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/freezer.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
33 * 4MB minimal write chunk size
35 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
38 * Passed into wb_writeback(), essentially a subset of writeback_control
40 struct wb_writeback_work
{
42 struct super_block
*sb
;
43 unsigned long *older_than_this
;
44 enum writeback_sync_modes sync_mode
;
45 unsigned int tagged_writepages
:1;
46 unsigned int for_kupdate
:1;
47 unsigned int range_cyclic
:1;
48 unsigned int for_background
:1;
49 enum wb_reason reason
; /* why was writeback initiated? */
51 struct list_head list
; /* pending work list */
52 struct completion
*done
; /* set if the caller waits */
56 * writeback_in_progress - determine whether there is writeback in progress
57 * @bdi: the device's backing_dev_info structure.
59 * Determine whether there is writeback waiting to be handled against a
62 int writeback_in_progress(struct backing_dev_info
*bdi
)
64 return test_bit(BDI_writeback_running
, &bdi
->state
);
66 EXPORT_SYMBOL(writeback_in_progress
);
68 static inline struct backing_dev_info
*inode_to_bdi(struct inode
*inode
)
70 struct super_block
*sb
= inode
->i_sb
;
72 if (strcmp(sb
->s_type
->name
, "bdev") == 0)
73 return inode
->i_mapping
->backing_dev_info
;
78 static inline struct inode
*wb_inode(struct list_head
*head
)
80 return list_entry(head
, struct inode
, i_wb_list
);
84 * Include the creation of the trace points after defining the
85 * wb_writeback_work structure and inline functions so that the definition
86 * remains local to this file.
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/writeback.h>
91 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
92 static void bdi_wakeup_flusher(struct backing_dev_info
*bdi
)
95 wake_up_process(bdi
->wb
.task
);
98 * The bdi thread isn't there, wake up the forker thread which
99 * will create and run it.
101 wake_up_process(default_backing_dev_info
.wb
.task
);
105 static void bdi_queue_work(struct backing_dev_info
*bdi
,
106 struct wb_writeback_work
*work
)
108 trace_writeback_queue(bdi
, work
);
110 spin_lock_bh(&bdi
->wb_lock
);
111 list_add_tail(&work
->list
, &bdi
->work_list
);
113 trace_writeback_nothread(bdi
, work
);
114 bdi_wakeup_flusher(bdi
);
115 spin_unlock_bh(&bdi
->wb_lock
);
119 __bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
120 bool range_cyclic
, enum wb_reason reason
)
122 struct wb_writeback_work
*work
;
125 * This is WB_SYNC_NONE writeback, so if allocation fails just
126 * wakeup the thread for old dirty data writeback
128 work
= kzalloc(sizeof(*work
), GFP_ATOMIC
);
131 trace_writeback_nowork(bdi
);
132 wake_up_process(bdi
->wb
.task
);
137 work
->sync_mode
= WB_SYNC_NONE
;
138 work
->nr_pages
= nr_pages
;
139 work
->range_cyclic
= range_cyclic
;
140 work
->reason
= reason
;
142 bdi_queue_work(bdi
, work
);
146 * bdi_start_writeback - start writeback
147 * @bdi: the backing device to write from
148 * @nr_pages: the number of pages to write
149 * @reason: reason why some writeback work was initiated
152 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
153 * started when this function returns, we make no guarantees on
154 * completion. Caller need not hold sb s_umount semaphore.
157 void bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
158 enum wb_reason reason
)
160 __bdi_start_writeback(bdi
, nr_pages
, true, reason
);
164 * bdi_start_background_writeback - start background writeback
165 * @bdi: the backing device to write from
168 * This makes sure WB_SYNC_NONE background writeback happens. When
169 * this function returns, it is only guaranteed that for given BDI
170 * some IO is happening if we are over background dirty threshold.
171 * Caller need not hold sb s_umount semaphore.
173 void bdi_start_background_writeback(struct backing_dev_info
*bdi
)
176 * We just wake up the flusher thread. It will perform background
177 * writeback as soon as there is no other work to do.
179 trace_writeback_wake_background(bdi
);
180 spin_lock_bh(&bdi
->wb_lock
);
181 bdi_wakeup_flusher(bdi
);
182 spin_unlock_bh(&bdi
->wb_lock
);
186 * Remove the inode from the writeback list it is on.
188 void inode_wb_list_del(struct inode
*inode
)
190 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
192 spin_lock(&bdi
->wb
.list_lock
);
193 list_del_init(&inode
->i_wb_list
);
194 spin_unlock(&bdi
->wb
.list_lock
);
198 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
199 * furthest end of its superblock's dirty-inode list.
201 * Before stamping the inode's ->dirtied_when, we check to see whether it is
202 * already the most-recently-dirtied inode on the b_dirty list. If that is
203 * the case then the inode must have been redirtied while it was being written
204 * out and we don't reset its dirtied_when.
