2 * kernel/stop_machine.c
4 * Copyright (C) 2008, 2005 IBM Corporation.
6 * Copyright (C) 2010 SUSE Linux Products GmbH
9 * This file is released under the GPLv2 and any later version.
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 #include <linux/sched/wake_q.h>
27 * Structure to determine completion condition and record errors. May
28 * be shared by works on different cpus.
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 */
36 /* the actual stopper, one per every possible cpu, enabled on online cpus */
38 struct task_struct *thread;
41 bool enabled; /* is this stopper enabled? */
42 struct list_head works; /* list of pending works */
44 struct cpu_stop_work stop_work; /* for stop_cpus */
47 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
48 static bool stop_machine_initialized = false;
50 /* static data for stop_cpus */
51 static DEFINE_MUTEX(stop_cpus_mutex);
52 static bool stop_cpus_in_progress;
54 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
56 memset(done, 0, sizeof(*done));
57 atomic_set(&done->nr_todo, nr_todo);
58 init_completion(&done->completion);
61 /* signal completion unless @done is NULL */
62 static void cpu_stop_signal_done(struct cpu_stop_done *done)
64 if (atomic_dec_and_test(&done->nr_todo))
65 complete(&done->completion);
68 static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
69 struct cpu_stop_work *work,
70 struct wake_q_head *wakeq)
72 list_add_tail(&work->list, &stopper->works);
73 wake_q_add(wakeq, stopper->thread);
76 /* queue @work to @stopper. if offline, @work is completed immediately */
77 static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
79 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
84 raw_spin_lock_irqsave(&stopper->lock, flags);
85 enabled = stopper->enabled;
87 __cpu_stop_queue_work(stopper, work, &wakeq);
89 cpu_stop_signal_done(work->done);
90 raw_spin_unlock_irqrestore(&stopper->lock, flags);
98 * stop_one_cpu - stop a cpu
100 * @fn: function to execute
101 * @arg: argument to @fn
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
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.
118 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
119 * otherwise, the return value of @fn.
121 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
123 struct cpu_stop_done done;
124 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
126 cpu_stop_init_done(&done, 1);
127 if (!cpu_stop_queue_work(cpu, &work))
130 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
131 * cycle by doing a preemption:
134 wait_for_completion(&done.completion);
138 /* This controls the threads on each CPU. */
139 enum multi_stop_state {
140 /* Dummy starting state for thread. */
142 /* Awaiting everyone to be scheduled. */
144 /* Disable interrupts. */
145 MULTI_STOP_DISABLE_IRQ,
146 /* Run the function */
152 struct multi_stop_data {
155 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
156 unsigned int num_threads;
157 const struct cpumask *active_cpus;
159 enum multi_stop_state state;
163 static void set_state(struct multi_stop_data *msdata,
164 enum multi_stop_state newstate)
166 /* Reset ack counter. */
167 atomic_set(&msdata->thread_ack, msdata->num_threads);
169 msdata->state = newstate;
172 /* Last one to ack a state moves to the next state. */
173 static void ack_state(struct multi_stop_data *msdata)
175 if (atomic_dec_and_test(&msdata->thread_ack))
176 set_state(msdata, msdata->state + 1);
179 /* This is the cpu_stop function which stops the CPU. */
180 static int multi_cpu_stop(void *data)
182 struct multi_stop_data *msdata = data;
183 enum multi_stop_state curstate = MULTI_STOP_NONE;
184 int cpu = smp_processor_id(), err = 0;
189 * When called from stop_machine_from_inactive_cpu(), irq might
190 * already be disabled. Save the state and restore it on exit.
192 local_save_flags(flags);
194 if (!msdata->active_cpus)
195 is_active = cpu == cpumask_first(cpu_online_mask);
197 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
199 /* Simple state machine */
201 /* Chill out and ensure we re-read multi_stop_state. */
203 if (msdata->state != curstate) {
204 curstate = msdata->state;
206 case MULTI_STOP_DISABLE_IRQ:
212 err = msdata->fn(msdata->data);
218 } else if (curstate > MULTI_STOP_PREPARE) {
220 * At this stage all other CPUs we depend on must spin
221 * in the same loop. Any reason for hard-lockup should
222 * be detected and reported on their side.
224 touch_nmi_watchdog();
226 } while (curstate != MULTI_STOP_EXIT);
228 local_irq_restore(flags);
232 static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
233 int cpu2, struct cpu_stop_work *work2)
235 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
236 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
237 DEFINE_WAKE_Q(wakeq);
240 raw_spin_lock_irq(&stopper1->lock);
241 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
244 if (!stopper1->enabled || !stopper2->enabled)
247 * Ensure that if we race with __stop_cpus() the stoppers won't get
248 * queued up in reverse order leading to system deadlock.
