Loading...
1/*
2 * kernel/stop_machine.c
3 *
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>
8 *
9 * This file is released under the GPLv2 and any later version.
10 */
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
25/*
26 * Structure to determine completion condition and record errors. May
27 * be shared by works on different cpus.
28 */
29struct 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 */
33};
34
35/* the actual stopper, one per every possible cpu, enabled on online cpus */
36struct cpu_stopper {
37 struct task_struct *thread;
38
39 spinlock_t lock;
40 bool enabled; /* is this stopper enabled? */
41 struct list_head works; /* list of pending works */
42
43 struct cpu_stop_work stop_work; /* for stop_cpus */
44};
45
46static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
47static bool stop_machine_initialized = false;
48
49/* static data for stop_cpus */
50static DEFINE_MUTEX(stop_cpus_mutex);
51static bool stop_cpus_in_progress;
52
53static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
54{
55 memset(done, 0, sizeof(*done));
56 atomic_set(&done->nr_todo, nr_todo);
57 init_completion(&done->completion);
58}
59
60/* signal completion unless @done is NULL */
61static void cpu_stop_signal_done(struct cpu_stop_done *done)
62{
63 if (atomic_dec_and_test(&done->nr_todo))
64 complete(&done->completion);
65}
66
67static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
68 struct cpu_stop_work *work)
69{
70 list_add_tail(&work->list, &stopper->works);
71 wake_up_process(stopper->thread);
72}
73
74/* queue @work to @stopper. if offline, @work is completed immediately */
75static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
76{
77 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
78 unsigned long flags;
79 bool enabled;
80
81 spin_lock_irqsave(&stopper->lock, flags);
82 enabled = stopper->enabled;
83 if (enabled)
84 __cpu_stop_queue_work(stopper, work);
85 else if (work->done)
86 cpu_stop_signal_done(work->done);
87 spin_unlock_irqrestore(&stopper->lock, flags);
88
89 return enabled;
90}
91
92/**
93 * stop_one_cpu - stop a cpu
94 * @cpu: cpu to stop
95 * @fn: function to execute
96 * @arg: argument to @fn
97 *
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
101 * complete.
102 *
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.
108 *
109 * CONTEXT:
110 * Might sleep.
111 *
112 * RETURNS:
113 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
114 * otherwise, the return value of @fn.
115 */
116int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
117{
118 struct cpu_stop_done done;
119 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
120
121 cpu_stop_init_done(&done, 1);
122 if (!cpu_stop_queue_work(cpu, &work))
123 return -ENOENT;
124 /*
125 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
126 * cycle by doing a preemption:
127 */
128 cond_resched();
129 wait_for_completion(&done.completion);
130 return done.ret;
131}
132
133/* This controls the threads on each CPU. */
134enum multi_stop_state {
135 /* Dummy starting state for thread. */
136 MULTI_STOP_NONE,
137 /* Awaiting everyone to be scheduled. */
138 MULTI_STOP_PREPARE,
139 /* Disable interrupts. */
140 MULTI_STOP_DISABLE_IRQ,
141 /* Run the function */
142 MULTI_STOP_RUN,
143 /* Exit */
144 MULTI_STOP_EXIT,
145};
146
147struct multi_stop_data {
148 cpu_stop_fn_t fn;
149 void *data;
150 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
151 unsigned int num_threads;
152 const struct cpumask *active_cpus;
153
154 enum multi_stop_state state;
155 atomic_t thread_ack;
156};
157
158static void set_state(struct multi_stop_data *msdata,
159 enum multi_stop_state newstate)
160{
161 /* Reset ack counter. */
162 atomic_set(&msdata->thread_ack, msdata->num_threads);
163 smp_wmb();
164 msdata->state = newstate;
165}
166
167/* Last one to ack a state moves to the next state. */
168static void ack_state(struct multi_stop_data *msdata)
169{
170 if (atomic_dec_and_test(&msdata->thread_ack))
171 set_state(msdata, msdata->state + 1);
172}
173
174/* This is the cpu_stop function which stops the CPU. */
175static int multi_cpu_stop(void *data)
176{
177 struct multi_stop_data *msdata = data;
178 enum multi_stop_state curstate = MULTI_STOP_NONE;
179 int cpu = smp_processor_id(), err = 0;
180 unsigned long flags;
181 bool is_active;
182
183 /*
184 * When called from stop_machine_from_inactive_cpu(), irq might
185 * already be disabled. Save the state and restore it on exit.