206 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
208 assert_spin_locked(&wb
->list_lock
);
209 if (!list_empty(&wb
->b_dirty
)) {
212 tail
= wb_inode(wb
->b_dirty
.next
);
213 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
214 inode
->dirtied_when
= jiffies
;
216 list_move(&inode
->i_wb_list
, &wb
->b_dirty
);
220 * requeue inode for re-scanning after bdi->b_io list is exhausted.
222 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
224 assert_spin_locked(&wb
->list_lock
);
225 list_move(&inode
->i_wb_list
, &wb
->b_more_io
);
228 static void inode_sync_complete(struct inode
*inode
)
230 inode
->i_state
&= ~I_SYNC
;
231 /* Waiters must see I_SYNC cleared before being woken up */
233 wake_up_bit(&inode
->i_state
, __I_SYNC
);
236 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
238 bool ret
= time_after(inode
->dirtied_when
, t
);
241 * For inodes being constantly redirtied, dirtied_when can get stuck.
242 * It _appears_ to be in the future, but is actually in distant past.
243 * This test is necessary to prevent such wrapped-around relative times
244 * from permanently stopping the whole bdi writeback.
246 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
252 * Move expired (dirtied after work->older_than_this) dirty inodes from
253 * @delaying_queue to @dispatch_queue.
255 static int move_expired_inodes(struct list_head
*delaying_queue
,
256 struct list_head
*dispatch_queue
,
257 struct wb_writeback_work
*work
)
260 struct list_head
*pos
, *node
;
261 struct super_block
*sb
= NULL
;
266 while (!list_empty(delaying_queue
)) {
267 inode
= wb_inode(delaying_queue
->prev
);
268 if (work
->older_than_this
&&
269 inode_dirtied_after(inode
, *work
->older_than_this
))
271 if (sb
&& sb
!= inode
->i_sb
)
274 list_move(&inode
->i_wb_list
, &tmp
);
278 /* just one sb in list, splice to dispatch_queue and we're done */
280 list_splice(&tmp
, dispatch_queue
);
284 /* Move inodes from one superblock together */
285 while (!list_empty(&tmp
)) {
286 sb
= wb_inode(tmp
.prev
)->i_sb
;
287 list_for_each_prev_safe(pos
, node
, &tmp
) {
288 inode
= wb_inode(pos
);
289 if (inode
->i_sb
== sb
)
290 list_move(&inode
->i_wb_list
, dispatch_queue
);
298 * Queue all expired dirty inodes for io, eldest first.
300 * newly dirtied b_dirty b_io b_more_io
301 * =============> gf edc BA
303 * newly dirtied b_dirty b_io b_more_io
304 * =============> g fBAedc
306 * +--> dequeue for IO
308 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
)
311 assert_spin_locked(&wb
->list_lock
);
312 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
313 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, work
);
314 trace_writeback_queue_io(wb
, work
, moved
);
317 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
319 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
))
320 return inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
325 * Wait for writeback on an inode to complete. Called with i_lock held.
326 * Caller must make sure inode cannot go away when we drop i_lock.
328 static void __inode_wait_for_writeback(struct inode
*inode
)
329 __releases(inode
->i_lock
)
330 __acquires(inode
->i_lock
)
332 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
333 wait_queue_head_t
*wqh
;
335 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
336 while (inode
->i_state
& I_SYNC
) {
337 spin_unlock(&inode
->i_lock
);
338 __wait_on_bit(wqh
, &wq
, inode_wait
, TASK_UNINTERRUPTIBLE
);
339 spin_lock(&inode
->i_lock
);
344 * Wait for writeback on an inode to complete. Caller must have inode pinned.
346 void inode_wait_for_writeback(struct inode
*inode
)
348 spin_lock(&inode
->i_lock
);
349 __inode_wait_for_writeback(inode
);
350 spin_unlock(&inode
->i_lock
);
354 * Sleep until I_SYNC is cleared. This function must be called with i_lock
355 * held and drops it. It is aimed for callers not holding any inode reference
356 * so once i_lock is dropped, inode can go away.
358 static void inode_sleep_on_writeback(struct inode
*inode
)
359 __releases(inode
->i_lock
)
362 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
365 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
366 sleep
= inode
->i_state
& I_SYNC
;
367 spin_unlock(&inode
->i_lock
);
370 finish_wait(wqh
, &wait
);
374 * Find proper writeback list for the inode depending on its current state and
375 * possibly also change of its state while we were doing writeback. Here we
376 * handle things such as livelock prevention or fairness of writeback among
377 * inodes. This function can be called only by flusher thread - noone else
378 * processes all inodes in writeback lists and requeueing inodes behind flusher
379 * thread's back can have unexpected consequences.
381 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
382 struct writeback_control
*wbc
)
384 if (inode
->i_state
& I_FREEING
)
388 * Sync livelock prevention. Each inode is tagged and synced in one
389 * shot. If still dirty, it will be redirty_tail()'ed below. Update
390 * the dirty time to prevent enqueue and sync it again.