250 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
251 * queued a work on cpu1 but not on cpu2, we hold both locks.
253 * It can be falsely true but it is safe to spin until it is cleared,
254 * queue_stop_cpus_work() does everything under preempt_disable().
257 if (unlikely(stop_cpus_in_progress))
261 __cpu_stop_queue_work(stopper1, work1, &wakeq);
262 __cpu_stop_queue_work(stopper2, work2, &wakeq);
264 * The waking up of stopper threads has to happen
265 * in the same scheduling context as the queueing.
266 * Otherwise, there is a possibility of one of the
267 * above stoppers being woken up by another CPU,
268 * and preempting us. This will cause us to n ot
269 * wake up the other stopper forever.
273 raw_spin_unlock(&stopper2->lock);
274 raw_spin_unlock_irq(&stopper1->lock);
276 if (unlikely(err == -EDEADLK)) {
277 while (stop_cpus_in_progress)
290 * stop_two_cpus - stops two cpus
291 * @cpu1: the cpu to stop
292 * @cpu2: the other cpu to stop
293 * @fn: function to execute
294 * @arg: argument to @fn
296 * Stops both the current and specified CPU and runs @fn on one of them.
298 * returns when both are completed.
300 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
302 struct cpu_stop_done done;
303 struct cpu_stop_work work1, work2;
304 struct multi_stop_data msdata;
306 msdata = (struct multi_stop_data){
310 .active_cpus = cpumask_of(cpu1),
313 work1 = work2 = (struct cpu_stop_work){
314 .fn = multi_cpu_stop,
319 cpu_stop_init_done(&done, 2);
320 set_state(&msdata, MULTI_STOP_PREPARE);
324 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
327 wait_for_completion(&done.completion);
332 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
334 * @fn: function to execute
335 * @arg: argument to @fn
336 * @work_buf: pointer to cpu_stop_work structure
338 * Similar to stop_one_cpu() but doesn't wait for completion. The
339 * caller is responsible for ensuring @work_buf is currently unused
340 * and will remain untouched until stopper starts executing @fn.
346 * true if cpu_stop_work was queued successfully and @fn will be called,
349 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
350 struct cpu_stop_work *work_buf)
352 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
353 return cpu_stop_queue_work(cpu, work_buf);
356 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
357 cpu_stop_fn_t fn, void *arg,
358 struct cpu_stop_done *done)
360 struct cpu_stop_work *work;
365 * Disable preemption while queueing to avoid getting
366 * preempted by a stopper which might wait for other stoppers
367 * to enter @fn which can lead to deadlock.
370 stop_cpus_in_progress = true;
371 for_each_cpu(cpu, cpumask) {
372 work = &per_cpu(cpu_stopper.stop_work, cpu);
376 if (cpu_stop_queue_work(cpu, work))
379 stop_cpus_in_progress = false;
385 static int __stop_cpus(const struct cpumask *cpumask,
386 cpu_stop_fn_t fn, void *arg)
388 struct cpu_stop_done done;
390 cpu_stop_init_done(&done, cpumask_weight(cpumask));
391 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
393 wait_for_completion(&done.completion);
398 * stop_cpus - stop multiple cpus
399 * @cpumask: cpus to stop
400 * @fn: function to execute
401 * @arg: argument to @fn
403 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
404 * @fn is run in a process context with the highest priority
405 * preempting any task on the cpu and monopolizing it. This function
406 * returns after all executions are complete.
408 * This function doesn't guarantee the cpus in @cpumask stay online
409 * till @fn completes. If some cpus go down in the middle, execution
410 * on the cpu may happen partially or fully on different cpus. @fn
411 * should either be ready for that or the caller should ensure that
412 * the cpus stay online until this function completes.
414 * All stop_cpus() calls are serialized making it safe for @fn to wait
415 * for all cpus to start executing it.
421 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
422 * @cpumask were offline; otherwise, 0 if all executions of @fn
423 * returned 0, any non zero return value if any returned non zero.
425 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
429 /* static works are used, process one request at a time */
430 mutex_lock(&stop_cpus_mutex);
431 ret = __stop_cpus(cpumask, fn, arg);
432 mutex_unlock(&stop_cpus_mutex);
437 * try_stop_cpus - try to stop multiple cpus
438 * @cpumask: cpus to stop
439 * @fn: function to execute
440 * @arg: argument to @fn
442 * Identical to stop_cpus() except that it fails with -EAGAIN if
443 * someone else is already using the facility.
449 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
450 * @fn(@arg) was not executed at all because all cpus in @cpumask were
451 * offline; otherwise, 0 if all executions of @fn returned 0, any non
452 * zero return value if any returned non zero.
