stop_machine.c

  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	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{
 87	list_add_tail(&work->list, &stopper->works);
 88}
 89
 90/* queue @work to @stopper.  if offline, @work is completed immediately */
 91static bool cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
 92{
 93	struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
 94	unsigned long flags;
 95	bool enabled;
 96
 97	preempt_disable();
 98	raw_spin_lock_irqsave(&stopper->lock, flags);
 99	enabled = stopper->enabled;
100	if (enabled)
101		__cpu_stop_queue_work(stopper, work);
102	else if (work->done)
103		cpu_stop_signal_done(work->done);
104	raw_spin_unlock_irqrestore(&stopper->lock, flags);
105
106	if (enabled)
107		wake_up_process(stopper->thread);
108	preempt_enable();
109
110	return enabled;
111}
112
113/**
114 * stop_one_cpu - stop a cpu
115 * @cpu: cpu to stop
116 * @fn: function to execute
117 * @arg: argument to @fn
118 *
119 * Execute @fn(@arg) on @cpu.  @fn is run in a process context with
120 * the highest priority preempting any task on the cpu and
121 * monopolizing it.  This function returns after the execution is
122 * complete.
123 *
124 * This function doesn't guarantee @cpu stays online till @fn
125 * completes.  If @cpu goes down in the middle, execution may happen
126 * partially or fully on different cpus.  @fn should either be ready
127 * for that or the caller should ensure that @cpu stays online until
128 * this function completes.
129 *
130 * CONTEXT:
131 * Might sleep.
132 *
133 * RETURNS:
134 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
135 * otherwise, the return value of @fn.
136 */
137int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
138{
139	struct cpu_stop_done done;
140	struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done, .caller = _RET_IP_ };
141
142	cpu_stop_init_done(&done, 1);
143	if (!cpu_stop_queue_work(cpu, &work))
144		return -ENOENT;
145	/*
146	 * In case @cpu == smp_proccessor_id() we can avoid a sleep+wakeup
147	 * cycle by doing a preemption:
148	 */
149	cond_resched();
150	wait_for_completion(&done.completion);
151	return done.ret;
152}
153
154/* This controls the threads on each CPU. */
155enum multi_stop_state {
156	/* Dummy starting state for thread. */
157	MULTI_STOP_NONE,
158	/* Awaiting everyone to be scheduled. */
159	MULTI_STOP_PREPARE,
160	/* Disable interrupts. */
161	MULTI_STOP_DISABLE_IRQ,
162	/* Run the function */
163	MULTI_STOP_RUN,
164	/* Exit */
165	MULTI_STOP_EXIT,
166};
167
168struct multi_stop_data {
169	cpu_stop_fn_t		fn;
170	void			*data;
171	/* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
172	unsigned int		num_threads;
173	const struct cpumask	*active_cpus;
174
175	enum multi_stop_state	state;
176	atomic_t		thread_ack;
177};
178
179static void set_state(struct multi_stop_data *msdata,
180		      enum multi_stop_state newstate)
181{
182	/* Reset ack counter. */
183	atomic_set(&msdata->thread_ack, msdata->num_threads);
184	smp_wmb();
185	WRITE_ONCE(msdata->state, newstate);
186}
187
188/* Last one to ack a state moves to the next state. */
189static void ack_state(struct multi_stop_data *msdata)
190{
191	if (atomic_dec_and_test(&msdata->thread_ack))
192		set_state(msdata, msdata->state + 1);
193}
194
195notrace void __weak stop_machine_yield(const struct cpumask *cpumask)
196{
197	cpu_relax();
198}
199
200/* This is the cpu_stop function which stops the CPU. */
201static int multi_cpu_stop(void *data)
202{
203	struct multi_stop_data *msdata = data;
204	enum multi_stop_state newstate, curstate = MULTI_STOP_NONE;
205	int cpu = smp_processor_id(), err = 0;
206	const struct cpumask *cpumask;
207	unsigned long flags;
208	bool is_active;
209
210	/*
211	 * When called from stop_machine_from_inactive_cpu(), irq might
212	 * already be disabled.  Save the state and restore it on exit.
