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 list_move(&worker->worker_list,
112 &worker->workers->idle_list);
114 spin_unlock_irqrestore(&worker->workers->lock, flags);
119 * helper function to move a thread off the idle list after new
120 * pending work is added.
122 static void check_busy_worker(struct btrfs_worker_thread *worker)
124 if (worker->idle && atomic_read(&worker->num_pending) >=
125 worker->workers->idle_thresh) {
127 spin_lock_irqsave(&worker->workers->lock, flags);
130 if (!list_empty(&worker->worker_list)) {
131 list_move_tail(&worker->worker_list,
132 &worker->workers->worker_list);
134 spin_unlock_irqrestore(&worker->workers->lock, flags);
138 static void check_pending_worker_creates(struct btrfs_worker_thread *worker)
140 struct btrfs_workers *workers = worker->workers;
141 struct worker_start *start;
145 if (!workers->atomic_start_pending)
148 start = kzalloc(sizeof(*start), GFP_NOFS);
152 start->work.func = start_new_worker_func;
153 start->queue = workers;
155 spin_lock_irqsave(&workers->lock, flags);
156 if (!workers->atomic_start_pending)
159 workers->atomic_start_pending = 0;
160 if (workers->num_workers + workers->num_workers_starting >=
161 workers->max_workers)
164 workers->num_workers_starting += 1;
165 spin_unlock_irqrestore(&workers->lock, flags);
166 btrfs_queue_worker(workers->atomic_worker_start, &start->work);
171 spin_unlock_irqrestore(&workers->lock, flags);
174 static noinline void run_ordered_completions(struct btrfs_workers *workers,
175 struct btrfs_work *work)
177 if (!workers->ordered)
180 set_bit(WORK_DONE_BIT, &work->flags);
182 spin_lock(&workers->order_lock);
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);
194 if (!test_bit(WORK_DONE_BIT, &work->flags))
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
202 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
205 spin_unlock(&workers->order_lock);
207 work->ordered_func(work);
209 /* now take the lock again and call the freeing code */
210 spin_lock(&workers->order_lock);
211 list_del(&work->order_list);
212 work->ordered_free(work);
215 spin_unlock(&workers->order_lock);
218 static void put_worker(struct btrfs_worker_thread *worker)
220 if (atomic_dec_and_test(&worker->refs))
224 static int try_worker_shutdown(struct btrfs_worker_thread *worker)
228 spin_lock_irq(&worker->lock);
229 spin_lock(&worker->workers->lock);
230 if (worker->workers->num_workers > 1 &&
233 !list_empty(&worker->worker_list) &&
234 list_empty(&worker->prio_pending) &&
235 list_empty(&worker->pending) &&
236 atomic_read(&worker->num_pending) == 0) {
238 list_del_init(&worker->worker_list);
239 worker->workers->num_workers--;
241 spin_unlock(&worker->workers->lock);
242 spin_unlock_irq(&worker->lock);
249 static struct btrfs_work *get_next_work(struct btrfs_worker_thread *worker,
250 struct list_head *prio_head,
251 struct list_head *head)
253 struct btrfs_work *work = NULL;
254 struct list_head *cur = NULL;
256 if(!list_empty(prio_head))
257 cur = prio_head->next;
260 if (!list_empty(&worker->prio_pending))
263 if (!list_empty(head))
270 spin_lock_irq(&worker->lock);
271 list_splice_tail_init(&worker->prio_pending, prio_head);
272 list_splice_tail_init(&worker->pending, head);
274 if (!list_empty(prio_head))
275 cur = prio_head->next;
276 else if (!list_empty(head))
278 spin_unlock_irq(&worker->lock);
284 work = list_entry(cur, struct btrfs_work, list);
291 * main loop for servicing work items
293 static int worker_loop(void *arg)
295 struct btrfs_worker_thread *worker = arg;
296 struct list_head head;
297 struct list_head prio_head;
298 struct btrfs_work *work;
300 INIT_LIST_HEAD(&head);
301 INIT_LIST_HEAD(&prio_head);
308 work = get_next_work(worker, &prio_head, &head);
312 list_del(&work->list);
313 clear_bit(WORK_QUEUED_BIT, &work->flags);
315 work->worker = worker;
319 atomic_dec(&worker->num_pending);
321 * unless this is an ordered work queue,
322 * 'work' was probably freed by func above.
