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 drop our item from the list */
210 spin_lock(&workers
->order_lock
);
211 list_del(&work
->order_list
);
212 spin_unlock(&workers
->order_lock
);
215 * we don't want to call the ordered free functions
216 * with the lock held though
218 work
->ordered_free(work
);
219 spin_lock(&workers
->order_lock
);
222 spin_unlock(&workers
->order_lock
);
225 static void put_worker(struct btrfs_worker_thread
*worker
)
227 if (atomic_dec_and_test(&worker
->refs
))
231 static int try_worker_shutdown(struct btrfs_worker_thread
*worker
)
235 spin_lock_irq(&worker
->lock
);
236 spin_lock(&worker
->workers
->lock
);
237 if (worker
->workers
->num_workers
> 1 &&
240 !list_empty(&worker
->worker_list
) &&
241 list_empty(&worker
->prio_pending
) &&
242 list_empty(&worker
->pending
) &&
243 atomic_read(&worker
->num_pending
) == 0) {
245 list_del_init(&worker
->worker_list
);
246 worker
->workers
->num_workers
--;
248 spin_unlock(&worker
->workers
->lock
);
249 spin_unlock_irq(&worker
->lock
);
256 static struct btrfs_work
*get_next_work(struct btrfs_worker_thread
*worker
,
257 struct list_head
*prio_head
,
258 struct list_head
*head
)
260 struct btrfs_work
*work
= NULL
;
261 struct list_head
*cur
= NULL
;
263 if(!list_empty(prio_head
))
264 cur
= prio_head
->next
;
267 if (!list_empty(&worker
->prio_pending
))
270 if (!list_empty(head
))
277 spin_lock_irq(&worker
->lock
);
278 list_splice_tail_init(&worker
->prio_pending
, prio_head
);
279 list_splice_tail_init(&worker
->pending
, head
);
281 if (!list_empty(prio_head
))
282 cur
= prio_head
->next
;
283 else if (!list_empty(head
))
285 spin_unlock_irq(&worker
->lock
);
291 work
= list_entry(cur
, struct btrfs_work
, list
);
298 * main loop for servicing work items
300 static int worker_loop(void *arg
)
302 struct btrfs_worker_thread
*worker
= arg
;
303 struct list_head head
;
304 struct list_head prio_head
;
305 struct btrfs_work
*work
;
307 INIT_LIST_HEAD(&head
);
308 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_freezable_should_stop(NULL
)) {
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_freezable_should_stop(NULL
))
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_freezable_should_stop(NULL
)) {
394 schedule_timeout(HZ
* 120);
395 if (!worker
->working
&&
396 try_worker_shutdown(worker
)) {
401 __set_current_state(TASK_RUNNING
);
403 } while (!kthread_freezable_should_stop(NULL
));
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 list_splice_init(&workers
->idle_list
, &workers
->worker_list
);
418 while (!list_empty(&workers
->worker_list
)) {
419 cur
= workers
->worker_list
.next
;
420 worker
= list_entry(cur
, struct btrfs_worker_thread
,
423 atomic_inc(&worker
->refs
);
424 workers
->num_workers
-= 1;
425 if (!list_empty(&worker
->worker_list
)) {
426 list_del_init(&worker
->worker_list
);
431 spin_unlock_irq(&workers
->lock
);
433 kthread_stop(worker
->task
);
434 spin_lock_irq(&workers
->lock
);
437 spin_unlock_irq(&workers
->lock
);
441 * simple init on struct btrfs_workers
443 void btrfs_init_workers(struct btrfs_workers
*workers
, char *name
, int max
,
444 struct btrfs_workers
*async_helper
)
446 workers
->num_workers
= 0;
447 workers
->num_workers_starting
= 0;
448 INIT_LIST_HEAD(&workers
->worker_list
);
449 INIT_LIST_HEAD(&workers
->idle_list
);
450 INIT_LIST_HEAD(&workers
->order_list
);
451 INIT_LIST_HEAD(&workers
->prio_order_list
);
452 spin_lock_init(&workers
->lock
);
453 spin_lock_init(&workers
->order_lock
);
454 workers
->max_workers
= max
;
455 workers
->idle_thresh
= 32;
456 workers
->name
= name
;
457 workers
->ordered
= 0;
458 workers
->atomic_start_pending
= 0;
459 workers
->atomic_worker_start
= async_helper
;
463 * starts new worker threads. This does not enforce the max worker
464 * count in case you need to temporarily go past it.