213	 */
214	local_save_flags(flags);
215
216	if (!msdata->active_cpus) {
217		cpumask = cpu_online_mask;
218		is_active = cpu == cpumask_first(cpumask);
219	} else {
220		cpumask = msdata->active_cpus;
221		is_active = cpumask_test_cpu(cpu, cpumask);
222	}
223
224	/* Simple state machine */
225	do {
226		/* Chill out and ensure we re-read multi_stop_state. */
227		stop_machine_yield(cpumask);
228		newstate = READ_ONCE(msdata->state);
229		if (newstate != curstate) {
230			curstate = newstate;
231			switch (curstate) {
232			case MULTI_STOP_DISABLE_IRQ:
233				local_irq_disable();
234				hard_irq_disable();
235				break;
236			case MULTI_STOP_RUN:
237				if (is_active)
238					err = msdata->fn(msdata->data);
239				break;
240			default:
241				break;
242			}
243			ack_state(msdata);
244		} else if (curstate > MULTI_STOP_PREPARE) {
245			/*
246			 * At this stage all other CPUs we depend on must spin
247			 * in the same loop. Any reason for hard-lockup should
248			 * be detected and reported on their side.
249			 */
250			touch_nmi_watchdog();
251			/* Also suppress RCU CPU stall warnings. */
252			rcu_momentary_eqs();
253		}
254	} while (curstate != MULTI_STOP_EXIT);
255
256	local_irq_restore(flags);
257	return err;
258}
259
260static int cpu_stop_queue_two_works(int cpu1, struct cpu_stop_work *work1,
261				    int cpu2, struct cpu_stop_work *work2)
262{
263	struct cpu_stopper *stopper1 = per_cpu_ptr(&cpu_stopper, cpu1);
264	struct cpu_stopper *stopper2 = per_cpu_ptr(&cpu_stopper, cpu2);
265	int err;
266
267retry:
268	/*
269	 * The waking up of stopper threads has to happen in the same
270	 * scheduling context as the queueing.  Otherwise, there is a
271	 * possibility of one of the above stoppers being woken up by another
272	 * CPU, and preempting us. This will cause us to not wake up the other
273	 * stopper forever.
274	 */
275	preempt_disable();
276	raw_spin_lock_irq(&stopper1->lock);
277	raw_spin_lock_nested(&stopper2->lock, SINGLE_DEPTH_NESTING);
278
279	if (!stopper1->enabled || !stopper2->enabled) {
280		err = -ENOENT;
281		goto unlock;
282	}
283
284	/*
285	 * Ensure that if we race with __stop_cpus() the stoppers won't get
286	 * queued up in reverse order leading to system deadlock.
287	 *
288	 * We can't miss stop_cpus_in_progress if queue_stop_cpus_work() has
289	 * queued a work on cpu1 but not on cpu2, we hold both locks.
290	 *
291	 * It can be falsely true but it is safe to spin until it is cleared,
292	 * queue_stop_cpus_work() does everything under preempt_disable().
293	 */
294	if (unlikely(stop_cpus_in_progress)) {
295		err = -EDEADLK;
296		goto unlock;
297	}
298
299	err = 0;
300	__cpu_stop_queue_work(stopper1, work1);
301	__cpu_stop_queue_work(stopper2, work2);
302
303unlock:
304	raw_spin_unlock(&stopper2->lock);
305	raw_spin_unlock_irq(&stopper1->lock);
306
307	if (unlikely(err == -EDEADLK)) {
308		preempt_enable();
309
310		while (stop_cpus_in_progress)
311			cpu_relax();
312
313		goto retry;
314	}
315
316	if (!err) {
317		wake_up_process(stopper1->thread);
318		wake_up_process(stopper2->thread);
319	}
320	preempt_enable();
321
322	return err;
323}
324/**
325 * stop_two_cpus - stops two cpus
326 * @cpu1: the cpu to stop
327 * @cpu2: the other cpu to stop
328 * @fn: function to execute
329 * @arg: argument to @fn
330 *
331 * Stops both the current and specified CPU and runs @fn on one of them.
332 *
333 * returns when both are completed.
