projects/linux·blob master · refs

710 lines · 18.4 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/bug.h>
11#include <linux/compiler.h>
12#include <linux/completion.h>
13#include <linux/cpu.h>
14#include <linux/init.h>
15#include <linux/kthread.h>
16#include <linux/export.h>
17#include <linux/percpu.h>
18#include <linux/sched.h>
19#include <linux/stop_machine.h>
20#include <linux/interrupt.h>
21#include <linux/kallsyms.h>
22#include <linux/smpboot.h>
23#include <linux/atomic.h>
24#include <linux/nmi.h>
25#include <linux/sched/wake_q.h>
26
27/*
28 * Structure to determine completion condition and record errors. May
29 * be shared by works on different cpus.
30 */
31struct cpu_stop_done {
32 atomic_t nr_todo; /* nr left to execute */
33 int ret; /* collected return value */
34 struct completion completion; /* fired if nr_todo reaches 0 */
35};
36
37/* the actual stopper, one per every possible cpu, enabled on online cpus */
38struct cpu_stopper {
39 struct task_struct *thread;
40
41 raw_spinlock_t lock;
42 bool enabled; /* is this stopper enabled? */
43 struct list_head works; /* list of pending works */
44
45 struct cpu_stop_work stop_work; /* for stop_cpus */
46 unsigned long caller;
47 cpu_stop_fn_t fn;
48};
49
50static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
51static bool stop_machine_initialized = false;
52
53void print_stop_info(const char *log_lvl, struct task_struct *task)
54{
55 /*
56 * If @task is a stopper task, it cannot migrate and task_cpu() is
57 * stable.
58 */
59 struct cpu_stopper *stopper = per_cpu_ptr(&cpu_stopper, task_cpu(task));
60
61 if (task != stopper->thread)
62 return;
63
64 printk("%sStopper: %pS <- %pS\n", log_lvl, stopper->fn, (void *)stopper->caller);
65}
66
67/* static data for stop_cpus */
68static DEFINE_MUTEX(stop_cpus_mutex);
69static bool stop_cpus_in_progress;
70
71static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
72{
73 memset(done, 0, sizeof(*done));
74 atomic_set(&done->nr_todo, nr_todo);
75 init_completion(&done->completion);
76}
77
78/* signal completion unless @done is NULL */
79static void cpu_stop_signal_done(struct cpu_stop_done *done)
80{
81 if (atomic_dec_and_test(&done->nr_todo))
82 complete(&done->completion);
83}
84
85static void __cpu_stop_queue_work(struct cpu_stopper *stopper,
86 struct cpu_stop_work *work)
87{
88 list_add_tail(&work->list, &stopper->works);
89}
90
91/* queue @work to @stopper. if offline, @work is completed immediately */
92static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
93{
94 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
95 unsigned long flags;
96 bool enabled;
97
98 preempt_disable();
99 raw_spin_lock_irqsave(&stopper->lock, flags);
100 enabled = stopper->enabled;
101 if (enabled)
102 __cpu_stop_queue_work(stopper, work);
103 else if (work->done)
104 cpu_stop_signal_done(work->done);
105 raw_spin_unlock_irqrestore(&stopper->lock, flags);
106
107 if (enabled)
108 wake_up_process(stopper->thread);
109 preempt_enable();
110
111 return enabled;
112}
113
114/**
115 * stop_one_cpu - stop a cpu
116 * @cpu: cpu to stop
117 * @fn: function to execute
118 * @arg: argument to @fn
119 *
120 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
121 * the highest priority preempting any task on the cpu and
122 * monopolizing it. This function returns after the execution is
123 * complete.
124 *
125 * This function doesn't guarantee @cpu stays online till @fn
126 * completes. If @cpu goes down in the middle, execution may happen
127 * partially or fully on different cpus. @fn should either be ready
128 * for that or the caller should ensure that @cpu stays online until
129 * this function completes.
130 *
131 * CONTEXT:
132 * Might sleep.
133 *
134 * RETURNS:
135 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
136 * otherwise, the return value of @fn.
