Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * kernel/stop_machine.c
4 *
5 * Copyright (C) 2008, 2005 IBM Corporation.
6 * Copyright (C) 2008, 2005 Rusty Russell rusty@rustcorp.com.au
7 * Copyright (C) 2010 SUSE Linux Products GmbH
8 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
9 */
10#include <linux/compiler.h>
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>
25
26/*
27 * Structure to determine completion condition and record errors. May
28 * be shared by works on different cpus.
29 */
30struct 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 */
34};
35
36/* the actual stopper, one per every possible cpu, enabled on online cpus */
37struct cpu_stopper {
38 struct task_struct *thread;
39
40 raw_spinlock_t lock;
41 bool enabled; /* is this stopper enabled? */
42 struct list_head works; /* list of pending works */
43
44 struct cpu_stop_work stop_work; /* for stop_cpus */
45};
46
47static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
48static bool stop_machine_initialized = false;
49
50/* static data for stop_cpus */
51static DEFINE_MUTEX(stop_cpus_mutex);
52static bool stop_cpus_in_progress;
53
54static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
55{
56 memset(done, 0, sizeof(*done));
57 atomic_set(&done->nr_todo, nr_todo);
58 init_completion(&done->completion);
59}
60
61/* signal completion unless @done is NULL */
62static void cpu_stop_signal_done(struct cpu_stop_done *done)
63{
64 if (atomic_dec_and_test(&done->nr_todo))
65 complete(&done->completion);
66}
67
68static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
69 struct cpu_stop_work *work,
70 struct wake_q_head *wakeq)
71{
72 list_add_tail(&work->list, &stopper->works);
73 wake_q_add(wakeq, stopper->thread);
74}
75
76/* queue @work to @stopper. if offline, @work is completed immediately */
77static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
78{
79 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
80 DEFINE_WAKE_Q(wakeq);
81 unsigned long flags;
82 bool enabled;
83
84 preempt_disable();
85 raw_spin_lock_irqsave(&stopper->lock, flags);
86 enabled = stopper->enabled;
87 if (enabled)
88 __cpu_stop_queue_work(stopper, work, &wakeq);
89 else if (work->done)
90 cpu_stop_signal_done(work->done);
91 raw_spin_unlock_irqrestore(&stopper->lock, flags);
92
93 wake_up_q(&wakeq);
94 preempt_enable();
95
96 return enabled;
97}
98
99/**
100 * stop_one_cpu - stop a cpu
101 * @cpu: cpu to stop
102 * @fn: function to execute
103 * @arg: argument to @fn
104 *
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
108 * complete.
109 *
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.
115 *
116 * CONTEXT:
117 * Might sleep.
118 *
119 * RETURNS:
120 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
121 * otherwise, the return value of @fn.
122 */
123int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
124{
125 struct cpu_stop_done done;
126 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
127
128 cpu_stop_init_done(&done, 1);
129 if (!cpu_stop_queue_work(cpu, &work))
130 return -ENOENT;
131 /*
132 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
133 * cycle by doing a preemption:
134 */
135 cond_resched();
136 wait_for_completion(&done.completion);
137 return done.ret;
138}
139
140/* This controls the threads on each CPU. */
141enum multi_stop_state {
142 /* Dummy starting state for thread. */
143 MULTI_STOP_NONE,
144 /* Awaiting everyone to be scheduled. */
145 MULTI_STOP_PREPARE,
146 /* Disable interrupts. */
147 MULTI_STOP_DISABLE_IRQ,
148 /* Run the function */
149 MULTI_STOP_RUN,
150 /* Exit */
151 MULTI_STOP_EXIT,
152};
153
154struct multi_stop_data {
155 cpu_stop_fn_t fn;
156 void *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;
160
161 enum multi_stop_state state;
162 atomic_t thread_ack;
163};
164
165static void set_state(struct multi_stop_data *msdata,
166 enum multi_stop_state newstate)
167{
168 /* Reset ack counter. */
169 atomic_set(&msdata->thread_ack, msdata->num_threads);
170 smp_wmb();
171 WRITE_ONCE(msdata->state, newstate);
172}
173
174/* Last one to ack a state moves to the next state. */
175static void ack_state(struct multi_stop_data *msdata)
176{
177 if (atomic_dec_and_test(&msdata->thread_ack))
178 set_state(msdata, msdata->state + 1);
179}
180
181void __weak stop_machine_yield(const struct cpumask *cpumask)
182{
183 cpu_relax();
184}
185
186/* This is the cpu_stop function which stops the CPU. */
187static int multi_cpu_stop(void *data)
188{
189 struct multi_stop_data *msdata = data;
190 enum multi_stop_state newstate, curstate = MULTI_STOP_NONE;
191 int cpu = smp_processor_id(), err = 0;
192 const struct cpumask *cpumask;
193 unsigned long flags;
194 bool is_active;
195
196 /*
197 * When called from stop_machine_from_inactive_cpu(), irq might
198 * already be disabled. Save the state and restore it on exit.
