4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
32 * 4MB minimal write chunk size
34 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
37 * Passed into wb_writeback(), essentially a subset of writeback_control
39 struct wb_writeback_work
{
41 struct super_block
*sb
;
42 unsigned long *older_than_this
;
43 enum writeback_sync_modes sync_mode
;
44 unsigned int tagged_writepages
:1;
45 unsigned int for_kupdate
:1;
46 unsigned int range_cyclic
:1;
47 unsigned int for_background
:1;
48 enum wb_reason reason
; /* why was writeback initiated? */
50 struct list_head list
; /* pending work list */
51 struct completion
*done
; /* set if the caller waits */
55 * writeback_in_progress - determine whether there is writeback in progress
56 * @bdi: the device's backing_dev_info structure.
58 * Determine whether there is writeback waiting to be handled against a
61 int writeback_in_progress(struct backing_dev_info
*bdi
)
63 return test_bit(BDI_writeback_running
, &bdi
->state
);
65 EXPORT_SYMBOL(writeback_in_progress
);
67 static inline struct backing_dev_info
*inode_to_bdi(struct inode
*inode
)
69 struct super_block
*sb
= inode
->i_sb
;
71 if (strcmp(sb
->s_type
->name
, "bdev") == 0)
72 return inode
->i_mapping
->backing_dev_info
;
77 static inline struct inode
*wb_inode(struct list_head
*head
)
79 return list_entry(head
, struct inode
, i_wb_list
);
83 * Include the creation of the trace points after defining the
84 * wb_writeback_work structure and inline functions so that the definition
85 * remains local to this file.
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/writeback.h>
90 static void bdi_queue_work(struct backing_dev_info
*bdi
,
91 struct wb_writeback_work
*work
)
93 trace_writeback_queue(bdi
, work
);
95 spin_lock_bh(&bdi
->wb_lock
);
96 list_add_tail(&work
->list
, &bdi
->work_list
);
97 spin_unlock_bh(&bdi
->wb_lock
);
99 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
103 __bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
104 bool range_cyclic
, enum wb_reason reason
)
106 struct wb_writeback_work
*work
;
109 * This is WB_SYNC_NONE writeback, so if allocation fails just
110 * wakeup the thread for old dirty data writeback
112 work
= kzalloc(sizeof(*work
), GFP_ATOMIC
);
114 trace_writeback_nowork(bdi
);
115 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
119 work
->sync_mode
= WB_SYNC_NONE
;
120 work
->nr_pages
= nr_pages
;
121 work
->range_cyclic
= range_cyclic
;
122 work
->reason
= reason
;
124 bdi_queue_work(bdi
, work
);
128 * bdi_start_writeback - start writeback
129 * @bdi: the backing device to write from
130 * @nr_pages: the number of pages to write
131 * @reason: reason why some writeback work was initiated
134 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
135 * started when this function returns, we make no guarantees on
136 * completion. Caller need not hold sb s_umount semaphore.
139 void bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
140 enum wb_reason reason
)
142 __bdi_start_writeback(bdi
, nr_pages
, true, reason
);
146 * bdi_start_background_writeback - start background writeback
147 * @bdi: the backing device to write from
150 * This makes sure WB_SYNC_NONE background writeback happens. When
151 * this function returns, it is only guaranteed that for given BDI
152 * some IO is happening if we are over background dirty threshold.
153 * Caller need not hold sb s_umount semaphore.
155 void bdi_start_background_writeback(struct backing_dev_info
*bdi
)
158 * We just wake up the flusher thread. It will perform background
159 * writeback as soon as there is no other work to do.
161 trace_writeback_wake_background(bdi
);
162 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
166 * Remove the inode from the writeback list it is on.
168 void inode_wb_list_del(struct inode
*inode
)
170 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
172 spin_lock(&bdi
->wb
.list_lock
);
173 list_del_init(&inode
->i_wb_list
);
174 spin_unlock(&bdi
->wb
.list_lock
);
178 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
179 * furthest end of its superblock's dirty-inode list.
181 * Before stamping the inode's ->dirtied_when, we check to see whether it is
182 * already the most-recently-dirtied inode on the b_dirty list. If that is
183 * the case then the inode must have been redirtied while it was being written
184 * out and we don't reset its dirtied_when.
186 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
188 assert_spin_locked(&wb
->list_lock
);
189 if (!list_empty(&wb
->b_dirty
)) {
192 tail
= wb_inode(wb
->b_dirty
.next
);
193 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
194 inode
->dirtied_when
= jiffies
;
196 list_move(&inode
->i_wb_list
, &wb
->b_dirty
);
200 * requeue inode for re-scanning after bdi->b_io list is exhausted.
