1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * kernel/stop_machine.c
5 * Copyright (C) 2008, 2005 IBM Corporation.
6 * Copyright (C) 2008, 2005 Rusty Russell rusty@rustcorp.com.au
7 * Copyright (C) 2010 SUSE Linux Products GmbH
8 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
10 #include <linux/completion.h>
11 #include <linux/cpu.h>
12 #include <linux/init.h>
13 #include <linux/kthread.h>
14 #include <linux/export.h>
15 #include <linux/percpu.h>
16 #include <linux/sched.h>
17 #include <linux/stop_machine.h>
18 #include <linux/interrupt.h>
19 #include <linux/kallsyms.h>
20 #include <linux/smpboot.h>
21 #include <linux/atomic.h>
22 #include <linux/nmi.h>
23 #include <linux/sched/wake_q.h>
26 * Structure to determine completion condition and record errors. May
27 * be shared by works on different cpus.
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 */
35 /* the actual stopper, one per every possible cpu, enabled on online cpus */
37 struct task_struct *thread;
40 bool enabled; /* is this stopper enabled? */
41 struct list_head works; /* list of pending works */
43 struct cpu_stop_work stop_work; /* for stop_cpus */
46 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
47 static bool stop_machine_initialized = false;
49 /* static data for stop_cpus */
50 static DEFINE_MUTEX(stop_cpus_mutex);
51 static bool stop_cpus_in_progress;
53 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
55 memset(done, 0, sizeof(*done));
56 atomic_set(&done->nr_todo, nr_todo);
57 init_completion(&done->completion);
60 /* signal completion unless @done is NULL */
61 static void cpu_stop_signal_done(struct cpu_stop_done *done)
63 if (atomic_dec_and_test(&done->nr_todo))
64 complete(&done->completion);
67 static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
68 struct cpu_stop_work *work,
69 struct wake_q_head *wakeq)
71 list_add_tail(&work->list, &stopper->works);
72 wake_q_add(wakeq, stopper->thread);
75 /* queue @work to @stopper. if offline, @work is completed immediately */
76 static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
78 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);
99 * stop_one_cpu - stop a cpu
101 * @fn: function to execute
102 * @arg: argument to @fn
104 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
105 * the highest priority preempting any task on the cpu and
106 * monopolizing it. This function returns after the execution is
109 * This function doesn't guarantee @cpu stays online till @fn
110 * completes. If @cpu goes down in the middle, execution may happen
111 * partially or fully on different cpus. @fn should either be ready
112 * for that or the caller should ensure that @cpu stays online until
113 * this function completes.
119 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
120 * otherwise, the return value of @fn.
122 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
124 struct cpu_stop_done done;
125 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
127 cpu_stop_init_done(&done, 1);
128 if (!cpu_stop_queue_work(cpu, &work))
131 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
132 * cycle by doing a preemption:
135 wait_for_completion(&done.completion);
139 /* This controls the threads on each CPU. */
140 enum multi_stop_state {
141 /* Dummy starting state for thread. */
143 /* Awaiting everyone to be scheduled. */
145 /* Disable interrupts. */
146 MULTI_STOP_DISABLE_IRQ,
147 /* Run the function */
153 struct multi_stop_data {
156 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
157 unsigned int num_threads;
158 const struct cpumask *active_cpus;
160 enum multi_stop_state state;
164 static void set_state(struct multi_stop_data *msdata,
165 enum multi_stop_state newstate)
167 /* Reset ack counter. */
168 atomic_set(&msdata->thread_ack, msdata->num_threads);
170 msdata->state = newstate;
173 /* Last one to ack a state moves to the next state. */
174 static void ack_state(struct multi_stop_data *msdata)
176 if (atomic_dec_and_test(&msdata->thread_ack))
177 set_state(msdata, msdata->state + 1);
180 /* This is the cpu_stop function which stops the CPU. */
181 static int multi_cpu_stop(void *data)
183 struct multi_stop_data *msdata = data;
184 enum multi_stop_state curstate = MULTI_STOP_NONE;
185 int cpu = smp_processor_id(), err = 0;
190 * When called from stop_machine_from_inactive_cpu(), irq might
191 * already be disabled. Save the state and restore it on exit.
193 local_save_flags(flags);
195 if (!msdata->active_cpus)
196 is_active = cpu == cpumask_first(cpu_online_mask);
198 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
200 /* Simple state machine */
202 /* Chill out and ensure we re-read multi_stop_state. */
204 if (msdata->state != curstate) {
205 curstate = msdata->state;
207 case MULTI_STOP_DISABLE_IRQ:
213 err = msdata->fn(msdata->data);
219 } else if (curstate > MULTI_STOP_PREPARE) {
221 * At this stage all other CPUs we depend on must spin
222 * in the same loop. Any reason for hard-lockup should
223 * be detected and reported on their side.