186 */
187 local_save_flags(flags);
188
189 if (!msdata->active_cpus)
190 is_active = cpu == cpumask_first(cpu_online_mask);
191 else
192 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
193
194 /* Simple state machine */
195 do {
196 /* Chill out and ensure we re-read multi_stop_state. */
197 cpu_relax_yield();
198 if (msdata->state != curstate) {
199 curstate = msdata->state;
200 switch (curstate) {
201 case MULTI_STOP_DISABLE_IRQ:
202 local_irq_disable();
203 hard_irq_disable();
204 break;
205 case MULTI_STOP_RUN:
206 if (is_active)
207 err = msdata->fn(msdata->data);
208 break;
209 default:
210 break;
211 }
212 ack_state(msdata);
213 } else if (curstate > MULTI_STOP_PREPARE) {
214 /*
215 * At this stage all other CPUs we depend on must spin
216 * in the same loop. Any reason for hard-lockup should
217 * be detected and reported on their side.
218 */
219 touch_nmi_watchdog();
220 }
221 } while (curstate != MULTI_STOP_EXIT);
222
223 local_irq_restore(flags);
224 return err;
225}
226
227static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
228 int cpu2, struct cpu_stop_work *work2)
229{
230 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
231 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
232 int err;
233retry:
234 spin_lock_irq(&stopper1->lock);
235 spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
236
237 err = -ENOENT;
238 if (!stopper1->enabled || !stopper2->enabled)
239 goto unlock;
240 /*
241 * Ensure that if we race with __stop_cpus() the stoppers won't get
242 * queued up in reverse order leading to system deadlock.
243 *
244 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
245 * queued a work on cpu1 but not on cpu2, we hold both locks.
246 *
247 * It can be falsely true but it is safe to spin until it is cleared,
248 * queue_stop_cpus_work() does everything under preempt_disable().
249 */
250 err = -EDEADLK;
251 if (unlikely(stop_cpus_in_progress))
252 goto unlock;
253
254 err = 0;
255 __cpu_stop_queue_work(stopper1, work1);
256 __cpu_stop_queue_work(stopper2, work2);
257unlock:
258 spin_unlock(&stopper2->lock);
259 spin_unlock_irq(&stopper1->lock);
260
261 if (unlikely(err == -EDEADLK)) {
262 while (stop_cpus_in_progress)
263 cpu_relax();
264 goto retry;
265 }
266 return err;
267}
268/**
269 * stop_two_cpus - stops two cpus
270 * @cpu1: the cpu to stop
271 * @cpu2: the other cpu to stop
272 * @fn: function to execute
273 * @arg: argument to @fn
274 *
275 * Stops both the current and specified CPU and runs @fn on one of them.
276 *
277 * returns when both are completed.
