projects/linux·blob v5.12-rc1 · refs

686 lines · 17.8 KiB

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 [email protected]
7 * Copyright (C) 2010 SUSE Linux Products GmbH
8 * Copyright (C) 2010 Tejun Heo <[email protected]>
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 unsigned long caller;
46 cpu_stop_fn_t fn;
47};
48
49static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
50static bool stop_machine_initialized = false;
51
52void print_stop_info(const char *log_lvl, struct task_struct *task)
53{
54 /*
55 * If @task is a stopper task, it cannot migrate and task_cpu() is
56 * stable.
57 */
58 struct cpu_stopper *stopper = per_cpu_ptr(&cpu_stopper, task_cpu(task));
59
60 if (task != stopper->thread)
61 return;
62
63 printk("%sStopper: %pS <- %pS\n", log_lvl, stopper->fn, (void *)stopper->caller);
64}
65
66/* static data for stop_cpus */
67static DEFINE_MUTEX(stop_cpus_mutex);
68static bool stop_cpus_in_progress;
69
70static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
71{
72 memset(done, 0, sizeof(*done));
73 atomic_set(&done->nr_todo, nr_todo);
74 init_completion(&done->completion);
75}
76
77/* signal completion unless @done is NULL */
78static void cpu_stop_signal_done(struct cpu_stop_done *done)
79{
80 if (atomic_dec_and_test(&done->nr_todo))
81 complete(&done->completion);
82}
83
84static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
85 struct cpu_stop_work *work,
86 struct wake_q_head *wakeq)
87{
88 list_add_tail(&work->list, &stopper->works);
89 wake_q_add(wakeq, stopper->thread);
90}
91
92/* queue @work to @stopper. if offline, @work is completed immediately */
93static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
94{
95 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
96 DEFINE_WAKE_Q(wakeq);
97 unsigned long flags;
98 bool enabled;
99
100 preempt_disable();
101 raw_spin_lock_irqsave(&stopper->lock, flags);
102 enabled = stopper->enabled;
103 if (enabled)
104 __cpu_stop_queue_work(stopper, work, &wakeq);
105 else if (work->done)
106 cpu_stop_signal_done(work->done);
107 raw_spin_unlock_irqrestore(&stopper->lock, flags);
108
109 wake_up_q(&wakeq);
110 preempt_enable();
111
112 return enabled;
113}
114
115/**
116 * stop_one_cpu - stop a cpu
117 * @cpu: cpu to stop
118 * @fn: function to execute
119 * @arg: argument to @fn
120 *
121 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
122 * the highest priority preempting any task on the cpu and
123 * monopolizing it. This function returns after the execution is
124 * complete.
125 *
126 * This function doesn't guarantee @cpu stays online till @fn
127 * completes. If @cpu goes down in the middle, execution may happen
128 * partially or fully on different cpus. @fn should either be ready
129 * for that or the caller should ensure that @cpu stays online until
130 * this function completes.
131 *
132 * CONTEXT:
133 * Might sleep.
134 *
135 * RETURNS:
136 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
137 * otherwise, the return value of @fn.
138 */
139int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
140{
141 struct cpu_stop_done done;
142 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done, .caller = _RET_IP_ };
143
144 cpu_stop_init_done(&done, 1);
145 if (!cpu_stop_queue_work(cpu, &work))
146 return -ENOENT;
147 /*
148 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
149 * cycle by doing a preemption:
150 */
151 cond_resched();
152 wait_for_completion(&done.completion);
153 return done.ret;
154}
155
156/* This controls the threads on each CPU. */
157enum multi_stop_state {
158 /* Dummy starting state for thread. */
159 MULTI_STOP_NONE,
160 /* Awaiting everyone to be scheduled. */
161 MULTI_STOP_PREPARE,
162 /* Disable interrupts. */
163 MULTI_STOP_DISABLE_IRQ,
164 /* Run the function */
165 MULTI_STOP_RUN,
166 /* Exit */
167 MULTI_STOP_EXIT,
168};
169
170struct multi_stop_data {
171 cpu_stop_fn_t fn;
172 void *data;
173 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
174 unsigned int num_threads;
175 const struct cpumask *active_cpus;
176
177 enum multi_stop_state state;
178 atomic_t thread_ack;
179};
180
181static void set_state(struct multi_stop_data *msdata,
182 enum multi_stop_state newstate)
183{
184 /* Reset ack counter. */
185 atomic_set(&msdata->thread_ack, msdata->num_threads);
186 smp_wmb();
187 WRITE_ONCE(msdata->state, newstate);
188}
189
190/* Last one to ack a state moves to the next state. */
191static void ack_state(struct multi_stop_data *msdata)
192{
193 if (atomic_dec_and_test(&msdata->thread_ack))
194 set_state(msdata, msdata->state + 1);
195}
196
197notrace void __weak stop_machine_yield(const struct cpumask *cpumask)
198{
199 cpu_relax();
200}
201
202/* This is the cpu_stop function which stops the CPU. */
203static int multi_cpu_stop(void *data)
204{
205 struct multi_stop_data *msdata = data;
206 enum multi_stop_state newstate, curstate = MULTI_STOP_NONE;
207 int cpu = smp_processor_id(), err = 0;
208 const struct cpumask *cpumask;
209 unsigned long flags;
210 bool is_active;
211
212 /*
213 * When called from stop_machine_from_inactive_cpu(), irq might
214 * already be disabled. Save the state and restore it on exit.
