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/module.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 <asm/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);
45 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
47 memset(done, 0, sizeof(*done));
48 atomic_set(&done->nr_todo, nr_todo);
49 init_completion(&done->completion);
52 /* signal completion unless @done is NULL */
53 static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
57 done->executed = true;
58 if (atomic_dec_and_test(&done->nr_todo))
59 complete(&done->completion);
63 /* queue @work to @stopper. if offline, @work is completed immediately */
64 static void cpu_stop_queue_work(struct cpu_stopper *stopper,
65 struct cpu_stop_work *work)
69 spin_lock_irqsave(&stopper->lock, flags);
71 if (stopper->enabled) {
72 list_add_tail(&work->list, &stopper->works);
73 wake_up_process(stopper->thread);
75 cpu_stop_signal_done(work->done, false);
77 spin_unlock_irqrestore(&stopper->lock, flags);
81 * stop_one_cpu - stop a cpu
83 * @fn: function to execute
84 * @arg: argument to @fn
86 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
87 * the highest priority preempting any task on the cpu and
88 * monopolizing it. This function returns after the execution is
91 * This function doesn't guarantee @cpu stays online till @fn
92 * completes. If @cpu goes down in the middle, execution may happen
93 * partially or fully on different cpus. @fn should either be ready
94 * for that or the caller should ensure that @cpu stays online until
95 * this function completes.
101 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
102 * otherwise, the return value of @fn.
104 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
106 struct cpu_stop_done done;
107 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
109 cpu_stop_init_done(&done, 1);
110 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), &work);
111 wait_for_completion(&done.completion);
112 return done.executed ? done.ret : -ENOENT;
116 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
118 * @fn: function to execute
119 * @arg: argument to @fn
121 * Similar to stop_one_cpu() but doesn't wait for completion. The
122 * caller is responsible for ensuring @work_buf is currently unused
123 * and will remain untouched until stopper starts executing @fn.
128 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
129 struct cpu_stop_work *work_buf)
131 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
132 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
135 /* static data for stop_cpus */
136 static DEFINE_MUTEX(stop_cpus_mutex);
137 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
139 int __stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
141 struct cpu_stop_work *work;
142 struct cpu_stop_done done;
145 /* initialize works and done */
146 for_each_cpu(cpu, cpumask) {
147 work = &per_cpu(stop_cpus_work, cpu);
152 cpu_stop_init_done(&done, cpumask_weight(cpumask));
155 * Disable preemption while queueing to avoid getting
156 * preempted by a stopper which might wait for other stoppers
157 * to enter @fn which can lead to deadlock.
160 for_each_cpu(cpu, cpumask)
161 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
162 &per_cpu(stop_cpus_work, cpu));
165 wait_for_completion(&done.completion);
166 return done.executed ? done.ret : -ENOENT;
170 * stop_cpus - stop multiple cpus
171 * @cpumask: cpus to stop
172 * @fn: function to execute
173 * @arg: argument to @fn
175 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
176 * @fn is run in a process context with the highest priority
177 * preempting any task on the cpu and monopolizing it. This function
178 * returns after all executions are complete.
180 * This function doesn't guarantee the cpus in @cpumask stay online
181 * till @fn completes. If some cpus go down in the middle, execution
182 * on the cpu may happen partially or fully on different cpus. @fn
183 * should either be ready for that or the caller should ensure that
184 * the cpus stay online until this function completes.
186 * All stop_cpus() calls are serialized making it safe for @fn to wait
187 * for all cpus to start executing it.
193 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
194 * @cpumask were offline; otherwise, 0 if all executions of @fn
195 * returned 0, any non zero return value if any returned non zero.
197 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
201 /* static works are used, process one request at a time */
202 mutex_lock(&stop_cpus_mutex);
203 ret = __stop_cpus(cpumask, fn, arg);
204 mutex_unlock(&stop_cpus_mutex);
209 * try_stop_cpus - try to stop multiple cpus
210 * @cpumask: cpus to stop
211 * @fn: function to execute
212 * @arg: argument to @fn
214 * Identical to stop_cpus() except that it fails with -EAGAIN if
215 * someone else is already using the facility.
221 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
222 * @fn(@arg) was not executed at all because all cpus in @cpumask were
223 * offline; otherwise, 0 if all executions of @fn returned 0, any non
224 * zero return value if any returned non zero.
