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 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 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
181/* This is the cpu_stop function which stops the CPU. */
182static int multi_cpu_stop(void *data)
183{
184 struct multi_stop_data *msdata = data;
185 enum multi_stop_state curstate = MULTI_STOP_NONE;
186 int cpu = smp_processor_id(), err = 0;
187 unsigned long flags;
188 bool is_active;
189
190 /*
191 * When called from stop_machine_from_inactive_cpu(), irq might
192 * already be disabled. Save the state and restore it on exit.
193 */
194 local_save_flags(flags);
195
196 if (!msdata->active_cpus)
197 is_active = cpu == cpumask_first(cpu_online_mask);
198 else
199 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
200
201 /* Simple state machine */
202 do {
203 /* Chill out and ensure we re-read multi_stop_state. */
204 cpu_relax_yield();
205 if (msdata->state != curstate) {
206 curstate = msdata->state;
207 switch (curstate) {
208 case MULTI_STOP_DISABLE_IRQ:
209 local_irq_disable();
210 hard_irq_disable();
211 break;
212 case MULTI_STOP_RUN:
213 if (is_active)
214 err = msdata->fn(msdata->data);
215 break;
216 default:
217 break;
218 }
219 ack_state(msdata);
220 } else if (curstate > MULTI_STOP_PREPARE) {
221 /*
222 * At this stage all other CPUs we depend on must spin
223 * in the same loop. Any reason for hard-lockup should
224 * be detected and reported on their side.
225 */
226 touch_nmi_watchdog();
227 }
228 } while (curstate != MULTI_STOP_EXIT);
229
230 local_irq_restore(flags);
231 return err;
232}
233
234static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
235 int cpu2, struct cpu_stop_work *work2)
236{
237 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
238 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
239 DEFINE_WAKE_Q(wakeq);
240 int err;
241
242retry:
243 /*
244 * The waking up of stopper threads has to happen in the same
245 * scheduling context as the queueing. Otherwise, there is a
246 * possibility of one of the above stoppers being woken up by another
247 * CPU, and preempting us. This will cause us to not wake up the other
248 * stopper forever.
249 */
250 preempt_disable();
251 raw_spin_lock_irq(&stopper1->lock);
252 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
253
254 if (!stopper1->enabled || !stopper2->enabled) {
255 err = -ENOENT;
256 goto unlock;
257 }
258
259 /*
260 * Ensure that if we race with __stop_cpus() the stoppers won't get
261 * queued up in reverse order leading to system deadlock.
262 *
263 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
264 * queued a work on cpu1 but not on cpu2, we hold both locks.
265 *
266 * It can be falsely true but it is safe to spin until it is cleared,
267 * queue_stop_cpus_work() does everything under preempt_disable().
268 */
269 if (unlikely(stop_cpus_in_progress)) {
270 err = -EDEADLK;
271 goto unlock;
272 }
273
274 err = 0;
275 __cpu_stop_queue_work(stopper1, work1, &wakeq);
276 __cpu_stop_queue_work(stopper2, work2, &wakeq);
277
278unlock:
279 raw_spin_unlock(&stopper2->lock);
280 raw_spin_unlock_irq(&stopper1->lock);
281
282 if (unlikely(err == -EDEADLK)) {
283 preempt_enable();
284
285 while (stop_cpus_in_progress)
286 cpu_relax();
287
288 goto retry;
289 }
290
291 wake_up_q(&wakeq);
292 preempt_enable();
293
294 return err;
295}
296/**
297 * stop_two_cpus - stops two cpus
298 * @cpu1: the cpu to stop
299 * @cpu2: the other cpu to stop
300 * @fn: function to execute
301 * @arg: argument to @fn
302 *
303 * Stops both the current and specified CPU and runs @fn on one of them.
304 *
305 * returns when both are completed.
306 */
307int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
308{
309 struct cpu_stop_done done;
310 struct cpu_stop_work work1, work2;
311 struct multi_stop_data msdata;
312
313 msdata = (struct multi_stop_data){
314 .fn = fn,
315 .data = arg,
316 .num_threads = 2,
317 .active_cpus = cpumask_of(cpu1),
318 };
319
320 work1 = work2 = (struct cpu_stop_work){
321 .fn = multi_cpu_stop,
322 .arg = &msdata,
323 .done = &done
324 };
325
326 cpu_stop_init_done(&done, 2);
327 set_state(&msdata, MULTI_STOP_PREPARE);
328
329 if (cpu1 > cpu2)
330 swap(cpu1, cpu2);
331 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
332 return -ENOENT;
333
334 wait_for_completion(&done.completion);
335 return done.ret;
336}
337
338/**
339 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
340 * @cpu: cpu to stop
341 * @fn: function to execute
342 * @arg: argument to @fn
343 * @work_buf: pointer to cpu_stop_work structure
344 *
345 * Similar to stop_one_cpu() but doesn't wait for completion. The
346 * caller is responsible for ensuring @work_buf is currently unused
347 * and will remain untouched until stopper starts executing @fn.
