blob: e98c70b7f0738a3cad2bb7e22695af3c13044e77 [file] [log] [blame]
lh9ed821d2023-04-07 01:36:19 -07001/*
2 * kernel/stop_machine.c
3 *
4 * Copyright (C) 2008, 2005 IBM Corporation.
5 * Copyright (C) 2008, 2005 Rusty Russell rusty@rustcorp.com.au
6 * Copyright (C) 2010 SUSE Linux Products GmbH
7 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
8 *
9 * This file is released under the GPLv2 and any later version.
10 */
11#include <linux/completion.h>
12#include <linux/cpu.h>
13#include <linux/init.h>
14#include <linux/kthread.h>
15#include <linux/export.h>
16#include <linux/percpu.h>
17#include <linux/sched.h>
18#include <linux/stop_machine.h>
19#include <linux/interrupt.h>
20#include <linux/kallsyms.h>
21
22#include <linux/atomic.h>
23
24/*
25 * Structure to determine completion condition and record errors. May
26 * be shared by works on different cpus.
27 */
28struct cpu_stop_done {
29 atomic_t nr_todo; /* nr left to execute */
30 bool executed; /* actually executed? */
31 int ret; /* collected return value */
32 struct task_struct *waiter; /* woken when nr_todo reaches 0 */
33};
34
35/* the actual stopper, one per every possible cpu, enabled on online cpus */
36struct cpu_stopper {
37 raw_spinlock_t lock;
38 bool enabled; /* is this stopper enabled? */
39 struct list_head works; /* list of pending works */
40 struct task_struct *thread; /* stopper thread */
41};
42
43static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
44static bool stop_machine_initialized = false;
45
46static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
47{
48 memset(done, 0, sizeof(*done));
49 atomic_set(&done->nr_todo, nr_todo);
50 done->waiter = current;
51}
52
53/* signal completion unless @done is NULL */
54static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
55{
56 if (done) {
57 if (executed)
58 done->executed = true;
59 if (atomic_dec_and_test(&done->nr_todo)) {
60 wake_up_process(done->waiter);
61 done->waiter = NULL;
62 }
63 }
64}
65
66/* queue @work to @stopper. if offline, @work is completed immediately */
67static void cpu_stop_queue_work(struct cpu_stopper *stopper,
68 struct cpu_stop_work *work)
69{
70 unsigned long flags;
71
72 raw_spin_lock_irqsave(&stopper->lock, flags);
73
74 if (stopper->enabled) {
75 list_add_tail(&work->list, &stopper->works);
76 wake_up_process(stopper->thread);
77 } else
78 cpu_stop_signal_done(work->done, false);
79
80 raw_spin_unlock_irqrestore(&stopper->lock, flags);
81}
82
83static void wait_for_stop_done(struct cpu_stop_done *done)
84{
85 set_current_state(TASK_UNINTERRUPTIBLE);
86 while (atomic_read(&done->nr_todo)) {
87 schedule();
88 set_current_state(TASK_UNINTERRUPTIBLE);
89 }
90 /*
91 * We need to wait until cpu_stop_signal_done() has cleared
92 * done->waiter.
93 */
94 while (done->waiter)
95 cpu_relax();
96 set_current_state(TASK_RUNNING);
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 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), &work);
130 wait_for_stop_done(&done);
131 return done.executed ? done.ret : -ENOENT;
132}
133
134/**
135 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
136 * @cpu: cpu to stop
137 * @fn: function to execute
138 * @arg: argument to @fn
139 *
140 * Similar to stop_one_cpu() but doesn't wait for completion. The
141 * caller is responsible for ensuring @work_buf is currently unused
142 * and will remain untouched until stopper starts executing @fn.
143 *
144 * CONTEXT:
145 * Don't care.
146 */
147void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
148 struct cpu_stop_work *work_buf)
149{
150 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
151 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu), work_buf);
152}
153
154/* static data for stop_cpus */
155static DEFINE_MUTEX(stop_cpus_mutex);
156static DEFINE_MUTEX(stopper_lock);
157static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
158
159static void queue_stop_cpus_work(const struct cpumask *cpumask,
160 cpu_stop_fn_t fn, void *arg,
161 struct cpu_stop_done *done, bool inactive)
162{
163 struct cpu_stop_work *work;
164 unsigned int cpu;
165
166 /* initialize works and done */
167 for_each_cpu(cpu, cpumask) {
168 work = &per_cpu(stop_cpus_work, cpu);
169 work->fn = fn;
170 work->arg = arg;
171 work->done = done;
172 }
173
174 /*
175 * Make sure that all work is queued on all cpus before we
176 * any of the cpus can execute it.