324 run_ordered_completions(worker->workers, work);
326 check_pending_worker_creates(worker);
330 spin_lock_irq(&worker->lock);
331 check_idle_worker(worker);
333 if (freezing(current)) {
335 spin_unlock_irq(&worker->lock);
338 spin_unlock_irq(&worker->lock);
339 if (!kthread_should_stop()) {
342 * we've dropped the lock, did someone else
346 if (!list_empty(&worker->pending) ||
347 !list_empty(&worker->prio_pending))
351 * this short schedule allows more work to
352 * come in without the queue functions
353 * needing to go through wake_up_process()
355 * worker->working is still 1, so nobody
356 * is going to try and wake us up
360 if (!list_empty(&worker->pending) ||
361 !list_empty(&worker->prio_pending))
364 if (kthread_should_stop())
367 /* still no more work?, sleep for real */
368 spin_lock_irq(&worker->lock);
369 set_current_state(TASK_INTERRUPTIBLE);
370 if (!list_empty(&worker->pending) ||
371 !list_empty(&worker->prio_pending)) {
372 spin_unlock_irq(&worker->lock);
373 set_current_state(TASK_RUNNING);
378 * this makes sure we get a wakeup when someone
379 * adds something new to the queue
382 spin_unlock_irq(&worker->lock);
384 if (!kthread_should_stop()) {
385 schedule_timeout(HZ * 120);
386 if (!worker->working &&
387 try_worker_shutdown(worker)) {
392 __set_current_state(TASK_RUNNING);
394 } while (!kthread_should_stop());
399 * this will wait for all the worker threads to shutdown
401 void btrfs_stop_workers(struct btrfs_workers *workers)
403 struct list_head *cur;
404 struct btrfs_worker_thread *worker;
407 spin_lock_irq(&workers->lock);
408 list_splice_init(&workers->idle_list, &workers->worker_list);
409 while (!list_empty(&workers->worker_list)) {
410 cur = workers->worker_list.next;
411 worker = list_entry(cur, struct btrfs_worker_thread,
414 atomic_inc(&worker->refs);
415 workers->num_workers -= 1;
416 if (!list_empty(&worker->worker_list)) {
417 list_del_init(&worker->worker_list);
422 spin_unlock_irq(&workers->lock);
424 kthread_stop(worker->task);
425 spin_lock_irq(&workers->lock);
428 spin_unlock_irq(&workers->lock);
432 * simple init on struct btrfs_workers
434 void btrfs_init_workers(struct btrfs_workers *workers, char *name, int max,
435 struct btrfs_workers *async_helper)
437 workers->num_workers = 0;
438 workers->num_workers_starting = 0;
439 INIT_LIST_HEAD(&workers->worker_list);
440 INIT_LIST_HEAD(&workers->idle_list);
441 INIT_LIST_HEAD(&workers->order_list);
442 INIT_LIST_HEAD(&workers->prio_order_list);
443 spin_lock_init(&workers->lock);
444 spin_lock_init(&workers->order_lock);
445 workers->max_workers = max;
446 workers->idle_thresh = 32;
447 workers->name = name;
448 workers->ordered = 0;
449 workers->atomic_start_pending = 0;
450 workers->atomic_worker_start = async_helper;
454 * starts new worker threads. This does not enforce the max worker
455 * count in case you need to temporarily go past it.
457 static int __btrfs_start_workers(struct btrfs_workers *workers)
459 struct btrfs_worker_thread *worker;
462 worker = kzalloc(sizeof(*worker), GFP_NOFS);
468 INIT_LIST_HEAD(&worker->pending);
469 INIT_LIST_HEAD(&worker->prio_pending);
470 INIT_LIST_HEAD(&worker->worker_list);
471 spin_lock_init(&worker->lock);
473 atomic_set(&worker->num_pending, 0);
474 atomic_set(&worker->refs, 1);
475 worker->workers = workers;
476 worker->task = kthread_run(worker_loop, worker,
477 "btrfs-%s-%d", workers->name,
478 workers->num_workers + 1);
479 if (IS_ERR(worker->task)) {
480 ret = PTR_ERR(worker->task);
484 spin_lock_irq(&workers->lock);
485 list_add_tail(&worker->worker_list, &workers->idle_list);
487 workers->num_workers++;
488 workers->num_workers_starting--;
489 WARN_ON(workers->num_workers_starting < 0);
490 spin_unlock_irq(&workers->lock);
494 spin_lock_irq(&workers->lock);
495 workers->num_workers_starting--;
496 spin_unlock_irq(&workers->lock);
500 int btrfs_start_workers(struct btrfs_workers *workers)
502 spin_lock_irq(&workers->lock);
503 workers->num_workers_starting++;
504 spin_unlock_irq(&workers->lock);
505 return __btrfs_start_workers(workers);
509 * run through the list and find a worker thread that doesn't have a lot
510 * to do right now. This can return null if we aren't yet at the thread
511 * count limit and all of the threads are busy.
