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 void __weak stop_machine_yield(const struct cpumask *cpumask)
185 /* This is the cpu_stop function which stops the CPU. */
186 static int multi_cpu_stop(void *data)
188 struct multi_stop_data *msdata = data;
189 enum multi_stop_state curstate = MULTI_STOP_NONE;
190 int cpu = smp_processor_id(), err = 0;
191 const struct cpumask *cpumask;
196 * When called from stop_machine_from_inactive_cpu(), irq might
197 * already be disabled. Save the state and restore it on exit.
199 local_save_flags(flags);
201 if (!msdata->active_cpus) {
202 cpumask = cpu_online_mask;
203 is_active = cpu == cpumask_first(cpumask);
205 cpumask = msdata->active_cpus;
206 is_active = cpumask_test_cpu(cpu, cpumask);
209 /* Simple state machine */
211 /* Chill out and ensure we re-read multi_stop_state. */
212 stop_machine_yield(cpumask);
213 if (msdata->state != curstate) {
214 curstate = msdata->state;
216 case MULTI_STOP_DISABLE_IRQ:
222 err = msdata->fn(msdata->data);
228 } else if (curstate > MULTI_STOP_PREPARE) {
230 * At this stage all other CPUs we depend on must spin
231 * in the same loop. Any reason for hard-lockup should
232 * be detected and reported on their side.
234 touch_nmi_watchdog();
236 } while (curstate != MULTI_STOP_EXIT);
238 local_irq_restore(flags);
242 static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
243 int cpu2, struct cpu_stop_work *work2)
245 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
246 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
247 DEFINE_WAKE_Q(wakeq);
252 * The waking up of stopper threads has to happen in the same
253 * scheduling context as the queueing. Otherwise, there is a
254 * possibility of one of the above stoppers being woken up by another
255 * CPU, and preempting us. This will cause us to not wake up the other
259 raw_spin_lock_irq(&stopper1->lock);
260 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
262 if (!stopper1->enabled || !stopper2->enabled) {
268 * Ensure that if we race with __stop_cpus() the stoppers won't get
269 * queued up in reverse order leading to system deadlock.
271 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
272 * queued a work on cpu1 but not on cpu2, we hold both locks.
274 * It can be falsely true but it is safe to spin until it is cleared,
275 * queue_stop_cpus_work() does everything under preempt_disable().
277 if (unlikely(stop_cpus_in_progress)) {
283 __cpu_stop_queue_work(stopper1, work1, &wakeq);
284 __cpu_stop_queue_work(stopper2, work2, &wakeq);
287 raw_spin_unlock(&stopper2->lock);
288 raw_spin_unlock_irq(&stopper1->lock);
290 if (unlikely(err == -EDEADLK)) {
293 while (stop_cpus_in_progress)
305 * stop_two_cpus - stops two cpus
306 * @cpu1: the cpu to stop
307 * @cpu2: the other cpu to stop
308 * @fn: function to execute
309 * @arg: argument to @fn
311 * Stops both the current and specified CPU and runs @fn on one of them.
313 * returns when both are completed.
315 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
317 struct cpu_stop_done done;
318 struct cpu_stop_work work1, work2;
319 struct multi_stop_data msdata;
321 msdata = (struct multi_stop_data){
325 .active_cpus = cpumask_of(cpu1),
328 work1 = work2 = (struct cpu_stop_work){
329 .fn = multi_cpu_stop,
334 cpu_stop_init_done(&done, 2);
335 set_state(&msdata, MULTI_STOP_PREPARE);
339 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
342 wait_for_completion(&done.completion);
347 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
349 * @fn: function to execute
350 * @arg: argument to @fn
351 * @work_buf: pointer to cpu_stop_work structure
353 * Similar to stop_one_cpu() but doesn't wait for completion. The
354 * caller is responsible for ensuring @work_buf is currently unused
355 * and will remain untouched until stopper starts executing @fn.
361 * true if cpu_stop_work was queued successfully and @fn will be called,
364 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
365 struct cpu_stop_work *work_buf)
367 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
368 return cpu_stop_queue_work(cpu, work_buf);
371 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
372 cpu_stop_fn_t fn, void *arg,
373 struct cpu_stop_done *done)
375 struct cpu_stop_work *work;
380 * Disable preemption while queueing to avoid getting
381 * preempted by a stopper which might wait for other stoppers
382 * to enter @fn which can lead to deadlock.
385 stop_cpus_in_progress = true;
387 for_each_cpu(cpu, cpumask) {
388 work = &per_cpu(cpu_stopper.stop_work, cpu);
392 if (cpu_stop_queue_work(cpu, work))
396 stop_cpus_in_progress = false;
402 static int __stop_cpus(const struct cpumask *cpumask,
403 cpu_stop_fn_t fn, void *arg)
405 struct cpu_stop_done done;
407 cpu_stop_init_done(&done, cpumask_weight(cpumask));
408 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
410 wait_for_completion(&done.completion);
415 * stop_cpus - stop multiple cpus
416 * @cpumask: cpus to stop
417 * @fn: function to execute
418 * @arg: argument to @fn
420 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
421 * @fn is run in a process context with the highest priority
422 * preempting any task on the cpu and monopolizing it. This function
423 * returns after all executions are complete.
425 * This function doesn't guarantee the cpus in @cpumask stay online
426 * till @fn completes. If some cpus go down in the middle, execution
427 * on the cpu may happen partially or fully on different cpus. @fn
428 * should either be ready for that or the caller should ensure that
429 * the cpus stay online until this function completes.
431 * All stop_cpus() calls are serialized making it safe for @fn to wait
432 * for all cpus to start executing it.