215 */
216 local_save_flags(flags);
217
218 if (!msdata->active_cpus) {
219 cpumask = cpu_online_mask;
220 is_active = cpu == cpumask_first(cpumask);
221 } else {
222 cpumask = msdata->active_cpus;
223 is_active = cpumask_test_cpu(cpu, cpumask);
224 }
225
226 /* Simple state machine */
227 do {
228 /* Chill out and ensure we re-read multi_stop_state. */
229 stop_machine_yield(cpumask);
230 newstate = READ_ONCE(msdata->state);
231 if (newstate != curstate) {
232 curstate = newstate;
233 switch (curstate) {
234 case MULTI_STOP_DISABLE_IRQ:
235 local_irq_disable();
236 hard_irq_disable();
237 break;
238 case MULTI_STOP_RUN:
239 if (is_active)
240 err = msdata->fn(msdata->data);
241 break;
242 default:
243 break;
244 }
245 ack_state(msdata);
246 } else if (curstate > MULTI_STOP_PREPARE) {
247 /*
248 * At this stage all other CPUs we depend on must spin
249 * in the same loop. Any reason for hard-lockup should
250 * be detected and reported on their side.
251 */
252 touch_nmi_watchdog();
253 }
254 rcu_momentary_dyntick_idle();
255 } while (curstate != MULTI_STOP_EXIT);
256
257 local_irq_restore(flags);
258 return err;
259}
260
261static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
262 int cpu2, struct cpu_stop_work *work2)
263{
264 struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
265 struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
266 DEFINE_WAKE_Q(wakeq);
267 int err;
268
269retry:
270 /*
271 * The waking up of stopper threads has to happen in the same
272 * scheduling context as the queueing. Otherwise, there is a
273 * possibility of one of the above stoppers being woken up by another
274 * CPU, and preempting us. This will cause us to not wake up the other
275 * stopper forever.
276 */
277 preempt_disable();
278 raw_spin_lock_irq(&stopper1->lock);
279 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
280
281 if (!stopper1->enabled || !stopper2->enabled) {
282 err = -ENOENT;
283 goto unlock;
284 }
285
286 /*
287 * Ensure that if we race with __stop_cpus() the stoppers won't get
288 * queued up in reverse order leading to system deadlock.
289 *
290 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
291 * queued a work on cpu1 but not on cpu2, we hold both locks.
292 *
293 * It can be falsely true but it is safe to spin until it is cleared,
294 * queue_stop_cpus_work() does everything under preempt_disable().
295 */
296 if (unlikely(stop_cpus_in_progress)) {
297 err = -EDEADLK;
298 goto unlock;
299 }
300
301 err = 0;
302 __cpu_stop_queue_work(stopper1, work1, &wakeq);
303 __cpu_stop_queue_work(stopper2, work2, &wakeq);
304
305unlock:
306 raw_spin_unlock(&stopper2->lock);
307 raw_spin_unlock_irq(&stopper1->lock);
308
309 if (unlikely(err == -EDEADLK)) {
310 preempt_enable();
311
312 while (stop_cpus_in_progress)
313 cpu_relax();
314
315 goto retry;
316 }
317
318 wake_up_q(&wakeq);
319 preempt_enable();
320
321 return err;
322}
323/**
324 * stop_two_cpus - stops two cpus
325 * @cpu1: the cpu to stop
326 * @cpu2: the other cpu to stop
327 * @fn: function to execute
328 * @arg: argument to @fn
329 *
330 * Stops both the current and specified CPU and runs @fn on one of them.
331 *
332 * returns when both are completed.
