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>
29 #include <linux/device.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 unsigned int for_sync
:1; /* sync(2) WB_SYNC_ALL writeback */
50 enum wb_reason reason
; /* why was writeback initiated? */
52 struct list_head list
; /* pending work list */
53 struct completion
*done
; /* set if the caller waits */
57 * writeback_in_progress - determine whether there is writeback in progress
58 * @bdi: the device's backing_dev_info structure.
60 * Determine whether there is writeback waiting to be handled against a
63 int writeback_in_progress(struct backing_dev_info
*bdi
)
65 return test_bit(BDI_writeback_running
, &bdi
->state
);
67 EXPORT_SYMBOL(writeback_in_progress
);
69 struct backing_dev_info
*inode_to_bdi(struct inode
*inode
)
71 struct super_block
*sb
;
74 return &noop_backing_dev_info
;
78 if (sb_is_blkdev_sb(sb
))
79 return blk_get_backing_dev_info(I_BDEV(inode
));
83 EXPORT_SYMBOL_GPL(inode_to_bdi
);
85 static inline struct inode
*wb_inode(struct list_head
*head
)
87 return list_entry(head
, struct inode
, i_wb_list
);
91 * Include the creation of the trace points after defining the
92 * wb_writeback_work structure and inline functions so that the definition
93 * remains local to this file.
95 #define CREATE_TRACE_POINTS
96 #include <trace/events/writeback.h>
98 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage
);
100 static void bdi_wakeup_thread(struct backing_dev_info
*bdi
)
102 spin_lock_bh(&bdi
->wb_lock
);
103 if (test_bit(BDI_registered
, &bdi
->state
))
104 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
105 spin_unlock_bh(&bdi
->wb_lock
);
108 static void bdi_queue_work(struct backing_dev_info
*bdi
,
109 struct wb_writeback_work
*work
)
111 trace_writeback_queue(bdi
, work
);
113 spin_lock_bh(&bdi
->wb_lock
);
114 if (!test_bit(BDI_registered
, &bdi
->state
)) {
116 complete(work
->done
);
119 list_add_tail(&work
->list
, &bdi
->work_list
);
120 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
122 spin_unlock_bh(&bdi
->wb_lock
);
126 __bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
127 bool range_cyclic
, enum wb_reason reason
)
129 struct wb_writeback_work
*work
;
132 * This is WB_SYNC_NONE writeback, so if allocation fails just
133 * wakeup the thread for old dirty data writeback
135 work
= kzalloc(sizeof(*work
), GFP_ATOMIC
);
137 trace_writeback_nowork(bdi
);
138 bdi_wakeup_thread(bdi
);
142 work
->sync_mode
= WB_SYNC_NONE
;
143 work
->nr_pages
= nr_pages
;
144 work
->range_cyclic
= range_cyclic
;
145 work
->reason
= reason
;
147 bdi_queue_work(bdi
, work
);
151 * bdi_start_writeback - start writeback
152 * @bdi: the backing device to write from
153 * @nr_pages: the number of pages to write
154 * @reason: reason why some writeback work was initiated
157 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
158 * started when this function returns, we make no guarantees on
159 * completion. Caller need not hold sb s_umount semaphore.
162 void bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
163 enum wb_reason reason
)
165 __bdi_start_writeback(bdi
, nr_pages
, true, reason
);
169 * bdi_start_background_writeback - start background writeback
170 * @bdi: the backing device to write from
173 * This makes sure WB_SYNC_NONE background writeback happens. When
174 * this function returns, it is only guaranteed that for given BDI
175 * some IO is happening if we are over background dirty threshold.
176 * Caller need not hold sb s_umount semaphore.
178 void bdi_start_background_writeback(struct backing_dev_info
*bdi
)
181 * We just wake up the flusher thread. It will perform background
182 * writeback as soon as there is no other work to do.
184 trace_writeback_wake_background(bdi
);
185 bdi_wakeup_thread(bdi
);
189 * Remove the inode from the writeback list it is on.