454 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
458 /* static works are used, process one request at a time */
459 if (!mutex_trylock(&stop_cpus_mutex))
461 ret = __stop_cpus(cpumask, fn, arg);
462 mutex_unlock(&stop_cpus_mutex);
466 static int cpu_stop_should_run(unsigned int cpu)
468 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
472 raw_spin_lock_irqsave(&stopper->lock, flags);
473 run = !list_empty(&stopper->works);
474 raw_spin_unlock_irqrestore(&stopper->lock, flags);
478 static void cpu_stopper_thread(unsigned int cpu)
480 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
481 struct cpu_stop_work *work;
485 raw_spin_lock_irq(&stopper->lock);
486 if (!list_empty(&stopper->works)) {
487 work = list_first_entry(&stopper->works,
488 struct cpu_stop_work, list);
489 list_del_init(&work->list);
491 raw_spin_unlock_irq(&stopper->lock);
494 cpu_stop_fn_t fn = work->fn;
495 void *arg = work->arg;
496 struct cpu_stop_done *done = work->done;
499 /* cpu stop callbacks must not sleep, make in_atomic() == T */
505 cpu_stop_signal_done(done);
508 WARN_ONCE(preempt_count(),
509 "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
514 void stop_machine_park(int cpu)
516 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
518 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
519 * the pending works before it parks, until then it is fine to queue
522 stopper->enabled = false;
523 kthread_park(stopper->thread);
526 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
528 static void cpu_stop_create(unsigned int cpu)
530 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
533 static void cpu_stop_park(unsigned int cpu)
535 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
537 WARN_ON(!list_empty(&stopper->works));
540 void stop_machine_unpark(int cpu)
542 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
544 stopper->enabled = true;
545 kthread_unpark(stopper->thread);
548 static struct smp_hotplug_thread cpu_stop_threads = {
549 .store = &cpu_stopper.thread,
550 .thread_should_run = cpu_stop_should_run,
551 .thread_fn = cpu_stopper_thread,
552 .thread_comm = "migration/%u",
553 .create = cpu_stop_create,
554 .park = cpu_stop_park,
558 static int __init cpu_stop_init(void)
562 for_each_possible_cpu(cpu) {
563 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
565 raw_spin_lock_init(&stopper->lock);
566 INIT_LIST_HEAD(&stopper->works);
569 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
570 stop_machine_unpark(raw_smp_processor_id());
571 stop_machine_initialized = true;
574 early_initcall(cpu_stop_init);
576 int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
577 const struct cpumask *cpus)
579 struct multi_stop_data msdata = {
582 .num_threads = num_online_cpus(),
586 lockdep_assert_cpus_held();
588 if (!stop_machine_initialized) {
590 * Handle the case where stop_machine() is called
591 * early in boot before stop_machine() has been
597 WARN_ON_ONCE(msdata.num_threads != 1);
599 local_irq_save(flags);
602 local_irq_restore(flags);
607 /* Set the initial state and stop all online cpus. */
608 set_state(&msdata, MULTI_STOP_PREPARE);
609 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
612 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
616 /* No CPUs can come up or down during this. */
618 ret = stop_machine_cpuslocked(fn, data, cpus);
622 EXPORT_SYMBOL_GPL(stop_machine);
625 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
626 * @fn: the function to run
627 * @data: the data ptr for the @fn()
628 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
630 * This is identical to stop_machine() but can be called from a CPU which
631 * is not active. The local CPU is in the process of hotplug (so no other
632 * CPU hotplug can start) and not marked active and doesn't have enough
635 * This function provides stop_machine() functionality for such state by
636 * using busy-wait for synchronization and executing @fn directly for local
640 * Local CPU is inactive. Temporarily stops all active CPUs.
643 * 0 if all executions of @fn returned 0, any non zero return value if any
646 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
647 const struct cpumask *cpus)
649 struct multi_stop_data msdata = { .fn = fn, .data = data,
650 .active_cpus = cpus };
651 struct cpu_stop_done done;
654 /* Local CPU must be inactive and CPU hotplug in progress. */
655 BUG_ON(cpu_active(raw_smp_processor_id()));
656 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
658 /* No proper task established and can't sleep - busy wait for lock. */
659 while (!mutex_trylock(&stop_cpus_mutex))
662 /* Schedule work on other CPUs and execute directly for local CPU */
663 set_state(&msdata, MULTI_STOP_PREPARE);
664 cpu_stop_init_done(&done, num_active_cpus());
665 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
667 ret = multi_cpu_stop(&msdata);
669 /* Busy wait for completion. */
670 while (!completion_done(&done.completion))
673 mutex_unlock(&stop_cpus_mutex);
674 return ret ?: done.ret;