348 *
349 * CONTEXT:
350 * Don't care.
351 *
352 * RETURNS:
353 * true if cpu_stop_work was queued successfully and @fn will be called,
354 * false otherwise.
355 */
356bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
357 struct cpu_stop_work *work_buf)
358{
359 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
360 return cpu_stop_queue_work(cpu, work_buf);
361}
362
363static bool queue_stop_cpus_work(const struct cpumask *cpumask,
364 cpu_stop_fn_t fn, void *arg,
365 struct cpu_stop_done *done)
366{
367 struct cpu_stop_work *work;
368 unsigned int cpu;
369 bool queued = false;
370
371 /*
372 * Disable preemption while queueing to avoid getting
373 * preempted by a stopper which might wait for other stoppers
374 * to enter @fn which can lead to deadlock.
375 */
376 preempt_disable();
377 stop_cpus_in_progress = true;
378 for_each_cpu(cpu, cpumask) {
379 work = &per_cpu(cpu_stopper.stop_work, cpu);
380 work->fn = fn;
381 work->arg = arg;
382 work->done = done;
383 if (cpu_stop_queue_work(cpu, work))
384 queued = true;
385 }
386 stop_cpus_in_progress = false;
387 preempt_enable();
388
389 return queued;
390}
391
392static int __stop_cpus(const struct cpumask *cpumask,
393 cpu_stop_fn_t fn, void *arg)
394{
395 struct cpu_stop_done done;
396
397 cpu_stop_init_done(&done, cpumask_weight(cpumask));
398 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
399 return -ENOENT;
400 wait_for_completion(&done.completion);
401 return done.ret;
402}
403
404/**
405 * stop_cpus - stop multiple cpus
406 * @cpumask: cpus to stop
407 * @fn: function to execute
408 * @arg: argument to @fn
409 *
410 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
411 * @fn is run in a process context with the highest priority
412 * preempting any task on the cpu and monopolizing it. This function
413 * returns after all executions are complete.
414 *
415 * This function doesn't guarantee the cpus in @cpumask stay online
416 * till @fn completes. If some cpus go down in the middle, execution
417 * on the cpu may happen partially or fully on different cpus. @fn
418 * should either be ready for that or the caller should ensure that
419 * the cpus stay online until this function completes.
420 *
421 * All stop_cpus() calls are serialized making it safe for @fn to wait
422 * for all cpus to start executing it.
423 *
424 * CONTEXT:
425 * Might sleep.
426 *
427 * RETURNS:
428 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
429 * @cpumask were offline; otherwise, 0 if all executions of @fn
430 * returned 0, any non zero return value if any returned non zero.
431 */
432int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
433{
434 int ret;
435
436 /* static works are used, process one request at a time */
437 mutex_lock(&stop_cpus_mutex);
438 ret = __stop_cpus(cpumask, fn, arg);
439 mutex_unlock(&stop_cpus_mutex);
440 return ret;
441}
442
443/**
444 * try_stop_cpus - try to stop multiple cpus
445 * @cpumask: cpus to stop
446 * @fn: function to execute
447 * @arg: argument to @fn
448 *
449 * Identical to stop_cpus() except that it fails with -EAGAIN if
450 * someone else is already using the facility.
451 *
452 * CONTEXT:
453 * Might sleep.
454 *
455 * RETURNS:
456 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
457 * @fn(@arg) was not executed at all because all cpus in @cpumask were
458 * offline; otherwise, 0 if all executions of @fn returned 0, any non
459 * zero return value if any returned non zero.
460 */
461int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
462{
463 int ret;
464
465 /* static works are used, process one request at a time */
466 if (!mutex_trylock(&stop_cpus_mutex))
467 return -EAGAIN;
468 ret = __stop_cpus(cpumask, fn, arg);
469 mutex_unlock(&stop_cpus_mutex);
470 return ret;
471}
472
473static int cpu_stop_should_run(unsigned int cpu)
474{
475 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
476 unsigned long flags;
477 int run;
478
479 raw_spin_lock_irqsave(&stopper->lock, flags);
480 run = !list_empty(&stopper->works);
481 raw_spin_unlock_irqrestore(&stopper->lock, flags);
482 return run;
483}
484
485static void cpu_stopper_thread(unsigned int cpu)
486{
487 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
488 struct cpu_stop_work *work;
489
490repeat:
491 work = NULL;
492 raw_spin_lock_irq(&stopper->lock);
493 if (!list_empty(&stopper->works)) {
494 work = list_first_entry(&stopper->works,
495 struct cpu_stop_work, list);
496 list_del_init(&work->list);
497 }
498 raw_spin_unlock_irq(&stopper->lock);
499
500 if (work) {
501 cpu_stop_fn_t fn = work->fn;
502 void *arg = work->arg;
503 struct cpu_stop_done *done = work->done;
504 int ret;
505
506 /* cpu stop callbacks must not sleep, make in_atomic() == T */
507 preempt_count_inc();
508 ret = fn(arg);
509 if (done) {
510 if (ret)
511 done->ret = ret;
512 cpu_stop_signal_done(done);
513 }
514 preempt_count_dec();
515 WARN_ONCE(preempt_count(),
516 "cpu_stop: %pf(%p) leaked preempt count\n", fn, arg);
517 goto repeat;
518 }
519}
520
521void stop_machine_park(int cpu)
522{
523 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
524 /*
525 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
526 * the pending works before it parks, until then it is fine to queue
527 * the new works.
