xref: /linux/fs/btrfs/async-thread.c (revision 4413e16d9d21673bb5048a2e542f1aaa00015c2e)
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 
19 #include <linux/kthread.h>
20 #include <linux/slab.h>
21 #include <linux/list.h>
22 #include <linux/spinlock.h>
23 #include <linux/freezer.h>
24 #include "async-thread.h"
25 
26 #define WORK_QUEUED_BIT 0
27 #define WORK_DONE_BIT 1
28 #define WORK_ORDER_DONE_BIT 2
29 #define WORK_HIGH_PRIO_BIT 3
30 
31 /*
32  * container for the kthread task pointer and the list of pending work
33  * One of these is allocated per thread.
34  */
35 struct btrfs_worker_thread {
36 	/* pool we belong to */
37 	struct btrfs_workers *workers;
38 
39 	/* list of struct btrfs_work that are waiting for service */
40 	struct list_head pending;
41 	struct list_head prio_pending;
42 
43 	/* list of worker threads from struct btrfs_workers */
44 	struct list_head worker_list;
45 
46 	/* kthread */
47 	struct task_struct *task;
48 
49 	/* number of things on the pending list */
50 	atomic_t num_pending;
51 
52 	/* reference counter for this struct */
53 	atomic_t refs;
54 
55 	unsigned long sequence;
56 
57 	/* protects the pending list. */
58 	spinlock_t lock;
59 
60 	/* set to non-zero when this thread is already awake and kicking */
61 	int working;
62 
63 	/* are we currently idle */
64 	int idle;
65 };
66 
67 static int __btrfs_start_workers(struct btrfs_workers *workers);
68 
69 /*
70  * btrfs_start_workers uses kthread_run, which can block waiting for memory
71  * for a very long time.  It will actually throttle on page writeback,
72  * and so it may not make progress until after our btrfs worker threads
73  * process all of the pending work structs in their queue
74  *
75  * This means we can't use btrfs_start_workers from inside a btrfs worker
76  * thread that is used as part of cleaning dirty memory, which pretty much
77  * involves all of the worker threads.
78  *
79  * Instead we have a helper queue who never has more than one thread
80  * where we scheduler thread start operations.  This worker_start struct
81  * is used to contain the work and hold a pointer to the queue that needs
82  * another worker.
83  */
84 struct worker_start {
85 	struct btrfs_work work;
86 	struct btrfs_workers *queue;
87 };
88 
89 static void start_new_worker_func(struct btrfs_work *work)
90 {
91 	struct worker_start *start;
92 	start = container_of(work, struct worker_start, work);
93 	__btrfs_start_workers(start->queue);
94 	kfree(start);
95 }
96 
97 /*
98  * helper function to move a thread onto the idle list after it
99  * has finished some requests.
100  */
101 static void check_idle_worker(struct btrfs_worker_thread *worker)
102 {
103 	if (!worker->idle && atomic_read(&worker->num_pending) <
104 	    worker->workers->idle_thresh / 2) {
105 		unsigned long flags;
106 		spin_lock_irqsave(&worker->workers->lock, flags);
107 		worker->idle = 1;
108 
109 		/* the list may be empty if the worker is just starting */
110 		if (!list_empty(&worker->worker_list)) {
111 			list_move(&worker->worker_list,
112 				 &worker->workers->idle_list);
113 		}
114 		spin_unlock_irqrestore(&worker->workers->lock, flags);
115 	}
116 }
117 
118 /*
119  * helper function to move a thread off the idle list after new
120  * pending work is added.
