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/compiler.h>
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);
85 raw_spin_lock_irqsave(&stopper->lock, flags);
86 enabled = stopper->enabled;
88 __cpu_stop_queue_work(stopper, work, &wakeq);
90 cpu_stop_signal_done(work->done);
91 raw_spin_unlock_irqrestore(&stopper->lock, flags);
100 * stop_one_cpu - stop a cpu
102 * @fn: function to execute
103 * @arg: argument to @fn
105 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
106 * the highest priority preempting any task on the cpu and
107 * monopolizing it. This function returns after the execution is
110 * This function doesn't guarantee @cpu stays online till @fn
111 * completes. If @cpu goes down in the middle, execution may happen
112 * partially or fully on different cpus. @fn should either be ready
113 * for that or the caller should ensure that @cpu stays online until
114 * this function completes.
120 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
121 * otherwise, the return value of @fn.
123 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
125 struct cpu_stop_done done;
126 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
128 cpu_stop_init_done(&done, 1);
129 if (!cpu_stop_queue_work(cpu, &work))
132 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
133 * cycle by doing a preemption:
136 wait_for_completion(&done.completion);
140 /* This controls the threads on each CPU. */
141 enum multi_stop_state {
142 /* Dummy starting state for thread. */
144 /* Awaiting everyone to be scheduled. */
146 /* Disable interrupts. */
147 MULTI_STOP_DISABLE_IRQ,
148 /* Run the function */
154 struct multi_stop_data {
157 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
158 unsigned int num_threads;
159 const struct cpumask *active_cpus;
161 enum multi_stop_state state;
165 static void set_state(struct multi_stop_data *msdata,
166 enum multi_stop_state newstate)
168 /* Reset ack counter. */
169 atomic_set(&msdata->thread_ack, msdata->num_threads);
171 WRITE_ONCE(msdata->state, newstate);
174 /* Last one to ack a state moves to the next state. */
175 static void ack_state(struct multi_stop_data *msdata)
177 if (atomic_dec_and_test(&msdata->thread_ack))
178 set_state(msdata, msdata->state + 1);
181 notrace void __weak stop_machine_yield(const struct cpumask *cpumask)
186 /* This is the cpu_stop function which stops the CPU. */
187 static int multi_cpu_stop(void *data)
189 struct multi_stop_data *msdata = data;
190 enum multi_stop_state newstate, curstate = MULTI_STOP_NONE;
191 int cpu = smp_processor_id(), err = 0;
192 const struct cpumask *cpumask;
197 * When called from stop_machine_from_inactive_cpu(), irq might
198 * already be disabled. Save the state and restore it on exit.
200 local_save_flags(flags);
202 if (!msdata->active_cpus) {
203 cpumask = cpu_online_mask;
204 is_active = cpu == cpumask_first(cpumask);
206 cpumask = msdata->active_cpus;
207 is_active = cpumask_test_cpu(cpu, cpumask);
210 /* Simple state machine */
212 /* Chill out and ensure we re-read multi_stop_state. */
213 stop_machine_yield(cpumask);
214 newstate = READ_ONCE(msdata->state);
215 if (newstate != curstate) {
218 case MULTI_STOP_DISABLE_IRQ:
224 err = msdata->fn(msdata->data);
230 } else if (curstate > MULTI_STOP_PREPARE) {
232 * At this stage all other CPUs we depend on must spin
233 * in the same loop. Any reason for hard-lockup should
234 * be detected and reported on their side.
236 touch_nmi_watchdog();
238 rcu_momentary_dyntick_idle();
239 } while (curstate != MULTI_STOP_EXIT);
241 local_irq_restore(flags);
245 static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
246 int cpu2, struct cpu_stop_work *work2)
248 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
249 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
250 DEFINE_WAKE_Q(wakeq);
255 * The waking up of stopper threads has to happen in the same
256 * scheduling context as the queueing. Otherwise, there is a
257 * possibility of one of the above stoppers being woken up by another
258 * CPU, and preempting us. This will cause us to not wake up the other
262 raw_spin_lock_irq(&stopper1->lock);
263 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
265 if (!stopper1->enabled || !stopper2->enabled) {
271 * Ensure that if we race with __stop_cpus() the stoppers won't get
272 * queued up in reverse order leading to system deadlock.