392 if ((inode
->i_state
& I_DIRTY
) &&
393 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
394 inode
->dirtied_when
= jiffies
;
396 if (wbc
->pages_skipped
) {
398 * writeback is not making progress due to locked
399 * buffers. Skip this inode for now.
401 redirty_tail(inode
, wb
);
405 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
407 * We didn't write back all the pages. nfs_writepages()
408 * sometimes bales out without doing anything.
410 if (wbc
->nr_to_write
<= 0) {
411 /* Slice used up. Queue for next turn. */
412 requeue_io(inode
, wb
);
415 * Writeback blocked by something other than
416 * congestion. Delay the inode for some time to
417 * avoid spinning on the CPU (100% iowait)
418 * retrying writeback of the dirty page/inode
419 * that cannot be performed immediately.
421 redirty_tail(inode
, wb
);
423 } else if (inode
->i_state
& I_DIRTY
) {
425 * Filesystems can dirty the inode during writeback operations,
426 * such as delayed allocation during submission or metadata
427 * updates after data IO completion.
429 redirty_tail(inode
, wb
);
431 /* The inode is clean. Remove from writeback lists. */
432 list_del_init(&inode
->i_wb_list
);
437 * Write out an inode and its dirty pages. Do not update the writeback list
438 * linkage. That is left to the caller. The caller is also responsible for
439 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
442 __writeback_single_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
443 struct writeback_control
*wbc
)
445 struct address_space
*mapping
= inode
->i_mapping
;
446 long nr_to_write
= wbc
->nr_to_write
;
450 WARN_ON(!(inode
->i_state
& I_SYNC
));
452 ret
= do_writepages(mapping
, wbc
);
455 * Make sure to wait on the data before writing out the metadata.
456 * This is important for filesystems that modify metadata on data
459 if (wbc
->sync_mode
== WB_SYNC_ALL
) {
460 int err
= filemap_fdatawait(mapping
);
466 * Some filesystems may redirty the inode during the writeback
467 * due to delalloc, clear dirty metadata flags right before
470 spin_lock(&inode
->i_lock
);
471 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
472 if (!mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
473 inode
->i_state
&= ~I_DIRTY_PAGES
;
474 dirty
= inode
->i_state
& I_DIRTY
;
475 inode
->i_state
&= ~(I_DIRTY_SYNC
| I_DIRTY_DATASYNC
);
476 spin_unlock(&inode
->i_lock
);
477 /* Don't write the inode if only I_DIRTY_PAGES was set */
478 if (dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
479 int err
= write_inode(inode
, wbc
);
483 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
488 * Write out an inode's dirty pages. Either the caller has an active reference
489 * on the inode or the inode has I_WILL_FREE set.
491 * This function is designed to be called for writing back one inode which
492 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
493 * and does more profound writeback list handling in writeback_sb_inodes().
496 writeback_single_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
497 struct writeback_control
*wbc
)
501 spin_lock(&inode
->i_lock
);
502 if (!atomic_read(&inode
->i_count
))
503 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
505 WARN_ON(inode
->i_state
& I_WILL_FREE
);
507 if (inode
->i_state
& I_SYNC
) {
508 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
511 * It's a data-integrity sync. We must wait. Since callers hold
512 * inode reference or inode has I_WILL_FREE set, it cannot go
515 __inode_wait_for_writeback(inode
);
517 WARN_ON(inode
->i_state
& I_SYNC
);
519 * Skip inode if it is clean. We don't want to mess with writeback
520 * lists in this function since flusher thread may be doing for example
521 * sync in parallel and if we move the inode, it could get skipped. So
522 * here we make sure inode is on some writeback list and leave it there
523 * unless we have completely cleaned the inode.
525 if (!(inode
->i_state
& I_DIRTY
))
527 inode
->i_state
|= I_SYNC
;
528 spin_unlock(&inode
->i_lock
);
530 ret
= __writeback_single_inode(inode
, wb
, wbc
);
532 spin_lock(&wb
->list_lock
);
533 spin_lock(&inode
->i_lock
);
535 * If inode is clean, remove it from writeback lists. Otherwise don't
536 * touch it. See comment above for explanation.
538 if (!(inode
->i_state
& I_DIRTY
))
539 list_del_init(&inode
->i_wb_list
);
540 spin_unlock(&wb
->list_lock
);
541 inode_sync_complete(inode
);
543 spin_unlock(&inode
->i_lock
);
547 static long writeback_chunk_size(struct backing_dev_info
*bdi
,
548 struct wb_writeback_work
*work
)
553 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
554 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
555 * here avoids calling into writeback_inodes_wb() more than once.
557 * The intended call sequence for WB_SYNC_ALL writeback is:
560 * writeback_sb_inodes() <== called only once
561 * write_cache_pages() <== called once for each inode
562 * (quickly) tag currently dirty pages
563 * (maybe slowly) sync all tagged pages
565 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
568 pages
= min(bdi
->avg_write_bandwidth
/ 2,
569 global_dirty_limit
/ DIRTY_SCOPE
);
570 pages
= min(pages
, work
->nr_pages
);
571 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
572 MIN_WRITEBACK_PAGES
);
579 * Write a portion of b_io inodes which belong to @sb.