191 void inode_wb_list_del(struct inode
*inode
)
193 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
195 spin_lock(&bdi
->wb
.list_lock
);
196 list_del_init(&inode
->i_wb_list
);
197 spin_unlock(&bdi
->wb
.list_lock
);
201 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
202 * furthest end of its superblock's dirty-inode list.
204 * Before stamping the inode's ->dirtied_when, we check to see whether it is
205 * already the most-recently-dirtied inode on the b_dirty list. If that is
206 * the case then the inode must have been redirtied while it was being written
207 * out and we don't reset its dirtied_when.
209 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
211 assert_spin_locked(&wb
->list_lock
);
212 if (!list_empty(&wb
->b_dirty
)) {
215 tail
= wb_inode(wb
->b_dirty
.next
);
216 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
217 inode
->dirtied_when
= jiffies
;
219 list_move(&inode
->i_wb_list
, &wb
->b_dirty
);
223 * requeue inode for re-scanning after bdi->b_io list is exhausted.
225 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
227 assert_spin_locked(&wb
->list_lock
);
228 list_move(&inode
->i_wb_list
, &wb
->b_more_io
);
231 static void inode_sync_complete(struct inode
*inode
)
233 inode
->i_state
&= ~I_SYNC
;
234 /* If inode is clean an unused, put it into LRU now... */
235 inode_add_lru(inode
);
236 /* Waiters must see I_SYNC cleared before being woken up */
238 wake_up_bit(&inode
->i_state
, __I_SYNC
);
241 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
243 bool ret
= time_after(inode
->dirtied_when
, t
);
246 * For inodes being constantly redirtied, dirtied_when can get stuck.
247 * It _appears_ to be in the future, but is actually in distant past.
248 * This test is necessary to prevent such wrapped-around relative times
249 * from permanently stopping the whole bdi writeback.
251 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
257 * Move expired (dirtied before work->older_than_this) dirty inodes from
258 * @delaying_queue to @dispatch_queue.
260 static int move_expired_inodes(struct list_head
*delaying_queue
,
261 struct list_head
*dispatch_queue
,
262 struct wb_writeback_work
*work
)
265 struct list_head
*pos
, *node
;
266 struct super_block
*sb
= NULL
;
271 while (!list_empty(delaying_queue
)) {
272 inode
= wb_inode(delaying_queue
->prev
);
273 if (work
->older_than_this
&&
274 inode_dirtied_after(inode
, *work
->older_than_this
))
276 list_move(&inode
->i_wb_list
, &tmp
);
278 if (sb_is_blkdev_sb(inode
->i_sb
))
280 if (sb
&& sb
!= inode
->i_sb
)
285 /* just one sb in list, splice to dispatch_queue and we're done */
287 list_splice(&tmp
, dispatch_queue
);
291 /* Move inodes from one superblock together */
292 while (!list_empty(&tmp
)) {
293 sb
= wb_inode(tmp
.prev
)->i_sb
;
294 list_for_each_prev_safe(pos
, node
, &tmp
) {
295 inode
= wb_inode(pos
);
296 if (inode
->i_sb
== sb
)
297 list_move(&inode
->i_wb_list
, dispatch_queue
);
305 * Queue all expired dirty inodes for io, eldest first.
307 * newly dirtied b_dirty b_io b_more_io
308 * =============> gf edc BA
310 * newly dirtied b_dirty b_io b_more_io
311 * =============> g fBAedc
313 * +--> dequeue for IO
315 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
)
318 assert_spin_locked(&wb
->list_lock
);
319 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
320 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, work
);
321 trace_writeback_queue_io(wb
, work
, moved
);
324 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
328 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
329 trace_writeback_write_inode_start(inode
, wbc
);
330 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
331 trace_writeback_write_inode(inode
, wbc
);
338 * Wait for writeback on an inode to complete. Called with i_lock held.
339 * Caller must make sure inode cannot go away when we drop i_lock.
