2 * kernel/stop_machine.c
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>
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>
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)
70 list_add_tail(&work->list, &stopper->works);
71 wake_up_process(stopper->thread);
74 /* queue @work to @stopper. if offline, @work is completed immediately */
75 static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
77 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
81 spin_lock_irqsave(&stopper->lock, flags);
82 enabled = stopper->enabled;
84 __cpu_stop_queue_work(stopper, work);
86 cpu_stop_signal_done(work->done);
87 spin_unlock_irqrestore(&stopper->lock, flags);
93 * stop_one_cpu - stop a cpu
95 * @fn: function to execute
96 * @arg: argument to @fn
98 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
99 * the highest priority preempting any task on the cpu and
100 * monopolizing it. This function returns after the execution is
103 * This function doesn't guarantee @cpu stays online till @fn
104 * completes. If @cpu goes down in the middle, execution may happen
105 * partially or fully on different cpus. @fn should either be ready
106 * for that or the caller should ensure that @cpu stays online until
107 * this function completes.
113 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
114 * otherwise, the return value of @fn.
116 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
118 struct cpu_stop_done done;
119 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
121 cpu_stop_init_done(&done, 1);
122 if (!cpu_stop_queue_work(cpu, &work))
124 wait_for_completion(&done.completion);
128 /* This controls the threads on each CPU. */
129 enum multi_stop_state {
130 /* Dummy starting state for thread. */
132 /* Awaiting everyone to be scheduled. */
134 /* Disable interrupts. */
135 MULTI_STOP_DISABLE_IRQ,
136 /* Run the function */
142 struct multi_stop_data {
145 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
146 unsigned int num_threads;
147 const struct cpumask *active_cpus;
149 enum multi_stop_state state;
153 static void set_state(struct multi_stop_data *msdata,
154 enum multi_stop_state newstate)
156 /* Reset ack counter. */
157 atomic_set(&msdata->thread_ack, msdata->num_threads);
159 msdata->state = newstate;
162 /* Last one to ack a state moves to the next state. */
163 static void ack_state(struct multi_stop_data *msdata)
165 if (atomic_dec_and_test(&msdata->thread_ack))
166 set_state(msdata, msdata->state + 1);
169 /* This is the cpu_stop function which stops the CPU. */
170 static int multi_cpu_stop(void *data)
172 struct multi_stop_data *msdata = data;
173 enum multi_stop_state curstate = MULTI_STOP_NONE;
174 int cpu = smp_processor_id(), err = 0;
179 * When called from stop_machine_from_inactive_cpu(), irq might
180 * already be disabled. Save the state and restore it on exit.
182 local_save_flags(flags);
184 if (!msdata->active_cpus)
185 is_active = cpu == cpumask_first(cpu_online_mask);
187 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
189 /* Simple state machine */
191 /* Chill out and ensure we re-read multi_stop_state. */
193 if (msdata->state != curstate) {
194 curstate = msdata->state;
196 case MULTI_STOP_DISABLE_IRQ:
202 err = msdata->fn(msdata->data);
208 } else if (curstate > MULTI_STOP_PREPARE) {
210 * At this stage all other CPUs we depend on must spin
211 * in the same loop. Any reason for hard-lockup should
212 * be detected and reported on their side.
214 touch_nmi_watchdog();
216 } while (curstate != MULTI_STOP_EXIT);
218 local_irq_restore(flags);
222 static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
223 int cpu2, struct cpu_stop_work *work2)
225 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
226 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
229 spin_lock_irq(&stopper1->lock);
230 spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
233 if (!stopper1->enabled || !stopper2->enabled)
236 * Ensure that if we race with __stop_cpus() the stoppers won't get
237 * queued up in reverse order leading to system deadlock.
239 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
240 * queued a work on cpu1 but not on cpu2, we hold both locks.
242 * It can be falsely true but it is safe to spin until it is cleared,
243 * queue_stop_cpus_work() does everything under preempt_disable().
