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/module.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
  22 #include <asm/atomic.h>
  23
  24 /*
  25  * Structure to determine completion condition and record errors.  May
  26  * be shared by works on different cpus.
  27  */
  28 struct cpu_stop_done {
  29         atomic_t                nr_todo;        /* nr left to execute */
  30         bool                    executed;       /* actually executed? */
  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         spinlock_t              lock;
  38         struct list_head        works;          /* list of pending works */
  39         struct task_struct      *thread;        /* stopper thread */
  40         bool                    enabled;        /* is this stopper enabled? */
  41 };
  42
  43 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
  44
  45 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
  46 {
  47         memset(done, 0, sizeof(*done));
  48         atomic_set(&done->nr_todo, nr_todo);
  49         init_completion(&done->completion);
  50 }
  51
  52 /* signal completion unless @done is NULL */
  53 static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
  54 {
  55         if (done) {
  56                 if (executed)
  57                         done->executed = true;
  58                 if (atomic_dec_and_test(&done->nr_todo))
  59                         complete(&done->completion);
  60         }
  61 }
  62
  63 /* queue @work to @stopper.  if offline, @work is completed immediately */
  64 static void cpu_stop_queue_work(struct cpu_stopper *stopper,
  65                                 struct cpu_stop_work *work)
  66 {
  67         unsigned long flags;
  68
  69         spin_lock_irqsave(&stopper->lock, flags);
  70
  71         if (stopper->enabled) {
  72                 list_add_tail(&work->list, &stopper->works);
  73                 wake_up_process(stopper->thread);
  74         } else
  75                 cpu_stop_signal_done(work->done, false);
  76
  77         spin_unlock_irqrestore(&stopper->lock, flags);
  78 }
  79
  80 /**
  81  * stop_one_cpu - stop a cpu
  82  * @cpu: cpu to stop
  83  * @fn: function to execute
  84  * @arg: argument to @fn
  85  *
  86  * Execute @fn(@arg) on @cpu.  @fn is run in a process context with
  87  * the highest priority preempting any task on the cpu and
  88  * monopolizing it.  This function returns after the execution is
  89  * complete.
  90  *
  91  * This function doesn't guarantee @cpu stays online till @fn
  92  * completes.  If @cpu goes down in the middle, execution may happen
  93  * partially or fully on different cpus.  @fn should either be ready
  94  * for that or the caller should ensure that @cpu stays online until
  95  * this function completes.
  96  *
  97  * CONTEXT:
  98  * Might sleep.
  99  *
 100  * RETURNS:
 101  * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
 102  * otherwise, the return value of @fn.
 103  */
 104 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
 105 {
 106         struct cpu_stop_done done;
 107         struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
 108
 109         cpu_stop_init_done(&done, 1);
 110         cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), &work);
 111         wait_for_completion(&done.completion);
 112         return done.executed ? done.ret : -ENOENT;
 113 }
 114
 115 /**
 116  * stop_one_cpu_nowait - stop a cpu but don't wait for completion
 117  * @cpu: cpu to stop
 118  * @fn: function to execute
 119  * @arg: argument to @fn
 120  *
 121  * Similar to stop_one_cpu() but doesn't wait for completion.  The
 122  * caller is responsible for ensuring @work_buf is currently unused
 123  * and will remain untouched until stopper starts executing @fn.
 124  *
 125  * CONTEXT:
 126  * Don't care.
 127  */
 128 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
 129                         struct cpu_stop_work *work_buf)
 130 {
 131         *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
 132         cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
 133 }
 134
 135 /* static data for stop_cpus */
 136 static DEFINE_MUTEX(stop_cpus_mutex);
 137 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
 138
 139 int __stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 140 {
 141         struct cpu_stop_work *work;
 142         struct cpu_stop_done done;
 143         unsigned int cpu;
 144
 145         /* initialize works and done */
 146         for_each_cpu(cpu, cpumask) {
 147                 work = &per_cpu(stop_cpus_work, cpu);
 148                 work->fn = fn;
 149                 work->arg = arg;
 150                 work->done = &done;
 151         }
 152         cpu_stop_init_done(&done, cpumask_weight(cpumask));
 153
 154         /*
 155          * Disable preemption while queueing to avoid getting
 156          * preempted by a stopper which might wait for other stoppers
 157          * to enter @fn which can lead to deadlock.