341 static void __inode_wait_for_writeback(struct inode
*inode
)
342 __releases(inode
->i_lock
)
343 __acquires(inode
->i_lock
)
345 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
346 wait_queue_head_t
*wqh
;
348 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
349 while (inode
->i_state
& I_SYNC
) {
350 spin_unlock(&inode
->i_lock
);
351 __wait_on_bit(wqh
, &wq
, bit_wait
,
352 TASK_UNINTERRUPTIBLE
);
353 spin_lock(&inode
->i_lock
);
358 * Wait for writeback on an inode to complete. Caller must have inode pinned.
360 void inode_wait_for_writeback(struct inode
*inode
)
362 spin_lock(&inode
->i_lock
);
363 __inode_wait_for_writeback(inode
);
364 spin_unlock(&inode
->i_lock
);
368 * Sleep until I_SYNC is cleared. This function must be called with i_lock
369 * held and drops it. It is aimed for callers not holding any inode reference
370 * so once i_lock is dropped, inode can go away.
372 static void inode_sleep_on_writeback(struct inode
*inode
)
373 __releases(inode
->i_lock
)
376 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
379 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
380 sleep
= inode
->i_state
& I_SYNC
;
381 spin_unlock(&inode
->i_lock
);
384 finish_wait(wqh
, &wait
);
388 * Find proper writeback list for the inode depending on its current state and
389 * possibly also change of its state while we were doing writeback. Here we
390 * handle things such as livelock prevention or fairness of writeback among
391 * inodes. This function can be called only by flusher thread - noone else
392 * processes all inodes in writeback lists and requeueing inodes behind flusher
393 * thread's back can have unexpected consequences.
395 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
396 struct writeback_control
*wbc
)
398 if (inode
->i_state
& I_FREEING
)
402 * Sync livelock prevention. Each inode is tagged and synced in one
403 * shot. If still dirty, it will be redirty_tail()'ed below. Update
404 * the dirty time to prevent enqueue and sync it again.
406 if ((inode
->i_state
& I_DIRTY
) &&
407 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
408 inode
->dirtied_when
= jiffies
;
410 if (wbc
->pages_skipped
) {
412 * writeback is not making progress due to locked
413 * buffers. Skip this inode for now.
415 redirty_tail(inode
, wb
);
419 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
421 * We didn't write back all the pages. nfs_writepages()
422 * sometimes bales out without doing anything.
424 if (wbc
->nr_to_write
<= 0) {
425 /* Slice used up. Queue for next turn. */
426 requeue_io(inode
, wb
);
429 * Writeback blocked by something other than
430 * congestion. Delay the inode for some time to
431 * avoid spinning on the CPU (100% iowait)
432 * retrying writeback of the dirty page/inode
433 * that cannot be performed immediately.
435 redirty_tail(inode
, wb
);
437 } else if (inode
->i_state
& I_DIRTY
) {
439 * Filesystems can dirty the inode during writeback operations,
440 * such as delayed allocation during submission or metadata
441 * updates after data IO completion.
443 redirty_tail(inode
, wb
);
445 /* The inode is clean. Remove from writeback lists. */
446 list_del_init(&inode
->i_wb_list
);
451 * Write out an inode and its dirty pages. Do not update the writeback list
452 * linkage. That is left to the caller. The caller is also responsible for
453 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
456 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
458 struct address_space
*mapping
= inode
->i_mapping
;
459 long nr_to_write
= wbc
->nr_to_write
;
463 WARN_ON(!(inode
->i_state
& I_SYNC
));
465 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
467 ret
= do_writepages(mapping
, wbc
);
470 * Make sure to wait on the data before writing out the metadata.
471 * This is important for filesystems that modify metadata on data
472 * I/O completion. We don't do it for sync(2) writeback because it has a
473 * separate, external IO completion path and ->sync_fs for guaranteeing
474 * inode metadata is written back correctly.