137 */
138int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
139{
140 struct cpu_stop_done done;
141 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done, .caller = _RET_IP_ };
142
143 cpu_stop_init_done(&done, 1);
144 if (!cpu_stop_queue_work(cpu, &work))
145 return -ENOENT;
146 /*
147 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
148 * cycle by doing a preemption:
149 */
150 cond_resched();
151 wait_for_completion(&done.completion);
152 return done.ret;
153}
154
155/* This controls the threads on each CPU. */
156enum multi_stop_state {
157 /* Dummy starting state for thread. */
158 MULTI_STOP_NONE,
159 /* Awaiting everyone to be scheduled. */
160 MULTI_STOP_PREPARE,
161 /* Disable interrupts. */
162 MULTI_STOP_DISABLE_IRQ,
163 /* Run the function */
164 MULTI_STOP_RUN,
165 /* Exit */
166 MULTI_STOP_EXIT,
167};
168
169struct multi_stop_data {
170 cpu_stop_fn_t fn;
171 void *data;
172 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
173 unsigned int num_threads;
174 const struct cpumask *active_cpus;
175
176 enum multi_stop_state state;
177 atomic_t thread_ack;
178};
179
180static void set_state(struct multi_stop_data *msdata,
181 enum multi_stop_state newstate)
182{
183 /* Reset ack counter. */
184 atomic_set(&msdata->thread_ack, msdata->num_threads);
185 smp_wmb();
186 WRITE_ONCE(msdata->state, newstate);
187}
188
189/* Last one to ack a state moves to the next state. */
190static void ack_state(struct multi_stop_data *msdata)
191{
192 if (atomic_dec_and_test(&msdata->thread_ack))
193 set_state(msdata, msdata->state + 1);
194}
195
196notrace void __weak stop_machine_yield(const struct cpumask *cpumask)
197{
198 cpu_relax();
199}
200
201/* This is the cpu_stop function which stops the CPU. */
202static int multi_cpu_stop(void *data)
203{
204 struct multi_stop_data *msdata = data;
205 enum multi_stop_state newstate, curstate = MULTI_STOP_NONE;
206 int cpu = smp_processor_id(), err = 0;
207 const struct cpumask *cpumask;
208 unsigned long flags;
209 bool is_active;
210
211 /*
212 * When called from stop_machine_from_inactive_cpu(), irq might
213 * already be disabled. Save the state and restore it on exit.
214 */
215 local_save_flags(flags);
216
217 if (!msdata->active_cpus) {
218 cpumask = cpu_online_mask;
219 is_active = cpu == cpumask_first(cpumask);
220 } else {
221 cpumask = msdata->active_cpus;
222 is_active = cpumask_test_cpu(cpu, cpumask);
223 }
224
225 /* Simple state machine */
226 do {
227 /* Chill out and ensure we re-read multi_stop_state. */
228 stop_machine_yield(cpumask);
229 newstate = READ_ONCE(msdata->state);
230 if (newstate != curstate) {
231 curstate = newstate;
232 switch (curstate) {
233 case MULTI_STOP_DISABLE_IRQ:
234 local_irq_disable();
235 hard_irq_disable();
236 break;
237 case MULTI_STOP_RUN:
238 if (is_active)
239 err = msdata->fn(msdata->data);
240 break;
241 default:
242 break;
243 }
244 ack_state(msdata);
245 } else if (curstate > MULTI_STOP_PREPARE) {
246 /*
247 * At this stage all other CPUs we depend on must spin
248 * in the same loop. Any reason for hard-lockup should
249 * be detected and reported on their side.
250 */
251 touch_nmi_watchdog();
252 /* Also suppress RCU CPU stall warnings. */
253 rcu_momentary_eqs();
254 }
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 int err;
267
268retry:
269 /*
270 * The waking up of stopper threads has to happen in the same
271 * scheduling context as the queueing. Otherwise, there is a
272 * possibility of one of the above stoppers being woken up by another
273 * CPU, and preempting us. This will cause us to not wake up the other
274 * stopper forever.
275 */
276 preempt_disable();
277 raw_spin_lock_irq(&stopper1->lock);
278 raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
279
280 if (!stopper1->enabled || !stopper2->enabled) {
281 err = -ENOENT;
282 goto unlock;
283 }
284
285 /*
286 * Ensure that if we race with __stop_cpus() the stoppers won't get
287 * queued up in reverse order leading to system deadlock.
288 *
289 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
290 * queued a work on cpu1 but not on cpu2, we hold both locks.
291 *
292 * It can be falsely true but it is safe to spin until it is cleared,
293 * queue_stop_cpus_work() does everything under preempt_disable().
