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/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 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 spin_lock_irqsave(&stopper->lock, flags);
85 enabled = stopper->enabled;
86 if (enabled)
87 __cpu_stop_queue_work(stopper, work, &wakeq);
88 else if (work->done)
89 cpu_stop_signal_done(work->done);
90 spin_unlock_irqrestore(&stopper->lock, flags);
91
92 wake_up_q(&wakeq);
93
94 return enabled;
95}
96
97/**
98 * stop_one_cpu - stop a cpu
99 * @cpu: cpu to stop
100 * @fn: function to execute
101 * @arg: argument to @fn
102 *
103 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
104 * the highest priority preempting any task on the cpu and
105 * monopolizing it. This function returns after the execution is
106 * complete.
107 *
108 * This function doesn't guarantee @cpu stays online till @fn
109 * completes. If @cpu goes down in the middle, execution may happen
110 * partially or fully on different cpus. @fn should either be ready
111 * for that or the caller should ensure that @cpu stays online until
112 * this function completes.
113 *
114 * CONTEXT:
115 * Might sleep.
116 *
117 * RETURNS:
118 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
119 * otherwise, the return value of @fn.
120 */
121int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
122{
123 struct cpu_stop_done done;
124 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
125
126 cpu_stop_init_done(&done, 1);
127 if (!cpu_stop_queue_work(cpu, &work))
128 return -ENOENT;
129 /*
130 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
131 * cycle by doing a preemption:
132 */
133 cond_resched();
134 wait_for_completion(&done.completion);
135 return done.ret;
136}
137
138/* This controls the threads on each CPU. */
139enum multi_stop_state {
140 /* Dummy starting state for thread. */
141 MULTI_STOP_NONE,
142 /* Awaiting everyone to be scheduled. */
143 MULTI_STOP_PREPARE,
144 /* Disable interrupts. */
145 MULTI_STOP_DISABLE_IRQ,
146 /* Run the function */
147 MULTI_STOP_RUN,
148 /* Exit */
149 MULTI_STOP_EXIT,
150};
151
152struct multi_stop_data {
153 cpu_stop_fn_t fn;
154 void *data;
155 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
156 unsigned int num_threads;
157 const struct cpumask *active_cpus;
158
159 enum multi_stop_state state;
160 atomic_t thread_ack;
161};
162
163static void set_state(struct multi_stop_data *msdata,
164 enum multi_stop_state newstate)
165{
166 /* Reset ack counter. */
167 atomic_set(&msdata->thread_ack, msdata->num_threads);
168 smp_wmb();
169 msdata->state = newstate;
170}
171
172/* Last one to ack a state moves to the next state. */
173static void ack_state(struct multi_stop_data *msdata)
174{
175 if (atomic_dec_and_test(&msdata->thread_ack))
176 set_state(msdata, msdata->state + 1);
177}
178
179/* This is the cpu_stop function which stops the CPU. */
180static int multi_cpu_stop(void *data)
181{
182 struct multi_stop_data *msdata = data;
183 enum multi_stop_state curstate = MULTI_STOP_NONE;
184 int cpu = smp_processor_id(), err = 0;
185 unsigned long flags;
186 bool is_active;
187
188 /*
189 * When called from stop_machine_from_inactive_cpu(), irq might
190 * already be disabled. Save the state and restore it on exit.
191 */
192 local_save_flags(flags);
193
194 if (!msdata->active_cpus)
195 is_active = cpu == cpumask_first(cpu_online_mask);
196 else
197 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
198
199 /* Simple state machine */
200 do {
201 /* Chill out and ensure we re-read multi_stop_state. */
202 cpu_relax_yield();
203 if (msdata->state != curstate) {
204 curstate = msdata->state;
205 switch (curstate) {
206 case MULTI_STOP_DISABLE_IRQ:
207 local_irq_disable();
208 hard_irq_disable();
209 break;
210 case MULTI_STOP_RUN:
211 if (is_active)
212 err = msdata->fn(msdata->data);
213 break;
214 default:
215 break;
216 }
217 ack_state(msdata);
218 } else if (curstate > MULTI_STOP_PREPARE) {
219 /*
220 * At this stage all other CPUs we depend on must spin
221 * in the same loop. Any reason for hard-lockup should
222 * be detected and reported on their side.