438 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
439 * @cpumask were offline; otherwise, 0 if all executions of @fn
440 * returned 0, any non zero return value if any returned non zero.
442 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
446 /* static works are used, process one request at a time */
447 mutex_lock(&stop_cpus_mutex);
448 ret = __stop_cpus(cpumask, fn, arg);
449 mutex_unlock(&stop_cpus_mutex);
454 * try_stop_cpus - try to stop multiple cpus
455 * @cpumask: cpus to stop
456 * @fn: function to execute
457 * @arg: argument to @fn
459 * Identical to stop_cpus() except that it fails with -EAGAIN if
460 * someone else is already using the facility.
466 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
467 * @fn(@arg) was not executed at all because all cpus in @cpumask were
468 * offline; otherwise, 0 if all executions of @fn returned 0, any non
469 * zero return value if any returned non zero.
471 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
475 /* static works are used, process one request at a time */
476 if (!mutex_trylock(&stop_cpus_mutex))
478 ret = __stop_cpus(cpumask, fn, arg);
479 mutex_unlock(&stop_cpus_mutex);
483 static int cpu_stop_should_run(unsigned int cpu)
485 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
489 raw_spin_lock_irqsave(&stopper->lock, flags);
490 run = !list_empty(&stopper->works);
491 raw_spin_unlock_irqrestore(&stopper->lock, flags);
495 static void cpu_stopper_thread(unsigned int cpu)
497 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
498 struct cpu_stop_work *work;
502 raw_spin_lock_irq(&stopper->lock);
503 if (!list_empty(&stopper->works)) {
504 work = list_first_entry(&stopper->works,
505 struct cpu_stop_work, list);
506 list_del_init(&work->list);
508 raw_spin_unlock_irq(&stopper->lock);
511 cpu_stop_fn_t fn = work->fn;
512 void *arg = work->arg;
513 struct cpu_stop_done *done = work->done;
516 /* cpu stop callbacks must not sleep, make in_atomic() == T */
522 cpu_stop_signal_done(done);
525 WARN_ONCE(preempt_count(),
526 "cpu_stop: %ps(%p) leaked preempt count\n", fn, arg);
531 void stop_machine_park(int cpu)
533 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
535 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
536 * the pending works before it parks, until then it is fine to queue
539 stopper->enabled = false;
540 kthread_park(stopper->thread);
543 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
545 static void cpu_stop_create(unsigned int cpu)
547 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
550 static void cpu_stop_park(unsigned int cpu)
552 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
554 WARN_ON(!list_empty(&stopper->works));
557 void stop_machine_unpark(int cpu)
559 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
561 stopper->enabled = true;
562 kthread_unpark(stopper->thread);
565 static struct smp_hotplug_thread cpu_stop_threads = {
566 .store = &cpu_stopper.thread,
567 .thread_should_run = cpu_stop_should_run,
568 .thread_fn = cpu_stopper_thread,
569 .thread_comm = "migration/%u",
570 .create = cpu_stop_create,
571 .park = cpu_stop_park,
575 static int __init cpu_stop_init(void)
579 for_each_possible_cpu(cpu) {
580 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
582 raw_spin_lock_init(&stopper->lock);
583 INIT_LIST_HEAD(&stopper->works);
586 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
587 stop_machine_unpark(raw_smp_processor_id());
588 stop_machine_initialized = true;
591 early_initcall(cpu_stop_init);
593 int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
594 const struct cpumask *cpus)
596 struct multi_stop_data msdata = {
599 .num_threads = num_online_cpus(),
603 lockdep_assert_cpus_held();
605 if (!stop_machine_initialized) {
607 * Handle the case where stop_machine() is called
608 * early in boot before stop_machine() has been
614 WARN_ON_ONCE(msdata.num_threads != 1);
616 local_irq_save(flags);
619 local_irq_restore(flags);
624 /* Set the initial state and stop all online cpus. */
625 set_state(&msdata, MULTI_STOP_PREPARE);
626 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
629 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
633 /* No CPUs can come up or down during this. */
635 ret = stop_machine_cpuslocked(fn, data, cpus);
639 EXPORT_SYMBOL_GPL(stop_machine);
642 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
643 * @fn: the function to run
644 * @data: the data ptr for the @fn()
645 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
647 * This is identical to stop_machine() but can be called from a CPU which
648 * is not active. The local CPU is in the process of hotplug (so no other
649 * CPU hotplug can start) and not marked active and doesn't have enough
652 * This function provides stop_machine() functionality for such state by
653 * using busy-wait for synchronization and executing @fn directly for local
657 * Local CPU is inactive. Temporarily stops all active CPUs.
660 * 0 if all executions of @fn returned 0, any non zero return value if any
663 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
664 const struct cpumask *cpus)
666 struct multi_stop_data msdata = { .fn = fn, .data = data,
667 .active_cpus = cpus };
668 struct cpu_stop_done done;
671 /* Local CPU must be inactive and CPU hotplug in progress. */
672 BUG_ON(cpu_active(raw_smp_processor_id()));
673 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
675 /* No proper task established and can't sleep - busy wait for lock. */
676 while (!mutex_trylock(&stop_cpus_mutex))
679 /* Schedule work on other CPUs and execute directly for local CPU */
680 set_state(&msdata, MULTI_STOP_PREPARE);
681 cpu_stop_init_done(&done, num_active_cpus());
682 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
684 ret = multi_cpu_stop(&msdata);
686 /* Busy wait for completion. */
687 while (!completion_done(&done.completion))
690 mutex_unlock(&stop_cpus_mutex);
691 return ret ?: done.ret;