246 if (unlikely(stop_cpus_in_progress))
250 __cpu_stop_queue_work(stopper1, work1);
251 __cpu_stop_queue_work(stopper2, work2);
253 spin_unlock(&stopper2->lock);
254 spin_unlock_irq(&stopper1->lock);
256 if (unlikely(err == -EDEADLK)) {
257 while (stop_cpus_in_progress)
264 * stop_two_cpus - stops two cpus
265 * @cpu1: the cpu to stop
266 * @cpu2: the other cpu to stop
267 * @fn: function to execute
268 * @arg: argument to @fn
270 * Stops both the current and specified CPU and runs @fn on one of them.
272 * returns when both are completed.
274 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
276 struct cpu_stop_done done;
277 struct cpu_stop_work work1, work2;
278 struct multi_stop_data msdata;
280 msdata = (struct multi_stop_data){
284 .active_cpus = cpumask_of(cpu1),
287 work1 = work2 = (struct cpu_stop_work){
288 .fn = multi_cpu_stop,
293 cpu_stop_init_done(&done, 2);
294 set_state(&msdata, MULTI_STOP_PREPARE);
298 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
301 wait_for_completion(&done.completion);
306 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
308 * @fn: function to execute
309 * @arg: argument to @fn
310 * @work_buf: pointer to cpu_stop_work structure
312 * Similar to stop_one_cpu() but doesn't wait for completion. The
313 * caller is responsible for ensuring @work_buf is currently unused
314 * and will remain untouched until stopper starts executing @fn.
320 * true if cpu_stop_work was queued successfully and @fn will be called,
323 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
324 struct cpu_stop_work *work_buf)
326 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
327 return cpu_stop_queue_work(cpu, work_buf);
330 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
331 cpu_stop_fn_t fn, void *arg,
332 struct cpu_stop_done *done)
334 struct cpu_stop_work *work;
339 * Disable preemption while queueing to avoid getting
340 * preempted by a stopper which might wait for other stoppers
341 * to enter @fn which can lead to deadlock.
344 stop_cpus_in_progress = true;
345 for_each_cpu(cpu, cpumask) {
346 work = &per_cpu(cpu_stopper.stop_work, cpu);
350 if (cpu_stop_queue_work(cpu, work))
353 stop_cpus_in_progress = false;
359 static int __stop_cpus(const struct cpumask *cpumask,
360 cpu_stop_fn_t fn, void *arg)
362 struct cpu_stop_done done;
364 cpu_stop_init_done(&done, cpumask_weight(cpumask));
365 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
367 wait_for_completion(&done.completion);
372 * stop_cpus - stop multiple cpus
373 * @cpumask: cpus to stop
374 * @fn: function to execute
375 * @arg: argument to @fn
377 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
378 * @fn is run in a process context with the highest priority
379 * preempting any task on the cpu and monopolizing it. This function
380 * returns after all executions are complete.
382 * This function doesn't guarantee the cpus in @cpumask stay online
383 * till @fn completes. If some cpus go down in the middle, execution
384 * on the cpu may happen partially or fully on different cpus. @fn
385 * should either be ready for that or the caller should ensure that
386 * the cpus stay online until this function completes.
388 * All stop_cpus() calls are serialized making it safe for @fn to wait
389 * for all cpus to start executing it.
395 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
396 * @cpumask were offline; otherwise, 0 if all executions of @fn
397 * returned 0, any non zero return value if any returned non zero.
399 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
403 /* static works are used, process one request at a time */
404 mutex_lock(&stop_cpus_mutex);
405 ret = __stop_cpus(cpumask, fn, arg);
406 mutex_unlock(&stop_cpus_mutex);
411 * try_stop_cpus - try to stop multiple cpus
412 * @cpumask: cpus to stop
413 * @fn: function to execute
414 * @arg: argument to @fn
416 * Identical to stop_cpus() except that it fails with -EAGAIN if
417 * someone else is already using the facility.
423 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
424 * @fn(@arg) was not executed at all because all cpus in @cpumask were
425 * offline; otherwise, 0 if all executions of @fn returned 0, any non
426 * zero return value if any returned non zero.