278 */
279int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
280{
281 struct cpu_stop_done done;
282 struct cpu_stop_work work1, work2;
283 struct multi_stop_data msdata;
284
285 msdata = (struct multi_stop_data){
286 .fn = fn,
287 .data = arg,
288 .num_threads = 2,
289 .active_cpus = cpumask_of(cpu1),
290 };
291
292 work1 = work2 = (struct cpu_stop_work){
293 .fn = multi_cpu_stop,
294 .arg = &msdata,
295 .done = &done
296 };
297
298 cpu_stop_init_done(&done, 2);
299 set_state(&msdata, MULTI_STOP_PREPARE);
300
301 if (cpu1 > cpu2)
302 swap(cpu1, cpu2);
303 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
304 return -ENOENT;
305
306 wait_for_completion(&done.completion);
307 return done.ret;
308}
309
310/**
311 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
312 * @cpu: cpu to stop
313 * @fn: function to execute
314 * @arg: argument to @fn
315 * @work_buf: pointer to cpu_stop_work structure
316 *
317 * Similar to stop_one_cpu() but doesn't wait for completion. The
318 * caller is responsible for ensuring @work_buf is currently unused
319 * and will remain untouched until stopper starts executing @fn.
320 *
321 * CONTEXT:
322 * Don't care.
323 *
324 * RETURNS:
325 * true if cpu_stop_work was queued successfully and @fn will be called,
326 * false otherwise.
327 */
328bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
329 struct cpu_stop_work *work_buf)
330{
331 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
332 return cpu_stop_queue_work(cpu, work_buf);
333}
334
335static bool queue_stop_cpus_work(const struct cpumask *cpumask,
336 cpu_stop_fn_t fn, void *arg,
337 struct cpu_stop_done *done)
338{
339 struct cpu_stop_work *work;
340 unsigned int cpu;
341 bool queued = false;
342
343 /*
344 * Disable preemption while queueing to avoid getting
345 * preempted by a stopper which might wait for other stoppers
346 * to enter @fn which can lead to deadlock.
347 */
348 preempt_disable();
349 stop_cpus_in_progress = true;
350 for_each_cpu(cpu, cpumask) {
351 work = &per_cpu(cpu_stopper.stop_work, cpu);
352 work->fn = fn;
353 work->arg = arg;
354 work->done = done;
355 if (cpu_stop_queue_work(cpu, work))
356 queued = true;
357 }
358 stop_cpus_in_progress = false;
359 preempt_enable();
360
361 return queued;
362}
363
364static int __stop_cpus(const struct cpumask *cpumask,
365 cpu_stop_fn_t fn, void *arg)
366{
367 struct cpu_stop_done done;
368
369 cpu_stop_init_done(&done, cpumask_weight(cpumask));
370 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
371 return -ENOENT;
372 wait_for_completion(&done.completion);
373 return done.ret;
374}
375
376/**
377 * stop_cpus - stop multiple cpus
378 * @cpumask: cpus to stop
379 * @fn: function to execute
380 * @arg: argument to @fn
381 *
382 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
383 * @fn is run in a process context with the highest priority
384 * preempting any task on the cpu and monopolizing it. This function
385 * returns after all executions are complete.
386 *
387 * This function doesn't guarantee the cpus in @cpumask stay online
388 * till @fn completes. If some cpus go down in the middle, execution
389 * on the cpu may happen partially or fully on different cpus. @fn
390 * should either be ready for that or the caller should ensure that
391 * the cpus stay online until this function completes.
392 *
393 * All stop_cpus() calls are serialized making it safe for @fn to wait
394 * for all cpus to start executing it.
395 *
396 * CONTEXT:
397 * Might sleep.
398 *
399 * RETURNS:
400 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
401 * @cpumask were offline; otherwise, 0 if all executions of @fn
402 * returned 0, any non zero return value if any returned non zero.
403 */
404int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
405{
406 int ret;
407
408 /* static works are used, process one request at a time */
409 mutex_lock(&stop_cpus_mutex);
410 ret = __stop_cpus(cpumask, fn, arg);
411 mutex_unlock(&stop_cpus_mutex);
412 return ret;
413}
414
415/**
416 * try_stop_cpus - try to stop multiple cpus
417 * @cpumask: cpus to stop
418 * @fn: function to execute
419 * @arg: argument to @fn
420 *
421 * Identical to stop_cpus() except that it fails with -EAGAIN if
422 * someone else is already using the facility.