223 */
224 touch_nmi_watchdog();
225 }
226 } while (curstate != MULTI_STOP_EXIT);
227
228 local_irq_restore(flags);
229 return err;
230}
231
232static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
233 int cpu2, struct cpu_stop_work *work2)
234{
235 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
236 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
237 DEFINE_WAKE_Q(wakeq);
238 int err;
239retry:
240 spin_lock_irq(&stopper1->lock);
241 spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
242
243 err = -ENOENT;
244 if (!stopper1->enabled || !stopper2->enabled)
245 goto unlock;
246 /*
247 * Ensure that if we race with __stop_cpus() the stoppers won't get
248 * queued up in reverse order leading to system deadlock.
249 *
250 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
251 * queued a work on cpu1 but not on cpu2, we hold both locks.
252 *
253 * It can be falsely true but it is safe to spin until it is cleared,
254 * queue_stop_cpus_work() does everything under preempt_disable().
255 */
256 err = -EDEADLK;
257 if (unlikely(stop_cpus_in_progress))
258 goto unlock;
259
260 err = 0;
261 __cpu_stop_queue_work(stopper1, work1, &wakeq);
262 __cpu_stop_queue_work(stopper2, work2, &wakeq);
263unlock:
264 spin_unlock(&stopper2->lock);
265 spin_unlock_irq(&stopper1->lock);
266
267 if (unlikely(err == -EDEADLK)) {
268 while (stop_cpus_in_progress)
269 cpu_relax();
270 goto retry;
271 }
272
273 wake_up_q(&wakeq);
274
275 return err;
276}
277/**
278 * stop_two_cpus - stops two cpus
279 * @cpu1: the cpu to stop
280 * @cpu2: the other cpu to stop
281 * @fn: function to execute
282 * @arg: argument to @fn
283 *
284 * Stops both the current and specified CPU and runs @fn on one of them.
285 *
286 * returns when both are completed.
287 */
288int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
289{
290 struct cpu_stop_done done;
291 struct cpu_stop_work work1, work2;
292 struct multi_stop_data msdata;
293
294 msdata = (struct multi_stop_data){
295 .fn = fn,
296 .data = arg,
297 .num_threads = 2,
298 .active_cpus = cpumask_of(cpu1),
299 };
300
301 work1 = work2 = (struct cpu_stop_work){
302 .fn = multi_cpu_stop,
303 .arg = &msdata,
304 .done = &done
305 };
306
307 cpu_stop_init_done(&done, 2);
308 set_state(&msdata, MULTI_STOP_PREPARE);
309
310 if (cpu1 > cpu2)
311 swap(cpu1, cpu2);
312 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
313 return -ENOENT;
314
315 wait_for_completion(&done.completion);
316 return done.ret;
317}
318
319/**
320 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
321 * @cpu: cpu to stop
322 * @fn: function to execute
323 * @arg: argument to @fn
324 * @work_buf: pointer to cpu_stop_work structure
325 *
326 * Similar to stop_one_cpu() but doesn't wait for completion. The
327 * caller is responsible for ensuring @work_buf is currently unused
328 * and will remain untouched until stopper starts executing @fn.
329 *
330 * CONTEXT:
331 * Don't care.
332 *
333 * RETURNS:
334 * true if cpu_stop_work was queued successfully and @fn will be called,
335 * false otherwise.
336 */
337bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
338 struct cpu_stop_work *work_buf)
339{
340 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
341 return cpu_stop_queue_work(cpu, work_buf);
342}
343
344static bool queue_stop_cpus_work(const struct cpumask *cpumask,
345 cpu_stop_fn_t fn, void *arg,
346 struct cpu_stop_done *done)
347{
348 struct cpu_stop_work *work;
349 unsigned int cpu;
350 bool queued = false;
351
352 /*
353 * Disable preemption while queueing to avoid getting
354 * preempted by a stopper which might wait for other stoppers
355 * to enter @fn which can lead to deadlock.