199 */
200 local_save_flags(flags);
201
202 if (!msdata->active_cpus) {
203 cpumask = cpu_online_mask;
204 is_active = cpu == cpumask_first(cpumask);
205 } else {
206 cpumask = msdata->active_cpus;
207 is_active = cpumask_test_cpu(cpu, cpumask);
208 }
209
210 /* Simple state machine */
211 do {
212 /* Chill out and ensure we re-read multi_stop_state. */
213 stop_machine_yield(cpumask);
214 newstate = READ_ONCE(msdata->state);
215 if (newstate != curstate) {
216 curstate = newstate;
217 switch (curstate) {
218 case MULTI_STOP_DISABLE_IRQ:
219 local_irq_disable();
220 hard_irq_disable();
221 break;
222 case MULTI_STOP_RUN:
223 if (is_active)
224 err = msdata->fn(msdata->data);
225 break;
226 default:
227 break;
228 }
229 ack_state(msdata);
230 } else if (curstate > MULTI_STOP_PREPARE) {
231 /*
232 * At this stage all other CPUs we depend on must spin
233 * in the same loop. Any reason for hard-lockup should
234 * be detected and reported on their side.
235 */
236 touch_nmi_watchdog();
237 }
238 } while (curstate != MULTI_STOP_EXIT);
239
240 local_irq_restore(flags);
241 return err;
242}
243
244static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
245 int cpu2, struct cpu_stop_work *work2)
246{
247 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
248 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
249 DEFINE_WAKE_Q(wakeq);
250 int err;
251
252retry:
253 /*
254 * The waking up of stopper threads has to happen in the same
255 * scheduling context as the queueing. Otherwise, there is a
256 * possibility of one of the above stoppers being woken up by another
257 * CPU, and preempting us. This will cause us to not wake up the other
258 * stopper forever.
259 */
260 preempt_disable();
261 raw_spin_lock_irq(&stopper1->lock);
262 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
263
264 if (!stopper1->enabled || !stopper2->enabled) {
265 err = -ENOENT;
266 goto unlock;
267 }
268
269 /*
270 * Ensure that if we race with __stop_cpus() the stoppers won't get
271 * queued up in reverse order leading to system deadlock.
272 *
273 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
274 * queued a work on cpu1 but not on cpu2, we hold both locks.
275 *
276 * It can be falsely true but it is safe to spin until it is cleared,
277 * queue_stop_cpus_work() does everything under preempt_disable().
278 */
279 if (unlikely(stop_cpus_in_progress)) {
280 err = -EDEADLK;
281 goto unlock;
282 }
283
284 err = 0;
285 __cpu_stop_queue_work(stopper1, work1, &wakeq);
286 __cpu_stop_queue_work(stopper2, work2, &wakeq);
287
288unlock:
289 raw_spin_unlock(&stopper2->lock);
290 raw_spin_unlock_irq(&stopper1->lock);
291
292 if (unlikely(err == -EDEADLK)) {
293 preempt_enable();
294
295 while (stop_cpus_in_progress)
296 cpu_relax();
297
298 goto retry;
299 }
300
301 wake_up_q(&wakeq);
302 preempt_enable();
303
304 return err;
305}
306/**
307 * stop_two_cpus - stops two cpus
308 * @cpu1: the cpu to stop
309 * @cpu2: the other cpu to stop
310 * @fn: function to execute
311 * @arg: argument to @fn
312 *
313 * Stops both the current and specified CPU and runs @fn on one of them.
314 *
315 * returns when both are completed.