528 */
529 stopper->enabled = false;
530 kthread_park(stopper->thread);
531}
532
533extern void sched_set_stop_task(int cpu, struct task_struct *stop);
534
535static void cpu_stop_create(unsigned int cpu)
536{
537 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
538}
539
540static void cpu_stop_park(unsigned int cpu)
541{
542 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
543
544 WARN_ON(!list_empty(&stopper->works));
545}
546
547void stop_machine_unpark(int cpu)
548{
549 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
550
551 stopper->enabled = true;
552 kthread_unpark(stopper->thread);
553}
554
555static struct smp_hotplug_thread cpu_stop_threads = {
556 .store = &cpu_stopper.thread,
557 .thread_should_run = cpu_stop_should_run,
558 .thread_fn = cpu_stopper_thread,
559 .thread_comm = "migration/%u",
560 .create = cpu_stop_create,
561 .park = cpu_stop_park,
562 .selfparking = true,
563};
564
565static int __init cpu_stop_init(void)
566{
567 unsigned int cpu;
568
569 for_each_possible_cpu(cpu) {
570 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
571
572 raw_spin_lock_init(&stopper->lock);
573 INIT_LIST_HEAD(&stopper->works);
574 }
575
576 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
577 stop_machine_unpark(raw_smp_processor_id());
578 stop_machine_initialized = true;
579 return 0;
580}
581early_initcall(cpu_stop_init);
582
583int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
584 const struct cpumask *cpus)
585{
586 struct multi_stop_data msdata = {
587 .fn = fn,
588 .data = data,
589 .num_threads = num_online_cpus(),
590 .active_cpus = cpus,
591 };
592
593 lockdep_assert_cpus_held();
594
595 if (!stop_machine_initialized) {
596 /*
597 * Handle the case where stop_machine() is called
598 * early in boot before stop_machine() has been
599 * initialized.
600 */
601 unsigned long flags;
602 int ret;
603
604 WARN_ON_ONCE(msdata.num_threads != 1);
605
606 local_irq_save(flags);
607 hard_irq_disable();
608 ret = (*fn)(data);
609 local_irq_restore(flags);
610
611 return ret;
612 }
613
614 /* Set the initial state and stop all online cpus. */
615 set_state(&msdata, MULTI_STOP_PREPARE);
616 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
617}
618
619int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
620{
621 int ret;
622
623 /* No CPUs can come up or down during this. */
624 cpus_read_lock();
625 ret = stop_machine_cpuslocked(fn, data, cpus);
626 cpus_read_unlock();
627 return ret;
628}
629EXPORT_SYMBOL_GPL(stop_machine);
630
631/**
632 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
633 * @fn: the function to run
634 * @data: the data ptr for the @fn()
635 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
636 *
637 * This is identical to stop_machine() but can be called from a CPU which
638 * is not active. The local CPU is in the process of hotplug (so no other
639 * CPU hotplug can start) and not marked active and doesn't have enough
640 * context to sleep.
641 *
642 * This function provides stop_machine() functionality for such state by
643 * using busy-wait for synchronization and executing @fn directly for local
644 * CPU.
645 *
646 * CONTEXT:
647 * Local CPU is inactive. Temporarily stops all active CPUs.
648 *
649 * RETURNS:
650 * 0 if all executions of @fn returned 0, any non zero return value if any
651 * returned non zero.
652 */
653int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
654 const struct cpumask *cpus)
655{
656 struct multi_stop_data msdata = { .fn = fn, .data = data,
657 .active_cpus = cpus };
658 struct cpu_stop_done done;
659 int ret;
660
661 /* Local CPU must be inactive and CPU hotplug in progress. */
662 BUG_ON(cpu_active(raw_smp_processor_id()));
663 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
664
665 /* No proper task established and can't sleep - busy wait for lock. */
666 while (!mutex_trylock(&stop_cpus_mutex))
667 cpu_relax();
668
669 /* Schedule work on other CPUs and execute directly for local CPU */
670 set_state(&msdata, MULTI_STOP_PREPARE);
671 cpu_stop_init_done(&done, num_active_cpus());
672 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
673 &done);
674 ret = multi_cpu_stop(&msdata);
675
676 /* Busy wait for completion. */
677 while (!completion_done(&done.completion))
678 cpu_relax();
679
680 mutex_unlock(&stop_cpus_mutex);
681 return ret ?: done.ret;
682}
683