581 * Return the number of pages and/or inodes written.
583 static long writeback_sb_inodes(struct super_block
*sb
,
584 struct bdi_writeback
*wb
,
585 struct wb_writeback_work
*work
)
587 struct writeback_control wbc
= {
588 .sync_mode
= work
->sync_mode
,
589 .tagged_writepages
= work
->tagged_writepages
,
590 .for_kupdate
= work
->for_kupdate
,
591 .for_background
= work
->for_background
,
592 .range_cyclic
= work
->range_cyclic
,
594 .range_end
= LLONG_MAX
,
596 unsigned long start_time
= jiffies
;
598 long wrote
= 0; /* count both pages and inodes */
600 while (!list_empty(&wb
->b_io
)) {
601 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
603 if (inode
->i_sb
!= sb
) {
606 * We only want to write back data for this
607 * superblock, move all inodes not belonging
608 * to it back onto the dirty list.
610 redirty_tail(inode
, wb
);
615 * The inode belongs to a different superblock.
616 * Bounce back to the caller to unpin this and
617 * pin the next superblock.
623 * Don't bother with new inodes or inodes being freed, first
624 * kind does not need periodic writeout yet, and for the latter
625 * kind writeout is handled by the freer.
627 spin_lock(&inode
->i_lock
);
628 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
629 spin_unlock(&inode
->i_lock
);
630 redirty_tail(inode
, wb
);
633 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
635 * If this inode is locked for writeback and we are not
636 * doing writeback-for-data-integrity, move it to
637 * b_more_io so that writeback can proceed with the
638 * other inodes on s_io.
640 * We'll have another go at writing back this inode
641 * when we completed a full scan of b_io.
643 spin_unlock(&inode
->i_lock
);
644 requeue_io(inode
, wb
);
645 trace_writeback_sb_inodes_requeue(inode
);
648 spin_unlock(&wb
->list_lock
);
651 * We already requeued the inode if it had I_SYNC set and we
652 * are doing WB_SYNC_NONE writeback. So this catches only the
655 if (inode
->i_state
& I_SYNC
) {
656 /* Wait for I_SYNC. This function drops i_lock... */
657 inode_sleep_on_writeback(inode
);
658 /* Inode may be gone, start again */
659 spin_lock(&wb
->list_lock
);
662 inode
->i_state
|= I_SYNC
;
663 spin_unlock(&inode
->i_lock
);
665 write_chunk
= writeback_chunk_size(wb
->bdi
, work
);
666 wbc
.nr_to_write
= write_chunk
;
667 wbc
.pages_skipped
= 0;
670 * We use I_SYNC to pin the inode in memory. While it is set
671 * evict_inode() will wait so the inode cannot be freed.
673 __writeback_single_inode(inode
, wb
, &wbc
);
675 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
676 wrote
+= write_chunk
- wbc
.nr_to_write
;
677 spin_lock(&wb
->list_lock
);
678 spin_lock(&inode
->i_lock
);
679 if (!(inode
->i_state
& I_DIRTY
))
681 requeue_inode(inode
, wb
, &wbc
);
682 inode_sync_complete(inode
);
683 spin_unlock(&inode
->i_lock
);
684 cond_resched_lock(&wb
->list_lock
);
686 * bail out to wb_writeback() often enough to check
687 * background threshold and other termination conditions.
690 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
692 if (work
->nr_pages
<= 0)
699 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
700 struct wb_writeback_work
*work
)
702 unsigned long start_time
= jiffies
;
705 while (!list_empty(&wb
->b_io
)) {
706 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
707 struct super_block
*sb
= inode
->i_sb
;
709 if (!grab_super_passive(sb
)) {
711 * grab_super_passive() may fail consistently due to
712 * s_umount being grabbed by someone else. Don't use
713 * requeue_io() to avoid busy retrying the inode/sb.
715 redirty_tail(inode
, wb
);
718 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
721 /* refer to the same tests at the end of writeback_sb_inodes */
723 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
725 if (work
->nr_pages
<= 0)
729 /* Leave any unwritten inodes on b_io */
733 long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
734 enum wb_reason reason
)
736 struct wb_writeback_work work
= {
737 .nr_pages
= nr_pages
,
738 .sync_mode
= WB_SYNC_NONE
,
743 spin_lock(&wb
->list_lock
);
744 if (list_empty(&wb
->b_io
))
746 __writeback_inodes_wb(wb
, &work
);
747 spin_unlock(&wb
->list_lock
);
749 return nr_pages
- work
.nr_pages
;
752 static bool over_bground_thresh(struct backing_dev_info
*bdi
)
754 unsigned long background_thresh
, dirty_thresh
;
756 global_dirty_limits(&background_thresh
, &dirty_thresh
);
758 if (global_page_state(NR_FILE_DIRTY
) +
759 global_page_state(NR_UNSTABLE_NFS
) > background_thresh
)
762 if (bdi_stat(bdi
, BDI_RECLAIMABLE
) >
763 bdi_dirty_limit(bdi
, background_thresh
))
770 * Called under wb->list_lock. If there are multiple wb per bdi,
771 * only the flusher working on the first wb should do it.