316 */
317int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
318{
319 struct cpu_stop_done done;
320 struct cpu_stop_work work1, work2;
321 struct multi_stop_data msdata;
322
323 msdata = (struct multi_stop_data){
324 .fn = fn,
325 .data = arg,
326 .num_threads = 2,
327 .active_cpus = cpumask_of(cpu1),
328 };
329
330 work1 = work2 = (struct cpu_stop_work){
331 .fn = multi_cpu_stop,
332 .arg = &msdata,
333 .done = &done
334 };
335
336 cpu_stop_init_done(&done, 2);
337 set_state(&msdata, MULTI_STOP_PREPARE);
338
339 if (cpu1 > cpu2)
340 swap(cpu1, cpu2);
341 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
342 return -ENOENT;
343
344 wait_for_completion(&done.completion);
345 return done.ret;
346}
347
348/**
349 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
350 * @cpu: cpu to stop
351 * @fn: function to execute
352 * @arg: argument to @fn
353 * @work_buf: pointer to cpu_stop_work structure
354 *
355 * Similar to stop_one_cpu() but doesn't wait for completion. The
356 * caller is responsible for ensuring @work_buf is currently unused
357 * and will remain untouched until stopper starts executing @fn.
358 *
359 * CONTEXT:
360 * Don't care.
361 *
362 * RETURNS:
363 * true if cpu_stop_work was queued successfully and @fn will be called,
364 * false otherwise.
365 */
366bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
367 struct cpu_stop_work *work_buf)
368{
369 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
370 return cpu_stop_queue_work(cpu, work_buf);
371}
372
373static bool queue_stop_cpus_work(const struct cpumask *cpumask,
374 cpu_stop_fn_t fn, void *arg,
375 struct cpu_stop_done *done)
376{
377 struct cpu_stop_work *work;
378 unsigned int cpu;
379 bool queued = false;
380
381 /*
382 * Disable preemption while queueing to avoid getting
383 * preempted by a stopper which might wait for other stoppers
384 * to enter @fn which can lead to deadlock.
385 */
386 preempt_disable();
387 stop_cpus_in_progress = true;
388 barrier();
389 for_each_cpu(cpu, cpumask) {
390 work = &per_cpu(cpu_stopper.stop_work, cpu);
391 work->fn = fn;
392 work->arg = arg;
393 work->done = done;
394 if (cpu_stop_queue_work(cpu, work))
395 queued = true;
396 }
397 barrier();
398 stop_cpus_in_progress = false;
399 preempt_enable();
400
401 return queued;
402}
403
404static int __stop_cpus(const struct cpumask *cpumask,
405 cpu_stop_fn_t fn, void *arg)
406{
407 struct cpu_stop_done done;
408
409 cpu_stop_init_done(&done, cpumask_weight(cpumask));
410 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
411 return -ENOENT;
412 wait_for_completion(&done.completion);
413 return done.ret;
414}
415
416/**
417 * stop_cpus - stop multiple cpus
418 * @cpumask: cpus to stop
419 * @fn: function to execute
420 * @arg: argument to @fn
421 *
422 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
423 * @fn is run in a process context with the highest priority
424 * preempting any task on the cpu and monopolizing it. This function
425 * returns after all executions are complete.
426 *
427 * This function doesn't guarantee the cpus in @cpumask stay online
428 * till @fn completes. If some cpus go down in the middle, execution
429 * on the cpu may happen partially or fully on different cpus. @fn
430 * should either be ready for that or the caller should ensure that
431 * the cpus stay online until this function completes.
432 *
433 * All stop_cpus() calls are serialized making it safe for @fn to wait
434 * for all cpus to start executing it.
435 *
436 * CONTEXT:
437 * Might sleep.
438 *
439 * RETURNS:
440 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
441 * @cpumask were offline; otherwise, 0 if all executions of @fn
442 * returned 0, any non zero return value if any returned non zero.
443 */
444int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
445{
446 int ret;
447
448 /* static works are used, process one request at a time */
449 mutex_lock(&stop_cpus_mutex);
450 ret = __stop_cpus(cpumask, fn, arg);
451 mutex_unlock(&stop_cpus_mutex);
452 return ret;
453}
454
455/**
456 * try_stop_cpus - try to stop multiple cpus
457 * @cpumask: cpus to stop
458 * @fn: function to execute
459 * @arg: argument to @fn
460 *
461 * Identical to stop_cpus() except that it fails with -EAGAIN if
462 * someone else is already using the facility.
463 *
464 * CONTEXT:
465 * Might sleep.