333 */
334int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
335{
336 struct cpu_stop_done done;
337 struct cpu_stop_work work1, work2;
338 struct multi_stop_data msdata;
339
340 msdata = (struct multi_stop_data){
341 .fn = fn,
342 .data = arg,
343 .num_threads = 2,
344 .active_cpus = cpumask_of(cpu1),
345 };
346
347 work1 = work2 = (struct cpu_stop_work){
348 .fn = multi_cpu_stop,
349 .arg = &msdata,
350 .done = &done,
351 .caller = _RET_IP_,
352 };
353
354 cpu_stop_init_done(&done, 2);
355 set_state(&msdata, MULTI_STOP_PREPARE);
356
357 if (cpu1 > cpu2)
358 swap(cpu1, cpu2);
359 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
360 return -ENOENT;
361
362 wait_for_completion(&done.completion);
363 return done.ret;
364}
365
366/**
367 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
368 * @cpu: cpu to stop
369 * @fn: function to execute
370 * @arg: argument to @fn
371 * @work_buf: pointer to cpu_stop_work structure
372 *
373 * Similar to stop_one_cpu() but doesn't wait for completion. The
374 * caller is responsible for ensuring @work_buf is currently unused
375 * and will remain untouched until stopper starts executing @fn.
376 *
377 * CONTEXT:
378 * Don't care.
379 *
380 * RETURNS:
381 * true if cpu_stop_work was queued successfully and @fn will be called,
382 * false otherwise.
383 */
384bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
385 struct cpu_stop_work *work_buf)
386{
387 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, .caller = _RET_IP_, };
388 return cpu_stop_queue_work(cpu, work_buf);
389}
390
391static bool queue_stop_cpus_work(const struct cpumask *cpumask,
392 cpu_stop_fn_t fn, void *arg,
393 struct cpu_stop_done *done)
394{
395 struct cpu_stop_work *work;
396 unsigned int cpu;
397 bool queued = false;
398
399 /*
400 * Disable preemption while queueing to avoid getting
401 * preempted by a stopper which might wait for other stoppers
402 * to enter @fn which can lead to deadlock.
403 */
404 preempt_disable();
405 stop_cpus_in_progress = true;
406 barrier();
407 for_each_cpu(cpu, cpumask) {
408 work = &per_cpu(cpu_stopper.stop_work, cpu);
409 work->fn = fn;
410 work->arg = arg;
411 work->done = done;
412 if (cpu_stop_queue_work(cpu, work))
413 queued = true;
414 }
415 barrier();
416 stop_cpus_in_progress = false;
417 preempt_enable();
418
419 return queued;
420}
421
422static int __stop_cpus(const struct cpumask *cpumask,
423 cpu_stop_fn_t fn, void *arg)
424{
425 struct cpu_stop_done done;
426
427 cpu_stop_init_done(&done, cpumask_weight(cpumask));
428 if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
429 return -ENOENT;
430 wait_for_completion(&done.completion);
431 return done.ret;
432}
433
434/**
435 * stop_cpus - stop multiple cpus
436 * @cpumask: cpus to stop
437 * @fn: function to execute
438 * @arg: argument to @fn
439 *
440 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
441 * @fn is run in a process context with the highest priority
442 * preempting any task on the cpu and monopolizing it. This function
443 * returns after all executions are complete.
444 *
445 * This function doesn't guarantee the cpus in @cpumask stay online
446 * till @fn completes. If some cpus go down in the middle, execution
447 * on the cpu may happen partially or fully on different cpus. @fn
448 * should either be ready for that or the caller should ensure that
449 * the cpus stay online until this function completes.
450 *
451 * All stop_cpus() calls are serialized making it safe for @fn to wait
452 * for all cpus to start executing it.
453 *
454 * CONTEXT:
455 * Might sleep.
456 *
457 * RETURNS:
458 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
459 * @cpumask were offline; otherwise, 0 if all executions of @fn
460 * returned 0, any non zero return value if any returned non zero.
461 */
462static int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
463{
464 int ret;
465
466 /* static works are used, process one request at a time */
467 mutex_lock(&stop_cpus_mutex);
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 stopper->caller = work->caller;
508 stopper->fn = fn;
509 preempt_count_inc();
510 ret = fn(arg);
511 if (done) {
512 if (ret)
513 done->ret = ret;
514 cpu_stop_signal_done(done);
515 }
516 preempt_count_dec();
517 stopper->fn = NULL;
518 stopper->caller = 0;
519 WARN_ONCE(preempt_count(),
520 "cpu_stop: %ps(%p) leaked preempt count\n", fn, arg);
521 goto repeat;
522 }
523}
524
525void stop_machine_park(int cpu)
526{
527 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
528 /*
529 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
530 * the pending works before it parks, until then it is fine to queue
531 * the new works.