773 static void wb_update_bandwidth(struct bdi_writeback
*wb
,
774 unsigned long start_time
)
776 __bdi_update_bandwidth(wb
->bdi
, 0, 0, 0, 0, 0, start_time
);
780 * Explicit flushing or periodic writeback of "old" data.
782 * Define "old": the first time one of an inode's pages is dirtied, we mark the
783 * dirtying-time in the inode's address_space. So this periodic writeback code
784 * just walks the superblock inode list, writing back any inodes which are
785 * older than a specific point in time.
787 * Try to run once per dirty_writeback_interval. But if a writeback event
788 * takes longer than a dirty_writeback_interval interval, then leave a
791 * older_than_this takes precedence over nr_to_write. So we'll only write back
792 * all dirty pages if they are all attached to "old" mappings.
794 static long wb_writeback(struct bdi_writeback
*wb
,
795 struct wb_writeback_work
*work
)
797 unsigned long wb_start
= jiffies
;
798 long nr_pages
= work
->nr_pages
;
799 unsigned long oldest_jif
;
803 oldest_jif
= jiffies
;
804 work
->older_than_this
= &oldest_jif
;
806 spin_lock(&wb
->list_lock
);
809 * Stop writeback when nr_pages has been consumed
811 if (work
->nr_pages
<= 0)
815 * Background writeout and kupdate-style writeback may
816 * run forever. Stop them if there is other work to do
817 * so that e.g. sync can proceed. They'll be restarted
818 * after the other works are all done.
820 if ((work
->for_background
|| work
->for_kupdate
) &&
821 !list_empty(&wb
->bdi
->work_list
))
825 * For background writeout, stop when we are below the
826 * background dirty threshold
828 if (work
->for_background
&& !over_bground_thresh(wb
->bdi
))
832 * Kupdate and background works are special and we want to
833 * include all inodes that need writing. Livelock avoidance is
834 * handled by these works yielding to any other work so we are
837 if (work
->for_kupdate
) {
838 oldest_jif
= jiffies
-
839 msecs_to_jiffies(dirty_expire_interval
* 10);
840 } else if (work
->for_background
)
841 oldest_jif
= jiffies
;
843 trace_writeback_start(wb
->bdi
, work
);
844 if (list_empty(&wb
->b_io
))
847 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
849 progress
= __writeback_inodes_wb(wb
, work
);
850 trace_writeback_written(wb
->bdi
, work
);
852 wb_update_bandwidth(wb
, wb_start
);
855 * Did we write something? Try for more
857 * Dirty inodes are moved to b_io for writeback in batches.
858 * The completion of the current batch does not necessarily
859 * mean the overall work is done. So we keep looping as long
860 * as made some progress on cleaning pages or inodes.
865 * No more inodes for IO, bail
867 if (list_empty(&wb
->b_more_io
))
870 * Nothing written. Wait for some inode to
871 * become available for writeback. Otherwise
872 * we'll just busyloop.
874 if (!list_empty(&wb
->b_more_io
)) {
875 trace_writeback_wait(wb
->bdi
, work
);
876 inode
= wb_inode(wb
->b_more_io
.prev
);
877 spin_lock(&inode
->i_lock
);
878 spin_unlock(&wb
->list_lock
);
879 /* This function drops i_lock... */
880 inode_sleep_on_writeback(inode
);
881 spin_lock(&wb
->list_lock
);
884 spin_unlock(&wb
->list_lock
);
886 return nr_pages
- work
->nr_pages
;
890 * Return the next wb_writeback_work struct that hasn't been processed yet.
892 static struct wb_writeback_work
*
893 get_next_work_item(struct backing_dev_info
*bdi
)
895 struct wb_writeback_work
*work
= NULL
;
897 spin_lock_bh(&bdi
->wb_lock
);
898 if (!list_empty(&bdi
->work_list
)) {
899 work
= list_entry(bdi
->work_list
.next
,
900 struct wb_writeback_work
, list
);
901 list_del_init(&work
->list
);
903 spin_unlock_bh(&bdi
->wb_lock
);
908 * Add in the number of potentially dirty inodes, because each inode
909 * write can dirty pagecache in the underlying blockdev.