513 static struct btrfs_worker_thread *next_worker(struct btrfs_workers *workers)
515 struct btrfs_worker_thread *worker;
516 struct list_head *next;
519 enforce_min = (workers->num_workers + workers->num_workers_starting) <
520 workers->max_workers;
523 * if we find an idle thread, don't move it to the end of the
524 * idle list. This improves the chance that the next submission
525 * will reuse the same thread, and maybe catch it while it is still
528 if (!list_empty(&workers->idle_list)) {
529 next = workers->idle_list.next;
530 worker = list_entry(next, struct btrfs_worker_thread,
534 if (enforce_min || list_empty(&workers->worker_list))
538 * if we pick a busy task, move the task to the end of the list.
539 * hopefully this will keep things somewhat evenly balanced.
540 * Do the move in batches based on the sequence number. This groups
541 * requests submitted at roughly the same time onto the same worker.
543 next = workers->worker_list.next;
544 worker = list_entry(next, struct btrfs_worker_thread, worker_list);
547 if (worker->sequence % workers->idle_thresh == 0)
548 list_move_tail(next, &workers->worker_list);
553 * selects a worker thread to take the next job. This will either find
554 * an idle worker, start a new worker up to the max count, or just return
555 * one of the existing busy workers.
557 static struct btrfs_worker_thread *find_worker(struct btrfs_workers *workers)
559 struct btrfs_worker_thread *worker;
561 struct list_head *fallback;
564 spin_lock_irqsave(&workers->lock, flags);
566 worker = next_worker(workers);
569 if (workers->num_workers + workers->num_workers_starting >=
570 workers->max_workers) {
572 } else if (workers->atomic_worker_start) {
573 workers->atomic_start_pending = 1;
576 workers->num_workers_starting++;
577 spin_unlock_irqrestore(&workers->lock, flags);
578 /* we're below the limit, start another worker */
579 ret = __btrfs_start_workers(workers);
580 spin_lock_irqsave(&workers->lock, flags);
591 * we have failed to find any workers, just
592 * return the first one we can find.
594 if (!list_empty(&workers->worker_list))
595 fallback = workers->worker_list.next;
596 if (!list_empty(&workers->idle_list))
597 fallback = workers->idle_list.next;
599 worker = list_entry(fallback,
600 struct btrfs_worker_thread, worker_list);
603 * this makes sure the worker doesn't exit before it is placed
604 * onto a busy/idle list
606 atomic_inc(&worker->num_pending);
607 spin_unlock_irqrestore(&workers->lock, flags);
612 * btrfs_requeue_work just puts the work item back on the tail of the list
613 * it was taken from. It is intended for use with long running work functions
614 * that make some progress and want to give the cpu up for others.
616 void btrfs_requeue_work(struct btrfs_work *work)
618 struct btrfs_worker_thread *worker = work->worker;
622 if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
625 spin_lock_irqsave(&worker->lock, flags);
626 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
627 list_add_tail(&work->list, &worker->prio_pending);
629 list_add_tail(&work->list, &worker->pending);
630 atomic_inc(&worker->num_pending);
632 /* by definition we're busy, take ourselves off the idle
636 spin_lock(&worker->workers->lock);
638 list_move_tail(&worker->worker_list,
639 &worker->workers->worker_list);
640 spin_unlock(&worker->workers->lock);
642 if (!worker->working) {
648 wake_up_process(worker->task);
649 spin_unlock_irqrestore(&worker->lock, flags);
652 void btrfs_set_work_high_prio(struct btrfs_work *work)
654 set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
658 * places a struct btrfs_work into the pending queue of one of the kthreads
660 void btrfs_queue_worker(struct btrfs_workers *workers, struct btrfs_work *work)
662 struct btrfs_worker_thread *worker;
666 /* don't requeue something already on a list */
667 if (test_and_set_bit(WORK_QUEUED_BIT, &work->flags))
670 worker = find_worker(workers);
671 if (workers->ordered) {
673 * you're not allowed to do ordered queues from an
676 spin_lock(&workers->order_lock);
677 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags)) {
678 list_add_tail(&work->order_list,
679 &workers->prio_order_list);
681 list_add_tail(&work->order_list, &workers->order_list);
683 spin_unlock(&workers->order_lock);
685 INIT_LIST_HEAD(&work->order_list);
688 spin_lock_irqsave(&worker->lock, flags);
690 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags))
691 list_add_tail(&work->list, &worker->prio_pending);
693 list_add_tail(&work->list, &worker->pending);
694 check_busy_worker(worker);
697 * avoid calling into wake_up_process if this thread has already
700 if (!worker->working)
705 wake_up_process(worker->task);
706 spin_unlock_irqrestore(&worker->lock, flags);