226 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
230 /* static works are used, process one request at a time */
231 if (!mutex_trylock(&stop_cpus_mutex))
233 ret = __stop_cpus(cpumask, fn, arg);
234 mutex_unlock(&stop_cpus_mutex);
238 static int cpu_stopper_thread(void *data)
240 struct cpu_stopper *stopper = data;
241 struct cpu_stop_work *work;
245 set_current_state(TASK_INTERRUPTIBLE); /* mb paired w/ kthread_stop */
247 if (kthread_should_stop()) {
248 __set_current_state(TASK_RUNNING);
253 spin_lock_irq(&stopper->lock);
254 if (!list_empty(&stopper->works)) {
255 work = list_first_entry(&stopper->works,
256 struct cpu_stop_work, list);
257 list_del_init(&work->list);
259 spin_unlock_irq(&stopper->lock);
262 cpu_stop_fn_t fn = work->fn;
263 void *arg = work->arg;
264 struct cpu_stop_done *done = work->done;
265 char ksym_buf[KSYM_NAME_LEN] __maybe_unused;
267 __set_current_state(TASK_RUNNING);
269 /* cpu stop callbacks are not allowed to sleep */
276 /* restore preemption and check it's still balanced */
278 WARN_ONCE(preempt_count(),
279 "cpu_stop: %s(%p) leaked preempt count\n",
280 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
283 cpu_stop_signal_done(done, true);
290 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
292 /* manage stopper for a cpu, mostly lifted from sched migration thread mgmt */
293 static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
294 unsigned long action, void *hcpu)
296 unsigned int cpu = (unsigned long)hcpu;
297 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
298 struct task_struct *p;
300 switch (action & ~CPU_TASKS_FROZEN) {
302 BUG_ON(stopper->thread || stopper->enabled ||
303 !list_empty(&stopper->works));
304 p = kthread_create_on_node(cpu_stopper_thread,
307 "migration/%d", cpu);
309 return notifier_from_errno(PTR_ERR(p));
311 kthread_bind(p, cpu);
312 sched_set_stop_task(cpu, p);
317 /* strictly unnecessary, as first user will wake it */
318 wake_up_process(stopper->thread);
320 spin_lock_irq(&stopper->lock);
321 stopper->enabled = true;
322 spin_unlock_irq(&stopper->lock);
325 #ifdef CONFIG_HOTPLUG_CPU
326 case CPU_UP_CANCELED:
329 struct cpu_stop_work *work;
331 sched_set_stop_task(cpu, NULL);
332 /* kill the stopper */
333 kthread_stop(stopper->thread);
334 /* drain remaining works */
335 spin_lock_irq(&stopper->lock);
336 list_for_each_entry(work, &stopper->works, list)
337 cpu_stop_signal_done(work->done, false);
338 stopper->enabled = false;
339 spin_unlock_irq(&stopper->lock);
340 /* release the stopper */
341 put_task_struct(stopper->thread);
342 stopper->thread = NULL;
352 * Give it a higher priority so that cpu stopper is available to other
353 * cpu notifiers. It currently shares the same priority as sched
354 * migration_notifier.
356 static struct notifier_block __cpuinitdata cpu_stop_cpu_notifier = {
357 .notifier_call = cpu_stop_cpu_callback,
361 static int __init cpu_stop_init(void)
363 void *bcpu = (void *)(long)smp_processor_id();
367 for_each_possible_cpu(cpu) {
368 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
370 spin_lock_init(&stopper->lock);
371 INIT_LIST_HEAD(&stopper->works);
374 /* start one for the boot cpu */
375 err = cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_UP_PREPARE,
377 BUG_ON(err != NOTIFY_OK);
378 cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_ONLINE, bcpu);
379 register_cpu_notifier(&cpu_stop_cpu_notifier);
383 early_initcall(cpu_stop_init);
385 #ifdef CONFIG_STOP_MACHINE
387 /* This controls the threads on each CPU. */
388 enum stopmachine_state {
389 /* Dummy starting state for thread. */
391 /* Awaiting everyone to be scheduled. */
393 /* Disable interrupts. */
394 STOPMACHINE_DISABLE_IRQ,
395 /* Run the function */
401 struct stop_machine_data {
404 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
405 unsigned int num_threads;
406 const struct cpumask *active_cpus;
408 enum stopmachine_state state;
412 static void set_state(struct stop_machine_data *smdata,
413 enum stopmachine_state newstate)
415 /* Reset ack counter. */
416 atomic_set(&smdata->thread_ack, smdata->num_threads);
418 smdata->state = newstate;
421 /* Last one to ack a state moves to the next state. */
422 static void ack_state(struct stop_machine_data *smdata)
424 if (atomic_dec_and_test(&smdata->thread_ack))
425 set_state(smdata, smdata->state + 1);
428 /* This is the cpu_stop function which stops the CPU. */
429 static int stop_machine_cpu_stop(void *data)
431 struct stop_machine_data *smdata = data;
432 enum stopmachine_state curstate = STOPMACHINE_NONE;
433 int cpu = smp_processor_id(), err = 0;
436 if (!smdata->active_cpus)
437 is_active = cpu == cpumask_first(cpu_online_mask);
439 is_active = cpumask_test_cpu(cpu, smdata->active_cpus);
441 /* Simple state machine */
443 /* Chill out and ensure we re-read stopmachine_state. */
445 if (smdata->state != curstate) {
446 curstate = smdata->state;
448 case STOPMACHINE_DISABLE_IRQ:
452 case STOPMACHINE_RUN:
454 err = smdata->fn(smdata->data);
461 } while (curstate != STOPMACHINE_EXIT);
467 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
469 struct stop_machine_data smdata = { .fn = fn, .data = data,
470 .num_threads = num_online_cpus(),
471 .active_cpus = cpus };
473 /* Set the initial state and stop all online cpus. */
474 set_state(&smdata, STOPMACHINE_PREPARE);
475 return stop_cpus(cpu_online_mask, stop_machine_cpu_stop, &smdata);
478 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
482 /* No CPUs can come up or down during this. */
484 ret = __stop_machine(fn, data, cpus);
488 EXPORT_SYMBOL_GPL(stop_machine);
490 #endif /* CONFIG_STOP_MACHINE */