177 */
178 if (!inactive) {
179 mutex_lock(&stopper_lock);
180 } else {
181 while (!mutex_trylock(&stopper_lock))
182 cpu_relax();
183 }
184 for_each_cpu(cpu, cpumask)
185 cpu_stop_queue_work(&per_cpu(cpu_stopper, cpu),
186 &per_cpu(stop_cpus_work, cpu));
187 mutex_unlock(&stopper_lock);
188}
189
190static int __stop_cpus(const struct cpumask *cpumask,
191 cpu_stop_fn_t fn, void *arg)
192{
193 struct cpu_stop_done done;
194
195 cpu_stop_init_done(&done, cpumask_weight(cpumask));
196 queue_stop_cpus_work(cpumask, fn, arg, &done, false);
197 wait_for_stop_done(&done);
198 return done.executed ? done.ret : -ENOENT;
199}
200
201/**
202 * stop_cpus - stop multiple cpus
203 * @cpumask: cpus to stop
204 * @fn: function to execute
205 * @arg: argument to @fn
206 *
207 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
208 * @fn is run in a process context with the highest priority
209 * preempting any task on the cpu and monopolizing it. This function
210 * returns after all executions are complete.
211 *
212 * This function doesn't guarantee the cpus in @cpumask stay online
213 * till @fn completes. If some cpus go down in the middle, execution
214 * on the cpu may happen partially or fully on different cpus. @fn
215 * should either be ready for that or the caller should ensure that
216 * the cpus stay online until this function completes.
217 *
218 * All stop_cpus() calls are serialized making it safe for @fn to wait
219 * for all cpus to start executing it.
220 *
221 * CONTEXT:
222 * Might sleep.
223 *
224 * RETURNS:
225 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
226 * @cpumask were offline; otherwise, 0 if all executions of @fn
227 * returned 0, any non zero return value if any returned non zero.
228 */
229int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
230{
231 int ret;
232
233 /* static works are used, process one request at a time */
234 mutex_lock(&stop_cpus_mutex);
235 ret = __stop_cpus(cpumask, fn, arg);
236 mutex_unlock(&stop_cpus_mutex);
237 return ret;
238}
239
240/**
241 * try_stop_cpus - try to stop multiple cpus
242 * @cpumask: cpus to stop
243 * @fn: function to execute
244 * @arg: argument to @fn
245 *
246 * Identical to stop_cpus() except that it fails with -EAGAIN if
247 * someone else is already using the facility.
248 *
249 * CONTEXT:
250 * Might sleep.
251 *
252 * RETURNS:
253 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
254 * @fn(@arg) was not executed at all because all cpus in @cpumask were
255 * offline; otherwise, 0 if all executions of @fn returned 0, any non
256 * zero return value if any returned non zero.
257 */
258int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
259{
260 int ret;
261
262 /* static works are used, process one request at a time */
263 if (!mutex_trylock(&stop_cpus_mutex))
264 return -EAGAIN;
265 ret = __stop_cpus(cpumask, fn, arg);
266 mutex_unlock(&stop_cpus_mutex);
267 return ret;
268}
269
270static int cpu_stopper_thread(void *data)
271{
272 struct cpu_stopper *stopper = data;
273 struct cpu_stop_work *work;
274 int ret;
275
276repeat:
277 set_current_state(TASK_INTERRUPTIBLE); /* mb paired w/ kthread_stop */
278
279 if (kthread_should_stop()) {
280 __set_current_state(TASK_RUNNING);
281 return 0;
282 }
283
284 work = NULL;
285 raw_spin_lock_irq(&stopper->lock);
286 if (!list_empty(&stopper->works)) {
287 work = list_first_entry(&stopper->works,
288 struct cpu_stop_work, list);
289 list_del_init(&work->list);
290 }
291 raw_spin_unlock_irq(&stopper->lock);
292
293 if (work) {
294 cpu_stop_fn_t fn = work->fn;
295 void *arg = work->arg;
296 struct cpu_stop_done *done = work->done;
297 char ksym_buf[KSYM_NAME_LEN] __maybe_unused;
298
299 __set_current_state(TASK_RUNNING);
300
301 /*
302 * Wait until the stopper finished scheduling on all
303 * cpus
304 */
305 mutex_lock(&stopper_lock);
306 /*
307 * Let other cpu threads continue as well
308 */
309 mutex_unlock(&stopper_lock);
310
311 /* cpu stop callbacks are not allowed to sleep */
312 preempt_disable();
313
314 ret = fn(arg);
315 if (ret)
316 done->ret = ret;
317
318 /* restore preemption and check it's still balanced */
319 preempt_enable();
320 WARN_ONCE(preempt_count(),
321 "cpu_stop: %s(%p) leaked preempt count\n",
322 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
323 ksym_buf), arg);
324
325 /*
326 * Make sure that the wakeup and setting done->waiter
327 * to NULL is atomic.