 158          */
 159         preempt_disable();
 160         for_each_cpu(cpu, cpumask)
 161                 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
 162                                     &per_cpu(stop_cpus_work, cpu));
 163         preempt_enable();
 164
 165         wait_for_completion(&done.completion);
 166         return done.executed ? done.ret : -ENOENT;
 167 }
 168
 169 /**
 170  * stop_cpus - stop multiple cpus
 171  * @cpumask: cpus to stop
 172  * @fn: function to execute
 173  * @arg: argument to @fn
 174  *
 175  * Execute @fn(@arg) on online cpus in @cpumask.  On each target cpu,
 176  * @fn is run in a process context with the highest priority
 177  * preempting any task on the cpu and monopolizing it.  This function
 178  * returns after all executions are complete.
 179  *
 180  * This function doesn't guarantee the cpus in @cpumask stay online
 181  * till @fn completes.  If some cpus go down in the middle, execution
 182  * on the cpu may happen partially or fully on different cpus.  @fn
 183  * should either be ready for that or the caller should ensure that
 184  * the cpus stay online until this function completes.
 185  *
 186  * All stop_cpus() calls are serialized making it safe for @fn to wait
 187  * for all cpus to start executing it.
 188  *
 189  * CONTEXT:
 190  * Might sleep.
 191  *
 192  * RETURNS:
 193  * -ENOENT if @fn(@arg) was not executed at all because all cpus in
 194  * @cpumask were offline; otherwise, 0 if all executions of @fn
 195  * returned 0, any non zero return value if any returned non zero.
 196  */
 197 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 198 {
 199         int ret;
 200
 201         /* static works are used, process one request at a time */
 202         mutex_lock(&stop_cpus_mutex);
 203         ret = __stop_cpus(cpumask, fn, arg);
 204         mutex_unlock(&stop_cpus_mutex);
 205         return ret;
 206 }
 207
 208 /**
 209  * try_stop_cpus - try to stop multiple cpus
 210  * @cpumask: cpus to stop
 211  * @fn: function to execute
 212  * @arg: argument to @fn
 213  *
 214  * Identical to stop_cpus() except that it fails with -EAGAIN if
 215  * someone else is already using the facility.
 216  *
 217  * CONTEXT:
 218  * Might sleep.
 219  *
 220  * RETURNS:
 221  * -EAGAIN if someone else is already stopping cpus, -ENOENT if
 222  * @fn(@arg) was not executed at all because all cpus in @cpumask were
 223  * offline; otherwise, 0 if all executions of @fn returned 0, any non
 224  * zero return value if any returned non zero.