202 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
204 assert_spin_locked(&wb
->list_lock
);
205 list_move(&inode
->i_wb_list
, &wb
->b_more_io
);
208 static void inode_sync_complete(struct inode
*inode
)
210 inode
->i_state
&= ~I_SYNC
;
211 /* If inode is clean an unused, put it into LRU now... */
212 inode_add_lru(inode
);
213 /* Waiters must see I_SYNC cleared before being woken up */
215 wake_up_bit(&inode
->i_state
, __I_SYNC
);
218 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
220 bool ret
= time_after(inode
->dirtied_when
, t
);
223 * For inodes being constantly redirtied, dirtied_when can get stuck.
224 * It _appears_ to be in the future, but is actually in distant past.
225 * This test is necessary to prevent such wrapped-around relative times
226 * from permanently stopping the whole bdi writeback.
228 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
234 * Move expired (dirtied before work->older_than_this) dirty inodes from
235 * @delaying_queue to @dispatch_queue.
237 static int move_expired_inodes(struct list_head
*delaying_queue
,
238 struct list_head
*dispatch_queue
,
239 struct wb_writeback_work
*work
)
242 struct list_head
*pos
, *node
;
243 struct super_block
*sb
= NULL
;
248 while (!list_empty(delaying_queue
)) {
249 inode
= wb_inode(delaying_queue
->prev
);
250 if (work
->older_than_this
&&
251 inode_dirtied_after(inode
, *work
->older_than_this
))
253 if (sb
&& sb
!= inode
->i_sb
)
256 list_move(&inode
->i_wb_list
, &tmp
);
260 /* just one sb in list, splice to dispatch_queue and we're done */
262 list_splice(&tmp
, dispatch_queue
);
266 /* Move inodes from one superblock together */
267 while (!list_empty(&tmp
)) {
268 sb
= wb_inode(tmp
.prev
)->i_sb
;
269 list_for_each_prev_safe(pos
, node
, &tmp
) {
270 inode
= wb_inode(pos
);
271 if (inode
->i_sb
== sb
)
272 list_move(&inode
->i_wb_list
, dispatch_queue
);
280 * Queue all expired dirty inodes for io, eldest first.
282 * newly dirtied b_dirty b_io b_more_io
283 * =============> gf edc BA
285 * newly dirtied b_dirty b_io b_more_io
286 * =============> g fBAedc
288 * +--> dequeue for IO
290 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
)
293 assert_spin_locked(&wb
->list_lock
);
294 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
295 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, work
);
296 trace_writeback_queue_io(wb
, work
, moved
);
299 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
303 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
304 trace_writeback_write_inode_start(inode
, wbc
);
305 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
306 trace_writeback_write_inode(inode
, wbc
);
313 * Wait for writeback on an inode to complete. Called with i_lock held.
314 * Caller must make sure inode cannot go away when we drop i_lock.
316 static void __inode_wait_for_writeback(struct inode
*inode
)
317 __releases(inode
->i_lock
)
318 __acquires(inode
->i_lock
)
320 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
321 wait_queue_head_t
*wqh
;
323 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
324 while (inode
->i_state
& I_SYNC
) {
325 spin_unlock(&inode
->i_lock
);
326 __wait_on_bit(wqh
, &wq
, inode_wait
, TASK_UNINTERRUPTIBLE
);
327 spin_lock(&inode
->i_lock
);
332 * Wait for writeback on an inode to complete. Caller must have inode pinned.
334 void inode_wait_for_writeback(struct inode
*inode
)
336 spin_lock(&inode
->i_lock
);
337 __inode_wait_for_writeback(inode
);
338 spin_unlock(&inode
->i_lock
);
342 * Sleep until I_SYNC is cleared. This function must be called with i_lock
343 * held and drops it. It is aimed for callers not holding any inode reference
344 * so once i_lock is dropped, inode can go away.
346 static void inode_sleep_on_writeback(struct inode
*inode
)
347 __releases(inode
->i_lock
)
350 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
353 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
354 sleep
= inode
->i_state
& I_SYNC
;
355 spin_unlock(&inode
->i_lock
);
358 finish_wait(wqh
, &wait
);
362 * Find proper writeback list for the inode depending on its current state and
363 * possibly also change of its state while we were doing writeback. Here we
364 * handle things such as livelock prevention or fairness of writeback among
365 * inodes. This function can be called only by flusher thread - noone else
366 * processes all inodes in writeback lists and requeueing inodes behind flusher
367 * thread's back can have unexpected consequences.
369 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
370 struct writeback_control
*wbc
)
372 if (inode
->i_state
& I_FREEING
)
376 * Sync livelock prevention. Each inode is tagged and synced in one
377 * shot. If still dirty, it will be redirty_tail()'ed below. Update
378 * the dirty time to prevent enqueue and sync it again.
380 if ((inode
->i_state
& I_DIRTY
) &&
381 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
382 inode
->dirtied_when
= jiffies
;
384 if (wbc
->pages_skipped
) {
386 * writeback is not making progress due to locked
387 * buffers. Skip this inode for now.