225 touch_nmi_watchdog();
227 } while (curstate != MULTI_STOP_EXIT);
229 local_irq_restore(flags);
233 static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
234 int cpu2, struct cpu_stop_work *work2)
236 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
237 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
238 DEFINE_WAKE_Q(wakeq);
243 * The waking up of stopper threads has to happen in the same
244 * scheduling context as the queueing. Otherwise, there is a
245 * possibility of one of the above stoppers being woken up by another
246 * CPU, and preempting us. This will cause us to not wake up the other
250 raw_spin_lock_irq(&stopper1->lock);
251 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
253 if (!stopper1->enabled || !stopper2->enabled) {
259 * Ensure that if we race with __stop_cpus() the stoppers won't get
260 * queued up in reverse order leading to system deadlock.
262 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
263 * queued a work on cpu1 but not on cpu2, we hold both locks.
265 * It can be falsely true but it is safe to spin until it is cleared,
266 * queue_stop_cpus_work() does everything under preempt_disable().
268 if (unlikely(stop_cpus_in_progress)) {
274 __cpu_stop_queue_work(stopper1, work1, &wakeq);
275 __cpu_stop_queue_work(stopper2, work2, &wakeq);
278 raw_spin_unlock(&stopper2->lock);
279 raw_spin_unlock_irq(&stopper1->lock);
281 if (unlikely(err == -EDEADLK)) {
284 while (stop_cpus_in_progress)
296 * stop_two_cpus - stops two cpus
297 * @cpu1: the cpu to stop
298 * @cpu2: the other cpu to stop
299 * @fn: function to execute
300 * @arg: argument to @fn
302 * Stops both the current and specified CPU and runs @fn on one of them.
304 * returns when both are completed.
306 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
308 struct cpu_stop_done done;
309 struct cpu_stop_work work1, work2;
310 struct multi_stop_data msdata;
312 msdata = (struct multi_stop_data){
316 .active_cpus = cpumask_of(cpu1),
319 work1 = work2 = (struct cpu_stop_work){
320 .fn = multi_cpu_stop,
325 cpu_stop_init_done(&done, 2);
326 set_state(&msdata, MULTI_STOP_PREPARE);
330 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
333 wait_for_completion(&done.completion);
338 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
340 * @fn: function to execute
341 * @arg: argument to @fn
342 * @work_buf: pointer to cpu_stop_work structure
344 * Similar to stop_one_cpu() but doesn't wait for completion. The
345 * caller is responsible for ensuring @work_buf is currently unused
346 * and will remain untouched until stopper starts executing @fn.
352 * true if cpu_stop_work was queued successfully and @fn will be called,
355 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
356 struct cpu_stop_work *work_buf)
358 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
359 return cpu_stop_queue_work(cpu, work_buf);
362 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
363 cpu_stop_fn_t fn, void *arg,
364 struct cpu_stop_done *done)
366 struct cpu_stop_work *work;
371 * Disable preemption while queueing to avoid getting
372 * preempted by a stopper which might wait for other stoppers
373 * to enter @fn which can lead to deadlock.
376 stop_cpus_in_progress = true;
377 for_each_cpu(cpu, cpumask) {
378 work = &per_cpu(cpu_stopper.stop_work, cpu);
382 if (cpu_stop_queue_work(cpu, work))
385 stop_cpus_in_progress = false;
391 static int __stop_cpus(const struct cpumask *cpumask,
392 cpu_stop_fn_t fn, void *arg)
394 struct cpu_stop_done done;
396 cpu_stop_init_done(&done, cpumask_weight(cpumask));
397 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
399 wait_for_completion(&done.completion);
404 * stop_cpus - stop multiple cpus
405 * @cpumask: cpus to stop
406 * @fn: function to execute
407 * @arg: argument to @fn
409 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
410 * @fn is run in a process context with the highest priority
411 * preempting any task on the cpu and monopolizing it. This function
412 * returns after all executions are complete.
414 * This function doesn't guarantee the cpus in @cpumask stay online
415 * till @fn completes. If some cpus go down in the middle, execution
416 * on the cpu may happen partially or fully on different cpus. @fn
417 * should either be ready for that or the caller should ensure that
418 * the cpus stay online until this function completes.
420 * All stop_cpus() calls are serialized making it safe for @fn to wait
421 * for all cpus to start executing it.
427 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
428 * @cpumask were offline; otherwise, 0 if all executions of @fn
429 * returned 0, any non zero return value if any returned non zero.
431 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
435 /* static works are used, process one request at a time */
436 mutex_lock(&stop_cpus_mutex);
437 ret = __stop_cpus(cpumask, fn, arg);
438 mutex_unlock(&stop_cpus_mutex);
443 * try_stop_cpus - try to stop multiple cpus
444 * @cpumask: cpus to stop
445 * @fn: function to execute
446 * @arg: argument to @fn
448 * Identical to stop_cpus() except that it fails with -EAGAIN if
449 * someone else is already using the facility.
