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>
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 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 /* This is the cpu_stop function which stops the CPU. */
182 static int multi_cpu_stop(void *data
)
184 struct multi_stop_data
*msdata
= data
;
185 enum multi_stop_state curstate
= MULTI_STOP_NONE
;
186 int cpu
= smp_processor_id(), err
= 0;
191 * When called from stop_machine_from_inactive_cpu(), irq might
192 * already be disabled. Save the state and restore it on exit.
194 local_save_flags(flags
);
196 if (!msdata
->active_cpus
)
197 is_active
= cpu
== cpumask_first(cpu_online_mask
);
199 is_active
= cpumask_test_cpu(cpu
, msdata
->active_cpus
);
201 /* Simple state machine */
203 /* Chill out and ensure we re-read multi_stop_state. */
205 if (msdata
->state
!= curstate
) {
206 curstate
= msdata
->state
;
208 case MULTI_STOP_DISABLE_IRQ
:
214 err
= msdata
->fn(msdata
->data
);
220 } else if (curstate
> MULTI_STOP_PREPARE
) {
222 * At this stage all other CPUs we depend on must spin
223 * in the same loop. Any reason for hard-lockup should
224 * be detected and reported on their side.
226 touch_nmi_watchdog();
228 } while (curstate
!= MULTI_STOP_EXIT
);
230 local_irq_restore(flags
);
234 static int cpu_stop_queue_two_works(int cpu1
, struct cpu_stop_work
*work1
,
235 int cpu2
, struct cpu_stop_work
*work2
)
237 struct cpu_stopper
*stopper1
= per_cpu_ptr(&cpu_stopper
, cpu1
);
238 struct cpu_stopper
*stopper2
= per_cpu_ptr(&cpu_stopper
, cpu2
);
239 DEFINE_WAKE_Q(wakeq
);
242 raw_spin_lock_irq(&stopper1
->lock
);
243 raw_spin_lock_nested(&stopper2
->lock
, SINGLE_DEPTH_NESTING
);
246 if (!stopper1
->enabled
|| !stopper2
->enabled
)
249 * Ensure that if we race with __stop_cpus() the stoppers won't get
250 * queued up in reverse order leading to system deadlock.
252 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
253 * queued a work on cpu1 but not on cpu2, we hold both locks.
255 * It can be falsely true but it is safe to spin until it is cleared,
256 * queue_stop_cpus_work() does everything under preempt_disable().
259 if (unlikely(stop_cpus_in_progress
))
263 __cpu_stop_queue_work(stopper1
, work1
, &wakeq
);
264 __cpu_stop_queue_work(stopper2
, work2
, &wakeq
);
266 * The waking up of stopper threads has to happen
267 * in the same scheduling context as the queueing.
268 * Otherwise, there is a possibility of one of the
269 * above stoppers being woken up by another CPU,
270 * and preempting us. This will cause us to n ot
271 * wake up the other stopper forever.
275 raw_spin_unlock(&stopper2
->lock
);
276 raw_spin_unlock_irq(&stopper1
->lock
);
278 if (unlikely(err
== -EDEADLK
)) {
279 while (stop_cpus_in_progress
)
292 * stop_two_cpus - stops two cpus
293 * @cpu1: the cpu to stop
294 * @cpu2: the other cpu to stop
295 * @fn: function to execute
296 * @arg: argument to @fn
298 * Stops both the current and specified CPU and runs @fn on one of them.
300 * returns when both are completed.
302 int stop_two_cpus(unsigned int cpu1
, unsigned int cpu2
, cpu_stop_fn_t fn
, void *arg
)
304 struct cpu_stop_done done
;
305 struct cpu_stop_work work1
, work2
;
306 struct multi_stop_data msdata
;
308 msdata
= (struct multi_stop_data
){
312 .active_cpus
= cpumask_of(cpu1
),
315 work1
= work2
= (struct cpu_stop_work
){
316 .fn
= multi_cpu_stop
,
321 cpu_stop_init_done(&done
, 2);
322 set_state(&msdata
, MULTI_STOP_PREPARE
);
326 if (cpu_stop_queue_two_works(cpu1
, &work1
, cpu2
, &work2
))
329 wait_for_completion(&done
.completion
);
334 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
336 * @fn: function to execute
337 * @arg: argument to @fn
338 * @work_buf: pointer to cpu_stop_work structure
340 * Similar to stop_one_cpu() but doesn't wait for completion. The
341 * caller is responsible for ensuring @work_buf is currently unused
342 * and will remain untouched until stopper starts executing @fn.