911 static unsigned long get_nr_dirty_pages(void)
913 return global_page_state(NR_FILE_DIRTY
) +
914 global_page_state(NR_UNSTABLE_NFS
) +
915 get_nr_dirty_inodes();
918 static long wb_check_background_flush(struct bdi_writeback
*wb
)
920 if (over_bground_thresh(wb
->bdi
)) {
922 struct wb_writeback_work work
= {
923 .nr_pages
= LONG_MAX
,
924 .sync_mode
= WB_SYNC_NONE
,
927 .reason
= WB_REASON_BACKGROUND
,
930 return wb_writeback(wb
, &work
);
936 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
938 unsigned long expired
;
942 * When set to zero, disable periodic writeback
944 if (!dirty_writeback_interval
)
947 expired
= wb
->last_old_flush
+
948 msecs_to_jiffies(dirty_writeback_interval
* 10);
949 if (time_before(jiffies
, expired
))
952 wb
->last_old_flush
= jiffies
;
953 nr_pages
= get_nr_dirty_pages();
956 struct wb_writeback_work work
= {
957 .nr_pages
= nr_pages
,
958 .sync_mode
= WB_SYNC_NONE
,
961 .reason
= WB_REASON_PERIODIC
,
964 return wb_writeback(wb
, &work
);
971 * Retrieve work items and do the writeback they describe
973 long wb_do_writeback(struct bdi_writeback
*wb
, int force_wait
)
975 struct backing_dev_info
*bdi
= wb
->bdi
;
976 struct wb_writeback_work
*work
;
979 set_bit(BDI_writeback_running
, &wb
->bdi
->state
);
980 while ((work
= get_next_work_item(bdi
)) != NULL
) {
982 * Override sync mode, in case we must wait for completion
983 * because this thread is exiting now.
986 work
->sync_mode
= WB_SYNC_ALL
;
988 trace_writeback_exec(bdi
, work
);
990 wrote
+= wb_writeback(wb
, work
);
993 * Notify the caller of completion if this is a synchronous
994 * work item, otherwise just free it.
997 complete(work
->done
);
1003 * Check for periodic writeback, kupdated() style
1005 wrote
+= wb_check_old_data_flush(wb
);
1006 wrote
+= wb_check_background_flush(wb
);
1007 clear_bit(BDI_writeback_running
, &wb
->bdi
->state
);
1013 * Handle writeback of dirty data for the device backed by this bdi. Also
1014 * wakes up periodically and does kupdated style flushing.
1016 int bdi_writeback_thread(void *data
)
1018 struct bdi_writeback
*wb
= data
;
1019 struct backing_dev_info
*bdi
= wb
->bdi
;
1022 current
->flags
|= PF_SWAPWRITE
;
1024 wb
->last_active
= jiffies
;
1027 * Our parent may run at a different priority, just set us to normal
1029 set_user_nice(current
, 0);
1031 trace_writeback_thread_start(bdi
);
1033 while (!kthread_freezable_should_stop(NULL
)) {
1035 * Remove own delayed wake-up timer, since we are already awake
1036 * and we'll take care of the preriodic write-back.
1038 del_timer(&wb
->wakeup_timer
);
1040 pages_written
= wb_do_writeback(wb
, 0);
1042 trace_writeback_pages_written(pages_written
);
1045 wb
->last_active
= jiffies
;
1047 set_current_state(TASK_INTERRUPTIBLE
);
1048 if (!list_empty(&bdi
->work_list
) || kthread_should_stop()) {
1049 __set_current_state(TASK_RUNNING
);
1053 if (wb_has_dirty_io(wb
) && dirty_writeback_interval
)
1054 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval
* 10));
1057 * We have nothing to do, so can go sleep without any
1058 * timeout and save power. When a work is queued or
1059 * something is made dirty - we will be woken up.
1065 /* Flush any work that raced with us exiting */
1066 if (!list_empty(&bdi
->work_list
))
1067 wb_do_writeback(wb
, 1);
1069 trace_writeback_thread_stop(bdi
);
1075 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1078 void wakeup_flusher_threads(long nr_pages
, enum wb_reason reason
)
1080 struct backing_dev_info
*bdi
;
1083 nr_pages
= global_page_state(NR_FILE_DIRTY
) +
1084 global_page_state(NR_UNSTABLE_NFS
);
1088 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
1089 if (!bdi_has_dirty_io(bdi
))
1091 __bdi_start_writeback(bdi
, nr_pages
, false, reason
);
1096 static noinline
void block_dump___mark_inode_dirty(struct inode
*inode
)
1098 if (inode
->i_ino
|| strcmp(inode
->i_sb
->s_id
, "bdev")) {
1099 struct dentry
*dentry
;
1100 const char *name
= "?";
1102 dentry
= d_find_alias(inode
);
1104 spin_lock(&dentry
->d_lock
);
1105 name
= (const char *) dentry
->d_name
.name
;
1108 "%s(%d): dirtied inode %lu (%s) on %s\n",
1109 current
->comm
, task_pid_nr(current
), inode
->i_ino
,
1110 name
, inode
->i_sb
->s_id
);
1112 spin_unlock(&dentry
->d_lock
);
1119 * __mark_inode_dirty - internal function
1120 * @inode: inode to mark
1121 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1122 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1123 * mark_inode_dirty_sync.
1125 * Put the inode on the super block's dirty list.
1127 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1128 * dirty list only if it is hashed or if it refers to a blockdev.
1129 * If it was not hashed, it will never be added to the dirty list
1130 * even if it is later hashed, as it will have been marked dirty already.