356 */
357 preempt_disable();
358 stop_cpus_in_progress = true;
359 for_each_cpu(cpu, cpumask) {
360 work = &per_cpu(cpu_stopper.stop_work, cpu);
361 work->fn = fn;
362 work->arg = arg;
363 work->done = done;
364 if (cpu_stop_queue_work(cpu, work))
365 queued = true;
366 }
367 stop_cpus_in_progress = false;
368 preempt_enable();
369
370 return queued;
371}
372
373static int __stop_cpus(const struct cpumask *cpumask,
374 cpu_stop_fn_t fn, void *arg)
375{
376 struct cpu_stop_done done;
377
378 cpu_stop_init_done(&done, cpumask_weight(cpumask));
379 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
380 return -ENOENT;
381 wait_for_completion(&done.completion);
382 return done.ret;
383}
384
385/**
386 * stop_cpus - stop multiple cpus
387 * @cpumask: cpus to stop
388 * @fn: function to execute
389 * @arg: argument to @fn
390 *
391 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
392 * @fn is run in a process context with the highest priority
393 * preempting any task on the cpu and monopolizing it. This function
394 * returns after all executions are complete.
395 *
396 * This function doesn't guarantee the cpus in @cpumask stay online
397 * till @fn completes. If some cpus go down in the middle, execution
398 * on the cpu may happen partially or fully on different cpus. @fn
399 * should either be ready for that or the caller should ensure that
400 * the cpus stay online until this function completes.
401 *
402 * All stop_cpus() calls are serialized making it safe for @fn to wait
403 * for all cpus to start executing it.
404 *
405 * CONTEXT:
406 * Might sleep.
407 *
408 * RETURNS:
409 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
410 * @cpumask were offline; otherwise, 0 if all executions of @fn
411 * returned 0, any non zero return value if any returned non zero.
412 */
413int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
414{
415 int ret;
416
417 /* static works are used, process one request at a time */
418 mutex_lock(&stop_cpus_mutex);
419 ret = __stop_cpus(cpumask, fn, arg);
420 mutex_unlock(&stop_cpus_mutex);
421 return ret;
422}
423
424/**
425 * try_stop_cpus - try to stop multiple cpus
426 * @cpumask: cpus to stop
427 * @fn: function to execute
428 * @arg: argument to @fn
429 *
430 * Identical to stop_cpus() except that it fails with -EAGAIN if
431 * someone else is already using the facility.
432 *
433 * CONTEXT:
434 * Might sleep.
435 *
436 * RETURNS:
437 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
438 * @fn(@arg) was not executed at all because all cpus in @cpumask were
439 * offline; otherwise, 0 if all executions of @fn returned 0, any non
440 * zero return value if any returned non zero.
441 */
442int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
443{
444 int ret;
445
446 /* static works are used, process one request at a time */
447 if (!mutex_trylock(&stop_cpus_mutex))
448 return -EAGAIN;
449 ret = __stop_cpus(cpumask, fn, arg);
450 mutex_unlock(&stop_cpus_mutex);
451 return ret;
452}
453
454static int cpu_stop_should_run(unsigned int cpu)
455{
456 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
457 unsigned long flags;
458 int run;
459
460 spin_lock_irqsave(&stopper->lock, flags);
461 run = !list_empty(&stopper->works);
462 spin_unlock_irqrestore(&stopper->lock, flags);
463 return run;
464}
465
466static void cpu_stopper_thread(unsigned int cpu)
467{
468 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
469 struct cpu_stop_work *work;
470
471repeat:
472 work = NULL;
473 spin_lock_irq(&stopper->lock);
474 if (!list_empty(&stopper->works)) {
475 work = list_first_entry(&stopper->works,
476 struct cpu_stop_work, list);
477 list_del_init(&work->list);
478 }
479 spin_unlock_irq(&stopper->lock);
480
481 if (work) {
482 cpu_stop_fn_t fn = work->fn;
483 void *arg = work->arg;
484 struct cpu_stop_done *done = work->done;
485 int ret;
486
487 /* cpu stop callbacks must not sleep, make in_atomic() == T */
488 preempt_count_inc();
489 ret = fn(arg);
490 if (done) {
491 if (ret)
492 done->ret = ret;
493 cpu_stop_signal_done(done);
494 }
495 preempt_count_dec();
496 WARN_ONCE(preempt_count(),
497 "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
498 goto repeat;
499 }
500}
501
502void stop_machine_park(int cpu)
503{
504 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
505 /*
506 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
507 * the pending works before it parks, until then it is fine to queue
508 * the new works.