455 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
456 * @fn(@arg) was not executed at all because all cpus in @cpumask were
457 * offline; otherwise, 0 if all executions of @fn returned 0, any non
458 * zero return value if any returned non zero.
460 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
464 /* static works are used, process one request at a time */
465 if (!mutex_trylock(&stop_cpus_mutex))
467 ret = __stop_cpus(cpumask, fn, arg);
468 mutex_unlock(&stop_cpus_mutex);
472 static int cpu_stop_should_run(unsigned int cpu)
474 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
478 raw_spin_lock_irqsave(&stopper->lock, flags);
479 run = !list_empty(&stopper->works);
480 raw_spin_unlock_irqrestore(&stopper->lock, flags);
484 static void cpu_stopper_thread(unsigned int cpu)
486 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
487 struct cpu_stop_work *work;
491 raw_spin_lock_irq(&stopper->lock);
492 if (!list_empty(&stopper->works)) {
493 work = list_first_entry(&stopper->works,
494 struct cpu_stop_work, list);
495 list_del_init(&work->list);
497 raw_spin_unlock_irq(&stopper->lock);
500 cpu_stop_fn_t fn = work->fn;
501 void *arg = work->arg;
502 struct cpu_stop_done *done = work->done;
505 /* cpu stop callbacks must not sleep, make in_atomic() == T */
511 cpu_stop_signal_done(done);
514 WARN_ONCE(preempt_count(),
515 "cpu_stop: %ps(%p) leaked preempt count\n", fn, arg);
520 void stop_machine_park(int cpu)
522 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
524 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
525 * the pending works before it parks, until then it is fine to queue
528 stopper->enabled = false;
529 kthread_park(stopper->thread);
532 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
534 static void cpu_stop_create(unsigned int cpu)
536 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
539 static void cpu_stop_park(unsigned int cpu)
541 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
543 WARN_ON(!list_empty(&stopper->works));
546 void stop_machine_unpark(int cpu)
548 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
550 stopper->enabled = true;
551 kthread_unpark(stopper->thread);
554 static struct smp_hotplug_thread cpu_stop_threads = {
555 .store = &cpu_stopper.thread,
556 .thread_should_run = cpu_stop_should_run,
557 .thread_fn = cpu_stopper_thread,
558 .thread_comm = "migration/%u",
559 .create = cpu_stop_create,
560 .park = cpu_stop_park,
564 static int __init cpu_stop_init(void)
568 for_each_possible_cpu(cpu) {
569 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
571 raw_spin_lock_init(&stopper->lock);
572 INIT_LIST_HEAD(&stopper->works);
575 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
576 stop_machine_unpark(raw_smp_processor_id());
577 stop_machine_initialized = true;
580 early_initcall(cpu_stop_init);
582 int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
583 const struct cpumask *cpus)
585 struct multi_stop_data msdata = {
588 .num_threads = num_online_cpus(),
592 lockdep_assert_cpus_held();
594 if (!stop_machine_initialized) {
596 * Handle the case where stop_machine() is called
597 * early in boot before stop_machine() has been
603 WARN_ON_ONCE(msdata.num_threads != 1);
605 local_irq_save(flags);
608 local_irq_restore(flags);
613 /* Set the initial state and stop all online cpus. */
614 set_state(&msdata, MULTI_STOP_PREPARE);
615 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
618 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
622 /* No CPUs can come up or down during this. */
624 ret = stop_machine_cpuslocked(fn, data, cpus);
628 EXPORT_SYMBOL_GPL(stop_machine);
631 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
632 * @fn: the function to run
633 * @data: the data ptr for the @fn()
634 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
636 * This is identical to stop_machine() but can be called from a CPU which
637 * is not active. The local CPU is in the process of hotplug (so no other
638 * CPU hotplug can start) and not marked active and doesn't have enough
641 * This function provides stop_machine() functionality for such state by
642 * using busy-wait for synchronization and executing @fn directly for local
646 * Local CPU is inactive. Temporarily stops all active CPUs.
649 * 0 if all executions of @fn returned 0, any non zero return value if any
652 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
653 const struct cpumask *cpus)
655 struct multi_stop_data msdata = { .fn = fn, .data = data,
656 .active_cpus = cpus };
657 struct cpu_stop_done done;
660 /* Local CPU must be inactive and CPU hotplug in progress. */
661 BUG_ON(cpu_active(raw_smp_processor_id()));
662 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
664 /* No proper task established and can't sleep - busy wait for lock. */
665 while (!mutex_trylock(&stop_cpus_mutex))
668 /* Schedule work on other CPUs and execute directly for local CPU */
669 set_state(&msdata, MULTI_STOP_PREPARE);
670 cpu_stop_init_done(&done, num_active_cpus());
671 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
673 ret = multi_cpu_stop(&msdata);
675 /* Busy wait for completion. */
676 while (!completion_done(&done.completion))
679 mutex_unlock(&stop_cpus_mutex);
680 return ret ?: done.ret;