466 *
467 * RETURNS:
468 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
469 * @fn(@arg) was not executed at all because all cpus in @cpumask were
470 * offline; otherwise, 0 if all executions of @fn returned 0, any non
471 * zero return value if any returned non zero.
472 */
473int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
474{
475 int ret;
476
477 /* static works are used, process one request at a time */
478 if (!mutex_trylock(&stop_cpus_mutex))
479 return -EAGAIN;
480 ret = __stop_cpus(cpumask, fn, arg);
481 mutex_unlock(&stop_cpus_mutex);
482 return ret;
483}
484
485static int cpu_stop_should_run(unsigned int cpu)
486{
487 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
488 unsigned long flags;
489 int run;
490
491 raw_spin_lock_irqsave(&stopper->lock, flags);
492 run = !list_empty(&stopper->works);
493 raw_spin_unlock_irqrestore(&stopper->lock, flags);
494 return run;
495}
496
497static void cpu_stopper_thread(unsigned int cpu)
498{
499 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
500 struct cpu_stop_work *work;
501
502repeat:
503 work = NULL;
504 raw_spin_lock_irq(&stopper->lock);
505 if (!list_empty(&stopper->works)) {
506 work = list_first_entry(&stopper->works,
507 struct cpu_stop_work, list);
508 list_del_init(&work->list);
509 }
510 raw_spin_unlock_irq(&stopper->lock);
511
512 if (work) {
513 cpu_stop_fn_t fn = work->fn;
514 void *arg = work->arg;
515 struct cpu_stop_done *done = work->done;
516 int ret;
517
518 /* cpu stop callbacks must not sleep, make in_atomic() == T */
519 preempt_count_inc();
520 ret = fn(arg);
521 if (done) {
522 if (ret)
523 done->ret = ret;
524 cpu_stop_signal_done(done);
525 }
526 preempt_count_dec();
527 WARN_ONCE(preempt_count(),
528 "cpu_stop: %ps(%p) leaked preempt count\n", fn, arg);
529 goto repeat;
530 }
531}
532
533void stop_machine_park(int cpu)
534{
535 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
536 /*
537 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
538 * the pending works before it parks, until then it is fine to queue
539 * the new works.
540 */
541 stopper->enabled = false;
542 kthread_park(stopper->thread);
543}
544
545extern void sched_set_stop_task(int cpu, struct task_struct *stop);
546
547static void cpu_stop_create(unsigned int cpu)
548{
549 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
550}
551
552static void cpu_stop_park(unsigned int cpu)
553{
554 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
555
556 WARN_ON(!list_empty(&stopper->works));
557}
558
559void stop_machine_unpark(int cpu)
560{
561 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
562
563 stopper->enabled = true;
564 kthread_unpark(stopper->thread);
565}
566
567static struct smp_hotplug_thread cpu_stop_threads = {
568 .store = &cpu_stopper.thread,
569 .thread_should_run = cpu_stop_should_run,
570 .thread_fn = cpu_stopper_thread,
571 .thread_comm = "migration/%u",
572 .create = cpu_stop_create,
573 .park = cpu_stop_park,
574 .selfparking = true,
575};
576
577static int __init cpu_stop_init(void)
578{
579 unsigned int cpu;
580
581 for_each_possible_cpu(cpu) {
582 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
583
584 raw_spin_lock_init(&stopper->lock);
585 INIT_LIST_HEAD(&stopper->works);
586 }
587
588 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
589 stop_machine_unpark(raw_smp_processor_id());
590 stop_machine_initialized = true;
591 return 0;
592}
593early_initcall(cpu_stop_init);
594
595int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
596 const struct cpumask *cpus)
597{
598 struct multi_stop_data msdata = {
599 .fn = fn,
600 .data = data,
601 .num_threads = num_online_cpus(),
602 .active_cpus = cpus,
603 };
604
605 lockdep_assert_cpus_held();
606
607 if (!stop_machine_initialized) {
608 /*
609 * Handle the case where stop_machine() is called
610 * early in boot before stop_machine() has been
611 * initialized.