 225  */
 226 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
 227 {
 228         int ret;
 229
 230         /* static works are used, process one request at a time */
 231         if (!mutex_trylock(&stop_cpus_mutex))
 232                 return -EAGAIN;
 233         ret = __stop_cpus(cpumask, fn, arg);
 234         mutex_unlock(&stop_cpus_mutex);
 235         return ret;
 236 }
 237
 238 static int cpu_stopper_thread(void *data)
 239 {
 240         struct cpu_stopper *stopper = data;
 241         struct cpu_stop_work *work;
 242         int ret;
 243
 244 repeat:
 245         set_current_state(TASK_INTERRUPTIBLE);  /* mb paired w/ kthread_stop */
 246
 247         if (kthread_should_stop()) {
 248                 __set_current_state(TASK_RUNNING);
 249                 return 0;
 250         }
 251
 252         work = NULL;
 253         spin_lock_irq(&stopper->lock);
 254         if (!list_empty(&stopper->works)) {
 255                 work = list_first_entry(&stopper->works,
 256                                         struct cpu_stop_work, list);
 257                 list_del_init(&work->list);
 258         }
 259         spin_unlock_irq(&stopper->lock);
 260
 261         if (work) {
 262                 cpu_stop_fn_t fn = work->fn;
 263                 void *arg = work->arg;
 264                 struct cpu_stop_done *done = work->done;
 265                 char ksym_buf[KSYM_NAME_LEN];
 266
 267                 __set_current_state(TASK_RUNNING);
 268
 269                 /* cpu stop callbacks are not allowed to sleep */
 270                 preempt_disable();
 271
 272                 ret = fn(arg);
 273                 if (ret)
 274                         done->ret = ret;
 275
 276                 /* restore preemption and check it's still balanced */
 277                 preempt_enable();
 278                 WARN_ONCE(preempt_count(),
 279                           "cpu_stop: %s(%p) leaked preempt count\n",
 280                           kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
 281                                           ksym_buf), arg);
 282
 283                 cpu_stop_signal_done(done, true);
 284         } else
 285                 schedule();
 286
 287         goto repeat;
 288 }
 289
 290 /* manage stopper for a cpu, mostly lifted from sched migration thread mgmt */
 291 static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
 292                                            unsigned long action, void *hcpu)
 293 {
 294         struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
 295         unsigned int cpu = (unsigned long)hcpu;
 296         struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 297         struct task_struct *p;
 298
 299         switch (action & ~CPU_TASKS_FROZEN) {
 300         case CPU_UP_PREPARE:
 301                 BUG_ON(stopper->thread || stopper->enabled ||
 302                        !list_empty(&stopper->works));
 303                 p = kthread_create(cpu_stopper_thread, stopper, "migration/%d",
 304                                    cpu);
 305                 if (IS_ERR(p))
 306                         return NOTIFY_BAD;
 307                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
 308                 get_task_struct(p);
 309                 stopper->thread = p;
 310                 break;
 311
 312         case CPU_ONLINE:
 313                 kthread_bind(stopper->thread, cpu);
 314                 /* strictly unnecessary, as first user will wake it */
 315                 wake_up_process(stopper->thread);
 316                 /* mark enabled */
 317                 spin_lock_irq(&stopper->lock);
 318                 stopper->enabled = true;
 319                 spin_unlock_irq(&stopper->lock);
 320                 break;
 321
 322 #ifdef CONFIG_HOTPLUG_CPU
 323         case CPU_UP_CANCELED:
 324         case CPU_DEAD:
 325         {
 326                 struct cpu_stop_work *work;
 327
 328                 /* kill the stopper */
 329                 kthread_stop(stopper->thread);
 330                 /* drain remaining works */
 331                 spin_lock_irq(&stopper->lock);
 332                 list_for_each_entry(work, &stopper->works, list)
 333                         cpu_stop_signal_done(work->done, false);
 334                 stopper->enabled = false;
 335                 spin_unlock_irq(&stopper->lock);
 336                 /* release the stopper */
 337                 put_task_struct(stopper->thread);
 338                 stopper->thread = NULL;
 339                 break;
 340         }
 341 #endif
 342         }
 343
 344         return NOTIFY_OK;
 345 }
 346
 347 /*
 348  * Give it a higher priority so that cpu stopper is available to other
 349  * cpu notifiers.  It currently shares the same priority as sched
 350  * migration_notifier.