428 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
432 /* static works are used, process one request at a time */
433 if (!mutex_trylock(&stop_cpus_mutex))
435 ret = __stop_cpus(cpumask, fn, arg);
436 mutex_unlock(&stop_cpus_mutex);
440 static int cpu_stop_should_run(unsigned int cpu)
442 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
446 spin_lock_irqsave(&stopper->lock, flags);
447 run = !list_empty(&stopper->works);
448 spin_unlock_irqrestore(&stopper->lock, flags);
452 static void cpu_stopper_thread(unsigned int cpu)
454 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
455 struct cpu_stop_work *work;
459 spin_lock_irq(&stopper->lock);
460 if (!list_empty(&stopper->works)) {
461 work = list_first_entry(&stopper->works,
462 struct cpu_stop_work, list);
463 list_del_init(&work->list);
465 spin_unlock_irq(&stopper->lock);
468 cpu_stop_fn_t fn = work->fn;
469 void *arg = work->arg;
470 struct cpu_stop_done *done = work->done;
473 /* cpu stop callbacks must not sleep, make in_atomic() == T */
479 cpu_stop_signal_done(done);
482 WARN_ONCE(preempt_count(),
483 "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
488 void stop_machine_park(int cpu)
490 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
492 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
493 * the pending works before it parks, until then it is fine to queue
496 stopper->enabled = false;
497 kthread_park(stopper->thread);
500 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
502 static void cpu_stop_create(unsigned int cpu)
504 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
507 static void cpu_stop_park(unsigned int cpu)
509 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
511 WARN_ON(!list_empty(&stopper->works));
514 void stop_machine_unpark(int cpu)
516 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
518 stopper->enabled = true;
519 kthread_unpark(stopper->thread);
522 static struct smp_hotplug_thread cpu_stop_threads = {
523 .store = &cpu_stopper.thread,
524 .thread_should_run = cpu_stop_should_run,
525 .thread_fn = cpu_stopper_thread,
526 .thread_comm = "migration/%u",
527 .create = cpu_stop_create,
528 .park = cpu_stop_park,
532 static int __init cpu_stop_init(void)
536 for_each_possible_cpu(cpu) {
537 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
539 spin_lock_init(&stopper->lock);
540 INIT_LIST_HEAD(&stopper->works);
543 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
544 stop_machine_unpark(raw_smp_processor_id());
545 stop_machine_initialized = true;
548 early_initcall(cpu_stop_init);
550 static int __stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
552 struct multi_stop_data msdata = {
555 .num_threads = num_online_cpus(),
559 if (!stop_machine_initialized) {
561 * Handle the case where stop_machine() is called
562 * early in boot before stop_machine() has been
568 WARN_ON_ONCE(msdata.num_threads != 1);
570 local_irq_save(flags);
573 local_irq_restore(flags);
578 /* Set the initial state and stop all online cpus. */
579 set_state(&msdata, MULTI_STOP_PREPARE);
580 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
583 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
587 /* No CPUs can come up or down during this. */
589 ret = __stop_machine(fn, data, cpus);
593 EXPORT_SYMBOL_GPL(stop_machine);
596 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
597 * @fn: the function to run
598 * @data: the data ptr for the @fn()
599 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
601 * This is identical to stop_machine() but can be called from a CPU which
602 * is not active. The local CPU is in the process of hotplug (so no other
603 * CPU hotplug can start) and not marked active and doesn't have enough
606 * This function provides stop_machine() functionality for such state by
607 * using busy-wait for synchronization and executing @fn directly for local
611 * Local CPU is inactive. Temporarily stops all active CPUs.
614 * 0 if all executions of @fn returned 0, any non zero return value if any
617 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
618 const struct cpumask *cpus)
620 struct multi_stop_data msdata = { .fn = fn, .data = data,
621 .active_cpus = cpus };
622 struct cpu_stop_done done;
625 /* Local CPU must be inactive and CPU hotplug in progress. */
626 BUG_ON(cpu_active(raw_smp_processor_id()));
627 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
629 /* No proper task established and can't sleep - busy wait for lock. */
630 while (!mutex_trylock(&stop_cpus_mutex))
633 /* Schedule work on other CPUs and execute directly for local CPU */
634 set_state(&msdata, MULTI_STOP_PREPARE);
635 cpu_stop_init_done(&done, num_active_cpus());
636 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
638 ret = multi_cpu_stop(&msdata);
640 /* Busy wait for completion. */
641 while (!completion_done(&done.completion))
644 mutex_unlock(&stop_cpus_mutex);
645 return ret ?: done.ret;