423 *
424 * CONTEXT:
425 * Might sleep.
426 *
427 * RETURNS:
428 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
429 * @fn(@arg) was not executed at all because all cpus in @cpumask were
430 * offline; otherwise, 0 if all executions of @fn returned 0, any non
431 * zero return value if any returned non zero.
432 */
433int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
434{
435 int ret;
436
437 /* static works are used, process one request at a time */
438 if (!mutex_trylock(&stop_cpus_mutex))
439 return -EAGAIN;
440 ret = __stop_cpus(cpumask, fn, arg);
441 mutex_unlock(&stop_cpus_mutex);
442 return ret;
443}
444
445static int cpu_stop_should_run(unsigned int cpu)
446{
447 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
448 unsigned long flags;
449 int run;
450
451 spin_lock_irqsave(&stopper->lock, flags);
452 run = !list_empty(&stopper->works);
453 spin_unlock_irqrestore(&stopper->lock, flags);
454 return run;
455}
456
457static void cpu_stopper_thread(unsigned int cpu)
458{
459 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
460 struct cpu_stop_work *work;
461
462repeat:
463 work = NULL;
464 spin_lock_irq(&stopper->lock);
465 if (!list_empty(&stopper->works)) {
466 work = list_first_entry(&stopper->works,
467 struct cpu_stop_work, list);
468 list_del_init(&work->list);
469 }
470 spin_unlock_irq(&stopper->lock);
471
472 if (work) {
473 cpu_stop_fn_t fn = work->fn;
474 void *arg = work->arg;
475 struct cpu_stop_done *done = work->done;
476 int ret;
477
478 /* cpu stop callbacks must not sleep, make in_atomic() == T */
479 preempt_count_inc();
480 ret = fn(arg);
481 if (done) {
482 if (ret)
483 done->ret = ret;
484 cpu_stop_signal_done(done);
485 }
486 preempt_count_dec();
487 WARN_ONCE(preempt_count(),
488 "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
489 goto repeat;
490 }
491}
492
493void stop_machine_park(int cpu)
494{
495 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
496 /*
497 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
498 * the pending works before it parks, until then it is fine to queue
499 * the new works.
500 */
501 stopper->enabled = false;
502 kthread_park(stopper->thread);
503}
504
505extern void sched_set_stop_task(int cpu, struct task_struct *stop);
506
507static void cpu_stop_create(unsigned int cpu)
508{
509 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
510}
511
512static void cpu_stop_park(unsigned int cpu)
513{
514 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
515
516 WARN_ON(!list_empty(&stopper->works));
517}
518
519void stop_machine_unpark(int cpu)
520{
521 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
522
523 stopper->enabled = true;
524 kthread_unpark(stopper->thread);
525}
526
527static struct smp_hotplug_thread cpu_stop_threads = {
528 .store = &cpu_stopper.thread,
529 .thread_should_run = cpu_stop_should_run,
530 .thread_fn = cpu_stopper_thread,
531 .thread_comm = "migration/%u",
532 .create = cpu_stop_create,
533 .park = cpu_stop_park,
534 .selfparking = true,
535};
536
537static int __init cpu_stop_init(void)
538{
539 unsigned int cpu;
540
541 for_each_possible_cpu(cpu) {
542 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
543
544 spin_lock_init(&stopper->lock);
545 INIT_LIST_HEAD(&stopper->works);
546 }
547
548 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
549 stop_machine_unpark(raw_smp_processor_id());
550 stop_machine_initialized = true;
551 return 0;
552}
553early_initcall(cpu_stop_init);
554
555static int __stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
556{
557 struct multi_stop_data msdata = {
558 .fn = fn,
559 .data = data,
560 .num_threads = num_online_cpus(),
561 .active_cpus = cpus,
562 };
563
564 if (!stop_machine_initialized) {
565 /*
566 * Handle the case where stop_machine() is called
567 * early in boot before stop_machine() has been
568 * initialized.
