2 * Copyright (C) 2007 Oracle. All rights reserved.
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.
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.
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.
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"
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
32 * container for the kthread task pointer and the list of pending work
33 * One of these is allocated per thread.
35 struct btrfs_worker_thread {
36 /* pool we belong to */
37 struct btrfs_workers *workers;
39 /* list of struct btrfs_work that are waiting for service */
40 struct list_head pending;
41 struct list_head prio_pending;
43 /* list of worker threads from struct btrfs_workers */
44 struct list_head worker_list;
47 struct task_struct *task;
49 /* number of things on the pending list */
52 /* reference counter for this struct */
55 unsigned long sequence;
57 /* protects the pending list. */
60 /* set to non-zero when this thread is already awake and kicking */
63 /* are we currently idle */
67 static int __btrfs_start_workers(struct btrfs_workers *workers);
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
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.
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
85 struct btrfs_work work;
86 struct btrfs_workers *queue;
89 static void start_new_worker_func(struct btrfs_work *work)
91 struct worker_start *start;
92 start = container_of(work, struct worker_start, work);
93 __btrfs_start_workers(start->queue);
98 * helper function to move a thread onto the idle list after it
99 * has finished some requests.
101 static void check_idle_worker(struct btrfs_worker_thread *worker)
103 if (!worker->idle && atomic_read(&worker->num_pending) <
104 worker->workers->idle_thresh / 2) {
106 spin_lock_irqsave(&worker->workers->lock, flags);
109 /* the list may be empty if the worker is just starting */
110 if (!list_empty(&worker->worker_list) &&
111 !worker->workers->stopping) {
112 list_move(&worker->worker_list,
113 &worker->workers->idle_list);
115 spin_unlock_irqrestore(&worker->workers->lock, flags);
120 * helper function to move a thread off the idle list after new
121 * pending work is added.
123 static void check_busy_worker(struct btrfs_worker_thread *worker)
125 if (worker->idle && atomic_read(&worker->num_pending) >=
126 worker->workers->idle_thresh) {
128 spin_lock_irqsave(&worker->workers->lock, flags);
131 if (!list_empty(&worker->worker_list) &&
132 !worker->workers->stopping) {
133 list_move_tail(&worker->worker_list,
134 &worker->workers->worker_list);
136 spin_unlock_irqrestore(&worker->workers->lock, flags);
140 static void check_pending_worker_creates(struct btrfs_worker_thread *worker)
142 struct btrfs_workers *workers = worker->workers;
143 struct worker_start *start;
147 if (!workers->atomic_start_pending)
150 start = kzalloc(sizeof(*start), GFP_NOFS);
154 start->work.func = start_new_worker_func;
155 start->queue = workers;
157 spin_lock_irqsave(&workers->lock, flags);
158 if (!workers->atomic_start_pending)
161 workers->atomic_start_pending = 0;
162 if (workers->num_workers + workers->num_workers_starting >=
163 workers->max_workers)
166 workers->num_workers_starting += 1;
167 spin_unlock_irqrestore(&workers->lock, flags);
168 btrfs_queue_worker(workers->atomic_worker_start, &start->work);
173 spin_unlock_irqrestore(&workers->lock, flags);
176 static noinline void run_ordered_completions(struct btrfs_workers *workers,
177 struct btrfs_work *work)
179 if (!workers->ordered)
182 set_bit(WORK_DONE_BIT, &work->flags);
184 spin_lock(&workers->order_lock);
187 if (!list_empty(&workers->prio_order_list)) {
188 work = list_entry(workers->prio_order_list.next,
189 struct btrfs_work, order_list);
190 } else if (!list_empty(&workers->order_list)) {
191 work = list_entry(workers->order_list.next,
192 struct btrfs_work, order_list);