294 */
295 if (unlikely(stop_cpus_in_progress)) {
296 err = -EDEADLK;
297 goto unlock;
298 }
299
300 err = 0;
301 __cpu_stop_queue_work(stopper1, work1);
302 __cpu_stop_queue_work(stopper2, work2);
303
304unlock:
305 raw_spin_unlock(&stopper2->lock);
306 raw_spin_unlock_irq(&stopper1->lock);
307
308 if (unlikely(err == -EDEADLK)) {
309 preempt_enable();
310
311 while (stop_cpus_in_progress)
312 cpu_relax();
313
314 goto retry;
315 }
316
317 if (!err) {
318 wake_up_process(stopper1->thread);
319 wake_up_process(stopper2->thread);
320 }
321 preempt_enable();
322
323 return err;
324}
325/**
326 * stop_two_cpus - stops two cpus
327 * @cpu1: the cpu to stop
328 * @cpu2: the other cpu to stop
329 * @fn: function to execute
330 * @arg: argument to @fn
331 *
332 * Stops both the current and specified CPU and runs @fn on one of them.
333 *
334 * returns when both are completed.
335 */
336int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
337{
338 struct cpu_stop_done done;
339 struct cpu_stop_work work1, work2;
340 struct multi_stop_data msdata;
341
342 msdata = (struct multi_stop_data){
343 .fn = fn,
344 .data = arg,
345 .num_threads = 2,
346 .active_cpus = cpumask_of(cpu1),
347 };
348
349 work1 = work2 = (struct cpu_stop_work){
350 .fn = multi_cpu_stop,
351 .arg = &msdata,
352 .done = &done,
353 .caller = _RET_IP_,
354 };
355
356 cpu_stop_init_done(&done, 2);
357 set_state(&msdata, MULTI_STOP_PREPARE);
358
359 if (cpu1 > cpu2)
360 swap(cpu1, cpu2);
361 if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
362 return -ENOENT;
363
364 wait_for_completion(&done.completion);
365 return done.ret;
366}
367
368/**
369 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
370 * @cpu: cpu to stop
371 * @fn: function to execute
372 * @arg: argument to @fn
373 * @work_buf: pointer to cpu_stop_work structure
374 *
375 * Similar to stop_one_cpu() but doesn't wait for completion. The
376 * caller is responsible for ensuring @work_buf is currently unused
377 * and will remain untouched until stopper starts executing @fn.
378 *
379 * CONTEXT:
380 * Don't care, but the caller must ensure @cpu's stopper stays enabled
381 * until the work is queued, e.g. by preempt_disable().
382 */
383void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
384 struct cpu_stop_work *work_buf)
385{
386 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, .caller = _RET_IP_, };
387 WARN_ON_ONCE(!cpu_stop_queue_work(cpu, work_buf));
388}
389
390static bool queue_stop_cpus_work(const struct cpumask *cpumask,
391 cpu_stop_fn_t fn, void *arg,
392 struct cpu_stop_done *done)
393{
394 struct cpu_stop_work *work;
395 unsigned int cpu;
396 bool queued = false;
397
398 /*
399 * Disable preemption while queueing to avoid getting
400 * preempted by a stopper which might wait for other stoppers
401 * to enter @fn which can lead to deadlock.
402 */
403 preempt_disable();
404 stop_cpus_in_progress = true;
405 barrier();
406 for_each_cpu(cpu, cpumask) {
407 work = &per_cpu(cpu_stopper.stop_work, cpu);
408 work->fn = fn;
409 work->arg = arg;
410 work->done = done;
411 work->caller = _RET_IP_;
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
537static void cpu_stop_create(unsigned int cpu)
538{
539 sched_set_stop_task(cpu, per_cpu(cpu_stopper.thread, cpu));
540}
541
542static void cpu_stop_park(unsigned int cpu)
543{
544 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
545
546 WARN_ON(!list_empty(&stopper->works));
547}
548
549void stop_machine_unpark(int cpu)
550{
551 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
552
553 stopper->enabled = true;
554 kthread_unpark(stopper->thread);
555}
556
557static struct smp_hotplug_thread cpu_stop_threads = {
558 .store = &cpu_stopper.thread,
559 .thread_should_run = cpu_stop_should_run,
560 .thread_fn = cpu_stopper_thread,
561 .thread_comm = "migration/%u",
562 .create = cpu_stop_create,
563 .park = cpu_stop_park,
564 .selfparking = true,
565};
566
567static int __init cpu_stop_init(void)
568{
569 unsigned int cpu;
570
571 for_each_possible_cpu(cpu) {
572 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
573
574 raw_spin_lock_init(&stopper->lock);
575 INIT_LIST_HEAD(&stopper->works);
576 }
577
578 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
579 stop_machine_unpark(raw_smp_processor_id());
580 stop_machine_initialized = true;
581 return 0;
582}
583early_initcall(cpu_stop_init);
584
585int stop_machine_cpuslocked(cpu_stop_fn_t fn, void *data,
586 const struct cpumask *cpus)
587{
588 struct multi_stop_data msdata = {
589 .fn = fn,
590 .data = data,
591 .num_threads = num_online_cpus(),
592 .active_cpus = cpus,
593 };
594
595 lockdep_assert_cpus_held();
596
597 if (!stop_machine_initialized) {
598 /*
599 * Handle the case where stop_machine() is called
600 * early in boot before stop_machine() has been
601 * initialized.