532 */
533 stopper->enabled = false;
534 kthread_park(stopper->thread);
535}
536
537extern void sched_set_stop_task(int cpu, struct task_struct *stop);
538
539static void cpu_stop_create(unsigned int cpu)
540{
541 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
542}
543
544static void cpu_stop_park(unsigned int cpu)
545{
546 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
547
548 WARN_ON(!list_empty(&stopper->works));
549}
550
551void stop_machine_unpark(int cpu)
552{
553 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
554
555 stopper->enabled = true;
556 kthread_unpark(stopper->thread);
557}
558
559static struct smp_hotplug_thread cpu_stop_threads = {
560 .store = &cpu_stopper.thread,
561 .thread_should_run = cpu_stop_should_run,
562 .thread_fn = cpu_stopper_thread,
563 .thread_comm = "migration/%u",
564 .create = cpu_stop_create,
565 .park = cpu_stop_park,
566 .selfparking = true,
567};
568
569static int __init cpu_stop_init(void)
570{
571 unsigned int cpu;
572
573 for_each_possible_cpu(cpu) {
574 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
575
576 raw_spin_lock_init(&stopper->lock);
577 INIT_LIST_HEAD(&stopper->works);
578 }
579
580 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
581 stop_machine_unpark(raw_smp_processor_id());
582 stop_machine_initialized = true;
583 return 0;
584}
585early_initcall(cpu_stop_init);
586
587int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
588 const struct cpumask *cpus)
589{
590 struct multi_stop_data msdata = {
591 .fn = fn,
592 .data = data,
593 .num_threads = num_online_cpus(),
594 .active_cpus = cpus,
595 };
596
597 lockdep_assert_cpus_held();
598
599 if (!stop_machine_initialized) {
600 /*
601 * Handle the case where stop_machine() is called
602 * early in boot before stop_machine() has been
603 * initialized.
604 */
605 unsigned long flags;
606 int ret;
607
608 WARN_ON_ONCE(msdata.num_threads != 1);
609
610 local_irq_save(flags);
611 hard_irq_disable();
612 ret = (*fn)(data);
613 local_irq_restore(flags);
614
615 return ret;
616 }
617
618 /* Set the initial state and stop all online cpus. */
619 set_state(&msdata, MULTI_STOP_PREPARE);
620 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
621}
622
623int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
624{
625 int ret;
626
627 /* No CPUs can come up or down during this. */
628 cpus_read_lock();
629 ret = stop_machine_cpuslocked(fn, data, cpus);
630 cpus_read_unlock();
631 return ret;
632}
633EXPORT_SYMBOL_GPL(stop_machine);
634
635/**
636 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
637 * @fn: the function to run
638 * @data: the data ptr for the @fn()
639 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
640 *
641 * This is identical to stop_machine() but can be called from a CPU which
642 * is not active. The local CPU is in the process of hotplug (so no other
643 * CPU hotplug can start) and not marked active and doesn't have enough
644 * context to sleep.
645 *
646 * This function provides stop_machine() functionality for such state by
647 * using busy-wait for synchronization and executing @fn directly for local
648 * CPU.
649 *
650 * CONTEXT:
651 * Local CPU is inactive. Temporarily stops all active CPUs.
652 *
653 * RETURNS:
654 * 0 if all executions of @fn returned 0, any non zero return value if any
655 * returned non zero.
656 */
657int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
658 const struct cpumask *cpus)
659{
660 struct multi_stop_data msdata = { .fn = fn, .data = data,
661 .active_cpus = cpus };
662 struct cpu_stop_done done;
663 int ret;
664
665 /* Local CPU must be inactive and CPU hotplug in progress. */
666 BUG_ON(cpu_active(raw_smp_processor_id()));
667 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
668
669 /* No proper task established and can't sleep - busy wait for lock. */
670 while (!mutex_trylock(&stop_cpus_mutex))
671 cpu_relax();
672
673 /* Schedule work on other CPUs and execute directly for local CPU */
674 set_state(&msdata, MULTI_STOP_PREPARE);
675 cpu_stop_init_done(&done, num_active_cpus());
676 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
677 &done);
678 ret = multi_cpu_stop(&msdata);
679
680 /* Busy wait for completion. */
681 while (!completion_done(&done.completion))
682 cpu_relax();
683
684 mutex_unlock(&stop_cpus_mutex);
685 return ret ?: done.ret;
686}