389 redirty_tail(inode
, wb
);
393 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
395 * We didn't write back all the pages. nfs_writepages()
396 * sometimes bales out without doing anything.
398 if (wbc
->nr_to_write
<= 0) {
399 /* Slice used up. Queue for next turn. */
400 requeue_io(inode
, wb
);
403 * Writeback blocked by something other than
404 * congestion. Delay the inode for some time to
405 * avoid spinning on the CPU (100% iowait)
406 * retrying writeback of the dirty page/inode
407 * that cannot be performed immediately.
409 redirty_tail(inode
, wb
);
411 } else if (inode
->i_state
& I_DIRTY
) {
413 * Filesystems can dirty the inode during writeback operations,
414 * such as delayed allocation during submission or metadata
415 * updates after data IO completion.
417 redirty_tail(inode
, wb
);
419 /* The inode is clean. Remove from writeback lists. */
420 list_del_init(&inode
->i_wb_list
);
425 * Write out an inode and its dirty pages. Do not update the writeback list
426 * linkage. That is left to the caller. The caller is also responsible for
427 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
430 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
432 struct address_space
*mapping
= inode
->i_mapping
;
433 long nr_to_write
= wbc
->nr_to_write
;
437 WARN_ON(!(inode
->i_state
& I_SYNC
));
439 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
441 ret
= do_writepages(mapping
, wbc
);
444 * Make sure to wait on the data before writing out the metadata.
445 * This is important for filesystems that modify metadata on data
448 if (wbc
->sync_mode
== WB_SYNC_ALL
) {
449 int err
= filemap_fdatawait(mapping
);
455 * Some filesystems may redirty the inode during the writeback
456 * due to delalloc, clear dirty metadata flags right before
459 spin_lock(&inode
->i_lock
);
460 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
461 if (!mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
462 inode
->i_state
&= ~I_DIRTY_PAGES
;
463 dirty
= inode
->i_state
& I_DIRTY
;
464 inode
->i_state
&= ~(I_DIRTY_SYNC
| I_DIRTY_DATASYNC
);
465 spin_unlock(&inode
->i_lock
);
466 /* Don't write the inode if only I_DIRTY_PAGES was set */
467 if (dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
468 int err
= write_inode(inode
, wbc
);
472 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
477 * Write out an inode's dirty pages. Either the caller has an active reference
478 * on the inode or the inode has I_WILL_FREE set.
480 * This function is designed to be called for writing back one inode which
481 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
482 * and does more profound writeback list handling in writeback_sb_inodes().
485 writeback_single_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
486 struct writeback_control
*wbc
)
490 spin_lock(&inode
->i_lock
);
491 if (!atomic_read(&inode
->i_count
))
492 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
494 WARN_ON(inode
->i_state
& I_WILL_FREE
);
496 if (inode
->i_state
& I_SYNC
) {
497 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
500 * It's a data-integrity sync. We must wait. Since callers hold
501 * inode reference or inode has I_WILL_FREE set, it cannot go
504 __inode_wait_for_writeback(inode
);
506 WARN_ON(inode
->i_state
& I_SYNC
);
508 * Skip inode if it is clean and we have no outstanding writeback in
509 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
510 * function since flusher thread may be doing for example sync in
511 * parallel and if we move the inode, it could get skipped. So here we
512 * make sure inode is on some writeback list and leave it there unless
513 * we have completely cleaned the inode.
515 if (!(inode
->i_state
& I_DIRTY
) &&
516 (wbc
->sync_mode
!= WB_SYNC_ALL
||
517 !mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
)))
519 inode
->i_state
|= I_SYNC
;
520 spin_unlock(&inode
->i_lock
);
522 ret
= __writeback_single_inode(inode
, wbc
);
524 spin_lock(&wb
->list_lock
);
525 spin_lock(&inode
->i_lock
);
527 * If inode is clean, remove it from writeback lists. Otherwise don't
528 * touch it. See comment above for explanation.
530 if (!(inode
->i_state
& I_DIRTY
))
531 list_del_init(&inode
->i_wb_list
);
532 spin_unlock(&wb
->list_lock
);
533 inode_sync_complete(inode
);
535 spin_unlock(&inode
->i_lock
);
539 static long writeback_chunk_size(struct backing_dev_info
*bdi
,
540 struct wb_writeback_work
*work
)
545 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
546 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
547 * here avoids calling into writeback_inodes_wb() more than once.
549 * The intended call sequence for WB_SYNC_ALL writeback is:
552 * writeback_sb_inodes() <== called only once
553 * write_cache_pages() <== called once for each inode
554 * (quickly) tag currently dirty pages
555 * (maybe slowly) sync all tagged pages
557 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
560 pages
= min(bdi
->avg_write_bandwidth
/ 2,
561 global_dirty_limit
/ DIRTY_SCOPE
);
562 pages
= min(pages
, work
->nr_pages
);
563 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
564 MIN_WRITEBACK_PAGES
);
571 * Write a portion of b_io inodes which belong to @sb.