476 if (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
) {
477 int err
= filemap_fdatawait(mapping
);
483 * Some filesystems may redirty the inode during the writeback
484 * due to delalloc, clear dirty metadata flags right before
487 spin_lock(&inode
->i_lock
);
489 dirty
= inode
->i_state
& I_DIRTY
;
490 inode
->i_state
&= ~I_DIRTY
;
493 * Paired with smp_mb() in __mark_inode_dirty(). This allows
494 * __mark_inode_dirty() to test i_state without grabbing i_lock -
495 * either they see the I_DIRTY bits cleared or we see the dirtied
498 * I_DIRTY_PAGES is always cleared together above even if @mapping
499 * still has dirty pages. The flag is reinstated after smp_mb() if
500 * necessary. This guarantees that either __mark_inode_dirty()
501 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
505 if (mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
506 inode
->i_state
|= I_DIRTY_PAGES
;
508 spin_unlock(&inode
->i_lock
);
510 /* Don't write the inode if only I_DIRTY_PAGES was set */
511 if (dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
512 int err
= write_inode(inode
, wbc
);
516 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
521 * Write out an inode's dirty pages. Either the caller has an active reference
522 * on the inode or the inode has I_WILL_FREE set.
524 * This function is designed to be called for writing back one inode which
525 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
526 * and does more profound writeback list handling in writeback_sb_inodes().
529 writeback_single_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
530 struct writeback_control
*wbc
)
534 spin_lock(&inode
->i_lock
);
535 if (!atomic_read(&inode
->i_count
))
536 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
538 WARN_ON(inode
->i_state
& I_WILL_FREE
);
540 if (inode
->i_state
& I_SYNC
) {
541 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
544 * It's a data-integrity sync. We must wait. Since callers hold
545 * inode reference or inode has I_WILL_FREE set, it cannot go
548 __inode_wait_for_writeback(inode
);
550 WARN_ON(inode
->i_state
& I_SYNC
);
552 * Skip inode if it is clean and we have no outstanding writeback in
553 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
554 * function since flusher thread may be doing for example sync in
555 * parallel and if we move the inode, it could get skipped. So here we
556 * make sure inode is on some writeback list and leave it there unless
557 * we have completely cleaned the inode.
559 if (!(inode
->i_state
& I_DIRTY
) &&
560 (wbc
->sync_mode
!= WB_SYNC_ALL
||
561 !mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
)))
563 inode
->i_state
|= I_SYNC
;
564 spin_unlock(&inode
->i_lock
);
566 ret
= __writeback_single_inode(inode
, wbc
);
568 spin_lock(&wb
->list_lock
);
569 spin_lock(&inode
->i_lock
);
571 * If inode is clean, remove it from writeback lists. Otherwise don't
572 * touch it. See comment above for explanation.
574 if (!(inode
->i_state
& I_DIRTY
))
575 list_del_init(&inode
->i_wb_list
);
576 spin_unlock(&wb
->list_lock
);
577 inode_sync_complete(inode
);
579 spin_unlock(&inode
->i_lock
);
583 static long writeback_chunk_size(struct backing_dev_info
*bdi
,
584 struct wb_writeback_work
*work
)
589 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
590 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
591 * here avoids calling into writeback_inodes_wb() more than once.
593 * The intended call sequence for WB_SYNC_ALL writeback is:
596 * writeback_sb_inodes() <== called only once
597 * write_cache_pages() <== called once for each inode
598 * (quickly) tag currently dirty pages
599 * (maybe slowly) sync all tagged pages
601 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
604 pages
= min(bdi
->avg_write_bandwidth
/ 2,
605 global_dirty_limit
/ DIRTY_SCOPE
);
606 pages
= min(pages
, work
->nr_pages
);
607 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
608 MIN_WRITEBACK_PAGES
);
615 * Write a portion of b_io inodes which belong to @sb.
617 * Return the number of pages and/or inodes written.
