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/lglock.h>
24 #include <linux/nmi.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;
51 * Avoids a race between stop_two_cpus and global stop_cpus, where
52 * the stoppers could get queued up in reverse order, leading to
53 * system deadlock. Using an lglock means stop_two_cpus remains
56 DEFINE_STATIC_LGLOCK(stop_cpus_lock);
58 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
60 memset(done, 0, sizeof(*done));
61 atomic_set(&done->nr_todo, nr_todo);
62 init_completion(&done->completion);
65 /* signal completion unless @done is NULL */
66 static void cpu_stop_signal_done(struct cpu_stop_done *done)
68 if (atomic_dec_and_test(&done->nr_todo))
69 complete(&done->completion);
72 static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
73 struct cpu_stop_work *work)
75 list_add_tail(&work->list, &stopper->works);
76 wake_up_process(stopper->thread);
79 /* queue @work to @stopper. if offline, @work is completed immediately */
80 static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
82 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
86 spin_lock_irqsave(&stopper->lock, flags);
87 enabled = stopper->enabled;
89 __cpu_stop_queue_work(stopper, work);
91 cpu_stop_signal_done(work->done);
92 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);
239 lg_double_lock(&stop_cpus_lock, cpu1, cpu2);
240 spin_lock_irq(&stopper1->lock);
241 spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
244 if (!stopper1->enabled || !stopper2->enabled)
248 __cpu_stop_queue_work(stopper1, work1);
249 __cpu_stop_queue_work(stopper2, work2);
251 spin_unlock(&stopper2->lock);
252 spin_unlock_irq(&stopper1->lock);
253 lg_double_unlock(&stop_cpus_lock, cpu1, cpu2);
258 * stop_two_cpus - stops two cpus
259 * @cpu1: the cpu to stop
260 * @cpu2: the other cpu to stop
261 * @fn: function to execute
262 * @arg: argument to @fn
264 * Stops both the current and specified CPU and runs @fn on one of them.
266 * returns when both are completed.
268 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
270 struct cpu_stop_done done;
271 struct cpu_stop_work work1, work2;
272 struct multi_stop_data msdata;
274 msdata = (struct multi_stop_data){
278 .active_cpus = cpumask_of(cpu1),
281 work1 = work2 = (struct cpu_stop_work){
282 .fn = multi_cpu_stop,
287 cpu_stop_init_done(&done, 2);
288 set_state(&msdata, MULTI_STOP_PREPARE);
292 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
295 wait_for_completion(&done.completion);
300 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
302 * @fn: function to execute
303 * @arg: argument to @fn
304 * @work_buf: pointer to cpu_stop_work structure
306 * Similar to stop_one_cpu() but doesn't wait for completion. The
307 * caller is responsible for ensuring @work_buf is currently unused
308 * and will remain untouched until stopper starts executing @fn.
314 * true if cpu_stop_work was queued successfully and @fn will be called,
317 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
318 struct cpu_stop_work *work_buf)
320 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
321 return cpu_stop_queue_work(cpu, work_buf);
324 /* static data for stop_cpus */
325 static DEFINE_MUTEX(stop_cpus_mutex);
327 static bool queue_stop_cpus_work(const struct cpumask *cpumask,
328 cpu_stop_fn_t fn, void *arg,
329 struct cpu_stop_done *done)
331 struct cpu_stop_work *work;
336 * Disable preemption while queueing to avoid getting
337 * preempted by a stopper which might wait for other stoppers
338 * to enter @fn which can lead to deadlock.
340 lg_global_lock(&stop_cpus_lock);
341 for_each_cpu(cpu, cpumask) {
342 work = &per_cpu(cpu_stopper.stop_work, cpu);
346 if (cpu_stop_queue_work(cpu, work))
349 lg_global_unlock(&stop_cpus_lock);
354 static int __stop_cpus(const struct cpumask *cpumask,
355 cpu_stop_fn_t fn, void *arg)
357 struct cpu_stop_done done;
359 cpu_stop_init_done(&done, cpumask_weight(cpumask));
360 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
362 wait_for_completion(&done.completion);
367 * stop_cpus - stop multiple cpus
368 * @cpumask: cpus to stop
369 * @fn: function to execute
370 * @arg: argument to @fn
372 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
373 * @fn is run in a process context with the highest priority
374 * preempting any task on the cpu and monopolizing it. This function
375 * returns after all executions are complete.
377 * This function doesn't guarantee the cpus in @cpumask stay online
378 * till @fn completes. If some cpus go down in the middle, execution
379 * on the cpu may happen partially or fully on different cpus. @fn
380 * should either be ready for that or the caller should ensure that
381 * the cpus stay online until this function completes.
383 * All stop_cpus() calls are serialized making it safe for @fn to wait
384 * for all cpus to start executing it.
390 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
391 * @cpumask were offline; otherwise, 0 if all executions of @fn
392 * returned 0, any non zero return value if any returned non zero.