569 */
570 unsigned long flags;
571 int ret;
572
573 WARN_ON_ONCE(msdata.num_threads != 1);
574
575 local_irq_save(flags);
576 hard_irq_disable();
577 ret = (*fn)(data);
578 local_irq_restore(flags);
579
580 return ret;
581 }
582
583 /* Set the initial state and stop all online cpus. */
584 set_state(&msdata, MULTI_STOP_PREPARE);
585 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
586}
587
588int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
589{
590 int ret;
591
592 /* No CPUs can come up or down during this. */
593 get_online_cpus();
594 ret = __stop_machine(fn, data, cpus);
595 put_online_cpus();
596 return ret;
597}
598EXPORT_SYMBOL_GPL(stop_machine);
599
600/**
601 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
602 * @fn: the function to run
603 * @data: the data ptr for the @fn()
604 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
605 *
606 * This is identical to stop_machine() but can be called from a CPU which
607 * is not active. The local CPU is in the process of hotplug (so no other
608 * CPU hotplug can start) and not marked active and doesn't have enough
609 * context to sleep.
610 *
611 * This function provides stop_machine() functionality for such state by
612 * using busy-wait for synchronization and executing @fn directly for local
613 * CPU.
614 *
615 * CONTEXT:
616 * Local CPU is inactive. Temporarily stops all active CPUs.
617 *
618 * RETURNS:
619 * 0 if all executions of @fn returned 0, any non zero return value if any
620 * returned non zero.
621 */
622int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
623 const struct cpumask *cpus)
624{
625 struct multi_stop_data msdata = { .fn = fn, .data = data,
626 .active_cpus = cpus };
627 struct cpu_stop_done done;
628 int ret;
629
630 /* Local CPU must be inactive and CPU hotplug in progress. */
631 BUG_ON(cpu_active(raw_smp_processor_id()));
632 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
633
634 /* No proper task established and can't sleep - busy wait for lock. */
635 while (!mutex_trylock(&stop_cpus_mutex))
636 cpu_relax();
637
638 /* Schedule work on other CPUs and execute directly for local CPU */
639 set_state(&msdata, MULTI_STOP_PREPARE);
640 cpu_stop_init_done(&done, num_active_cpus());
641 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
642 &done);
643 ret = multi_cpu_stop(&msdata);
644
645 /* Busy wait for completion. */
646 while (!completion_done(&done.completion))
647 cpu_relax();
648
649 mutex_unlock(&stop_cpus_mutex);
650 return ret ?: done.ret;
651}
1/*
2 * kernel/stop_machine.c
3 *
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>
8 *
9 * This file is released under the GPLv2 and any later version.
10 */
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
22#include <linux/atomic.h>
23
24/*
25 * Structure to determine completion condition and record errors. May
26 * be shared by works on different cpus.
27 */
28struct 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 */
33};
34
35/* the actual stopper, one per every possible cpu, enabled on online cpus */
36struct cpu_stopper {
37 spinlock_t lock;
38 bool enabled; /* is this stopper enabled? */
39 struct list_head works; /* list of pending works */
40 struct task_struct *thread; /* stopper thread */
41};
42
43static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
44static bool stop_machine_initialized = false;
45
46static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
47{
48 memset(done, 0, sizeof(*done));
49 atomic_set(&done->nr_todo, nr_todo);
50 init_completion(&done->completion);
51}
52
53/* signal completion unless @done is NULL */
54static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
55{
56 if (done) {
57 if (executed)
58 done->executed = true;
59 if (atomic_dec_and_test(&done->nr_todo))
60 complete(&done->completion);
61 }
62}
63
64/* queue @work to @stopper. if offline, @work is completed immediately */
65static void cpu_stop_queue_work(struct cpu_stopper *stopper,
66 struct cpu_stop_work *work)
67{
68 unsigned long flags;
69
70 spin_lock_irqsave(&stopper->lock, flags);
71
72 if (stopper->enabled) {
73 list_add_tail(&work->list, &stopper->works);
74 wake_up_process(stopper->thread);
75 } else
76 cpu_stop_signal_done(work->done, false);
77
78 spin_unlock_irqrestore(&stopper->lock, flags);
79}
80
81/**
82 * stop_one_cpu - stop a cpu
83 * @cpu: cpu to stop
84 * @fn: function to execute
85 * @arg: argument to @fn
86 *
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
90 * complete.