334 */
335int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
336{
337	struct cpu_stop_done done;
338	struct cpu_stop_work work1, work2;
339	struct multi_stop_data msdata;
340
341	msdata = (struct multi_stop_data){
342		.fn = fn,
343		.data = arg,
344		.num_threads = 2,
345		.active_cpus = cpumask_of(cpu1),
346	};
347
348	work1 = work2 = (struct cpu_stop_work){
349		.fn = multi_cpu_stop,
350		.arg = &msdata,
351		.done = &done,
352		.caller = _RET_IP_,
353	};
354
355	cpu_stop_init_done(&done, 2);
356	set_state(&msdata, MULTI_STOP_PREPARE);
357
358	if (cpu1 > cpu2)
359		swap(cpu1, cpu2);
360	if (cpu_stop_queue_two_works(cpu1, &work1, cpu2, &work2))
361		return -ENOENT;
362
363	wait_for_completion(&done.completion);
364	return done.ret;
365}
366
367/**
368 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
369 * @cpu: cpu to stop
370 * @fn: function to execute
371 * @arg: argument to @fn
372 * @work_buf: pointer to cpu_stop_work structure
373 *
374 * Similar to stop_one_cpu() but doesn't wait for completion.  The
375 * caller is responsible for ensuring @work_buf is currently unused
376 * and will remain untouched until stopper starts executing @fn.
377 *
378 * CONTEXT:
379 * Don't care.
380 *
381 * RETURNS:
382 * true if cpu_stop_work was queued successfully and @fn will be called,
383 * false otherwise.
384 */
385bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
386			struct cpu_stop_work *work_buf)
387{
388	*work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, .caller = _RET_IP_, };
389	return cpu_stop_queue_work(cpu, work_buf);
390}
391
392static bool queue_stop_cpus_work(const struct cpumask *cpumask,
393				 cpu_stop_fn_t fn, void *arg,
394				 struct cpu_stop_done *done)
395{
396	struct cpu_stop_work *work;
397	unsigned int cpu;
398	bool queued = false;
399
400	/*
401	 * Disable preemption while queueing to avoid getting
402	 * preempted by a stopper which might wait for other stoppers
403	 * to enter @fn which can lead to deadlock.
404	 */
405	preempt_disable();
406	stop_cpus_in_progress = true;
407	barrier();
408	for_each_cpu(cpu, cpumask) {
409		work = &per_cpu(cpu_stopper.stop_work, cpu);
410		work->fn = fn;
411		work->arg = arg;
412		work->done = done;
413		work->caller = _RET_IP_;
414		if (cpu_stop_queue_work(cpu, work))
415			queued = true;
416	}
417	barrier();
418	stop_cpus_in_progress = false;
419	preempt_enable();
420
421	return queued;
422}
423
424static int __stop_cpus(const struct cpumask *cpumask,
425		       cpu_stop_fn_t fn, void *arg)
426{
427	struct cpu_stop_done done;
428
429	cpu_stop_init_done(&done, cpumask_weight(cpumask));
430	if (!queue_stop_cpus_work(cpumask, fn, arg, &done))
431		return -ENOENT;
432	wait_for_completion(&done.completion);
433	return done.ret;
434}
435
436/**
437 * stop_cpus - stop multiple cpus
438 * @cpumask: cpus to stop
439 * @fn: function to execute
440 * @arg: argument to @fn
441 *
442 * Execute @fn(@arg) on online cpus in @cpumask.  On each target cpu,
443 * @fn is run in a process context with the highest priority
444 * preempting any task on the cpu and monopolizing it.  This function
445 * returns after all executions are complete.
446 *
447 * This function doesn't guarantee the cpus in @cpumask stay online
448 * till @fn completes.  If some cpus go down in the middle, execution
449 * on the cpu may happen partially or fully on different cpus.  @fn
450 * should either be ready for that or the caller should ensure that
451 * the cpus stay online until this function completes.
452 *
453 * All stop_cpus() calls are serialized making it safe for @fn to wait
454 * for all cpus to start executing it.
455 *
456 * CONTEXT:
457 * Might sleep.
458 *
459 * RETURNS:
460 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
461 * @cpumask were offline; otherwise, 0 if all executions of @fn
462 * returned 0, any non zero return value if any returned non zero.