1132 * In short, make sure you hash any inodes _before_ you start marking
1135 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1136 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1137 * the kernel-internal blockdev inode represents the dirtying time of the
1138 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1139 * page->mapping->host, so the page-dirtying time is recorded in the internal
1142 void __mark_inode_dirty(struct inode
*inode
, int flags
)
1144 struct super_block
*sb
= inode
->i_sb
;
1145 struct backing_dev_info
*bdi
= NULL
;
1148 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1149 * dirty the inode itself
1151 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
1152 if (sb
->s_op
->dirty_inode
)
1153 sb
->s_op
->dirty_inode(inode
, flags
);
1157 * make sure that changes are seen by all cpus before we test i_state
1162 /* avoid the locking if we can */
1163 if ((inode
->i_state
& flags
) == flags
)
1166 if (unlikely(block_dump
))
1167 block_dump___mark_inode_dirty(inode
);
1169 spin_lock(&inode
->i_lock
);
1170 if ((inode
->i_state
& flags
) != flags
) {
1171 const int was_dirty
= inode
->i_state
& I_DIRTY
;
1173 inode
->i_state
|= flags
;
1176 * If the inode is being synced, just update its dirty state.
1177 * The unlocker will place the inode on the appropriate
1178 * superblock list, based upon its state.
1180 if (inode
->i_state
& I_SYNC
)
1181 goto out_unlock_inode
;
1184 * Only add valid (hashed) inodes to the superblock's
1185 * dirty list. Add blockdev inodes as well.
1187 if (!S_ISBLK(inode
->i_mode
)) {
1188 if (inode_unhashed(inode
))
1189 goto out_unlock_inode
;
1191 if (inode
->i_state
& I_FREEING
)
1192 goto out_unlock_inode
;
1195 * If the inode was already on b_dirty/b_io/b_more_io, don't
1196 * reposition it (that would break b_dirty time-ordering).
1199 bool wakeup_bdi
= false;
1200 bdi
= inode_to_bdi(inode
);
1202 if (bdi_cap_writeback_dirty(bdi
)) {
1203 WARN(!test_bit(BDI_registered
, &bdi
->state
),
1204 "bdi-%s not registered\n", bdi
->name
);
1207 * If this is the first dirty inode for this
1208 * bdi, we have to wake-up the corresponding
1209 * bdi thread to make sure background
1210 * write-back happens later.
1212 if (!wb_has_dirty_io(&bdi
->wb
))
1216 spin_unlock(&inode
->i_lock
);
1217 spin_lock(&bdi
->wb
.list_lock
);
1218 inode
->dirtied_when
= jiffies
;
1219 list_move(&inode
->i_wb_list
, &bdi
->wb
.b_dirty
);
1220 spin_unlock(&bdi
->wb
.list_lock
);
1223 bdi_wakeup_thread_delayed(bdi
);
1228 spin_unlock(&inode
->i_lock
);
1231 EXPORT_SYMBOL(__mark_inode_dirty
);
1233 static void wait_sb_inodes(struct super_block
*sb
)
1235 struct inode
*inode
, *old_inode
= NULL
;
1238 * We need to be protected against the filesystem going from
1239 * r/o to r/w or vice versa.
1241 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1243 spin_lock(&inode_sb_list_lock
);
1246 * Data integrity sync. Must wait for all pages under writeback,
1247 * because there may have been pages dirtied before our sync
1248 * call, but which had writeout started before we write it out.
1249 * In which case, the inode may not be on the dirty list, but
1250 * we still have to wait for that writeout.
1252 list_for_each_entry(inode
, &sb
->s_inodes
, i_sb_list
) {
1253 struct address_space
*mapping
= inode
->i_mapping
;
1255 spin_lock(&inode
->i_lock
);
1256 if ((inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) ||
1257 (mapping
->nrpages
== 0)) {
1258 spin_unlock(&inode
->i_lock
);
1262 spin_unlock(&inode
->i_lock
);
1263 spin_unlock(&inode_sb_list_lock
);
1266 * We hold a reference to 'inode' so it couldn't have been
1267 * removed from s_inodes list while we dropped the
1268 * inode_sb_list_lock. We cannot iput the inode now as we can
1269 * be holding the last reference and we cannot iput it under
1270 * inode_sb_list_lock. So we keep the reference and iput it
1276 filemap_fdatawait(mapping
);
1280 spin_lock(&inode_sb_list_lock
);
1282 spin_unlock(&inode_sb_list_lock
);
1287 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1288 * @sb: the superblock
1289 * @nr: the number of pages to write
1290 * @reason: reason why some writeback work initiated
1292 * Start writeback on some inodes on this super_block. No guarantees are made
1293 * on how many (if any) will be written, and this function does not wait
1294 * for IO completion of submitted IO.