573 * Return the number of pages and/or inodes written.
575 static long writeback_sb_inodes(struct super_block
*sb
,
576 struct bdi_writeback
*wb
,
577 struct wb_writeback_work
*work
)
579 struct writeback_control wbc
= {
580 .sync_mode
= work
->sync_mode
,
581 .tagged_writepages
= work
->tagged_writepages
,
582 .for_kupdate
= work
->for_kupdate
,
583 .for_background
= work
->for_background
,
584 .range_cyclic
= work
->range_cyclic
,
586 .range_end
= LLONG_MAX
,
588 unsigned long start_time
= jiffies
;
590 long wrote
= 0; /* count both pages and inodes */
592 while (!list_empty(&wb
->b_io
)) {
593 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
595 if (inode
->i_sb
!= sb
) {
598 * We only want to write back data for this
599 * superblock, move all inodes not belonging
600 * to it back onto the dirty list.
602 redirty_tail(inode
, wb
);
607 * The inode belongs to a different superblock.
608 * Bounce back to the caller to unpin this and
609 * pin the next superblock.
615 * Don't bother with new inodes or inodes being freed, first
616 * kind does not need periodic writeout yet, and for the latter
617 * kind writeout is handled by the freer.
619 spin_lock(&inode
->i_lock
);
620 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
621 spin_unlock(&inode
->i_lock
);
622 redirty_tail(inode
, wb
);
625 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
627 * If this inode is locked for writeback and we are not
628 * doing writeback-for-data-integrity, move it to
629 * b_more_io so that writeback can proceed with the
630 * other inodes on s_io.
632 * We'll have another go at writing back this inode
633 * when we completed a full scan of b_io.
635 spin_unlock(&inode
->i_lock
);
636 requeue_io(inode
, wb
);
637 trace_writeback_sb_inodes_requeue(inode
);
640 spin_unlock(&wb
->list_lock
);
643 * We already requeued the inode if it had I_SYNC set and we
644 * are doing WB_SYNC_NONE writeback. So this catches only the
647 if (inode
->i_state
& I_SYNC
) {
648 /* Wait for I_SYNC. This function drops i_lock... */
649 inode_sleep_on_writeback(inode
);
650 /* Inode may be gone, start again */
651 spin_lock(&wb
->list_lock
);
654 inode
->i_state
|= I_SYNC
;
655 spin_unlock(&inode
->i_lock
);
657 write_chunk
= writeback_chunk_size(wb
->bdi
, work
);
658 wbc
.nr_to_write
= write_chunk
;
659 wbc
.pages_skipped
= 0;
662 * We use I_SYNC to pin the inode in memory. While it is set
663 * evict_inode() will wait so the inode cannot be freed.
665 __writeback_single_inode(inode
, &wbc
);
667 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
668 wrote
+= write_chunk
- wbc
.nr_to_write
;
669 spin_lock(&wb
->list_lock
);
670 spin_lock(&inode
->i_lock
);
671 if (!(inode
->i_state
& I_DIRTY
))
673 requeue_inode(inode
, wb
, &wbc
);
674 inode_sync_complete(inode
);
675 spin_unlock(&inode
->i_lock
);
676 cond_resched_lock(&wb
->list_lock
);
678 * bail out to wb_writeback() often enough to check
679 * background threshold and other termination conditions.
682 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
684 if (work
->nr_pages
<= 0)
691 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
692 struct wb_writeback_work
*work
)
694 unsigned long start_time
= jiffies
;
697 while (!list_empty(&wb
->b_io
)) {
698 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
699 struct super_block
*sb
= inode
->i_sb
;
701 if (!grab_super_passive(sb
)) {
703 * grab_super_passive() may fail consistently due to
704 * s_umount being grabbed by someone else. Don't use
705 * requeue_io() to avoid busy retrying the inode/sb.
707 redirty_tail(inode
, wb
);
710 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
713 /* refer to the same tests at the end of writeback_sb_inodes */
715 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
717 if (work
->nr_pages
<= 0)
721 /* Leave any unwritten inodes on b_io */
725 long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
726 enum wb_reason reason
)
728 struct wb_writeback_work work
= {
729 .nr_pages
= nr_pages
,
730 .sync_mode
= WB_SYNC_NONE
,
735 spin_lock(&wb
->list_lock
);
736 if (list_empty(&wb
->b_io
))
738 __writeback_inodes_wb(wb
, &work
);
739 spin_unlock(&wb
->list_lock
);
741 return nr_pages
- work
.nr_pages
;
744 static bool over_bground_thresh(struct backing_dev_info
*bdi
)
746 unsigned long background_thresh
, dirty_thresh
;
748 global_dirty_limits(&background_thresh
, &dirty_thresh
);
750 if (global_page_state(NR_FILE_DIRTY
) +
751 global_page_state(NR_UNSTABLE_NFS
) > background_thresh
)
754 if (bdi_stat(bdi
, BDI_RECLAIMABLE
) >
755 bdi_dirty_limit(bdi
, background_thresh
))
762 * Called under wb->list_lock. If there are multiple wb per bdi,
763 * only the flusher working on the first wb should do it.