602 */
603 unsigned long flags;
604 int ret;
605
606 WARN_ON_ONCE(msdata.num_threads != 1);
607
608 local_irq_save(flags);
609 hard_irq_disable();
610 ret = (*fn)(data);
611 local_irq_restore(flags);
612
613 return ret;
614 }
615
616 /* Set the initial state and stop all online cpus. */
617 set_state(&msdata, MULTI_STOP_PREPARE);
618 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
619}
620
621int stop_machine(cpu_stop_fn_t fn, void *data, const struct cpumask *cpus)
622{
623 int ret;
624
625 /* No CPUs can come up or down during this. */
626 cpus_read_lock();
627 ret = stop_machine_cpuslocked(fn, data, cpus);
628 cpus_read_unlock();
629 return ret;
630}
631EXPORT_SYMBOL_GPL(stop_machine);
632
633#ifdef CONFIG_SCHED_SMT
634/*
635 * INTEL_IFS is the only user of this API. That selftest can
636 * only be compiled if SMP=y. On x86 it selects SCHED_SMT.
637 * Keep the ifdefs for now.
638 */
639int stop_core_cpuslocked(unsigned int cpu, cpu_stop_fn_t fn, void *data)
640{
641 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
642
643 struct multi_stop_data msdata = {
644 .fn = fn,
645 .data = data,
646 .num_threads = cpumask_weight(smt_mask),
647 .active_cpus = smt_mask,
648 };
649
650 lockdep_assert_cpus_held();
651
652 /* Set the initial state and stop all online cpus. */
653 set_state(&msdata, MULTI_STOP_PREPARE);
654 return stop_cpus(smt_mask, multi_cpu_stop, &msdata);
655}
656EXPORT_SYMBOL_GPL(stop_core_cpuslocked);
657#endif
658
659/**
660 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
661 * @fn: the function to run
662 * @data: the data ptr for the @fn()
663 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
664 *
665 * This is identical to stop_machine() but can be called from a CPU which
666 * is not active. The local CPU is in the process of hotplug (so no other
667 * CPU hotplug can start) and not marked active and doesn't have enough
668 * context to sleep.
669 *
670 * This function provides stop_machine() functionality for such state by
671 * using busy-wait for synchronization and executing @fn directly for local
672 * CPU.
673 *
674 * CONTEXT:
675 * Local CPU is inactive. Temporarily stops all active CPUs.
676 *
677 * RETURNS:
678 * 0 if all executions of @fn returned 0, any non zero return value if any
679 * returned non zero.
680 */
681int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
682 const struct cpumask *cpus)
683{
684 struct multi_stop_data msdata = { .fn = fn, .data = data,
685 .active_cpus = cpus };
686 struct cpu_stop_done done;
687 int ret;
688
689 /* Local CPU must be inactive and CPU hotplug in progress. */
690 BUG_ON(cpu_active(raw_smp_processor_id()));
691 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
692
693 /* No proper task established and can't sleep - busy wait for lock. */
694 while (!mutex_trylock(&stop_cpus_mutex))
695 cpu_relax();
696
697 /* Schedule work on other CPUs and execute directly for local CPU */
698 set_state(&msdata, MULTI_STOP_PREPARE);
699 cpu_stop_init_done(&done, num_active_cpus());
700 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
701 &done);
702 ret = multi_cpu_stop(&msdata);
703
704 /* Busy wait for completion. */
705 while (!completion_done(&done.completion))
706 cpu_relax();
707
708 mutex_unlock(&stop_cpus_mutex);
709 return ret ?: done.ret;
710}