1296 void writeback_inodes_sb_nr(struct super_block
*sb
,
1298 enum wb_reason reason
)
1300 DECLARE_COMPLETION_ONSTACK(done
);
1301 struct wb_writeback_work work
= {
1303 .sync_mode
= WB_SYNC_NONE
,
1304 .tagged_writepages
= 1,
1310 if (sb
->s_bdi
== &noop_backing_dev_info
)
1312 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1313 bdi_queue_work(sb
->s_bdi
, &work
);
1314 wait_for_completion(&done
);
1316 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
1319 * writeback_inodes_sb - writeback dirty inodes from given super_block
1320 * @sb: the superblock
1321 * @reason: reason why some writeback work was initiated
1323 * Start writeback on some inodes on this super_block. No guarantees are made
1324 * on how many (if any) will be written, and this function does not wait
1325 * for IO completion of submitted IO.
1327 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1329 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1331 EXPORT_SYMBOL(writeback_inodes_sb
);
1334 * writeback_inodes_sb_if_idle - start writeback if none underway
1335 * @sb: the superblock
1336 * @reason: reason why some writeback work was initiated
1338 * Invoke writeback_inodes_sb if no writeback is currently underway.
1339 * Returns 1 if writeback was started, 0 if not.
1341 int writeback_inodes_sb_if_idle(struct super_block
*sb
, enum wb_reason reason
)
1343 if (!writeback_in_progress(sb
->s_bdi
)) {
1344 down_read(&sb
->s_umount
);
1345 writeback_inodes_sb(sb
, reason
);
1346 up_read(&sb
->s_umount
);
1351 EXPORT_SYMBOL(writeback_inodes_sb_if_idle
);
1354 * writeback_inodes_sb_nr_if_idle - start writeback if none underway
1355 * @sb: the superblock
1356 * @nr: the number of pages to write
1357 * @reason: reason why some writeback work was initiated
1359 * Invoke writeback_inodes_sb if no writeback is currently underway.
1360 * Returns 1 if writeback was started, 0 if not.
1362 int writeback_inodes_sb_nr_if_idle(struct super_block
*sb
,
1364 enum wb_reason reason
)
1366 if (!writeback_in_progress(sb
->s_bdi
)) {
1367 down_read(&sb
->s_umount
);
1368 writeback_inodes_sb_nr(sb
, nr
, reason
);
1369 up_read(&sb
->s_umount
);
1374 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle
);
1377 * sync_inodes_sb - sync sb inode pages
1378 * @sb: the superblock
1380 * This function writes and waits on any dirty inode belonging to this
1383 void sync_inodes_sb(struct super_block
*sb
)
1385 DECLARE_COMPLETION_ONSTACK(done
);
1386 struct wb_writeback_work work
= {
1388 .sync_mode
= WB_SYNC_ALL
,
1389 .nr_pages
= LONG_MAX
,
1392 .reason
= WB_REASON_SYNC
,
1395 /* Nothing to do? */
1396 if (sb
->s_bdi
== &noop_backing_dev_info
)
1398 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1400 bdi_queue_work(sb
->s_bdi
, &work
);
1401 wait_for_completion(&done
);
1405 EXPORT_SYMBOL(sync_inodes_sb
);
1408 * write_inode_now - write an inode to disk
1409 * @inode: inode to write to disk
1410 * @sync: whether the write should be synchronous or not
1412 * This function commits an inode to disk immediately if it is dirty. This is
1413 * primarily needed by knfsd.
1415 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1417 int write_inode_now(struct inode
*inode
, int sync
)
1419 struct bdi_writeback
*wb
= &inode_to_bdi(inode
)->wb
;
1420 struct writeback_control wbc
= {
1421 .nr_to_write
= LONG_MAX
,
1422 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1424 .range_end
= LLONG_MAX
,
1427 if (!mapping_cap_writeback_dirty(inode
->i_mapping
))
1428 wbc
.nr_to_write
= 0;
1431 return writeback_single_inode(inode
, wb
, &wbc
);
1433 EXPORT_SYMBOL(write_inode_now
);
1436 * sync_inode - write an inode and its pages to disk.
1437 * @inode: the inode to sync
1438 * @wbc: controls the writeback mode
1440 * sync_inode() will write an inode and its pages to disk. It will also
1441 * correctly update the inode on its superblock's dirty inode lists and will
1442 * update inode->i_state.
1444 * The caller must have a ref on the inode.
1446 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1448 return writeback_single_inode(inode
, &inode_to_bdi(inode
)->wb
, wbc
);
1450 EXPORT_SYMBOL(sync_inode
);
1453 * sync_inode_metadata - write an inode to disk
1454 * @inode: the inode to sync
1455 * @wait: wait for I/O to complete.
1457 * Write an inode to disk and adjust its dirty state after completion.
1459 * Note: only writes the actual inode, no associated data or other metadata.
1461 int sync_inode_metadata(struct inode
*inode
, int wait
)
1463 struct writeback_control wbc
= {
1464 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1465 .nr_to_write
= 0, /* metadata-only */
1468 return sync_inode(inode
, &wbc
);
1470 EXPORT_SYMBOL(sync_inode_metadata
);