274 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
275 * queued a work on cpu1 but not on cpu2, we hold both locks.
277 * It can be falsely true but it is safe to spin until it is cleared,
278 * queue_stop_cpus_work() does everything under preempt_disable().
280 if (unlikely(stop_cpus_in_progress)) {
286 __cpu_stop_queue_work(stopper1, work1, &wakeq);
287 __cpu_stop_queue_work(stopper2, work2, &wakeq);
290 raw_spin_unlock(&stopper2->lock);
291 raw_spin_unlock_irq(&stopper1->lock);
293 if (unlikely(err == -EDEADLK)) {
296 while (stop_cpus_in_progress)
308 * stop_two_cpus - stops two cpus
309 * @cpu1: the cpu to stop
310 * @cpu2: the other cpu to stop
311 * @fn: function to execute
312 * @arg: argument to @fn
314 * Stops both the current and specified CPU and runs @fn on one of them.
316 * returns when both are completed.
318 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
320 struct cpu_stop_done done;
321 struct cpu_stop_work work1, work2;
322 struct multi_stop_data msdata;
324 msdata = (struct multi_stop_data){
328 .active_cpus = cpumask_of(cpu1),
331 work1 = work2 = (struct cpu_stop_work){
332 .fn = multi_cpu_stop,
337 cpu_stop_init_done(&done, 2);
338 set_state(&msdata, MULTI_STOP_PREPARE);
342 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
345 wait_for_completion(&done.completion);
350 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
352 * @fn: function to execute
353 * @arg: argument to @fn
354 * @work_buf: pointer to cpu_stop_work structure
356 * Similar to stop_one_cpu() but doesn't wait for completion. The
357 * caller is responsible for ensuring @work_buf is currently unused
358 * and will remain untouched until stopper starts executing @fn.
364 * true if cpu_stop_work was queued successfully and @fn will be called,
367 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
368 struct cpu_stop_work *work_buf)
370 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
371 return cpu_stop_queue_work(cpu, work_buf);
374 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
375 cpu_stop_fn_t fn, void *arg,
376 struct cpu_stop_done *done)
378 struct cpu_stop_work *work;
383 * Disable preemption while queueing to avoid getting
384 * preempted by a stopper which might wait for other stoppers
385 * to enter @fn which can lead to deadlock.
388 stop_cpus_in_progress = true;
390 for_each_cpu(cpu, cpumask) {
391 work = &per_cpu(cpu_stopper.stop_work, cpu);
395 if (cpu_stop_queue_work(cpu, work))
399 stop_cpus_in_progress = false;
405 static int __stop_cpus(const struct cpumask *cpumask,
406 cpu_stop_fn_t fn, void *arg)
408 struct cpu_stop_done done;
410 cpu_stop_init_done(&done, cpumask_weight(cpumask));
411 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
413 wait_for_completion(&done.completion);
418 * stop_cpus - stop multiple cpus
419 * @cpumask: cpus to stop
420 * @fn: function to execute
421 * @arg: argument to @fn
423 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
424 * @fn is run in a process context with the highest priority
425 * preempting any task on the cpu and monopolizing it. This function
426 * returns after all executions are complete.
428 * This function doesn't guarantee the cpus in @cpumask stay online
429 * till @fn completes. If some cpus go down in the middle, execution
430 * on the cpu may happen partially or fully on different cpus. @fn
431 * should either be ready for that or the caller should ensure that
432 * the cpus stay online until this function completes.
434 * All stop_cpus() calls are serialized making it safe for @fn to wait
435 * for all cpus to start executing it.
441 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
442 * @cpumask were offline; otherwise, 0 if all executions of @fn
443 * returned 0, any non zero return value if any returned non zero.