348 * true if cpu_stop_work was queued successfully and @fn will be called,
351 bool stop_one_cpu_nowait(unsigned int cpu
, cpu_stop_fn_t fn
, void *arg
,
352 struct cpu_stop_work
*work_buf
)
354 *work_buf
= (struct cpu_stop_work
){ .fn
= fn
, .arg
= arg
, };
355 return cpu_stop_queue_work(cpu
, work_buf
);
358 static bool queue_stop_cpus_work(const struct cpumask
*cpumask
,
359 cpu_stop_fn_t fn
, void *arg
,
360 struct cpu_stop_done
*done
)
362 struct cpu_stop_work
*work
;
367 * Disable preemption while queueing to avoid getting
368 * preempted by a stopper which might wait for other stoppers
369 * to enter @fn which can lead to deadlock.
372 stop_cpus_in_progress
= true;
373 for_each_cpu(cpu
, cpumask
) {
374 work
= &per_cpu(cpu_stopper
.stop_work
, cpu
);
378 if (cpu_stop_queue_work(cpu
, work
))
381 stop_cpus_in_progress
= false;
387 static int __stop_cpus(const struct cpumask
*cpumask
,
388 cpu_stop_fn_t fn
, void *arg
)
390 struct cpu_stop_done done
;
392 cpu_stop_init_done(&done
, cpumask_weight(cpumask
));
393 if (!queue_stop_cpus_work(cpumask
, fn
, arg
, &done
))
395 wait_for_completion(&done
.completion
);
400 * stop_cpus - stop multiple cpus
401 * @cpumask: cpus to stop
402 * @fn: function to execute
403 * @arg: argument to @fn
405 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
406 * @fn is run in a process context with the highest priority
407 * preempting any task on the cpu and monopolizing it. This function
408 * returns after all executions are complete.
410 * This function doesn't guarantee the cpus in @cpumask stay online
411 * till @fn completes. If some cpus go down in the middle, execution
412 * on the cpu may happen partially or fully on different cpus. @fn
413 * should either be ready for that or the caller should ensure that
414 * the cpus stay online until this function completes.
416 * All stop_cpus() calls are serialized making it safe for @fn to wait
417 * for all cpus to start executing it.
423 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
424 * @cpumask were offline; otherwise, 0 if all executions of @fn
425 * returned 0, any non zero return value if any returned non zero.
427 int stop_cpus(const struct cpumask
*cpumask
, cpu_stop_fn_t fn
, void *arg
)
431 /* static works are used, process one request at a time */
432 mutex_lock(&stop_cpus_mutex
);
433 ret
= __stop_cpus(cpumask
, fn
, arg
);
434 mutex_unlock(&stop_cpus_mutex
);
439 * try_stop_cpus - try to stop multiple cpus
440 * @cpumask: cpus to stop
441 * @fn: function to execute
442 * @arg: argument to @fn
444 * Identical to stop_cpus() except that it fails with -EAGAIN if
445 * someone else is already using the facility.
451 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
452 * @fn(@arg) was not executed at all because all cpus in @cpumask were
453 * offline; otherwise, 0 if all executions of @fn returned 0, any non
454 * zero return value if any returned non zero.
456 int try_stop_cpus(const struct cpumask
*cpumask
, cpu_stop_fn_t fn
, void *arg
)
460 /* static works are used, process one request at a time */
461 if (!mutex_trylock(&stop_cpus_mutex
))
463 ret
= __stop_cpus(cpumask
, fn
, arg
);
464 mutex_unlock(&stop_cpus_mutex
);
468 static int cpu_stop_should_run(unsigned int cpu
)
470 struct cpu_stopper
*stopper
= &per_cpu(cpu_stopper
, cpu
);
474 raw_spin_lock_irqsave(&stopper
->lock
, flags
);
475 run
= !list_empty(&stopper
->works
);
476 raw_spin_unlock_irqrestore(&stopper
->lock
, flags
);
480 static void cpu_stopper_thread(unsigned int cpu
)
482 struct cpu_stopper
*stopper
= &per_cpu(cpu_stopper
, cpu
);
483 struct cpu_stop_work
*work
;
487 raw_spin_lock_irq(&stopper
->lock
);
488 if (!list_empty(&stopper
->works
)) {
489 work
= list_first_entry(&stopper
->works
,
490 struct cpu_stop_work
, list
);
491 list_del_init(&work
->list
);
493 raw_spin_unlock_irq(&stopper
->lock
);
496 cpu_stop_fn_t fn
= work
->fn
;
497 void *arg
= work
->arg
;
498 struct cpu_stop_done
*done
= work
->done
;
501 /* cpu stop callbacks must not sleep, make in_atomic() == T */
507 cpu_stop_signal_done(done
);
510 WARN_ONCE(preempt_count(),