 351  */
 352 static struct notifier_block __cpuinitdata cpu_stop_cpu_notifier = {
 353         .notifier_call  = cpu_stop_cpu_callback,
 354         .priority       = 10,
 355 };
 356
 357 static int __init cpu_stop_init(void)
 358 {
 359         void *bcpu = (void *)(long)smp_processor_id();
 360         unsigned int cpu;
 361         int err;
 362
 363         for_each_possible_cpu(cpu) {
 364                 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 365
 366                 spin_lock_init(&stopper->lock);
 367                 INIT_LIST_HEAD(&stopper->works);
 368         }
 369
 370         /* start one for the boot cpu */
 371         err = cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_UP_PREPARE,
 372                                     bcpu);
 373         BUG_ON(err == NOTIFY_BAD);
 374         cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_ONLINE, bcpu);
 375         register_cpu_notifier(&cpu_stop_cpu_notifier);
 376
 377         return 0;
 378 }
 379 early_initcall(cpu_stop_init);
 380
 381 #ifdef CONFIG_STOP_MACHINE
 382
 383 /* This controls the threads on each CPU. */
 384 enum stopmachine_state {
 385         /* Dummy starting state for thread. */
 386         STOPMACHINE_NONE,
 387         /* Awaiting everyone to be scheduled. */
 388         STOPMACHINE_PREPARE,
 389         /* Disable interrupts. */
 390         STOPMACHINE_DISABLE_IRQ,
 391         /* Run the function */
 392         STOPMACHINE_RUN,
 393         /* Exit */
 394         STOPMACHINE_EXIT,
 395 };
 396
 397 struct stop_machine_data {
 398         int                     (*fn)(void *);
 399         void                    *data;
 400         /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
 401         unsigned int            num_threads;
 402         const struct cpumask    *active_cpus;
 403
 404         enum stopmachine_state  state;
 405         atomic_t                thread_ack;
 406 };
 407
 408 static void set_state(struct stop_machine_data *smdata,
 409                       enum stopmachine_state newstate)
 410 {
 411         /* Reset ack counter. */
 412         atomic_set(&smdata->thread_ack, smdata->num_threads);
 413         smp_wmb();
 414         smdata->state = newstate;
 415 }
 416
 417 /* Last one to ack a state moves to the next state. */
 418 static void ack_state(struct stop_machine_data *smdata)
 419 {
 420         if (atomic_dec_and_test(&smdata->thread_ack))
 421                 set_state(smdata, smdata->state + 1);
 422 }
 423
 424 /* This is the cpu_stop function which stops the CPU. */
 425 static int stop_machine_cpu_stop(void *data)
 426 {
 427         struct stop_machine_data *smdata = data;
 428         enum stopmachine_state curstate = STOPMACHINE_NONE;
 429         int cpu = smp_processor_id(), err = 0;
 430         bool is_active;
 431
 432         if (!smdata->active_cpus)
 433                 is_active = cpu == cpumask_first(cpu_online_mask);
 434         else
 435                 is_active = cpumask_test_cpu(cpu, smdata->active_cpus);
 436
 437         /* Simple state machine */
 438         do {
 439                 /* Chill out and ensure we re-read stopmachine_state. */
 440                 cpu_relax();
 441                 if (smdata->state != curstate) {
 442                         curstate = smdata->state;
 443                         switch (curstate) {
 444                         case STOPMACHINE_DISABLE_IRQ:
 445                                 local_irq_disable();
 446                                 hard_irq_disable();
 447                                 break;
 448                         case STOPMACHINE_RUN:
 449                                 if (is_active)
 450                                         err = smdata->fn(smdata->data);
 451                                 break;
 452                         default:
 453                                 break;
 454                         }
 455                         ack_state(smdata);
 456                 }
 457         } while (curstate != STOPMACHINE_EXIT);
 458
 459         local_irq_enable();
 460         return err;
 461 }
 462
 463 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
 464 {
 465         struct stop_machine_data smdata = { .fn = fn, .data = data,
 466                                             .num_threads = num_online_cpus(),
 467                                             .active_cpus = cpus };
 468
 469         /* Set the initial state and stop all online cpus. */
 470         set_state(&smdata, STOPMACHINE_PREPARE);
 471         return stop_cpus(cpu_online_mask, stop_machine_cpu_stop, &smdata);
 472 }
 473
 474 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
 475 {
 476         int ret;
 477
 478         /* No CPUs can come up or down during this. */
 479         get_online_cpus();
 480         ret = __stop_machine(fn, data, cpus);
 481         put_online_cpus();
 482         return ret;
 483 }
 484 EXPORT_SYMBOL_GPL(stop_machine);
 485
 486 #endif  /* CONFIG_STOP_MACHINE */