765 static void wb_update_bandwidth(struct bdi_writeback
*wb
,
766 unsigned long start_time
)
768 __bdi_update_bandwidth(wb
->bdi
, 0, 0, 0, 0, 0, start_time
);
772 * Explicit flushing or periodic writeback of "old" data.
774 * Define "old": the first time one of an inode's pages is dirtied, we mark the
775 * dirtying-time in the inode's address_space. So this periodic writeback code
776 * just walks the superblock inode list, writing back any inodes which are
777 * older than a specific point in time.
779 * Try to run once per dirty_writeback_interval. But if a writeback event
780 * takes longer than a dirty_writeback_interval interval, then leave a
783 * older_than_this takes precedence over nr_to_write. So we'll only write back
784 * all dirty pages if they are all attached to "old" mappings.
786 static long wb_writeback(struct bdi_writeback
*wb
,
787 struct wb_writeback_work
*work
)
789 unsigned long wb_start
= jiffies
;
790 long nr_pages
= work
->nr_pages
;
791 unsigned long oldest_jif
;
795 oldest_jif
= jiffies
;
796 work
->older_than_this
= &oldest_jif
;
798 spin_lock(&wb
->list_lock
);
801 * Stop writeback when nr_pages has been consumed
803 if (work
->nr_pages
<= 0)
807 * Background writeout and kupdate-style writeback may
808 * run forever. Stop them if there is other work to do
809 * so that e.g. sync can proceed. They'll be restarted
810 * after the other works are all done.
812 if ((work
->for_background
|| work
->for_kupdate
) &&
813 !list_empty(&wb
->bdi
->work_list
))
817 * For background writeout, stop when we are below the
818 * background dirty threshold
820 if (work
->for_background
&& !over_bground_thresh(wb
->bdi
))
824 * Kupdate and background works are special and we want to
825 * include all inodes that need writing. Livelock avoidance is
826 * handled by these works yielding to any other work so we are
829 if (work
->for_kupdate
) {
830 oldest_jif
= jiffies
-
831 msecs_to_jiffies(dirty_expire_interval
* 10);
832 } else if (work
->for_background
)
833 oldest_jif
= jiffies
;
835 trace_writeback_start(wb
->bdi
, work
);
836 if (list_empty(&wb
->b_io
))
839 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
841 progress
= __writeback_inodes_wb(wb
, work
);
842 trace_writeback_written(wb
->bdi
, work
);
844 wb_update_bandwidth(wb
, wb_start
);
847 * Did we write something? Try for more
849 * Dirty inodes are moved to b_io for writeback in batches.
850 * The completion of the current batch does not necessarily
851 * mean the overall work is done. So we keep looping as long
852 * as made some progress on cleaning pages or inodes.
857 * No more inodes for IO, bail
859 if (list_empty(&wb
->b_more_io
))
862 * Nothing written. Wait for some inode to
863 * become available for writeback. Otherwise
864 * we'll just busyloop.
866 if (!list_empty(&wb
->b_more_io
)) {
867 trace_writeback_wait(wb
->bdi
, work
);
868 inode
= wb_inode(wb
->b_more_io
.prev
);
869 spin_lock(&inode
->i_lock
);
870 spin_unlock(&wb
->list_lock
);
871 /* This function drops i_lock... */
872 inode_sleep_on_writeback(inode
);
873 spin_lock(&wb
->list_lock
);
876 spin_unlock(&wb
->list_lock
);
878 return nr_pages
- work
->nr_pages
;
882 * Return the next wb_writeback_work struct that hasn't been processed yet.
884 static struct wb_writeback_work
*
885 get_next_work_item(struct backing_dev_info
*bdi
)
887 struct wb_writeback_work
*work
= NULL
;
889 spin_lock_bh(&bdi
->wb_lock
);
890 if (!list_empty(&bdi
->work_list
)) {
891 work
= list_entry(bdi
->work_list
.next
,
892 struct wb_writeback_work
, list
);
893 list_del_init(&work
->list
);
895 spin_unlock_bh(&bdi
->wb_lock
);
900 * Add in the number of potentially dirty inodes, because each inode
901 * write can dirty pagecache in the underlying blockdev.