196 if (!test_bit(WORK_DONE_BIT, &work->flags))
199 /* we are going to call the ordered done function, but
200 * we leave the work item on the list as a barrier so
201 * that later work items that are done don't have their
202 * functions called before this one returns
204 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
207 spin_unlock(&workers->order_lock);
209 work->ordered_func(work);
211 /* now take the lock again and drop our item from the list */
212 spin_lock(&workers->order_lock);
213 list_del(&work->order_list);
214 spin_unlock(&workers->order_lock);
217 * we don't want to call the ordered free functions
218 * with the lock held though
220 work->ordered_free(work);
221 spin_lock(&workers->order_lock);
224 spin_unlock(&workers->order_lock);
227 static void put_worker(struct btrfs_worker_thread *worker)
229 if (atomic_dec_and_test(&worker->refs))
233 static int try_worker_shutdown(struct btrfs_worker_thread *worker)
237 spin_lock_irq(&worker->lock);
238 spin_lock(&worker->workers->lock);
239 if (worker->workers->num_workers > 1 &&
242 !list_empty(&worker->worker_list) &&
243 list_empty(&worker->prio_pending) &&
244 list_empty(&worker->pending) &&
245 atomic_read(&worker->num_pending) == 0) {
247 list_del_init(&worker->worker_list);
248 worker->workers->num_workers--;
250 spin_unlock(&worker->workers->lock);
251 spin_unlock_irq(&worker->lock);
258 static struct btrfs_work *get_next_work(struct btrfs_worker_thread *worker,
259 struct list_head *prio_head,
260 struct list_head *head)
262 struct btrfs_work *work = NULL;
263 struct list_head *cur = NULL;
265 if (!list_empty(prio_head))
266 cur = prio_head->next;
269 if (!list_empty(&worker->prio_pending))
272 if (!list_empty(head))
279 spin_lock_irq(&worker->lock);
280 list_splice_tail_init(&worker->prio_pending, prio_head);
281 list_splice_tail_init(&worker->pending, head);
283 if (!list_empty(prio_head))
284 cur = prio_head->next;
285 else if (!list_empty(head))
287 spin_unlock_irq(&worker->lock);
293 work = list_entry(cur, struct btrfs_work, list);
300 * main loop for servicing work items
302 static int worker_loop(void *arg)
304 struct btrfs_worker_thread *worker = arg;
305 struct list_head head;
306 struct list_head prio_head;
307 struct btrfs_work *work;
309 INIT_LIST_HEAD(&head);
310 INIT_LIST_HEAD(&prio_head);
317 work = get_next_work(worker, &prio_head, &head);
321 list_del(&work->list);
322 clear_bit(WORK_QUEUED_BIT, &work->flags);
324 work->worker = worker;
328 atomic_dec(&worker->num_pending);
330 * unless this is an ordered work queue,
331 * 'work' was probably freed by func above.
333 run_ordered_completions(worker->workers, work);
335 check_pending_worker_creates(worker);
339 spin_lock_irq(&worker->lock);
340 check_idle_worker(worker);
342 if (freezing(current)) {
344 spin_unlock_irq(&worker->lock);
347 spin_unlock_irq(&worker->lock);
348 if (!kthread_should_stop()) {
351 * we've dropped the lock, did someone else
355 if (!list_empty(&worker->pending) ||
356 !list_empty(&worker->prio_pending))
360 * this short schedule allows more work to
361 * come in without the queue functions
362 * needing to go through wake_up_process()
364 * worker->working is still 1, so nobody
365 * is going to try and wake us up
369 if (!list_empty(&worker->pending) ||
370 !list_empty(&worker->prio_pending))
373 if (kthread_should_stop())
376 /* still no more work?, sleep for real */
377 spin_lock_irq(&worker->lock);
378 set_current_state(TASK_INTERRUPTIBLE);
379 if (!list_empty(&worker->pending) ||
380 !list_empty(&worker->prio_pending)) {
381 spin_unlock_irq(&worker->lock);
382 set_current_state(TASK_RUNNING);
387 * this makes sure we get a wakeup when someone
388 * adds something new to the queue
391 spin_unlock_irq(&worker->lock);
393 if (!kthread_should_stop()) {
394 schedule_timeout(HZ * 120);
395 if (!worker->working &&