91 *
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.
97 *
98 * CONTEXT:
99 * Might sleep.
100 *
101 * RETURNS:
102 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
103 * otherwise, the return value of @fn.
104 */
105int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
106{
107 struct cpu_stop_done done;
108 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
109
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;
114}
115
116/**
117 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
118 * @cpu: cpu to stop
119 * @fn: function to execute
120 * @arg: argument to @fn
121 *
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.
125 *
126 * CONTEXT:
127 * Don't care.
128 */
129void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
130 struct cpu_stop_work *work_buf)
131{
132 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
133 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
134}
135
136/* static data for stop_cpus */
137static DEFINE_MUTEX(stop_cpus_mutex);
138static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
139
140static void queue_stop_cpus_work(const struct cpumask *cpumask,
141 cpu_stop_fn_t fn, void *arg,
142 struct cpu_stop_done *done)
143{
144 struct cpu_stop_work *work;
145 unsigned int cpu;
146
147 /* initialize works and done */
148 for_each_cpu(cpu, cpumask) {
149 work = &per_cpu(stop_cpus_work, cpu);
150 work->fn = fn;
151 work->arg = arg;
152 work->done = done;
153 }
154
155 /*
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.
159 */
160 preempt_disable();
161 for_each_cpu(cpu, cpumask)
162 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
163 &per_cpu(stop_cpus_work, cpu));
164 preempt_enable();
165}
166
167static int __stop_cpus(const struct cpumask *cpumask,
168 cpu_stop_fn_t fn, void *arg)
169{
170 struct cpu_stop_done done;
171
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;
176}
177
178/**
179 * stop_cpus - stop multiple cpus
180 * @cpumask: cpus to stop
181 * @fn: function to execute
182 * @arg: argument to @fn
183 *
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.
188 *
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.
194 *
195 * All stop_cpus() calls are serialized making it safe for @fn to wait
196 * for all cpus to start executing it.
197 *
198 * CONTEXT:
199 * Might sleep.
200 *
201 * RETURNS:
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.
205 */
206int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
207{
208 int ret;
209
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);
214 return ret;
215}
216
217/**
218 * try_stop_cpus - try to stop multiple cpus
219 * @cpumask: cpus to stop
220 * @fn: function to execute
221 * @arg: argument to @fn
222 *
223 * Identical to stop_cpus() except that it fails with -EAGAIN if
224 * someone else is already using the facility.
225 *
226 * CONTEXT:
227 * Might sleep.
228 *
229 * RETURNS:
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.