328 */
329 local_irq_disable();
330 cpu_stop_signal_done(done, true);
331 local_irq_enable();
332 } else
333 schedule();
334
335 goto repeat;
336}
337
338extern void sched_set_stop_task(int cpu, struct task_struct *stop);
339
340/* manage stopper for a cpu, mostly lifted from sched migration thread mgmt */
341static int __cpuinit cpu_stop_cpu_callback(struct notifier_block *nfb,
342 unsigned long action, void *hcpu)
343{
344 unsigned int cpu = (unsigned long)hcpu;
345 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
346 struct task_struct *p;
347
348 switch (action & ~CPU_TASKS_FROZEN) {
349 case CPU_UP_PREPARE:
350 BUG_ON(stopper->thread || stopper->enabled ||
351 !list_empty(&stopper->works));
352 p = kthread_create_on_node(cpu_stopper_thread,
353 stopper,
354 cpu_to_node(cpu),
355 "migration/%d", cpu);
356 if (IS_ERR(p))
357 return notifier_from_errno(PTR_ERR(p));
358 get_task_struct(p);
359 p->flags |= PF_STOMPER;
360 kthread_bind(p, cpu);
361 sched_set_stop_task(cpu, p);
362 stopper->thread = p;
363 break;
364
365 case CPU_ONLINE:
366 /* strictly unnecessary, as first user will wake it */
367 wake_up_process(stopper->thread);
368 /* mark enabled */
369 raw_spin_lock_irq(&stopper->lock);
370 stopper->enabled = true;
371 raw_spin_unlock_irq(&stopper->lock);
372 break;
373
374#ifdef CONFIG_HOTPLUG_CPU
375 case CPU_UP_CANCELED:
376 case CPU_POST_DEAD:
377 {
378 struct cpu_stop_work *work;
379
380 sched_set_stop_task(cpu, NULL);
381 /* kill the stopper */
382 kthread_stop(stopper->thread);
383 /* drain remaining works */
384 raw_spin_lock_irq(&stopper->lock);
385 list_for_each_entry(work, &stopper->works, list)
386 cpu_stop_signal_done(work->done, false);
387 stopper->enabled = false;
388 raw_spin_unlock_irq(&stopper->lock);
389 /* release the stopper */
390 put_task_struct(stopper->thread);
391 stopper->thread = NULL;
392 break;
393 }
394#endif
395 }
396
397 return NOTIFY_OK;
398}
399
400/*
401 * Give it a higher priority so that cpu stopper is available to other
402 * cpu notifiers. It currently shares the same priority as sched
403 * migration_notifier.
404 */
405static struct notifier_block __cpuinitdata cpu_stop_cpu_notifier = {
406 .notifier_call = cpu_stop_cpu_callback,
407 .priority = 10,
408};
409
410static int __init cpu_stop_init(void)
411{
412 void *bcpu = (void *)(long)smp_processor_id();
413 unsigned int cpu;
414 int err;
415
416 for_each_possible_cpu(cpu) {
417 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
418
419 raw_spin_lock_init(&stopper->lock);
420 INIT_LIST_HEAD(&stopper->works);
421 }
422
423 /* start one for the boot cpu */
424 err = cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_UP_PREPARE,
425 bcpu);
426 BUG_ON(err != NOTIFY_OK);
427 cpu_stop_cpu_callback(&cpu_stop_cpu_notifier, CPU_ONLINE, bcpu);
428 register_cpu_notifier(&cpu_stop_cpu_notifier);
429
430 stop_machine_initialized = true;
431
432 return 0;
433}
434early_initcall(cpu_stop_init);
435
436#ifdef CONFIG_STOP_MACHINE
437
438/* This controls the threads on each CPU. */
439enum stopmachine_state {
440 /* Dummy starting state for thread. */
441 STOPMACHINE_NONE,
442 /* Awaiting everyone to be scheduled. */
443 STOPMACHINE_PREPARE,
444 /* Disable interrupts. */
445 STOPMACHINE_DISABLE_IRQ,
446 /* Run the function */
447 STOPMACHINE_RUN,
448 /* Exit */
449 STOPMACHINE_EXIT,
450};
451
452struct stop_machine_data {
453 int (*fn)(void *);
454 void *data;
455 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
456 unsigned int num_threads;
457 const struct cpumask *active_cpus;
458
459 enum stopmachine_state state;
460 atomic_t thread_ack;
461};
462
463static void set_state(struct stop_machine_data *smdata,
464 enum stopmachine_state newstate)
465{
466 /* Reset ack counter. */
467 atomic_set(&smdata->thread_ack, smdata->num_threads);
468 smp_wmb();
469 smdata->state = newstate;
470}
471
472/* Last one to ack a state moves to the next state. */
473static void ack_state(struct stop_machine_data *smdata)
474{
475 if (atomic_dec_and_test(&smdata->thread_ack))
476 set_state(smdata, smdata->state + 1);
477}
478
479/* This is the cpu_stop function which stops the CPU. */
480static int stop_machine_cpu_stop(void *data)
481{
482 struct stop_machine_data *smdata = data;
483 enum stopmachine_state curstate = STOPMACHINE_NONE;
484 int cpu = smp_processor_id(), err = 0;
485 unsigned long flags;
486 bool is_active;
487
488 /*
489 * When called from stop_machine_from_inactive_cpu(), irq might
490 * already be disabled. Save the state and restore it on exit.