396 try_worker_shutdown(worker)) {
401 __set_current_state(TASK_RUNNING);
403 } while (!kthread_should_stop());
408 * this will wait for all the worker threads to shutdown
410 void btrfs_stop_workers(struct btrfs_workers *workers)
412 struct list_head *cur;
413 struct btrfs_worker_thread *worker;
416 spin_lock_irq(&workers->lock);
417 workers->stopping = 1;
418 list_splice_init(&workers->idle_list, &workers->worker_list);
419 while (!list_empty(&workers->worker_list)) {
420 cur = workers->worker_list.next;
421 worker = list_entry(cur, struct btrfs_worker_thread,
424 atomic_inc(&worker->refs);
425 workers->num_workers -= 1;
426 if (!list_empty(&worker->worker_list)) {
427 list_del_init(&worker->worker_list);
432 spin_unlock_irq(&workers->lock);
434 kthread_stop(worker->task);
435 spin_lock_irq(&workers->lock);
438 spin_unlock_irq(&workers->lock);
442 * simple init on struct btrfs_workers
444 void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max,
445 struct btrfs_workers *async_helper)
447 workers->num_workers = 0;
448 workers->num_workers_starting = 0;
449 INIT_LIST_HEAD(&workers->worker_list);
450 INIT_LIST_HEAD(&workers->idle_list);
451 INIT_LIST_HEAD(&workers->order_list);
452 INIT_LIST_HEAD(&workers->prio_order_list);
453 spin_lock_init(&workers->lock);
454 spin_lock_init(&workers->order_lock);
455 workers->max_workers = max;
456 workers->idle_thresh = 32;
457 workers->name = name;
458 workers->ordered = 0;
459 workers->atomic_start_pending = 0;
460 workers->atomic_worker_start = async_helper;
461 workers->stopping = 0;
465 * starts new worker threads. This does not enforce the max worker
466 * count in case you need to temporarily go past it.
468 static int __btrfs_start_workers(struct btrfs_workers *workers)
470 struct btrfs_worker_thread *worker;
473 worker = kzalloc(sizeof(*worker), GFP_NOFS);
479 INIT_LIST_HEAD(&worker->pending);
480 INIT_LIST_HEAD(&worker->prio_pending);
481 INIT_LIST_HEAD(&worker->worker_list);
482 spin_lock_init(&worker->lock);
484 atomic_set(&worker->num_pending, 0);
485 atomic_set(&worker->refs, 1);
486 worker->workers = workers;
487 worker->task = kthread_create(worker_loop, worker,
488 "btrfs-%s-%d", workers->name,
489 workers->num_workers + 1);
490 if (IS_ERR(worker->task)) {
491 ret = PTR_ERR(worker->task);
495 spin_lock_irq(&workers->lock);
496 if (workers->stopping) {
497 spin_unlock_irq(&workers->lock);
501 list_add_tail(&worker->worker_list, &workers->idle_list);
503 workers->num_workers++;
504 workers->num_workers_starting--;
505 WARN_ON(workers->num_workers_starting < 0);
506 spin_unlock_irq(&workers->lock);
508 wake_up_process(worker->task);
512 kthread_stop(worker->task);
515 spin_lock_irq(&workers->lock);
516 workers->num_workers_starting--;
517 spin_unlock_irq(&workers->lock);
521 int btrfs_start_workers(struct btrfs_workers *workers)
523 spin_lock_irq(&workers->lock);
524 workers->num_workers_starting++;
525 spin_unlock_irq(&workers->lock);
526 return __btrfs_start_workers(workers);
530 * run through the list and find a worker thread that doesn't have a lot
531 * to do right now. This can return null if we aren't yet at the thread
532 * count limit and all of the threads are busy.
534 static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
536 struct btrfs_worker_thread *worker;
537 struct list_head *next;
540 enforce_min = (workers->num_workers + workers->num_workers_starting) <
541 workers->max_workers;
544 * if we find an idle thread, don't move it to the end of the
545 * idle list. This improves the chance that the next submission
546 * will reuse the same thread, and maybe catch it while it is still
549 if (!list_empty(&workers->idle_list)) {
550 next = workers->idle_list.next;
551 worker = list_entry(next, struct btrfs_worker_thread,
555 if (enforce_min || list_empty(&workers->worker_list))
559 * if we pick a busy task, move the task to the end of the list.