903 static unsigned long get_nr_dirty_pages(void)
905 return global_page_state(NR_FILE_DIRTY
) +
906 global_page_state(NR_UNSTABLE_NFS
) +
907 get_nr_dirty_inodes();
910 static long wb_check_background_flush(struct bdi_writeback
*wb
)
912 if (over_bground_thresh(wb
->bdi
)) {
914 struct wb_writeback_work work
= {
915 .nr_pages
= LONG_MAX
,
916 .sync_mode
= WB_SYNC_NONE
,
919 .reason
= WB_REASON_BACKGROUND
,
922 return wb_writeback(wb
, &work
);
928 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
930 unsigned long expired
;
934 * When set to zero, disable periodic writeback
936 if (!dirty_writeback_interval
)
939 expired
= wb
->last_old_flush
+
940 msecs_to_jiffies(dirty_writeback_interval
* 10);
941 if (time_before(jiffies
, expired
))
944 wb
->last_old_flush
= jiffies
;
945 nr_pages
= get_nr_dirty_pages();
948 struct wb_writeback_work work
= {
949 .nr_pages
= nr_pages
,
950 .sync_mode
= WB_SYNC_NONE
,
953 .reason
= WB_REASON_PERIODIC
,
956 return wb_writeback(wb
, &work
);
963 * Retrieve work items and do the writeback they describe
965 long wb_do_writeback(struct bdi_writeback
*wb
, int force_wait
)
967 struct backing_dev_info
*bdi
= wb
->bdi
;
968 struct wb_writeback_work
*work
;
971 set_bit(BDI_writeback_running
, &wb
->bdi
->state
);
972 while ((work
= get_next_work_item(bdi
)) != NULL
) {
974 * Override sync mode, in case we must wait for completion
975 * because this thread is exiting now.
978 work
->sync_mode
= WB_SYNC_ALL
;
980 trace_writeback_exec(bdi
, work
);
982 wrote
+= wb_writeback(wb
, work
);
985 * Notify the caller of completion if this is a synchronous
986 * work item, otherwise just free it.
989 complete(work
->done
);
995 * Check for periodic writeback, kupdated() style
997 wrote
+= wb_check_old_data_flush(wb
);
998 wrote
+= wb_check_background_flush(wb
);
999 clear_bit(BDI_writeback_running
, &wb
->bdi
->state
);
1005 * Handle writeback of dirty data for the device backed by this bdi. Also
1006 * reschedules periodically and does kupdated style flushing.
1008 void bdi_writeback_workfn(struct work_struct
*work
)
1010 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
1011 struct bdi_writeback
, dwork
);
1012 struct backing_dev_info
*bdi
= wb
->bdi
;
1015 set_worker_desc("flush-%s", dev_name(bdi
->dev
));
1016 current
->flags
|= PF_SWAPWRITE
;
1018 if (likely(!current_is_workqueue_rescuer() ||
1019 list_empty(&bdi
->bdi_list
))) {
1021 * The normal path. Keep writing back @bdi until its
1022 * work_list is empty. Note that this path is also taken
1023 * if @bdi is shutting down even when we're running off the
1024 * rescuer as work_list needs to be drained.
1027 pages_written
= wb_do_writeback(wb
, 0);
1028 trace_writeback_pages_written(pages_written
);
1029 } while (!list_empty(&bdi
->work_list
));
1032 * bdi_wq can't get enough workers and we're running off
1033 * the emergency worker. Don't hog it. Hopefully, 1024 is
1034 * enough for efficient IO.
1036 pages_written
= writeback_inodes_wb(&bdi
->wb
, 1024,
1037 WB_REASON_FORKER_THREAD
);
1038 trace_writeback_pages_written(pages_written
);
1041 if (!list_empty(&bdi
->work_list
) ||
1042 (wb_has_dirty_io(wb
) && dirty_writeback_interval
))
1043 queue_delayed_work(bdi_wq
, &wb
->dwork
,
1044 msecs_to_jiffies(dirty_writeback_interval
* 10));
1046 current
->flags
&= ~PF_SWAPWRITE
;
1050 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1053 void wakeup_flusher_threads(long nr_pages
, enum wb_reason reason
)
1055 struct backing_dev_info
*bdi
;
1058 nr_pages
= global_page_state(NR_FILE_DIRTY
) +
1059 global_page_state(NR_UNSTABLE_NFS
);
1063 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
1064 if (!bdi_has_dirty_io(bdi
))
1066 __bdi_start_writeback(bdi
, nr_pages
, false, reason
);
1071 static noinline
void block_dump___mark_inode_dirty(struct inode
*inode
)
1073 if (inode
->i_ino
|| strcmp(inode
->i_sb
->s_id
, "bdev")) {
1074 struct dentry
*dentry
;
1075 const char *name
= "?";
1077 dentry
= d_find_alias(inode
);
1079 spin_lock(&dentry
->d_lock
);
1080 name
= (const char *) dentry
->d_name
.name
;
1083 "%s(%d): dirtied inode %lu (%s) on %s\n",
1084 current
->comm
, task_pid_nr(current
), inode
->i_ino
,
1085 name
, inode
->i_sb
->s_id
);
1087 spin_unlock(&dentry
->d_lock
);
1094 * __mark_inode_dirty - internal function
1095 * @inode: inode to mark
1096 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1097 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1098 * mark_inode_dirty_sync.