234 */
235int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
236{
237 int ret;
238
239 /* static works are used, process one request at a time */
240 if (!mutex_trylock(&stop_cpus_mutex))
241 return -EAGAIN;
242 ret = __stop_cpus(cpumask, fn, arg);
243 mutex_unlock(&stop_cpus_mutex);
244 return ret;
245}
246
247static int cpu_stopper_thread(void *data)
248{
249 struct cpu_stopper *stopper = data;
250 struct cpu_stop_work *work;
251 int ret;
252
253repeat:
254 set_current_state(TASK_INTERRUPTIBLE); /* mb paired w/ kthread_stop */
255
256 if (kthread_should_stop()) {
257 __set_current_state(TASK_RUNNING);
258 return 0;
259 }
260
261 work = NULL;
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);
267 }
268 spin_unlock_irq(&stopper->lock);
269
270 if (work) {
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;
275
276 __set_current_state(TASK_RUNNING);
277
278 /* cpu stop callbacks are not allowed to sleep */
279 preempt_disable();
280
281 ret = fn(arg);
282 if (ret)
283 done->ret = ret;
284
285 /* restore preemption and check it's still balanced */
286 preempt_enable();
287 WARN_ONCE(preempt_count(),
288 "cpu_stop: %s(%p) leaked preempt count\n",
289 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
290 ksym_buf), arg);
291
292 cpu_stop_signal_done(done, true);
293 } else
294 schedule();
295
296 goto repeat;
297}
298
299extern void sched_set_stop_task(int cpu, struct task_struct *stop);
300
301/* manage stopper for a cpu, mostly lifted from sched migration thread mgmt */
302static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
303 unsigned long action, void *hcpu)
304{
305 unsigned int cpu = (unsigned long)hcpu;
306 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
307 struct task_struct *p;
308
309 switch (action & ~CPU_TASKS_FROZEN) {
310 case CPU_UP_PREPARE:
311 BUG_ON(stopper->thread || stopper->enabled ||
312 !list_empty(&stopper->works));
313 p = kthread_create_on_node(cpu_stopper_thread,
314 stopper,
315 cpu_to_node(cpu),
316 "migration/%d", cpu);
317 if (IS_ERR(p))
318 return notifier_from_errno(PTR_ERR(p));
319 get_task_struct(p);
320 kthread_bind(p, cpu);
321 sched_set_stop_task(cpu, p);
322 stopper->thread = p;
323 break;
324
325 case CPU_ONLINE:
326 /* strictly unnecessary, as first user will wake it */
327 wake_up_process(stopper->thread);
328 /* mark enabled */
329 spin_lock_irq(&stopper->lock);
330 stopper->enabled = true;
331 spin_unlock_irq(&stopper->lock);
332 break;
333
334#ifdef CONFIG_HOTPLUG_CPU
335 case CPU_UP_CANCELED:
336 case CPU_POST_DEAD:
337 {
338 struct cpu_stop_work *work;
339
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;
352 break;
353 }
354#endif
355 }
356
357 return NOTIFY_OK;
358}
359
360/*
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.
364 */
365static struct notifier_block __cpuinitdata cpu_stop_cpu_notifier = {
366 .notifier_call = cpu_stop_cpu_callback,
367 .priority = 10,
368};
369
370static int __init cpu_stop_init(void)
371{
372 void *bcpu = (void *)(long)smp_processor_id();
373 unsigned int cpu;
374 int err;
375
376 for_each_possible_cpu(cpu) {
377 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
378
379 spin_lock_init(&stopper->lock);
380 INIT_LIST_HEAD(&stopper->works);
381 }
382
383 /* start one for the boot cpu */
384 err = cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_UP_PREPARE,
385 bcpu);
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);
389
390 stop_machine_initialized = true;
391
392 return 0;
393}
394early_initcall(cpu_stop_init);
395
396#ifdef CONFIG_STOP_MACHINE
397
398/* This controls the threads on each CPU. */
399enum stopmachine_state {
400 /* Dummy starting state for thread. */
401 STOPMACHINE_NONE,
402 /* Awaiting everyone to be scheduled. */
403 STOPMACHINE_PREPARE,
404 /* Disable interrupts. */
405 STOPMACHINE_DISABLE_IRQ,
406 /* Run the function */
407 STOPMACHINE_RUN,
408 /* Exit */
409 STOPMACHINE_EXIT,
410};
411
412struct stop_machine_data {
413 int (*fn)(void *);
414 void *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;
418
419 enum stopmachine_state state;
420 atomic_t thread_ack;
421};
422
423static void set_state(struct stop_machine_data *smdata,
424 enum stopmachine_state newstate)
425{
426 /* Reset ack counter. */
427 atomic_set(&smdata->thread_ack, smdata->num_threads);
428 smp_wmb();
429 smdata->state = newstate;
430}
431
432/* Last one to ack a state moves to the next state. */
433static void ack_state(struct stop_machine_data *smdata)
434{
435 if (atomic_dec_and_test(&smdata->thread_ack))
436 set_state(smdata, smdata->state + 1);
437}
438
439/* This is the cpu_stop function which stops the CPU. */
440static int stop_machine_cpu_stop(void *data)
441{
442 struct stop_machine_data *smdata = data;
443 enum stopmachine_state curstate = STOPMACHINE_NONE;
444 int cpu = smp_processor_id(), err = 0;
445 unsigned long flags;
446 bool is_active;
447
448 /*
449 * When called from stop_machine_from_inactive_cpu(), irq might
450 * already be disabled. Save the state and restore it on exit.