466 static int __btrfs_start_workers(struct btrfs_workers
*workers
)
468 struct btrfs_worker_thread
*worker
;
471 worker
= kzalloc(sizeof(*worker
), GFP_NOFS
);
477 INIT_LIST_HEAD(&worker
->pending
);
478 INIT_LIST_HEAD(&worker
->prio_pending
);
479 INIT_LIST_HEAD(&worker
->worker_list
);
480 spin_lock_init(&worker
->lock
);
482 atomic_set(&worker
->num_pending
, 0);
483 atomic_set(&worker
->refs
, 1);
484 worker
->workers
= workers
;
485 worker
->task
= kthread_run(worker_loop
, worker
,
486 "btrfs-%s-%d", workers
->name
,
487 workers
->num_workers
+ 1);
488 if (IS_ERR(worker
->task
)) {
489 ret
= PTR_ERR(worker
->task
);
493 spin_lock_irq(&workers
->lock
);
494 list_add_tail(&worker
->worker_list
, &workers
->idle_list
);
496 workers
->num_workers
++;
497 workers
->num_workers_starting
--;
498 WARN_ON(workers
->num_workers_starting
< 0);
499 spin_unlock_irq(&workers
->lock
);
503 spin_lock_irq(&workers
->lock
);
504 workers
->num_workers_starting
--;
505 spin_unlock_irq(&workers
->lock
);
509 int btrfs_start_workers(struct btrfs_workers
*workers
)
511 spin_lock_irq(&workers
->lock
);
512 workers
->num_workers_starting
++;
513 spin_unlock_irq(&workers
->lock
);
514 return __btrfs_start_workers(workers
);
518 * run through the list and find a worker thread that doesn't have a lot
519 * to do right now. This can return null if we aren't yet at the thread
520 * count limit and all of the threads are busy.
522 static struct btrfs_worker_thread
*next_worker(struct btrfs_workers
*workers
)
524 struct btrfs_worker_thread
*worker
;
525 struct list_head
*next
;
528 enforce_min
= (workers
->num_workers
+ workers
->num_workers_starting
) <
529 workers
->max_workers
;
532 * if we find an idle thread, don't move it to the end of the
533 * idle list. This improves the chance that the next submission
534 * will reuse the same thread, and maybe catch it while it is still
537 if (!list_empty(&workers
->idle_list
)) {
538 next
= workers
->idle_list
.next
;
539 worker
= list_entry(next
, struct btrfs_worker_thread
,
543 if (enforce_min
|| list_empty(&workers
->worker_list
))
547 * if we pick a busy task, move the task to the end of the list.
548 * hopefully this will keep things somewhat evenly balanced.
549 * Do the move in batches based on the sequence number. This groups
550 * requests submitted at roughly the same time onto the same worker.
552 next
= workers
->worker_list
.next
;
553 worker
= list_entry(next
, struct btrfs_worker_thread
, worker_list
);
556 if (worker
->sequence
% workers
->idle_thresh
== 0)
557 list_move_tail(next
, &workers
->worker_list
);
562 * selects a worker thread to take the next job. This will either find
563 * an idle worker, start a new worker up to the max count, or just return
564 * one of the existing busy workers.
