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>
22 #include <linux/atomic.h>
25 * Structure to determine completion condition and record errors. May
26 * be shared by works on different cpus.
28 struct cpu_stop_done {
29 atomic_t nr_todo; /* nr left to execute */
30 bool executed; /* actually executed? */
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 */
38 bool enabled; /* is this stopper enabled? */
39 struct list_head works; /* list of pending works */
40 struct task_struct *thread; /* stopper thread */
43 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
44 static bool stop_machine_initialized = false;
46 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
48 memset(done, 0, sizeof(*done));
49 atomic_set(&done->nr_todo, nr_todo);
50 init_completion(&done->completion);
53 /* signal completion unless @done is NULL */
54 static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
58 done->executed = true;
59 if (atomic_dec_and_test(&done->nr_todo))
60 complete(&done->completion);
64 /* queue @work to @stopper. if offline, @work is completed immediately */
65 static void cpu_stop_queue_work(struct cpu_stopper *stopper,
66 struct cpu_stop_work *work)
70 spin_lock_irqsave(&stopper->lock, flags);
72 if (stopper->enabled) {
73 list_add_tail(&work->list, &stopper->works);
74 wake_up_process(stopper->thread);
76 cpu_stop_signal_done(work->done, false);
78 spin_unlock_irqrestore(&stopper->lock, flags);
82 * stop_one_cpu - stop a cpu
84 * @fn: function to execute
85 * @arg: argument to @fn
87 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
88 * the highest priority preempting any task on the cpu and
89 * monopolizing it. This function returns after the execution is
92 * This function doesn't guarantee @cpu stays online till @fn
93 * completes. If @cpu goes down in the middle, execution may happen
94 * partially or fully on different cpus. @fn should either be ready
95 * for that or the caller should ensure that @cpu stays online until
96 * this function completes.
102 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
103 * otherwise, the return value of @fn.
105 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
107 struct cpu_stop_done done;
108 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
110 cpu_stop_init_done(&done, 1);
111 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), &work);
112 wait_for_completion(&done.completion);
113 return done.executed ? done.ret : -ENOENT;
117 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
119 * @fn: function to execute
120 * @arg: argument to @fn
122 * Similar to stop_one_cpu() but doesn't wait for completion. The
123 * caller is responsible for ensuring @work_buf is currently unused
124 * and will remain untouched until stopper starts executing @fn.
129 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
130 struct cpu_stop_work *work_buf)
132 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
133 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
136 /* static data for stop_cpus */
137 static DEFINE_MUTEX(stop_cpus_mutex);
138 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
140 static void queue_stop_cpus_work(const struct cpumask *cpumask,
141 cpu_stop_fn_t fn, void *arg,
142 struct cpu_stop_done *done)
144 struct cpu_stop_work *work;
147 /* initialize works and done */
148 for_each_cpu(cpu, cpumask) {
149 work = &per_cpu(stop_cpus_work, cpu);
156 * Disable preemption while queueing to avoid getting
157 * preempted by a stopper which might wait for other stoppers
158 * to enter @fn which can lead to deadlock.
161 for_each_cpu(cpu, cpumask)
162 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
163 &per_cpu(stop_cpus_work, cpu));
167 static int __stop_cpus(const struct cpumask *cpumask,
168 cpu_stop_fn_t fn, void *arg)
170 struct cpu_stop_done done;
172 cpu_stop_init_done(&done, cpumask_weight(cpumask));
173 queue_stop_cpus_work(cpumask, fn, arg, &done);
174 wait_for_completion(&done.completion);
175 return done.executed ? done.ret : -ENOENT;
179 * stop_cpus - stop multiple cpus
180 * @cpumask: cpus to stop
181 * @fn: function to execute
182 * @arg: argument to @fn
184 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
185 * @fn is run in a process context with the highest priority
186 * preempting any task on the cpu and monopolizing it. This function
187 * returns after all executions are complete.
189 * This function doesn't guarantee the cpus in @cpumask stay online
190 * till @fn completes. If some cpus go down in the middle, execution
191 * on the cpu may happen partially or fully on different cpus. @fn
192 * should either be ready for that or the caller should ensure that
193 * the cpus stay online until this function completes.
195 * All stop_cpus() calls are serialized making it safe for @fn to wait
196 * for all cpus to start executing it.
202 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
203 * @cpumask were offline; otherwise, 0 if all executions of @fn
204 * returned 0, any non zero return value if any returned non zero.