511 "cpu_stop: %pf(%p) leaked preempt count\n", fn
, arg
);
516 void stop_machine_park(int cpu
)
518 struct cpu_stopper
*stopper
= &per_cpu(cpu_stopper
, cpu
);
520 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
521 * the pending works before it parks, until then it is fine to queue
524 stopper
->enabled
= false;
525 kthread_park(stopper
->thread
);
528 extern void sched_set_stop_task(int cpu
, struct task_struct
*stop
);
530 static void cpu_stop_create(unsigned int cpu
)
532 sched_set_stop_task(cpu
, per_cpu(cpu_stopper
.thread
, cpu
));
535 static void cpu_stop_park(unsigned int cpu
)
537 struct cpu_stopper
*stopper
= &per_cpu(cpu_stopper
, cpu
);
539 WARN_ON(!list_empty(&stopper
->works
));
542 void stop_machine_unpark(int cpu
)
544 struct cpu_stopper
*stopper
= &per_cpu(cpu_stopper
, cpu
);
546 stopper
->enabled
= true;
547 kthread_unpark(stopper
->thread
);
550 static struct smp_hotplug_thread cpu_stop_threads
= {
551 .store
= &cpu_stopper
.thread
,
552 .thread_should_run
= cpu_stop_should_run
,
553 .thread_fn
= cpu_stopper_thread
,
554 .thread_comm
= "migration/%u",
555 .create
= cpu_stop_create
,
556 .park
= cpu_stop_park
,
560 static int __init
cpu_stop_init(void)
564 for_each_possible_cpu(cpu
) {
565 struct cpu_stopper
*stopper
= &per_cpu(cpu_stopper
, cpu
);
567 raw_spin_lock_init(&stopper
->lock
);
568 INIT_LIST_HEAD(&stopper
->works
);
571 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads
));
572 stop_machine_unpark(raw_smp_processor_id());
573 stop_machine_initialized
= true;
576 early_initcall(cpu_stop_init
);
578 int stop_machine_cpuslocked(cpu_stop_fn_t fn
, void *data
,
579 const struct cpumask
*cpus
)
581 struct multi_stop_data msdata
= {
584 .num_threads
= num_online_cpus(),
588 lockdep_assert_cpus_held();
590 if (!stop_machine_initialized
) {
592 * Handle the case where stop_machine() is called
593 * early in boot before stop_machine() has been
599 WARN_ON_ONCE(msdata
.num_threads
!= 1);
601 local_irq_save(flags
);
604 local_irq_restore(flags
);
609 /* Set the initial state and stop all online cpus. */
610 set_state(&msdata
, MULTI_STOP_PREPARE
);
611 return stop_cpus(cpu_online_mask
, multi_cpu_stop
, &msdata
);
614 int stop_machine(cpu_stop_fn_t fn
, void *data
, const struct cpumask
*cpus
)
618 /* No CPUs can come up or down during this. */
620 ret
= stop_machine_cpuslocked(fn
, data
, cpus
);
624 EXPORT_SYMBOL_GPL(stop_machine
);
627 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
628 * @fn: the function to run
629 * @data: the data ptr for the @fn()
630 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
632 * This is identical to stop_machine() but can be called from a CPU which
633 * is not active. The local CPU is in the process of hotplug (so no other
634 * CPU hotplug can start) and not marked active and doesn't have enough
637 * This function provides stop_machine() functionality for such state by
638 * using busy-wait for synchronization and executing @fn directly for local
642 * Local CPU is inactive. Temporarily stops all active CPUs.
645 * 0 if all executions of @fn returned 0, any non zero return value if any
648 int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn
, void *data
,
649 const struct cpumask
*cpus
)
651 struct multi_stop_data msdata
= { .fn
= fn
, .data
= data
,
652 .active_cpus
= cpus
};
653 struct cpu_stop_done done
;
656 /* Local CPU must be inactive and CPU hotplug in progress. */
657 BUG_ON(cpu_active(raw_smp_processor_id()));
658 msdata
.num_threads
= num_active_cpus() + 1; /* +1 for local */
660 /* No proper task established and can't sleep - busy wait for lock. */
661 while (!mutex_trylock(&stop_cpus_mutex
))
664 /* Schedule work on other CPUs and execute directly for local CPU */
665 set_state(&msdata
, MULTI_STOP_PREPARE
);
666 cpu_stop_init_done(&done
, num_active_cpus());
667 queue_stop_cpus_work(cpu_active_mask
, multi_cpu_stop
, &msdata
,
669 ret
= multi_cpu_stop(&msdata
);
671 /* Busy wait for completion. */
672 while (!completion_done(&done
.completion
))
675 mutex_unlock(&stop_cpus_mutex
);
676 return ret
?: done
.ret
;