121  */
122 static void check_busy_worker(struct btrfs_worker_thread *worker)
123 {
124 	if (worker->idle && atomic_read(&worker->num_pending) >=
125 	    worker->workers->idle_thresh) {
126 		unsigned long flags;
127 		spin_lock_irqsave(&worker->workers->lock, flags);
128 		worker->idle = 0;
129 
130 		if (!list_empty(&worker->worker_list)) {
131 			list_move_tail(&worker->worker_list,
132 				      &worker->workers->worker_list);
133 		}
134 		spin_unlock_irqrestore(&worker->workers->lock, flags);
135 	}
136 }
137 
138 static void check_pending_worker_creates(struct btrfs_worker_thread *worker)
139 {
140 	struct btrfs_workers *workers = worker->workers;
141 	struct worker_start *start;
142 	unsigned long flags;
143 
144 	rmb();
145 	if (!workers->atomic_start_pending)
146 		return;
147 
148 	start = kzalloc(sizeof(*start), GFP_NOFS);
149 	if (!start)
150 		return;
151 
152 	start->work.func = start_new_worker_func;
153 	start->queue = workers;
154 
155 	spin_lock_irqsave(&workers->lock, flags);
156 	if (!workers->atomic_start_pending)
157 		goto out;
158 
159 	workers->atomic_start_pending = 0;
160 	if (workers->num_workers + workers->num_workers_starting >=
161 	    workers->max_workers)
162 		goto out;
163 
164 	workers->num_workers_starting += 1;
165 	spin_unlock_irqrestore(&workers->lock, flags);
166 	btrfs_queue_worker(workers->atomic_worker_start, &start->work);
167 	return;
168 
169 out:
170 	kfree(start);
171 	spin_unlock_irqrestore(&workers->lock, flags);
172 }
173 
174 static noinline void run_ordered_completions(struct btrfs_workers *workers,
175 					    struct btrfs_work *work)
176 {
177 	if (!workers->ordered)
178 		return;
179 
180 	set_bit(WORK_DONE_BIT, &work->flags);
181 
182 	spin_lock(&workers->order_lock);
183 
184 	while (1) {
185 		if (!list_empty(&workers->prio_order_list)) {
186 			work = list_entry(workers->prio_order_list.next,
187 					  struct btrfs_work, order_list);
188 		} else if (!list_empty(&workers->order_list)) {
189 			work = list_entry(workers->order_list.next,
190 					  struct btrfs_work, order_list);
191 		} else {
192 			break;
193 		}
194 		if (!test_bit(WORK_DONE_BIT, &work->flags))
195 			break;
196 
197 		/* we are going to call the ordered done function, but
198 		 * we leave the work item on the list as a barrier so
199 		 * that later work items that are done don't have their
200 		 * functions called before this one returns
201 		 */
202 		if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
203 			break;
204 
205 		spin_unlock(&workers->order_lock);
206 
207 		work->ordered_func(work);
208 
209 		/* now take the lock again and drop our item from the list */
210 		spin_lock(&workers->order_lock);
211 		list_del(&work->order_list);
212 		spin_unlock(&workers->order_lock);
213 
214 		/*
215 		 * we don't want to call the ordered free functions
216 		 * with the lock held though
217 		 */
218 		work->ordered_free(work);
219 		spin_lock(&workers->order_lock);
220 	}
221 
222 	spin_unlock(&workers->order_lock);
223 }
224 
225 static void put_worker(struct btrfs_worker_thread *worker)
226 {
227 	if (atomic_dec_and_test(&worker->refs))
228 		kfree(worker);
229 }
230 
231 static int try_worker_shutdown(struct btrfs_worker_thread *worker)
232 {
233 	int freeit = 0;
234 
235 	spin_lock_irq(&worker->lock);
236 	spin_lock(&worker->workers->lock);
237 	if (worker->workers->num_workers > 1 &&
238 	    worker->idle &&
239 	    !worker->working &&
240 	    !list_empty(&worker->worker_list) &&
241 	    list_empty(&worker->prio_pending) &&
242 	    list_empty(&worker->pending) &&
243 	    atomic_read(&worker->num_pending) == 0) {
244 		freeit = 1;
245 		list_del_init(&worker->worker_list);
246 		worker->workers->num_workers--;
247 	}
248 	spin_unlock(&worker->workers->lock);
249 	spin_unlock_irq(&worker->lock);
250 
251 	if (freeit)
252 		put_worker(worker);
253 	return freeit;
254 }
255 
256 static struct btrfs_work *get_next_work(struct btrfs_worker_thread *worker,
257 					struct list_head *prio_head,
258 					struct list_head *head)
259 {
260 	struct btrfs_work *work = NULL;
261 	struct list_head *cur = NULL;
262 
263 	if(!list_empty(prio_head))
264 		cur = prio_head->next;
265 
266 	smp_mb();
267 	if (!list_empty(&worker->prio_pending))
268 		goto refill;
269 
270 	if (!list_empty(head))
271 		cur = head->next;
272 
273 	if (cur)
274 		goto out;
275 
276 refill:
277 	spin_lock_irq(&worker->lock);
278 	list_splice_tail_init(&worker->prio_pending, prio_head);
279 	list_splice_tail_init(&worker->pending, head);
280 
281 	if (!list_empty(prio_head))
282 		cur = prio_head->next;
283 	else if (!list_empty(head))
284 		cur = head->next;
285 	spin_unlock_irq(&worker->lock);
286 
287 	if (!cur)
288 		goto out_fail;
289 
290 out:
291 	work = list_entry(cur, struct btrfs_work, list);
292 
293 out_fail:
294 	return work;
295 }
296 
297 /*
298  * main loop for servicing work items
299  */
300 static int worker_loop(void *arg)
301 {
302 	struct btrfs_worker_thread *worker = arg;
303 	struct list_head head;
304 	struct list_head prio_head;
305 	struct btrfs_work *work;
306 
307 	INIT_LIST_HEAD(&head);
308 	INIT_LIST_HEAD(&prio_head);
309 
310 	do {
311 again:
312 		while (1) {
313 
314 
315 			work = get_next_work(worker, &prio_head, &head);
316 			if (!work)
317 				break;
318 
319 			list_del(&work->list);
320 			clear_bit(WORK_QUEUED_BIT, &work->flags);
321 
322 			work->worker = worker;
323 
324 			work->func(work);
325 
326 			atomic_dec(&worker->num_pending);
327 			/*
328 			 * unless this is an ordered work queue,
329 			 * 'work' was probably freed by func above.