612 */
613 unsigned long flags;
614 int ret;
615
616 WARN_ON_ONCE(msdata.num_threads != 1);
617
618 local_irq_save(flags);
619 hard_irq_disable();
620 ret = (*fn)(data);
621 local_irq_restore(flags);
622
623 return ret;
624 }
625
626 /* Set the initial state and stop all online cpus. */
627 set_state(&msdata, MULTI_STOP_PREPARE);
628 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
629}
630
631int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
632{
633 int ret;
634
635 /* No CPUs can come up or down during this. */
636 cpus_read_lock();
637 ret = stop_machine_cpuslocked(fn, data, cpus);
638 cpus_read_unlock();
639 return ret;
640}
641EXPORT_SYMBOL_GPL(stop_machine);
642
643/**
644 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
645 * @fn: the function to run
646 * @data: the data ptr for the @fn()
647 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
648 *
649 * This is identical to stop_machine() but can be called from a CPU which
650 * is not active. The local CPU is in the process of hotplug (so no other
651 * CPU hotplug can start) and not marked active and doesn't have enough
652 * context to sleep.
653 *
654 * This function provides stop_machine() functionality for such state by
655 * using busy-wait for synchronization and executing @fn directly for local
656 * CPU.
657 *
658 * CONTEXT:
659 * Local CPU is inactive. Temporarily stops all active CPUs.
660 *
661 * RETURNS:
662 * 0 if all executions of @fn returned 0, any non zero return value if any
663 * returned non zero.
664 */
665int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
666 const struct cpumask *cpus)
667{
668 struct multi_stop_data msdata = { .fn = fn, .data = data,
669 .active_cpus = cpus };
670 struct cpu_stop_done done;
671 int ret;
672
673 /* Local CPU must be inactive and CPU hotplug in progress. */
674 BUG_ON(cpu_active(raw_smp_processor_id()));
675 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
676
677 /* No proper task established and can't sleep - busy wait for lock. */
678 while (!mutex_trylock(&stop_cpus_mutex))
679 cpu_relax();
680
681 /* Schedule work on other CPUs and execute directly for local CPU */
682 set_state(&msdata, MULTI_STOP_PREPARE);
683 cpu_stop_init_done(&done, num_active_cpus());
684 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
685 &done);
686 ret = multi_cpu_stop(&msdata);
687
688 /* Busy wait for completion. */
689 while (!completion_done(&done.completion))
690 cpu_relax();
691
692 mutex_unlock(&stop_cpus_mutex);
693 return ret ?: done.ret;
694}
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/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>
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);
44
45static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
46{
47 memset(done, 0, sizeof(*done));
48 atomic_set(&done->nr_todo, nr_todo);
49 init_completion(&done->completion);
50}
51
52/* signal completion unless @done is NULL */
53static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
54{
55 if (done) {
56 if (executed)
57 done->executed = true;
58 if (atomic_dec_and_test(&done->nr_todo))
59 complete(&done->completion);
60 }
61}
62
63/* queue @work to @stopper. if offline, @work is completed immediately */
64static void cpu_stop_queue_work(struct cpu_stopper *stopper,
65 struct cpu_stop_work *work)
66{
67 unsigned long flags;
68
69 spin_lock_irqsave(&stopper->lock, flags);
70
71 if (stopper->enabled) {
72 list_add_tail(&work->list, &stopper->works);
73 wake_up_process(stopper->thread);
74 } else
75 cpu_stop_signal_done(work->done, false);
76
77 spin_unlock_irqrestore(&stopper->lock, flags);
78}
79
80/**
81 * stop_one_cpu - stop a cpu
82 * @cpu: cpu to stop
83 * @fn: function to execute
84 * @arg: argument to @fn
85 *
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
89 * complete.
90 *
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.
96 *
97 * CONTEXT:
98 * Might sleep.
99 *
100 * RETURNS:
101 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
102 * otherwise, the return value of @fn.
103 */
104int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
105{
106 struct cpu_stop_done done;
107 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
108
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;
113}
114
115/**
116 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
117 * @cpu: cpu to stop
118 * @fn: function to execute
119 * @arg: argument to @fn
120 *
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.
124 *
125 * CONTEXT:
126 * Don't care.
127 */
128void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
129 struct cpu_stop_work *work_buf)
130{
131 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
132 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
133}
134
135/* static data for stop_cpus */
136static DEFINE_MUTEX(stop_cpus_mutex);
137static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
138
139static void queue_stop_cpus_work(const struct cpumask *cpumask,
140 cpu_stop_fn_t fn, void *arg,
141 struct cpu_stop_done *done)
142{
143 struct cpu_stop_work *work;
144 unsigned int cpu;
145
146 /* initialize works and done */
147 for_each_cpu(cpu, cpumask) {
148 work = &per_cpu(stop_cpus_work, cpu);
149 work->fn = fn;
150 work->arg = arg;
151 work->done = done;
152 }
153
154 /*
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.