560 * hopefully this will keep things somewhat evenly balanced.
561 * Do the move in batches based on the sequence number. This groups
562 * requests submitted at roughly the same time onto the same worker.
564 next = workers->worker_list.next;
565 worker = list_entry(next, struct btrfs_worker_thread, worker_list);
568 if (worker->sequence % workers->idle_thresh == 0)
569 list_move_tail(next, &workers->worker_list);
574 * selects a worker thread to take the next job. This will either find
575 * an idle worker, start a new worker up to the max count, or just return
576 * one of the existing busy workers.
578 static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
580 struct btrfs_worker_thread *worker;
582 struct list_head *fallback;
585 spin_lock_irqsave(&workers->lock, flags);
587 worker = next_worker(workers);
590 if (workers->num_workers + workers->num_workers_starting >=
591 workers->max_workers) {
593 } else if (workers->atomic_worker_start) {
594 workers->atomic_start_pending = 1;
597 workers->num_workers_starting++;
598 spin_unlock_irqrestore(&workers->lock, flags);
599 /* we're below the limit, start another worker */
600 ret = __btrfs_start_workers(workers);
601 spin_lock_irqsave(&workers->lock, flags);
612 * we have failed to find any workers, just
613 * return the first one we can find.
615 if (!list_empty(&workers->worker_list))
616 fallback = workers->worker_list.next;
617 if (!list_empty(&workers->idle_list))
618 fallback = workers->idle_list.next;
620 worker = list_entry(fallback,
621 struct btrfs_worker_thread, worker_list);
624 * this makes sure the worker doesn't exit before it is placed
625 * onto a busy/idle list
627 atomic_inc(&worker->num_pending);
628 spin_unlock_irqrestore(&workers->lock, flags);
633 * btrfs_requeue_work just puts the work item back on the tail of the list
634 * it was taken from. It is intended for use with long running work functions
635 * that make some progress and want to give the cpu up for others.
637 void btrfs_requeue_work(struct btrfs_work *work)
639 struct btrfs_worker_thread *worker = work->worker;
643 if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
646 spin_lock_irqsave(&worker->lock, flags);
647 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
648 list_add_tail(&work->list, &worker->prio_pending);
650 list_add_tail(&work->list, &worker->pending);
651 atomic_inc(&worker->num_pending);
653 /* by definition we're busy, take ourselves off the idle
657 spin_lock(&worker->workers->lock);
659 list_move_tail(&worker->worker_list,
660 &worker->workers->worker_list);
661 spin_unlock(&worker->workers->lock);
663 if (!worker->working) {
669 wake_up_process(worker->task);
670 spin_unlock_irqrestore(&worker->lock, flags);
673 void btrfs_set_work_high_prio(struct btrfs_work *work)
675 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
679 * places a struct btrfs_work into the pending queue of one of the kthreads
681 void btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
683 struct btrfs_worker_thread *worker;
687 /* don't requeue something already on a list */
688 if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
691 worker = find_worker(workers);
692 if (workers->ordered) {
694 * you're not allowed to do ordered queues from an
697 spin_lock(&workers->order_lock);
698 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags)) {
699 list_add_tail(&work->order_list,
700 &workers->prio_order_list);
702 list_add_tail(&work->order_list, &workers->order_list);
704 spin_unlock(&workers->order_lock);
706 INIT_LIST_HEAD(&work->order_list);
709 spin_lock_irqsave(&worker->lock, flags);
711 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
712 list_add_tail(&work->list, &worker->prio_pending);
714 list_add_tail(&work->list, &worker->pending);
715 check_busy_worker(worker);
718 * avoid calling into wake_up_process if this thread has already
721 if (!worker->working)
726 wake_up_process(worker->task);
727 spin_unlock_irqrestore(&worker->lock, flags);