206 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
210 /* static works are used, process one request at a time */
211 mutex_lock(&stop_cpus_mutex);
212 ret = __stop_cpus(cpumask, fn, arg);
213 mutex_unlock(&stop_cpus_mutex);
218 * try_stop_cpus - try to stop multiple cpus
219 * @cpumask: cpus to stop
220 * @fn: function to execute
221 * @arg: argument to @fn
223 * Identical to stop_cpus() except that it fails with -EAGAIN if
224 * someone else is already using the facility.
230 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
231 * @fn(@arg) was not executed at all because all cpus in @cpumask were
232 * offline; otherwise, 0 if all executions of @fn returned 0, any non
233 * zero return value if any returned non zero.
235 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
239 /* static works are used, process one request at a time */
240 if (!mutex_trylock(&stop_cpus_mutex))
242 ret = __stop_cpus(cpumask, fn, arg);
243 mutex_unlock(&stop_cpus_mutex);
247 static int cpu_stopper_thread(void *data)
249 struct cpu_stopper *stopper = data;
250 struct cpu_stop_work *work;
254 set_current_state(TASK_INTERRUPTIBLE); /* mb paired w/ kthread_stop */
256 if (kthread_should_stop()) {
257 __set_current_state(TASK_RUNNING);
262 spin_lock_irq(&stopper->lock);
263 if (!list_empty(&stopper->works)) {
264 work = list_first_entry(&stopper->works,
265 struct cpu_stop_work, list);
266 list_del_init(&work->list);
268 spin_unlock_irq(&stopper->lock);
271 cpu_stop_fn_t fn = work->fn;
272 void *arg = work->arg;
273 struct cpu_stop_done *done = work->done;
274 char ksym_buf[KSYM_NAME_LEN] __maybe_unused;
276 __set_current_state(TASK_RUNNING);
278 /* cpu stop callbacks are not allowed to sleep */
285 /* restore preemption and check it's still balanced */
287 WARN_ONCE(preempt_count(),
288 "cpu_stop: %s(%p) leaked preempt count\n",
289 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
292 cpu_stop_signal_done(done, true);
299 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
301 /* manage stopper for a cpu, mostly lifted from sched migration thread mgmt */
302 static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
303 unsigned long action, void *hcpu)
305 unsigned int cpu = (unsigned long)hcpu;
306 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
307 struct task_struct *p;
309 switch (action & ~CPU_TASKS_FROZEN) {
311 BUG_ON(stopper->thread || stopper->enabled ||
312 !list_empty(&stopper->works));
313 p = kthread_create_on_node(cpu_stopper_thread,
316 "migration/%d", cpu);
318 return notifier_from_errno(PTR_ERR(p));
320 kthread_bind(p, cpu);
321 sched_set_stop_task(cpu, p);
326 /* strictly unnecessary, as first user will wake it */
327 wake_up_process(stopper->thread);
329 spin_lock_irq(&stopper->lock);
330 stopper->enabled = true;
331 spin_unlock_irq(&stopper->lock);
334 #ifdef CONFIG_HOTPLUG_CPU
335 case CPU_UP_CANCELED:
338 struct cpu_stop_work *work;
340 sched_set_stop_task(cpu, NULL);
341 /* kill the stopper */
342 kthread_stop(stopper->thread);
343 /* drain remaining works */
344 spin_lock_irq(&stopper->lock);
345 list_for_each_entry(work, &stopper->works, list)
346 cpu_stop_signal_done(work->done, false);
347 stopper->enabled = false;
348 spin_unlock_irq(&stopper->lock);
349 /* release the stopper */
350 put_task_struct(stopper->thread);
351 stopper->thread = NULL;
361 * Give it a higher priority so that cpu stopper is available to other
362 * cpu notifiers. It currently shares the same priority as sched
363 * migration_notifier.