158 */
159 preempt_disable();
160 for_each_cpu(cpu, cpumask)
161 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
162 &per_cpu(stop_cpus_work, cpu));
163 preempt_enable();
164}
165
166static int __stop_cpus(const struct cpumask *cpumask,
167 cpu_stop_fn_t fn, void *arg)
168{
169 struct cpu_stop_done done;
170
171 cpu_stop_init_done(&done, cpumask_weight(cpumask));
172 queue_stop_cpus_work(cpumask, fn, arg, &done);
173 wait_for_completion(&done.completion);
174 return done.executed ? done.ret : -ENOENT;
175}
176
177/**
178 * stop_cpus - stop multiple cpus
179 * @cpumask: cpus to stop
180 * @fn: function to execute
181 * @arg: argument to @fn
182 *
183 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
184 * @fn is run in a process context with the highest priority
185 * preempting any task on the cpu and monopolizing it. This function
186 * returns after all executions are complete.
187 *
188 * This function doesn't guarantee the cpus in @cpumask stay online
189 * till @fn completes. If some cpus go down in the middle, execution
190 * on the cpu may happen partially or fully on different cpus. @fn
191 * should either be ready for that or the caller should ensure that
192 * the cpus stay online until this function completes.
193 *
194 * All stop_cpus() calls are serialized making it safe for @fn to wait
195 * for all cpus to start executing it.
196 *
197 * CONTEXT:
198 * Might sleep.
199 *
200 * RETURNS:
201 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
202 * @cpumask were offline; otherwise, 0 if all executions of @fn
203 * returned 0, any non zero return value if any returned non zero.
204 */
205int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
206{
207 int ret;
208
209 /* static works are used, process one request at a time */
210 mutex_lock(&stop_cpus_mutex);
211 ret = __stop_cpus(cpumask, fn, arg);
212 mutex_unlock(&stop_cpus_mutex);
213 return ret;
214}
215
216/**
217 * try_stop_cpus - try to stop multiple cpus
218 * @cpumask: cpus to stop
219 * @fn: function to execute
220 * @arg: argument to @fn
221 *
222 * Identical to stop_cpus() except that it fails with -EAGAIN if
223 * someone else is already using the facility.
224 *
225 * CONTEXT:
226 * Might sleep.
227 *
228 * RETURNS:
229 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
230 * @fn(@arg) was not executed at all because all cpus in @cpumask were
231 * offline; otherwise, 0 if all executions of @fn returned 0, any non
232 * zero return value if any returned non zero.
233 */
234int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
235{
236 int ret;
237
238 /* static works are used, process one request at a time */
239 if (!mutex_trylock(&stop_cpus_mutex))
240 return -EAGAIN;
241 ret = __stop_cpus(cpumask, fn, arg);
242 mutex_unlock(&stop_cpus_mutex);
243 return ret;
244}
245
246static int cpu_stopper_thread(void *data)
247{
248 struct cpu_stopper *stopper = data;
249 struct cpu_stop_work *work;
250 int ret;
251
252repeat:
253 set_current_state(TASK_INTERRUPTIBLE); /* mb paired w/ kthread_stop */
254
255 if (kthread_should_stop()) {
256 __set_current_state(TASK_RUNNING);
257 return 0;
258 }
259
260 work = NULL;
261 spin_lock_irq(&stopper->lock);
262 if (!list_empty(&stopper->works)) {
263 work = list_first_entry(&stopper->works,
264 struct cpu_stop_work, list);
265 list_del_init(&work->list);
266 }
267 spin_unlock_irq(&stopper->lock);
268
269 if (work) {
270 cpu_stop_fn_t fn = work->fn;
271 void *arg = work->arg;
272 struct cpu_stop_done *done = work->done;
273 char ksym_buf[KSYM_NAME_LEN] __maybe_unused;
274
275 __set_current_state(TASK_RUNNING);
276
277 /* cpu stop callbacks are not allowed to sleep */
278 preempt_disable();
279
280 ret = fn(arg);
281 if (ret)
282 done->ret = ret;
283
284 /* restore preemption and check it's still balanced */
285 preempt_enable();
286 WARN_ONCE(preempt_count(),
287 "cpu_stop: %s(%p) leaked preempt count\n",
288 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
289 ksym_buf), arg);
290
291 cpu_stop_signal_done(done, true);
292 } else
293 schedule();
294
295 goto repeat;
296}
297
298extern void sched_set_stop_task(int cpu, struct task_struct *stop);
299