619 static long writeback_sb_inodes(struct super_block
*sb
,
620 struct bdi_writeback
*wb
,
621 struct wb_writeback_work
*work
)
623 struct writeback_control wbc
= {
624 .sync_mode
= work
->sync_mode
,
625 .tagged_writepages
= work
->tagged_writepages
,
626 .for_kupdate
= work
->for_kupdate
,
627 .for_background
= work
->for_background
,
628 .for_sync
= work
->for_sync
,
629 .range_cyclic
= work
->range_cyclic
,
631 .range_end
= LLONG_MAX
,
633 unsigned long start_time
= jiffies
;
635 long wrote
= 0; /* count both pages and inodes */
637 while (!list_empty(&wb
->b_io
)) {
638 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
640 if (inode
->i_sb
!= sb
) {
643 * We only want to write back data for this
644 * superblock, move all inodes not belonging
645 * to it back onto the dirty list.
647 redirty_tail(inode
, wb
);
652 * The inode belongs to a different superblock.
653 * Bounce back to the caller to unpin this and
654 * pin the next superblock.
660 * Don't bother with new inodes or inodes being freed, first
661 * kind does not need periodic writeout yet, and for the latter
662 * kind writeout is handled by the freer.
664 spin_lock(&inode
->i_lock
);
665 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
666 spin_unlock(&inode
->i_lock
);
667 redirty_tail(inode
, wb
);
670 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
672 * If this inode is locked for writeback and we are not
673 * doing writeback-for-data-integrity, move it to
674 * b_more_io so that writeback can proceed with the
675 * other inodes on s_io.
677 * We'll have another go at writing back this inode
678 * when we completed a full scan of b_io.
680 spin_unlock(&inode
->i_lock
);
681 requeue_io(inode
, wb
);
682 trace_writeback_sb_inodes_requeue(inode
);
685 spin_unlock(&wb
->list_lock
);
688 * We already requeued the inode if it had I_SYNC set and we
689 * are doing WB_SYNC_NONE writeback. So this catches only the
692 if (inode
->i_state
& I_SYNC
) {
693 /* Wait for I_SYNC. This function drops i_lock... */
694 inode_sleep_on_writeback(inode
);
695 /* Inode may be gone, start again */
696 spin_lock(&wb
->list_lock
);
699 inode
->i_state
|= I_SYNC
;
700 spin_unlock(&inode
->i_lock
);
702 write_chunk
= writeback_chunk_size(wb
->bdi
, work
);
703 wbc
.nr_to_write
= write_chunk
;
704 wbc
.pages_skipped
= 0;
707 * We use I_SYNC to pin the inode in memory. While it is set
708 * evict_inode() will wait so the inode cannot be freed.
710 __writeback_single_inode(inode
, &wbc
);
712 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
713 wrote
+= write_chunk
- wbc
.nr_to_write
;
714 spin_lock(&wb
->list_lock
);
715 spin_lock(&inode
->i_lock
);
716 if (!(inode
->i_state
& I_DIRTY
))
718 requeue_inode(inode
, wb
, &wbc
);
719 inode_sync_complete(inode
);
720 spin_unlock(&inode
->i_lock
);
721 cond_resched_lock(&wb
->list_lock
);
723 * bail out to wb_writeback() often enough to check
724 * background threshold and other termination conditions.
727 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
729 if (work
->nr_pages
<= 0)
736 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
737 struct wb_writeback_work
*work
)
739 unsigned long start_time
= jiffies
;
742 while (!list_empty(&wb
->b_io
)) {
743 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
744 struct super_block
*sb
= inode
->i_sb
;
746 if (!grab_super_passive(sb
)) {
748 * grab_super_passive() may fail consistently due to
749 * s_umount being grabbed by someone else. Don't use
750 * requeue_io() to avoid busy retrying the inode/sb.