1100 * Put the inode on the super block's dirty list.
1102 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1103 * dirty list only if it is hashed or if it refers to a blockdev.
1104 * If it was not hashed, it will never be added to the dirty list
1105 * even if it is later hashed, as it will have been marked dirty already.
1107 * In short, make sure you hash any inodes _before_ you start marking
1110 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1111 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1112 * the kernel-internal blockdev inode represents the dirtying time of the
1113 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1114 * page->mapping->host, so the page-dirtying time is recorded in the internal
1117 void __mark_inode_dirty(struct inode
*inode
, int flags
)
1119 struct super_block
*sb
= inode
->i_sb
;
1120 struct backing_dev_info
*bdi
= NULL
;
1123 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1124 * dirty the inode itself
1126 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
1127 trace_writeback_dirty_inode_start(inode
, flags
);
1129 if (sb
->s_op
->dirty_inode
)
1130 sb
->s_op
->dirty_inode(inode
, flags
);
1132 trace_writeback_dirty_inode(inode
, flags
);
1136 * make sure that changes are seen by all cpus before we test i_state
1141 /* avoid the locking if we can */
1142 if ((inode
->i_state
& flags
) == flags
)
1145 if (unlikely(block_dump
))
1146 block_dump___mark_inode_dirty(inode
);
1148 spin_lock(&inode
->i_lock
);
1149 if ((inode
->i_state
& flags
) != flags
) {
1150 const int was_dirty
= inode
->i_state
& I_DIRTY
;
1152 inode
->i_state
|= flags
;
1155 * If the inode is being synced, just update its dirty state.
1156 * The unlocker will place the inode on the appropriate
1157 * superblock list, based upon its state.
1159 if (inode
->i_state
& I_SYNC
)
1160 goto out_unlock_inode
;
1163 * Only add valid (hashed) inodes to the superblock's
1164 * dirty list. Add blockdev inodes as well.
1166 if (!S_ISBLK(inode
->i_mode
)) {
1167 if (inode_unhashed(inode
))
1168 goto out_unlock_inode
;
1170 if (inode
->i_state
& I_FREEING
)
1171 goto out_unlock_inode
;
1174 * If the inode was already on b_dirty/b_io/b_more_io, don't
1175 * reposition it (that would break b_dirty time-ordering).
1178 bool wakeup_bdi
= false;
1179 bdi
= inode_to_bdi(inode
);
1181 if (bdi_cap_writeback_dirty(bdi
)) {
1182 WARN(!test_bit(BDI_registered
, &bdi
->state
),
1183 "bdi-%s not registered\n", bdi
->name
);
1186 * If this is the first dirty inode for this
1187 * bdi, we have to wake-up the corresponding
1188 * bdi thread to make sure background
1189 * write-back happens later.
1191 if (!wb_has_dirty_io(&bdi
->wb
))
1195 spin_unlock(&inode
->i_lock
);
1196 spin_lock(&bdi
->wb
.list_lock
);
1197 inode
->dirtied_when
= jiffies
;
1198 list_move(&inode
->i_wb_list
, &bdi
->wb
.b_dirty
);
1199 spin_unlock(&bdi
->wb
.list_lock
);
1202 bdi_wakeup_thread_delayed(bdi
);
1207 spin_unlock(&inode
->i_lock
);
1210 EXPORT_SYMBOL(__mark_inode_dirty
);
1212 static void wait_sb_inodes(struct super_block
*sb
)
1214 struct inode
*inode
, *old_inode
= NULL
;
1217 * We need to be protected against the filesystem going from
1218 * r/o to r/w or vice versa.
1220 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1222 spin_lock(&inode_sb_list_lock
);
1225 * Data integrity sync. Must wait for all pages under writeback,
1226 * because there may have been pages dirtied before our sync
1227 * call, but which had writeout started before we write it out.
1228 * In which case, the inode may not be on the dirty list, but
1229 * we still have to wait for that writeout.
1231 list_for_each_entry(inode
, &sb
->s_inodes
, i_sb_list
) {
1232 struct address_space
*mapping
= inode
->i_mapping
;
1234 spin_lock(&inode
->i_lock
);
1235 if ((inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) ||
1236 (mapping
->nrpages
== 0)) {
1237 spin_unlock(&inode
->i_lock
);
1241 spin_unlock(&inode
->i_lock
);
1242 spin_unlock(&inode_sb_list_lock
);
1245 * We hold a reference to 'inode' so it couldn't have been
1246 * removed from s_inodes list while we dropped the
1247 * inode_sb_list_lock. We cannot iput the inode now as we can
1248 * be holding the last reference and we cannot iput it under
1249 * inode_sb_list_lock. So we keep the reference and iput it
1255 filemap_fdatawait(mapping
);
1259 spin_lock(&inode_sb_list_lock
);
1261 spin_unlock(&inode_sb_list_lock
);
1266 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1267 * @sb: the superblock
1268 * @nr: the number of pages to write
1269 * @reason: reason why some writeback work initiated
1271 * Start writeback on some inodes on this super_block. No guarantees are made
1272 * on how many (if any) will be written, and this function does not wait
1273 * for IO completion of submitted IO.