330 			 */
331 			run_ordered_completions(worker->workers, work);
332 
333 			check_pending_worker_creates(worker);
334 			cond_resched();
335 		}
336 
337 		spin_lock_irq(&worker->lock);
338 		check_idle_worker(worker);
339 
340 		if (freezing(current)) {
341 			worker->working = 0;
342 			spin_unlock_irq(&worker->lock);
343 			try_to_freeze();
344 		} else {
345 			spin_unlock_irq(&worker->lock);
346 			if (!kthread_should_stop()) {
347 				cpu_relax();
348 				/*
349 				 * we've dropped the lock, did someone else
350 				 * jump_in?
351 				 */
352 				smp_mb();
353 				if (!list_empty(&worker->pending) ||
354 				    !list_empty(&worker->prio_pending))
355 					continue;
356 
357 				/*
358 				 * this short schedule allows more work to
359 				 * come in without the queue functions
360 				 * needing to go through wake_up_process()
361 				 *
362 				 * worker->working is still 1, so nobody
363 				 * is going to try and wake us up
364 				 */
365 				schedule_timeout(1);
366 				smp_mb();
367 				if (!list_empty(&worker->pending) ||
368 				    !list_empty(&worker->prio_pending))
369 					continue;
370 
371 				if (kthread_should_stop())
372 					break;
373 
374 				/* still no more work?, sleep for real */
375 				spin_lock_irq(&worker->lock);
376 				set_current_state(TASK_INTERRUPTIBLE);
377 				if (!list_empty(&worker->pending) ||
378 				    !list_empty(&worker->prio_pending)) {
379 					spin_unlock_irq(&worker->lock);
380 					set_current_state(TASK_RUNNING);
381 					goto again;
382 				}
383 
384 				/*
385 				 * this makes sure we get a wakeup when someone
386 				 * adds something new to the queue
387 				 */
388 				worker->working = 0;
389 				spin_unlock_irq(&worker->lock);
390 
391 				if (!kthread_should_stop()) {
392 					schedule_timeout(HZ * 120);
393 					if (!worker->working &&
394 					    try_worker_shutdown(worker)) {
395 						return 0;
396 					}
397 				}
398 			}
399 			__set_current_state(TASK_RUNNING);
400 		}
401 	} while (!kthread_should_stop());
402 	return 0;
403 }
404 
405 /*
406  * this will wait for all the worker threads to shutdown
407  */
408 void btrfs_stop_workers(struct btrfs_workers *workers)
409 {
410 	struct list_head *cur;
411 	struct btrfs_worker_thread *worker;
412 	int can_stop;
413 
414 	spin_lock_irq(&workers->lock);
415 	list_splice_init(&workers->idle_list, &workers->worker_list);
416 	while (!list_empty(&workers->worker_list)) {
417 		cur = workers->worker_list.next;
418 		worker = list_entry(cur, struct btrfs_worker_thread,
419 				    worker_list);
420 
421 		atomic_inc(&worker->refs);
422 		workers->num_workers -= 1;
423 		if (!list_empty(&worker->worker_list)) {
424 			list_del_init(&worker->worker_list);
425 			put_worker(worker);
426 			can_stop = 1;
427 		} else
428 			can_stop = 0;
429 		spin_unlock_irq(&workers->lock);
430 		if (can_stop)
431 			kthread_stop(worker->task);
432 		spin_lock_irq(&workers->lock);
433 		put_worker(worker);
434 	}
435 	spin_unlock_irq(&workers->lock);
436 }
437 
438 /*
439  * simple init on struct btrfs_workers
440  */
441 void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max,
442 			struct btrfs_workers *async_helper)
443 {
444 	workers->num_workers = 0;
445 	workers->num_workers_starting = 0;
446 	INIT_LIST_HEAD(&workers->worker_list);
447 	INIT_LIST_HEAD(&workers->idle_list);
448 	INIT_LIST_HEAD(&workers->order_list);
449 	INIT_LIST_HEAD(&workers->prio_order_list);
450 	spin_lock_init(&workers->lock);
451 	spin_lock_init(&workers->order_lock);
452 	workers->max_workers = max;
453 	workers->idle_thresh = 32;
454 	workers->name = name;
455 	workers->ordered = 0;
456 	workers->atomic_start_pending = 0;
457 	workers->atomic_worker_start = async_helper;
458 }
459 
460 /*
461  * starts new worker threads.  This does not enforce the max worker
462  * count in case you need to temporarily go past it.