394 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
398 /* static works are used, process one request at a time */
399 mutex_lock(&stop_cpus_mutex);
400 ret = __stop_cpus(cpumask, fn, arg);
401 mutex_unlock(&stop_cpus_mutex);
406 * try_stop_cpus - try to stop multiple cpus
407 * @cpumask: cpus to stop
408 * @fn: function to execute
409 * @arg: argument to @fn
411 * Identical to stop_cpus() except that it fails with -EAGAIN if
412 * someone else is already using the facility.
418 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
419 * @fn(@arg) was not executed at all because all cpus in @cpumask were
420 * offline; otherwise, 0 if all executions of @fn returned 0, any non
421 * zero return value if any returned non zero.
423 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
427 /* static works are used, process one request at a time */
428 if (!mutex_trylock(&stop_cpus_mutex))
430 ret = __stop_cpus(cpumask, fn, arg);
431 mutex_unlock(&stop_cpus_mutex);
435 static int cpu_stop_should_run(unsigned int cpu)
437 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
441 spin_lock_irqsave(&stopper->lock, flags);
442 run = !list_empty(&stopper->works);
443 spin_unlock_irqrestore(&stopper->lock, flags);
447 static void cpu_stopper_thread(unsigned int cpu)
449 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
450 struct cpu_stop_work *work;
454 spin_lock_irq(&stopper->lock);
455 if (!list_empty(&stopper->works)) {
456 work = list_first_entry(&stopper->works,
457 struct cpu_stop_work, list);
458 list_del_init(&work->list);
460 spin_unlock_irq(&stopper->lock);
463 cpu_stop_fn_t fn = work->fn;
464 void *arg = work->arg;
465 struct cpu_stop_done *done = work->done;
468 /* cpu stop callbacks must not sleep, make in_atomic() == T */
474 cpu_stop_signal_done(done);
477 WARN_ONCE(preempt_count(),
478 "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
483 void stop_machine_park(int cpu)
485 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
487 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
488 * the pending works before it parks, until then it is fine to queue
491 stopper->enabled = false;
492 kthread_park(stopper->thread);
495 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
497 static void cpu_stop_create(unsigned int cpu)
499 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
502 static void cpu_stop_park(unsigned int cpu)
504 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
506 WARN_ON(!list_empty(&stopper->works));
509 void stop_machine_unpark(int cpu)
511 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
513 stopper->enabled = true;
514 kthread_unpark(stopper->thread);
517 static struct smp_hotplug_thread cpu_stop_threads = {
518 .store = &cpu_stopper.thread,
519 .thread_should_run = cpu_stop_should_run,
520 .thread_fn = cpu_stopper_thread,
521 .thread_comm = "migration/%u",
522 .create = cpu_stop_create,
523 .park = cpu_stop_park,
527 static int __init cpu_stop_init(void)
531 for_each_possible_cpu(cpu) {
532 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
534 spin_lock_init(&stopper->lock);
535 INIT_LIST_HEAD(&stopper->works);
538 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
539 stop_machine_unpark(raw_smp_processor_id());
540 stop_machine_initialized = true;
543 early_initcall(cpu_stop_init);
545 static int __stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
547 struct multi_stop_data msdata = {
550 .num_threads = num_online_cpus(),
554 if (!stop_machine_initialized) {
556 * Handle the case where stop_machine() is called
557 * early in boot before stop_machine() has been
563 WARN_ON_ONCE(msdata.num_threads != 1);
565 local_irq_save(flags);
568 local_irq_restore(flags);
573 /* Set the initial state and stop all online cpus. */
574 set_state(&msdata, MULTI_STOP_PREPARE);
575 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
578 int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
582 /* No CPUs can come up or down during this. */
584 ret = __stop_machine(fn, data, cpus);
588 EXPORT_SYMBOL_GPL(stop_machine);
591 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
592 * @fn: the function to run
593 * @data: the data ptr for the @fn()
594 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
596 * This is identical to stop_machine() but can be called from a CPU which
597 * is not active. The local CPU is in the process of hotplug (so no other
598 * CPU hotplug can start) and not marked active and doesn't have enough
601 * This function provides stop_machine() functionality for such state by
602 * using busy-wait for synchronization and executing @fn directly for local
606 * Local CPU is inactive. Temporarily stops all active CPUs.
609 * 0 if all executions of @fn returned 0, any non zero return value if any
612 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
613 const struct cpumask *cpus)
615 struct multi_stop_data msdata = { .fn = fn, .data = data,
616 .active_cpus = cpus };
617 struct cpu_stop_done done;
620 /* Local CPU must be inactive and CPU hotplug in progress. */
621 BUG_ON(cpu_active(raw_smp_processor_id()));
622 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
624 /* No proper task established and can't sleep - busy wait for lock. */
625 while (!mutex_trylock(&stop_cpus_mutex))
628 /* Schedule work on other CPUs and execute directly for local CPU */
629 set_state(&msdata, MULTI_STOP_PREPARE);
630 cpu_stop_init_done(&done, num_active_cpus());
631 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
633 ret = multi_cpu_stop(&msdata);
635 /* Busy wait for completion. */
636 while (!completion_done(&done.completion))
639 mutex_unlock(&stop_cpus_mutex);
640 return ret ?: done.ret;