365 static struct notifier_block __cpuinitdata cpu_stop_cpu_notifier = {
366 .notifier_call = cpu_stop_cpu_callback,
370 static int __init cpu_stop_init(void)
372 void *bcpu = (void *)(long)smp_processor_id();
376 for_each_possible_cpu(cpu) {
377 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
379 spin_lock_init(&stopper->lock);
380 INIT_LIST_HEAD(&stopper->works);
383 /* start one for the boot cpu */
384 err = cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_UP_PREPARE,
386 BUG_ON(err != NOTIFY_OK);
387 cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_ONLINE, bcpu);
388 register_cpu_notifier(&cpu_stop_cpu_notifier);
390 stop_machine_initialized = true;
394 early_initcall(cpu_stop_init);
396 #ifdef CONFIG_STOP_MACHINE
398 /* This controls the threads on each CPU. */
399 enum stopmachine_state {
400 /* Dummy starting state for thread. */
402 /* Awaiting everyone to be scheduled. */
404 /* Disable interrupts. */
405 STOPMACHINE_DISABLE_IRQ,
406 /* Run the function */
412 struct stop_machine_data {
415 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
416 unsigned int num_threads;
417 const struct cpumask *active_cpus;
419 enum stopmachine_state state;
423 static void set_state(struct stop_machine_data *smdata,
424 enum stopmachine_state newstate)
426 /* Reset ack counter. */
427 atomic_set(&smdata->thread_ack, smdata->num_threads);
429 smdata->state = newstate;
432 /* Last one to ack a state moves to the next state. */
433 static void ack_state(struct stop_machine_data *smdata)
435 if (atomic_dec_and_test(&smdata->thread_ack))
436 set_state(smdata, smdata->state + 1);
439 /* This is the cpu_stop function which stops the CPU. */
440 static int stop_machine_cpu_stop(void *data)
442 struct stop_machine_data *smdata = data;
443 enum stopmachine_state curstate = STOPMACHINE_NONE;
444 int cpu = smp_processor_id(), err = 0;
449 * When called from stop_machine_from_inactive_cpu(), irq might
450 * already be disabled. Save the state and restore it on exit.
452 local_save_flags(flags);
454 if (!smdata->active_cpus)
455 is_active = cpu == cpumask_first(cpu_online_mask);
457 is_active = cpumask_test_cpu(cpu, smdata->active_cpus);
459 /* Simple state machine */
461 /* Chill out and ensure we re-read stopmachine_state. */
463 if (smdata->state != curstate) {
464 curstate = smdata->state;
466 case STOPMACHINE_DISABLE_IRQ:
470 case STOPMACHINE_RUN:
472 err = smdata->fn(smdata->data);
479 } while (curstate != STOPMACHINE_EXIT);
481 local_irq_restore(flags);
485 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
487 struct stop_machine_data smdata = { .fn = fn, .data = data,
488 .num_threads = num_online_cpus(),
489 .active_cpus = cpus };
491 if (!stop_machine_initialized) {
493 * Handle the case where stop_machine() is called
494 * early in boot before stop_machine() has been
500 WARN_ON_ONCE(smdata.num_threads != 1);
502 local_irq_save(flags);
505 local_irq_restore(flags);
510 /* Set the initial state and stop all online cpus. */
511 set_state(&smdata, STOPMACHINE_PREPARE);
512 return stop_cpus(cpu_online_mask, stop_machine_cpu_stop, &smdata);
515 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
519 /* No CPUs can come up or down during this. */
521 ret = __stop_machine(fn, data, cpus);
525 EXPORT_SYMBOL_GPL(stop_machine);
528 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
529 * @fn: the function to run
530 * @data: the data ptr for the @fn()
531 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
533 * This is identical to stop_machine() but can be called from a CPU which
534 * is not active. The local CPU is in the process of hotplug (so no other
535 * CPU hotplug can start) and not marked active and doesn't have enough
538 * This function provides stop_machine() functionality for such state by
539 * using busy-wait for synchronization and executing @fn directly for local
543 * Local CPU is inactive. Temporarily stops all active CPUs.
546 * 0 if all executions of @fn returned 0, any non zero return value if any
549 int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
550 const struct cpumask *cpus)
552 struct stop_machine_data smdata = { .fn = fn, .data = data,
553 .active_cpus = cpus };
554 struct cpu_stop_done done;
557 /* Local CPU must be inactive and CPU hotplug in progress. */
558 BUG_ON(cpu_active(raw_smp_processor_id()));
559 smdata.num_threads = num_active_cpus() + 1; /* +1 for local */
561 /* No proper task established and can't sleep - busy wait for lock. */
562 while (!mutex_trylock(&stop_cpus_mutex))
565 /* Schedule work on other CPUs and execute directly for local CPU */
566 set_state(&smdata, STOPMACHINE_PREPARE);
567 cpu_stop_init_done(&done, num_active_cpus());
568 queue_stop_cpus_work(cpu_active_mask, stop_machine_cpu_stop, &smdata,
570 ret = stop_machine_cpu_stop(&smdata);
572 /* Busy wait for completion. */
573 while (!completion_done(&done.completion))
576 mutex_unlock(&stop_cpus_mutex);
577 return ret ?: done.ret;
580 #endif /* CONFIG_STOP_MACHINE */