300/* manage stopper for a cpu, mostly lifted from sched migration thread mgmt */
301static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
302 unsigned long action, void *hcpu)
303{
304 unsigned int cpu = (unsigned long)hcpu;
305 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
306 struct task_struct *p;
307
308 switch (action & ~CPU_TASKS_FROZEN) {
309 case CPU_UP_PREPARE:
310 BUG_ON(stopper->thread || stopper->enabled ||
311 !list_empty(&stopper->works));
312 p = kthread_create_on_node(cpu_stopper_thread,
313 stopper,
314 cpu_to_node(cpu),
315 "migration/%d", cpu);
316 if (IS_ERR(p))
317 return notifier_from_errno(PTR_ERR(p));
318 get_task_struct(p);
319 kthread_bind(p, cpu);
320 sched_set_stop_task(cpu, p);
321 stopper->thread = p;
322 break;
323
324 case CPU_ONLINE:
325 /* strictly unnecessary, as first user will wake it */
326 wake_up_process(stopper->thread);
327 /* mark enabled */
328 spin_lock_irq(&stopper->lock);
329 stopper->enabled = true;
330 spin_unlock_irq(&stopper->lock);
331 break;
332
333#ifdef CONFIG_HOTPLUG_CPU
334 case CPU_UP_CANCELED:
335 case CPU_POST_DEAD:
336 {
337 struct cpu_stop_work *work;
338
339 sched_set_stop_task(cpu, NULL);
340 /* kill the stopper */
341 kthread_stop(stopper->thread);
342 /* drain remaining works */
343 spin_lock_irq(&stopper->lock);
344 list_for_each_entry(work, &stopper->works, list)
345 cpu_stop_signal_done(work->done, false);
346 stopper->enabled = false;
347 spin_unlock_irq(&stopper->lock);
348 /* release the stopper */
349 put_task_struct(stopper->thread);
350 stopper->thread = NULL;
351 break;
352 }
353#endif
354 }
355
356 return NOTIFY_OK;
357}
358
359/*
360 * Give it a higher priority so that cpu stopper is available to other
361 * cpu notifiers. It currently shares the same priority as sched
362 * migration_notifier.
363 */
364static struct notifier_block __cpuinitdata cpu_stop_cpu_notifier = {
365 .notifier_call = cpu_stop_cpu_callback,
366 .priority = 10,
367};
368
369static int __init cpu_stop_init(void)
370{
371 void *bcpu = (void *)(long)smp_processor_id();
372 unsigned int cpu;
373 int err;
374
375 for_each_possible_cpu(cpu) {
376 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
377
378 spin_lock_init(&stopper->lock);
379 INIT_LIST_HEAD(&stopper->works);
380 }
381
382 /* start one for the boot cpu */
383 err = cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_UP_PREPARE,
384 bcpu);
385 BUG_ON(err != NOTIFY_OK);
386 cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_ONLINE, bcpu);
387 register_cpu_notifier(&cpu_stop_cpu_notifier);
388
389 return 0;
390}
391early_initcall(cpu_stop_init);
392
393#ifdef CONFIG_STOP_MACHINE
394
395/* This controls the threads on each CPU. */
396enum stopmachine_state {
397 /* Dummy starting state for thread. */
398 STOPMACHINE_NONE,
399 /* Awaiting everyone to be scheduled. */
400 STOPMACHINE_PREPARE,
401 /* Disable interrupts. */
402 STOPMACHINE_DISABLE_IRQ,
403 /* Run the function */
404 STOPMACHINE_RUN,
405 /* Exit */
406 STOPMACHINE_EXIT,
407};
408
409struct stop_machine_data {
410 int (*fn)(void *);
411 void *data;
412 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
413 unsigned int num_threads;
414 const struct cpumask *active_cpus;
415
416 enum stopmachine_state state;
417 atomic_t thread_ack;
418};
419
420static void set_state(struct stop_machine_data *smdata,
421 enum stopmachine_state newstate)
422{
423 /* Reset ack counter. */
424 atomic_set(&smdata->thread_ack, smdata->num_threads);
425 smp_wmb();
426 smdata->state = newstate;
427}
428
429/* Last one to ack a state moves to the next state. */
430static void ack_state(struct stop_machine_data *smdata)
431{
432 if (atomic_dec_and_test(&smdata->thread_ack))
433 set_state(smdata, smdata->state + 1);
434}
435
436/* This is the cpu_stop function which stops the CPU. */
437static int stop_machine_cpu_stop(void *data)
438{
439 struct stop_machine_data *smdata = data;
440 enum stopmachine_state curstate = STOPMACHINE_NONE;
441 int cpu = smp_processor_id(), err = 0;
442 unsigned long flags;
443 bool is_active;
444
445 /*
446 * When called from stop_machine_from_inactive_cpu(), irq might
447 * already be disabled. Save the state and restore it on exit.