509 */
510 stopper->enabled = false;
511 kthread_park(stopper->thread);
512}
513
514extern void sched_set_stop_task(int cpu, struct task_struct *stop);
515
516static void cpu_stop_create(unsigned int cpu)
517{
518 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
519}
520
521static void cpu_stop_park(unsigned int cpu)
522{
523 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
524
525 WARN_ON(!list_empty(&stopper->works));
526}
527
528void stop_machine_unpark(int cpu)
529{
530 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
531
532 stopper->enabled = true;
533 kthread_unpark(stopper->thread);
534}
535
536static struct smp_hotplug_thread cpu_stop_threads = {
537 .store = &cpu_stopper.thread,
538 .thread_should_run = cpu_stop_should_run,
539 .thread_fn = cpu_stopper_thread,
540 .thread_comm = "migration/%u",
541 .create = cpu_stop_create,
542 .park = cpu_stop_park,
543 .selfparking = true,
544};
545
546static int __init cpu_stop_init(void)
547{
548 unsigned int cpu;
549
550 for_each_possible_cpu(cpu) {
551 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
552
553 spin_lock_init(&stopper->lock);
554 INIT_LIST_HEAD(&stopper->works);
555 }
556
557 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
558 stop_machine_unpark(raw_smp_processor_id());
559 stop_machine_initialized = true;
560 return 0;
561}
562early_initcall(cpu_stop_init);
563
564int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
565 const struct cpumask *cpus)
566{
567 struct multi_stop_data msdata = {
568 .fn = fn,
569 .data = data,
570 .num_threads = num_online_cpus(),
571 .active_cpus = cpus,
572 };
573
574 lockdep_assert_cpus_held();
575
576 if (!stop_machine_initialized) {
577 /*
578 * Handle the case where stop_machine() is called
579 * early in boot before stop_machine() has been
580 * initialized.
581 */
582 unsigned long flags;
583 int ret;
584
585 WARN_ON_ONCE(msdata.num_threads != 1);
586
587 local_irq_save(flags);
588 hard_irq_disable();
589 ret = (*fn)(data);
590 local_irq_restore(flags);
591
592 return ret;
593 }
594
595 /* Set the initial state and stop all online cpus. */
596 set_state(&msdata, MULTI_STOP_PREPARE);
597 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
598}
599
600int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
601{
602 int ret;
603
604 /* No CPUs can come up or down during this. */
605 cpus_read_lock();
606 ret = stop_machine_cpuslocked(fn, data, cpus);
607 cpus_read_unlock();
608 return ret;
609}
610EXPORT_SYMBOL_GPL(stop_machine);
611
612/**
613 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
614 * @fn: the function to run
615 * @data: the data ptr for the @fn()
616 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
617 *
618 * This is identical to stop_machine() but can be called from a CPU which
619 * is not active. The local CPU is in the process of hotplug (so no other
620 * CPU hotplug can start) and not marked active and doesn't have enough
621 * context to sleep.
622 *
623 * This function provides stop_machine() functionality for such state by
624 * using busy-wait for synchronization and executing @fn directly for local
625 * CPU.
626 *
627 * CONTEXT:
628 * Local CPU is inactive. Temporarily stops all active CPUs.
629 *
630 * RETURNS:
631 * 0 if all executions of @fn returned 0, any non zero return value if any
632 * returned non zero.
633 */
634int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
635 const struct cpumask *cpus)
636{
637 struct multi_stop_data msdata = { .fn = fn, .data = data,
638 .active_cpus = cpus };
639 struct cpu_stop_done done;
640 int ret;
641
642 /* Local CPU must be inactive and CPU hotplug in progress. */
643 BUG_ON(cpu_active(raw_smp_processor_id()));
644 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
645
646 /* No proper task established and can't sleep - busy wait for lock. */
647 while (!mutex_trylock(&stop_cpus_mutex))
648 cpu_relax();
649
650 /* Schedule work on other CPUs and execute directly for local CPU */
651 set_state(&msdata, MULTI_STOP_PREPARE);
652 cpu_stop_init_done(&done, num_active_cpus());
653 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
654 &done);
655 ret = multi_cpu_stop(&msdata);
656
657 /* Busy wait for completion. */
658 while (!completion_done(&done.completion))
659 cpu_relax();
660
661 mutex_unlock(&stop_cpus_mutex);
662 return ret ?: done.ret;
663}