1275 void writeback_inodes_sb_nr(struct super_block
*sb
,
1277 enum wb_reason reason
)
1279 DECLARE_COMPLETION_ONSTACK(done
);
1280 struct wb_writeback_work work
= {
1282 .sync_mode
= WB_SYNC_NONE
,
1283 .tagged_writepages
= 1,
1289 if (sb
->s_bdi
== &noop_backing_dev_info
)
1291 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1292 bdi_queue_work(sb
->s_bdi
, &work
);
1293 wait_for_completion(&done
);
1295 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
1298 * writeback_inodes_sb - writeback dirty inodes from given super_block
1299 * @sb: the superblock
1300 * @reason: reason why some writeback work was initiated
1302 * Start writeback on some inodes on this super_block. No guarantees are made
1303 * on how many (if any) will be written, and this function does not wait
1304 * for IO completion of submitted IO.
1306 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1308 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1310 EXPORT_SYMBOL(writeback_inodes_sb
);
1313 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1314 * @sb: the superblock
1315 * @nr: the number of pages to write
1316 * @reason: the reason of writeback
1318 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1319 * Returns 1 if writeback was started, 0 if not.
1321 int try_to_writeback_inodes_sb_nr(struct super_block
*sb
,
1323 enum wb_reason reason
)
1325 if (writeback_in_progress(sb
->s_bdi
))
1328 if (!down_read_trylock(&sb
->s_umount
))
1331 writeback_inodes_sb_nr(sb
, nr
, reason
);
1332 up_read(&sb
->s_umount
);
1335 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr
);
1338 * try_to_writeback_inodes_sb - try to start writeback if none underway
1339 * @sb: the superblock
1340 * @reason: reason why some writeback work was initiated
1342 * Implement by try_to_writeback_inodes_sb_nr()
1343 * Returns 1 if writeback was started, 0 if not.
1345 int try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1347 return try_to_writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1349 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
1352 * sync_inodes_sb - sync sb inode pages
1353 * @sb: the superblock
1355 * This function writes and waits on any dirty inode belonging to this
1358 void sync_inodes_sb(struct super_block
*sb
)
1360 DECLARE_COMPLETION_ONSTACK(done
);
1361 struct wb_writeback_work work
= {
1363 .sync_mode
= WB_SYNC_ALL
,
1364 .nr_pages
= LONG_MAX
,
1367 .reason
= WB_REASON_SYNC
,
1370 /* Nothing to do? */
1371 if (sb
->s_bdi
== &noop_backing_dev_info
)
1373 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1375 bdi_queue_work(sb
->s_bdi
, &work
);
1376 wait_for_completion(&done
);
1380 EXPORT_SYMBOL(sync_inodes_sb
);
1383 * write_inode_now - write an inode to disk
1384 * @inode: inode to write to disk
1385 * @sync: whether the write should be synchronous or not
1387 * This function commits an inode to disk immediately if it is dirty. This is
1388 * primarily needed by knfsd.
1390 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1392 int write_inode_now(struct inode
*inode
, int sync
)
1394 struct bdi_writeback
*wb
= &inode_to_bdi(inode
)->wb
;
1395 struct writeback_control wbc
= {
1396 .nr_to_write
= LONG_MAX
,
1397 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1399 .range_end
= LLONG_MAX
,
1402 if (!mapping_cap_writeback_dirty(inode
->i_mapping
))
1403 wbc
.nr_to_write
= 0;
1406 return writeback_single_inode(inode
, wb
, &wbc
);
1408 EXPORT_SYMBOL(write_inode_now
);
1411 * sync_inode - write an inode and its pages to disk.
1412 * @inode: the inode to sync
1413 * @wbc: controls the writeback mode
1415 * sync_inode() will write an inode and its pages to disk. It will also
1416 * correctly update the inode on its superblock's dirty inode lists and will
1417 * update inode->i_state.
1419 * The caller must have a ref on the inode.
1421 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1423 return writeback_single_inode(inode
, &inode_to_bdi(inode
)->wb
, wbc
);
1425 EXPORT_SYMBOL(sync_inode
);
1428 * sync_inode_metadata - write an inode to disk
1429 * @inode: the inode to sync
1430 * @wait: wait for I/O to complete.
1432 * Write an inode to disk and adjust its dirty state after completion.
1434 * Note: only writes the actual inode, no associated data or other metadata.
1436 int sync_inode_metadata(struct inode
*inode
, int wait
)
1438 struct writeback_control wbc
= {
1439 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1440 .nr_to_write
= 0, /* metadata-only */
1443 return sync_inode(inode
, &wbc
);
1445 EXPORT_SYMBOL(sync_inode_metadata
);