445 static int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
449 /* static works are used, process one request at a time */
450 mutex_lock(&stop_cpus_mutex);
451 ret = __stop_cpus(cpumask, fn, arg);
452 mutex_unlock(&stop_cpus_mutex);
456 static int cpu_stop_should_run(unsigned int cpu)
458 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
462 raw_spin_lock_irqsave(&stopper->lock, flags);
463 run = !list_empty(&stopper->works);
464 raw_spin_unlock_irqrestore(&stopper->lock, flags);
468 static void cpu_stopper_thread(unsigned int cpu)
470 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
471 struct cpu_stop_work *work;
475 raw_spin_lock_irq(&stopper->lock);
476 if (!list_empty(&stopper->works)) {
477 work = list_first_entry(&stopper->works,
478 struct cpu_stop_work, list);
479 list_del_init(&work->list);
481 raw_spin_unlock_irq(&stopper->lock);
484 cpu_stop_fn_t fn = work->fn;
485 void *arg = work->arg;
486 struct cpu_stop_done *done = work->done;
489 /* cpu stop callbacks must not sleep, make in_atomic() == T */
495 cpu_stop_signal_done(done);
498 WARN_ONCE(preempt_count(),
499 "cpu_stop: %ps(%p) leaked preempt count\n", fn, arg);
504 void stop_machine_park(int cpu)
506 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
508 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
509 * the pending works before it parks, until then it is fine to queue
512 stopper->enabled = false;
513 kthread_park(stopper->thread);
516 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
518 static void cpu_stop_create(unsigned int cpu)
520 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
523 static void cpu_stop_park(unsigned int cpu)
525 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
527 WARN_ON(!list_empty(&stopper->works));
530 void stop_machine_unpark(int cpu)
532 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
534 stopper->enabled = true;
535 kthread_unpark(stopper->thread);
538 static struct smp_hotplug_thread cpu_stop_threads = {
539 .store = &cpu_stopper.thread,
540 .thread_should_run = cpu_stop_should_run,
541 .thread_fn = cpu_stopper_thread,
542 .thread_comm = "migration/%u",
543 .create = cpu_stop_create,
544 .park = cpu_stop_park,
548 static int __init cpu_stop_init(void)
552 for_each_possible_cpu(cpu) {
553 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
555 raw_spin_lock_init(&stopper->lock);
556 INIT_LIST_HEAD(&stopper->works);
559 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
560 stop_machine_unpark(raw_smp_processor_id());
561 stop_machine_initialized = true;
564 early_initcall(cpu_stop_init);
566 int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
567 const struct cpumask *cpus)
569 struct multi_stop_data msdata = {
572 .num_threads = num_online_cpus(),
576 lockdep_assert_cpus_held();
578 if (!stop_machine_initialized) {
580 * Handle the case where stop_machine() is called
581 * early in boot before stop_machine() has been
587 WARN_ON_ONCE(msdata.num_threads != 1);
589 local_irq_save(flags);
592 local_irq_restore(flags);
597 /* Set the initial state and stop all online cpus. */
598 set_state(&msdata, MULTI_STOP_PREPARE);
599 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
602 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
606 /* No CPUs can come up or down during this. */
608 ret = stop_machine_cpuslocked(fn, data, cpus);
612 EXPORT_SYMBOL_GPL(stop_machine);
615 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
616 * @fn: the function to run
617 * @data: the data ptr for the @fn()
618 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
620 * This is identical to stop_machine() but can be called from a CPU which
621 * is not active. The local CPU is in the process of hotplug (so no other
622 * CPU hotplug can start) and not marked active and doesn't have enough
625 * This function provides stop_machine() functionality for such state by
626 * using busy-wait for synchronization and executing @fn directly for local
630 * Local CPU is inactive. Temporarily stops all active CPUs.
633 * 0 if all executions of @fn returned 0, any non zero return value if any
636 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
637 const struct cpumask *cpus)
639 struct multi_stop_data msdata = { .fn = fn, .data = data,
640 .active_cpus = cpus };
641 struct cpu_stop_done done;
644 /* Local CPU must be inactive and CPU hotplug in progress. */
645 BUG_ON(cpu_active(raw_smp_processor_id()));
646 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
648 /* No proper task established and can't sleep - busy wait for lock. */
649 while (!mutex_trylock(&stop_cpus_mutex))
652 /* Schedule work on other CPUs and execute directly for local CPU */
653 set_state(&msdata, MULTI_STOP_PREPARE);
654 cpu_stop_init_done(&done, num_active_cpus());
655 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
657 ret = multi_cpu_stop(&msdata);
659 /* Busy wait for completion. */
660 while (!completion_done(&done.completion))
663 mutex_unlock(&stop_cpus_mutex);
664 return ret ?: done.ret;