448 */
449 local_save_flags(flags);
450
451 if (!smdata->active_cpus)
452 is_active = cpu == cpumask_first(cpu_online_mask);
453 else
454 is_active = cpumask_test_cpu(cpu, smdata->active_cpus);
455
456 /* Simple state machine */
457 do {
458 /* Chill out and ensure we re-read stopmachine_state. */
459 cpu_relax();
460 if (smdata->state != curstate) {
461 curstate = smdata->state;
462 switch (curstate) {
463 case STOPMACHINE_DISABLE_IRQ:
464 local_irq_disable();
465 hard_irq_disable();
466 break;
467 case STOPMACHINE_RUN:
468 if (is_active)
469 err = smdata->fn(smdata->data);
470 break;
471 default:
472 break;
473 }
474 ack_state(smdata);
475 }
476 } while (curstate != STOPMACHINE_EXIT);
477
478 local_irq_restore(flags);
479 return err;
480}
481
482int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
483{
484 struct stop_machine_data smdata = { .fn = fn, .data = data,
485 .num_threads = num_online_cpus(),
486 .active_cpus = cpus };
487
488 /* Set the initial state and stop all online cpus. */
489 set_state(&smdata, STOPMACHINE_PREPARE);
490 return stop_cpus(cpu_online_mask, stop_machine_cpu_stop, &smdata);
491}
492
493int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
494{
495 int ret;
496
497 /* No CPUs can come up or down during this. */
498 get_online_cpus();
499 ret = __stop_machine(fn, data, cpus);
500 put_online_cpus();
501 return ret;
502}
503EXPORT_SYMBOL_GPL(stop_machine);
504
505/**
506 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
507 * @fn: the function to run
508 * @data: the data ptr for the @fn()
509 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
510 *
511 * This is identical to stop_machine() but can be called from a CPU which
512 * is not active. The local CPU is in the process of hotplug (so no other
513 * CPU hotplug can start) and not marked active and doesn't have enough
514 * context to sleep.
515 *
516 * This function provides stop_machine() functionality for such state by
517 * using busy-wait for synchronization and executing @fn directly for local
518 * CPU.
519 *
520 * CONTEXT:
521 * Local CPU is inactive. Temporarily stops all active CPUs.
522 *
523 * RETURNS:
524 * 0 if all executions of @fn returned 0, any non zero return value if any
525 * returned non zero.
526 */
527int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
528 const struct cpumask *cpus)
529{
530 struct stop_machine_data smdata = { .fn = fn, .data = data,
531 .active_cpus = cpus };
532 struct cpu_stop_done done;
533 int ret;
534
535 /* Local CPU must be inactive and CPU hotplug in progress. */
536 BUG_ON(cpu_active(raw_smp_processor_id()));
537 smdata.num_threads = num_active_cpus() + 1; /* +1 for local */
538
539 /* No proper task established and can't sleep - busy wait for lock. */
540 while (!mutex_trylock(&stop_cpus_mutex))
541 cpu_relax();
542
543 /* Schedule work on other CPUs and execute directly for local CPU */
544 set_state(&smdata, STOPMACHINE_PREPARE);
545 cpu_stop_init_done(&done, num_active_cpus());
546 queue_stop_cpus_work(cpu_active_mask, stop_machine_cpu_stop, &smdata,
547 &done);
548 ret = stop_machine_cpu_stop(&smdata);
549
550 /* Busy wait for completion. */
551 while (!completion_done(&done.completion))
552 cpu_relax();
553
554 mutex_unlock(&stop_cpus_mutex);
555 return ret ?: done.ret;
556}
557
558#endif /* CONFIG_STOP_MACHINE */