463  */
464 static int __btrfs_start_workers(struct btrfs_workers *workers)
465 {
466 	struct btrfs_worker_thread *worker;
467 	int ret = 0;
468 
469 	worker = kzalloc(sizeof(*worker), GFP_NOFS);
470 	if (!worker) {
471 		ret = -ENOMEM;
472 		goto fail;
473 	}
474 
475 	INIT_LIST_HEAD(&worker->pending);
476 	INIT_LIST_HEAD(&worker->prio_pending);
477 	INIT_LIST_HEAD(&worker->worker_list);
478 	spin_lock_init(&worker->lock);
479 
480 	atomic_set(&worker->num_pending, 0);
481 	atomic_set(&worker->refs, 1);
482 	worker->workers = workers;
483 	worker->task = kthread_run(worker_loop, worker,
484 				   "btrfs-%s-%d", workers->name,
485 				   workers->num_workers + 1);
486 	if (IS_ERR(worker->task)) {
487 		ret = PTR_ERR(worker->task);
488 		kfree(worker);
489 		goto fail;
490 	}
491 	spin_lock_irq(&workers->lock);
492 	list_add_tail(&worker->worker_list, &workers->idle_list);
493 	worker->idle = 1;
494 	workers->num_workers++;
495 	workers->num_workers_starting--;
496 	WARN_ON(workers->num_workers_starting < 0);
497 	spin_unlock_irq(&workers->lock);
498 
499 	return 0;
500 fail:
501 	spin_lock_irq(&workers->lock);
502 	workers->num_workers_starting--;
503 	spin_unlock_irq(&workers->lock);
504 	return ret;
505 }
506 
507 int btrfs_start_workers(struct btrfs_workers *workers)
508 {
509 	spin_lock_irq(&workers->lock);
510 	workers->num_workers_starting++;
511 	spin_unlock_irq(&workers->lock);
512 	return __btrfs_start_workers(workers);
513 }
514 
515 /*
516  * run through the list and find a worker thread that doesn't have a lot
517  * to do right now.  This can return null if we aren't yet at the thread
518  * count limit and all of the threads are busy.
519  */
520 static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
521 {
522 	struct btrfs_worker_thread *worker;
523 	struct list_head *next;
524 	int enforce_min;
525 
526 	enforce_min = (workers->num_workers + workers->num_workers_starting) <
527 		workers->max_workers;
528 
529 	/*
530 	 * if we find an idle thread, don't move it to the end of the
531 	 * idle list.  This improves the chance that the next submission
532 	 * will reuse the same thread, and maybe catch it while it is still
533 	 * working
534 	 */
535 	if (!list_empty(&workers->idle_list)) {
536 		next = workers->idle_list.next;
537 		worker = list_entry(next, struct btrfs_worker_thread,
538 				    worker_list);
539 		return worker;
540 	}
541 	if (enforce_min || list_empty(&workers->worker_list))
542 		return NULL;
543 
544 	/*
545 	 * if we pick a busy task, move the task to the end of the list.
546 	 * hopefully this will keep things somewhat evenly balanced.
547 	 * Do the move in batches based on the sequence number.  This groups
548 	 * requests submitted at roughly the same time onto the same worker.
549 	 */
550 	next = workers->worker_list.next;
551 	worker = list_entry(next, struct btrfs_worker_thread, worker_list);
552 	worker->sequence++;
553 
554 	if (worker->sequence % workers->idle_thresh == 0)
555 		list_move_tail(next, &workers->worker_list);
556 	return worker;
557 }
558 
559 /*
560  * selects a worker thread to take the next job.  This will either find
561  * an idle worker, start a new worker up to the max count, or just return
562  * one of the existing busy workers.