463 */
464static int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
465{
466	int ret;
467
468	/* static works are used, process one request at a time */
469	mutex_lock(&stop_cpus_mutex);
470	ret = __stop_cpus(cpumask, fn, arg);
471	mutex_unlock(&stop_cpus_mutex);
472	return ret;
473}
474
475static int cpu_stop_should_run(unsigned int cpu)
476{
477	struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
478	unsigned long flags;
479	int run;
480
481	raw_spin_lock_irqsave(&stopper->lock, flags);
482	run = !list_empty(&stopper->works);
483	raw_spin_unlock_irqrestore(&stopper->lock, flags);
484	return run;
485}
486
487static void cpu_stopper_thread(unsigned int cpu)
488{
489	struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
490	struct cpu_stop_work *work;
491
492repeat:
493	work = NULL;
494	raw_spin_lock_irq(&stopper->lock);
495	if (!list_empty(&stopper->works)) {
496		work = list_first_entry(&stopper->works,
497					struct cpu_stop_work, list);
498		list_del_init(&work->list);
499	}
500	raw_spin_unlock_irq(&stopper->lock);
501
502	if (work) {
503		cpu_stop_fn_t fn = work->fn;
504		void *arg = work->arg;
505		struct cpu_stop_done *done = work->done;
506		int ret;
507
508		/* cpu stop callbacks must not sleep, make in_atomic() == T */
509		stopper->caller = work->caller;
510		stopper->fn = fn;
511		preempt_count_inc();
512		ret = fn(arg);
513		if (done) {
514			if (ret)
515				done->ret = ret;
516			cpu_stop_signal_done(done);
517		}
518		preempt_count_dec();
519		stopper->fn = NULL;
520		stopper->caller = 0;
521		WARN_ONCE(preempt_count(),
522			  "cpu_stop: %ps(%p) leaked preempt count\n", fn, arg);
523		goto repeat;
524	}
525}
526
527void stop_machine_park(int cpu)
528{
529	struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
530	/*
531	 * Lockless. cpu_stopper_thread() will take stopper->lock and flush
532	 * the pending works before it parks, until then it is fine to queue
533	 * the new works.
534	 */
535	stopper->enabled = false;
536	kthread_park(stopper->thread);
537}
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#ifdef CONFIG_SCHED_SMT
636int stop_core_cpuslocked(unsigned int cpu, cpu_stop_fn_t fn, void *data)
637{
638	const struct cpumask *smt_mask = cpu_smt_mask(cpu);
639
640	struct multi_stop_data msdata = {
641		.fn = fn,
642		.data = data,
643		.num_threads = cpumask_weight(smt_mask),
644		.active_cpus = smt_mask,
645	};
646
647	lockdep_assert_cpus_held();
648
649	/* Set the initial state and stop all online cpus. */
650	set_state(&msdata, MULTI_STOP_PREPARE);
651	return stop_cpus(smt_mask, multi_cpu_stop, &msdata);
652}
653EXPORT_SYMBOL_GPL(stop_core_cpuslocked);
654#endif
655
656/**
657 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
658 * @fn: the function to run
659 * @data: the data ptr for the @fn()
660 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
661 *
662 * This is identical to stop_machine() but can be called from a CPU which
663 * is not active.  The local CPU is in the process of hotplug (so no other
664 * CPU hotplug can start) and not marked active and doesn't have enough
665 * context to sleep.
666 *
667 * This function provides stop_machine() functionality for such state by
668 * using busy-wait for synchronization and executing @fn directly for local
669 * CPU.
670 *
671 * CONTEXT:
672 * Local CPU is inactive.  Temporarily stops all active CPUs.
673 *
674 * RETURNS:
675 * 0 if all executions of @fn returned 0, any non zero return value if any
676 * returned non zero.
677 */
678int stop_machine_from_inactive_cpu(cpu_stop_fn_t fn, void *data,
679				  const struct cpumask *cpus)
680{
681	struct multi_stop_data msdata = { .fn = fn, .data = data,
682					    .active_cpus = cpus };
683	struct cpu_stop_done done;
684	int ret;
685
686	/* Local CPU must be inactive and CPU hotplug in progress. */
687	BUG_ON(cpu_active(raw_smp_processor_id()));
688	msdata.num_threads = num_active_cpus() + 1;	/* +1 for local */
689
690	/* No proper task established and can't sleep - busy wait for lock. */
691	while (!mutex_trylock(&stop_cpus_mutex))
692		cpu_relax();
693
694	/* Schedule work on other CPUs and execute directly for local CPU */
695	set_state(&msdata, MULTI_STOP_PREPARE);
696	cpu_stop_init_done(&done, num_active_cpus());
697	queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
698			     &done);
699	ret = multi_cpu_stop(&msdata);
700
701	/* Busy wait for completion. */
702	while (!completion_done(&done.completion))
703		cpu_relax();
704
705	mutex_unlock(&stop_cpus_mutex);
706	return ret ?: done.ret;
707}