451 */
452 local_save_flags(flags);
453
454 if (!smdata->active_cpus)
455 is_active = cpu == cpumask_first(cpu_online_mask);
456 else
457 is_active = cpumask_test_cpu(cpu, smdata->active_cpus);
458
459 /* Simple state machine */
460 do {
461 /* Chill out and ensure we re-read stopmachine_state. */
462 cpu_relax();
463 if (smdata->state != curstate) {
464 curstate = smdata->state;
465 switch (curstate) {
466 case STOPMACHINE_DISABLE_IRQ:
467 local_irq_disable();
468 hard_irq_disable();
469 break;
470 case STOPMACHINE_RUN:
471 if (is_active)
472 err = smdata->fn(smdata->data);
473 break;
474 default:
475 break;
476 }
477 ack_state(smdata);
478 }
479 } while (curstate != STOPMACHINE_EXIT);
480
481 local_irq_restore(flags);
482 return err;
483}
484
485int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
486{
487 struct stop_machine_data smdata = { .fn = fn, .data = data,
488 .num_threads = num_online_cpus(),
489 .active_cpus = cpus };
490
491 if (!stop_machine_initialized) {
492 /*
493 * Handle the case where stop_machine() is called
494 * early in boot before stop_machine() has been
495 * initialized.
496 */
497 unsigned long flags;
498 int ret;
499
500 WARN_ON_ONCE(smdata.num_threads != 1);
501
502 local_irq_save(flags);
503 hard_irq_disable();
504 ret = (*fn)(data);
505 local_irq_restore(flags);
506
507 return ret;
508 }
509
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);
513}
514
515int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
516{
517 int ret;
518
519 /* No CPUs can come up or down during this. */
520 get_online_cpus();
521 ret = __stop_machine(fn, data, cpus);
522 put_online_cpus();
523 return ret;
524}
525EXPORT_SYMBOL_GPL(stop_machine);
526
527/**
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)
532 *
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
536 * context to sleep.
537 *
538 * This function provides stop_machine() functionality for such state by
539 * using busy-wait for synchronization and executing @fn directly for local
540 * CPU.
541 *
542 * CONTEXT:
543 * Local CPU is inactive. Temporarily stops all active CPUs.
544 *
545 * RETURNS:
546 * 0 if all executions of @fn returned 0, any non zero return value if any
547 * returned non zero.
548 */
549int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
550 const struct cpumask *cpus)
551{
552 struct stop_machine_data smdata = { .fn = fn, .data = data,
553 .active_cpus = cpus };
554 struct cpu_stop_done done;
555 int ret;
556
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 */
560
561 /* No proper task established and can't sleep - busy wait for lock. */
562 while (!mutex_trylock(&stop_cpus_mutex))
563 cpu_relax();
564
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,
569 &done);
570 ret = stop_machine_cpu_stop(&smdata);
571
572 /* Busy wait for completion. */
573 while (!completion_done(&done.completion))
574 cpu_relax();
575
576 mutex_unlock(&stop_cpus_mutex);
577 return ret ?: done.ret;
578}
579
580#endif /* CONFIG_STOP_MACHINE */