566 static struct btrfs_worker_thread
*find_worker(struct btrfs_workers
*workers
)
568 struct btrfs_worker_thread
*worker
;
570 struct list_head
*fallback
;
573 spin_lock_irqsave(&workers
->lock
, flags
);
575 worker
= next_worker(workers
);
578 if (workers
->num_workers
+ workers
->num_workers_starting
>=
579 workers
->max_workers
) {
581 } else if (workers
->atomic_worker_start
) {
582 workers
->atomic_start_pending
= 1;
585 workers
->num_workers_starting
++;
586 spin_unlock_irqrestore(&workers
->lock
, flags
);
587 /* we're below the limit, start another worker */
588 ret
= __btrfs_start_workers(workers
);
589 spin_lock_irqsave(&workers
->lock
, flags
);
600 * we have failed to find any workers, just
601 * return the first one we can find.
603 if (!list_empty(&workers
->worker_list
))
604 fallback
= workers
->worker_list
.next
;
605 if (!list_empty(&workers
->idle_list
))
606 fallback
= workers
->idle_list
.next
;
608 worker
= list_entry(fallback
,
609 struct btrfs_worker_thread
, worker_list
);
612 * this makes sure the worker doesn't exit before it is placed
613 * onto a busy/idle list
615 atomic_inc(&worker
->num_pending
);
616 spin_unlock_irqrestore(&workers
->lock
, flags
);
621 * btrfs_requeue_work just puts the work item back on the tail of the list
622 * it was taken from. It is intended for use with long running work functions
623 * that make some progress and want to give the cpu up for others.
625 void btrfs_requeue_work(struct btrfs_work
*work
)
627 struct btrfs_worker_thread
*worker
= work
->worker
;
631 if (test_and_set_bit(WORK_QUEUED_BIT
, &work
->flags
))
634 spin_lock_irqsave(&worker
->lock
, flags
);
635 if (test_bit(WORK_HIGH_PRIO_BIT
, &work
->flags
))
636 list_add_tail(&work
->list
, &worker
->prio_pending
);
638 list_add_tail(&work
->list
, &worker
->pending
);
639 atomic_inc(&worker
->num_pending
);
641 /* by definition we're busy, take ourselves off the idle
645 spin_lock(&worker
->workers
->lock
);
647 list_move_tail(&worker
->worker_list
,
648 &worker
->workers
->worker_list
);
649 spin_unlock(&worker
->workers
->lock
);
651 if (!worker
->working
) {
657 wake_up_process(worker
->task
);
658 spin_unlock_irqrestore(&worker
->lock
, flags
);
661 void btrfs_set_work_high_prio(struct btrfs_work
*work
)
663 set_bit(WORK_HIGH_PRIO_BIT
, &work
->flags
);
667 * places a struct btrfs_work into the pending queue of one of the kthreads
669 void btrfs_queue_worker(struct btrfs_workers
*workers
, struct btrfs_work
*work
)
671 struct btrfs_worker_thread
*worker
;
675 /* don't requeue something already on a list */
676 if (test_and_set_bit(WORK_QUEUED_BIT
, &work
->flags
))
679 worker
= find_worker(workers
);
680 if (workers
->ordered
) {
682 * you're not allowed to do ordered queues from an
685 spin_lock(&workers
->order_lock
);
686 if (test_bit(WORK_HIGH_PRIO_BIT
, &work
->flags
)) {
687 list_add_tail(&work
->order_list
,
688 &workers
->prio_order_list
);
690 list_add_tail(&work
->order_list
, &workers
->order_list
);
692 spin_unlock(&workers
->order_lock
);
694 INIT_LIST_HEAD(&work
->order_list
);
697 spin_lock_irqsave(&worker
->lock
, flags
);
699 if (test_bit(WORK_HIGH_PRIO_BIT
, &work
->flags
))
700 list_add_tail(&work
->list
, &worker
->prio_pending
);
702 list_add_tail(&work
->list
, &worker
->pending
);
703 check_busy_worker(worker
);
706 * avoid calling into wake_up_process if this thread has already
709 if (!worker
->working
)
714 wake_up_process(worker
->task
);
715 spin_unlock_irqrestore(&worker
->lock
, flags
);