->lock, flags);
 508 }
 509
 510 static void cpu_stop_unpark(unsigned int cpu)
 511 {
 512         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 513
 514         spin_lock_irq(&stopper->lock);
 515         stopper->enabled = true;
 516         spin_unlock_irq(&stopper->lock);
 517 }
 518
 519 static struct smp_hotplug_thread cpu_stop_threads = {
 520         .store                  = &cpu_stopper_task,
 521         .thread_should_run      = cpu_stop_should_run,
 522         .thread_fn              = cpu_stopper_thread,
 523         .thread_comm            = "migration/%u",
 524         .create                 = cpu_stop_create,
 525         .setup                  = cpu_stop_unpark,
 526         .park                   = cpu_stop_park,
 527         .pre_unpark             = cpu_stop_unpark,
 528         .selfparking            = true,
 529 };
 530
 531 static int __init cpu_stop_init(void)
 532 {
 533         unsigned int cpu;
 534
 535         for_each_possible_cpu(cpu) {
 536                 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 537
 538                 spin_lock_init(&stopper->lock);
 539                 INIT_LIST_HEAD(&stopper->works);
 540         }
 541
 542         BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
 543         stop_machine_initialized = true;
 544         return 0;
 545 }
 546 early_initcall(cpu_stop_init);
 547
 548 #ifdef CONFIG_STOP_MACHINE
 549
 550 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
 551 {
 552         struct multi_stop_data msdata = {
 553                 .fn = fn,
 554                 .data = data,
 555                 .num_threads = num_online_cpus(),
 556                 .active_cpus = cpus,
 557         };
 558
 559         if (!stop_machine_initialized) {
 560                 /*
 561                  * Handle the case where stop_machine() is called
 562                  * early in boot before stop_machine() has been
 563                  * initialized.
 564                  */
 565                 unsigned long flags;
 566                 int ret;
 567
 568                 WARN_ON_ONCE(msdata.num_threads != 1);
 569
 570                 local_irq_save(flags);
 571                 hard_irq_disable();
 572                 ret = (*fn)(data);
 573                 local_irq_restore(flags);
 574
 575                 return ret;
 576         }
 577
 578         /* Set the initial state and stop all online cpus. */
 579         set_state(&msdata, MULTI_STOP_PREPARE);
 580         return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
 581 }
 582
 583 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
 584 {
 585         int ret;
 586
 587         /* No CPUs can come up or down during this. */
 588         get_online_cpus();
 589         ret = __stop_machine(fn, data, cpus);
 590         put_online_cpus();
 591         return ret;
 592 }
 593 EXPORT_SYMBOL_GPL(stop_machine);
 594
 595 /**
 596  * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
 597  * @fn: the function to run
 598  * @data: the data ptr for the @fn()
 599  * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
 600  *
 601  * This is identical to stop_machine() but can be called from a CPU which
 602  * is not active.  The local CPU is in the process of hotplug (so no other
 603  * CPU hotplug can start) and not marked active and doesn't have enough
 604  * context to sleep.
 605  *
 606  * This function provides stop_machine() functionality for such state by
 607  * using busy-wait for synchronization and executing @fn directly for local
 608  * CPU.
 609  *
 610  * CONTEXT:
 611  * Local CPU is inactive.  Temporarily stops all active CPUs.
 612  *
 613  * RETURNS:
 614  * 0 if all executions of @fn returned 0, any non zero return value if any
 615  * returned non zero.
 616  */
 617 int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
 618                                   const struct cpumask *cpus)
 619 {
 620         struct multi_stop_data msdata = { .fn = fn, .data = data,
 621                                             .active_cpus = cpus };
 622         struct cpu_stop_done done;
 623         int ret;
 624
 625         /* Local CPU must be inactive and CPU hotplug in progress. */
 626         BUG_ON(cpu_active(raw_smp_processor_id()));
 627         msdata.num_threads = num_active_cpus() + 1;     /* +1 for local */
 628
 629         /* No proper task established and can't sleep - busy wait for lock. */
 630         while (!mutex_trylock(&stop_cpus_mutex))
 631                 cpu_relax();
 632
 633         /* Schedule work on other CPUs and execute directly for local CPU */
 634         set_state(&msdata, MULTI_STOP_PREPARE);
 635         cpu_stop_init_done(&done, num_active_cpus());
 636         queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
 637                              &done);
 638         ret = multi_cpu_stop(&msdata);
 639
 640         /* Busy wait for completion. */
 641         while (!completion_done(&done.completion))
 642                 cpu_relax();
 643
 644         mutex_unlock(&stop_cpus_mutex);
 645         return ret ?: done.ret;
 646 }
 647
 648 #endif  /* CONFIG_STOP_MACHINE */