491 */
492 local_save_flags(flags);
493
494 if (!smdata->active_cpus)
495 is_active = cpu == cpumask_first(cpu_online_mask);
496 else
497 is_active = cpumask_test_cpu(cpu, smdata->active_cpus);
498
499 /* Simple state machine */
500 do {
501 /* Chill out and ensure we re-read stopmachine_state. */
502 cpu_relax();
503 if (smdata->state != curstate) {
504 curstate = smdata->state;
505 switch (curstate) {
506 case STOPMACHINE_DISABLE_IRQ:
507 local_irq_disable();
508 hard_irq_disable();
509 break;
510 case STOPMACHINE_RUN:
511 if (is_active)
512 err = smdata->fn(smdata->data);
513 break;
514 default:
515 break;
516 }
517 ack_state(smdata);
518 }
519 } while (curstate != STOPMACHINE_EXIT);
520
521 local_irq_restore(flags);
522 return err;
523}
524
525int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
526{
527 struct stop_machine_data smdata = { .fn = fn, .data = data,
528 .num_threads = num_online_cpus(),
529 .active_cpus = cpus };
530
531 if (!stop_machine_initialized) {
532 /*
533 * Handle the case where stop_machine() is called
534 * early in boot before stop_machine() has been
535 * initialized.
536 */
537 unsigned long flags;
538 int ret;
539
540 WARN_ON_ONCE(smdata.num_threads != 1);
541
542 local_irq_save(flags);
543 hard_irq_disable();
544 ret = (*fn)(data);
545 local_irq_restore(flags);
546
547 return ret;
548 }
549
550 /* Set the initial state and stop all online cpus. */
551 set_state(&smdata, STOPMACHINE_PREPARE);
552 return stop_cpus(cpu_online_mask, stop_machine_cpu_stop, &smdata);
553}
554
555int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
556{
557 int ret;
558
559 /* No CPUs can come up or down during this. */
560 get_online_cpus();
561 ret = __stop_machine(fn, data, cpus);
562 put_online_cpus();
563 return ret;
564}
565EXPORT_SYMBOL_GPL(stop_machine);
566
567/**
568 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
569 * @fn: the function to run
570 * @data: the data ptr for the @fn()
571 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
572 *
573 * This is identical to stop_machine() but can be called from a CPU which
574 * is not active. The local CPU is in the process of hotplug (so no other
575 * CPU hotplug can start) and not marked active and doesn't have enough
576 * context to sleep.
577 *
578 * This function provides stop_machine() functionality for such state by
579 * using busy-wait for synchronization and executing @fn directly for local
580 * CPU.
581 *
582 * CONTEXT:
583 * Local CPU is inactive. Temporarily stops all active CPUs.
584 *
585 * RETURNS:
586 * 0 if all executions of @fn returned 0, any non zero return value if any
587 * returned non zero.
588 */
589int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
590 const struct cpumask *cpus)
591{
592 struct stop_machine_data smdata = { .fn = fn, .data = data,
593 .active_cpus = cpus };
594 struct cpu_stop_done done;
595 int ret;
596
597 /* Local CPU must be inactive and CPU hotplug in progress. */
598 BUG_ON(cpu_active(raw_smp_processor_id()));
599 smdata.num_threads = num_active_cpus() + 1; /* +1 for local */
600
601 /* No proper task established and can't sleep - busy wait for lock. */
602 while (!mutex_trylock(&stop_cpus_mutex))
603 cpu_relax();
604
605 /* Schedule work on other CPUs and execute directly for local CPU */
606 set_state(&smdata, STOPMACHINE_PREPARE);
607 cpu_stop_init_done(&done, num_active_cpus());
608 queue_stop_cpus_work(cpu_active_mask, stop_machine_cpu_stop, &smdata,
609 &done, true);
610 ret = stop_machine_cpu_stop(&smdata);
611
612 /* Busy wait for completion. */
613 while (atomic_read(&done.nr_todo))
614 cpu_relax();
615
616 mutex_unlock(&stop_cpus_mutex);
617 return ret ?: done.ret;
618}
619
620#endif /* CONFIG_STOP_MACHINE */