563  */
564 static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
565 {
566 	struct btrfs_worker_thread *worker;
567 	unsigned long flags;
568 	struct list_head *fallback;
569 	int ret;
570 
571 	spin_lock_irqsave(&workers->lock, flags);
572 again:
573 	worker = next_worker(workers);
574 
575 	if (!worker) {
576 		if (workers->num_workers + workers->num_workers_starting >=
577 		    workers->max_workers) {
578 			goto fallback;
579 		} else if (workers->atomic_worker_start) {
580 			workers->atomic_start_pending = 1;
581 			goto fallback;
582 		} else {
583 			workers->num_workers_starting++;
584 			spin_unlock_irqrestore(&workers->lock, flags);
585 			/* we're below the limit, start another worker */
586 			ret = __btrfs_start_workers(workers);
587 			spin_lock_irqsave(&workers->lock, flags);
588 			if (ret)
589 				goto fallback;
590 			goto again;
591 		}
592 	}
593 	goto found;
594 
595 fallback:
596 	fallback = NULL;
597 	/*
598 	 * we have failed to find any workers, just
599 	 * return the first one we can find.
600 	 */
601 	if (!list_empty(&workers->worker_list))
602 		fallback = workers->worker_list.next;
603 	if (!list_empty(&workers->idle_list))
604 		fallback = workers->idle_list.next;
605 	BUG_ON(!fallback);
606 	worker = list_entry(fallback,
607 		  struct btrfs_worker_thread, worker_list);
608 found:
609 	/*
610 	 * this makes sure the worker doesn't exit before it is placed
611 	 * onto a busy/idle list
612 	 */
613 	atomic_inc(&worker->num_pending);
614 	spin_unlock_irqrestore(&workers->lock, flags);
615 	return worker;
616 }
617 
618 /*
619  * btrfs_requeue_work just puts the work item back on the tail of the list
620  * it was taken from.  It is intended for use with long running work functions
621  * that make some progress and want to give the cpu up for others.
622  */
623 void btrfs_requeue_work(struct btrfs_work *work)
624 {
625 	struct btrfs_worker_thread *worker = work->worker;
626 	unsigned long flags;
627 	int wake = 0;
628 
629 	if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
630 		return;
631 
632 	spin_lock_irqsave(&worker->lock, flags);
633 	if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
634 		list_add_tail(&work->list, &worker->prio_pending);
635 	else
636 		list_add_tail(&work->list, &worker->pending);
637 	atomic_inc(&worker->num_pending);
638 
639 	/* by definition we're busy, take ourselves off the idle
640 	 * list
641 	 */
642 	if (worker->idle) {
643 		spin_lock(&worker->workers->lock);
644 		worker->idle = 0;
645 		list_move_tail(&worker->worker_list,
646 			      &worker->workers->worker_list);
647 		spin_unlock(&worker->workers->lock);
648 	}
649 	if (!worker->working) {
650 		wake = 1;
651 		worker->working = 1;
652 	}
653 
654 	if (wake)
655 		wake_up_process(worker->task);
656 	spin_unlock_irqrestore(&worker->lock, flags);
657 }
658 
659 void btrfs_set_work_high_prio(struct btrfs_work *work)
660 {
661 	set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
662 }
663 
664 /*
665  * places a struct btrfs_work into the pending queue of one of the kthreads
666  */
667 void btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
668 {
669 	struct btrfs_worker_thread *worker;
670 	unsigned long flags;
671 	int wake = 0;
672 
673 	/* don't requeue something already on a list */
674 	if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
675 		return;
676 
677 	worker = find_worker(workers);
678 	if (workers->ordered) {
679 		/*
680 		 * you're not allowed to do ordered queues from an
681 		 * interrupt handler
682 		 */
683 		spin_lock(&workers->order_lock);
684 		if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags)) {
685 			list_add_tail(&work->order_list,
686 				      &workers->prio_order_list);
687 		} else {
688 			list_add_tail(&work->order_list, &workers->order_list);
689 		}
690 		spin_unlock(&workers->order_lock);
691 	} else {
692 		INIT_LIST_HEAD(&work->order_list);
693 	}
694 
695 	spin_lock_irqsave(&worker->lock, flags);
696 
697 	if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
698 		list_add_tail(&work->list, &worker->prio_pending);
699 	else
700 		list_add_tail(&work->list, &worker->pending);
701 	check_busy_worker(worker);
702 
703 	/*
704 	 * avoid calling into wake_up_process if this thread has already
705 	 * been kicked
706 	 */
707 	if (!worker->working)
708 		wake = 1;
709 	worker->working = 1;
710 
711 	if (wake)
712 		wake_up_process(worker->task);
713 	spin_unlock_irqrestore(&worker->lock, flags);
714 }
715