752 redirty_tail(inode
, wb
);
755 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
758 /* refer to the same tests at the end of writeback_sb_inodes */
760 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
762 if (work
->nr_pages
<= 0)
766 /* Leave any unwritten inodes on b_io */
770 static long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
771 enum wb_reason reason
)
773 struct wb_writeback_work work
= {
774 .nr_pages
= nr_pages
,
775 .sync_mode
= WB_SYNC_NONE
,
780 spin_lock(&wb
->list_lock
);
781 if (list_empty(&wb
->b_io
))
783 __writeback_inodes_wb(wb
, &work
);
784 spin_unlock(&wb
->list_lock
);
786 return nr_pages
- work
.nr_pages
;
789 static bool over_bground_thresh(struct backing_dev_info
*bdi
)
791 unsigned long background_thresh
, dirty_thresh
;
793 global_dirty_limits(&background_thresh
, &dirty_thresh
);
795 if (global_page_state(NR_FILE_DIRTY
) +
796 global_page_state(NR_UNSTABLE_NFS
) > background_thresh
)
799 if (bdi_stat(bdi
, BDI_RECLAIMABLE
) >
800 bdi_dirty_limit(bdi
, background_thresh
))
807 * Called under wb->list_lock. If there are multiple wb per bdi,
808 * only the flusher working on the first wb should do it.
810 static void wb_update_bandwidth(struct bdi_writeback
*wb
,
811 unsigned long start_time
)
813 __bdi_update_bandwidth(wb
->bdi
, 0, 0, 0, 0, 0, start_time
);
817 * Explicit flushing or periodic writeback of "old" data.
819 * Define "old": the first time one of an inode's pages is dirtied, we mark the
820 * dirtying-time in the inode's address_space. So this periodic writeback code
821 * just walks the superblock inode list, writing back any inodes which are
822 * older than a specific point in time.
824 * Try to run once per dirty_writeback_interval. But if a writeback event
825 * takes longer than a dirty_writeback_interval interval, then leave a
828 * older_than_this takes precedence over nr_to_write. So we'll only write back
829 * all dirty pages if they are all attached to "old" mappings.
831 static long wb_writeback(struct bdi_writeback
*wb
,
832 struct wb_writeback_work
*work
)
834 unsigned long wb_start
= jiffies
;
835 long nr_pages
= work
->nr_pages
;
836 unsigned long oldest_jif
;
840 oldest_jif
= jiffies
;
841 work
->older_than_this
= &oldest_jif
;
843 spin_lock(&wb
->list_lock
);
846 * Stop writeback when nr_pages has been consumed
848 if (work
->nr_pages
<= 0)
852 * Background writeout and kupdate-style writeback may
853 * run forever. Stop them if there is other work to do
854 * so that e.g. sync can proceed. They'll be restarted
855 * after the other works are all done.
857 if ((work
->for_background
|| work
->for_kupdate
) &&
858 !list_empty(&wb
->bdi
->work_list
))
862 * For background writeout, stop when we are below the
863 * background dirty threshold
865 if (work
->for_background
&& !over_bground_thresh(wb
->bdi
))
869 * Kupdate and background works are special and we want to
870 * include all inodes that need writing. Livelock avoidance is
871 * handled by these works yielding to any other work so we are
874 if (work
->for_kupdate
) {
875 oldest_jif
= jiffies
-
876 msecs_to_jiffies(dirty_expire_interval
* 10);
877 } else if (work
->for_background
)
878 oldest_jif
= jiffies
;
880 trace_writeback_start(wb
->bdi
, work
);
881 if (list_empty(&wb
->b_io
))
884 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
886 progress
= __writeback_inodes_wb(wb
, work
);
887 trace_writeback_written(wb
->bdi
, work
);
889 wb_update_bandwidth(wb
, wb_start
);
892 * Did we write something? Try for more
894 * Dirty inodes are moved to b_io for writeback in batches.
895 * The completion of the current batch does not necessarily
896 * mean the overall work is done. So we keep looping as long
897 * as made some progress on cleaning pages or inodes.
902 * No more inodes for IO, bail
904 if (list_empty(&wb
->b_more_io
))
907 * Nothing written. Wait for some inode to
908 * become available for writeback. Otherwise
909 * we'll just busyloop.
911 if (!list_empty(&wb
->b_more_io
)) {
912 trace_writeback_wait(wb
->bdi
, work
);
913 inode
= wb_inode(wb
->b_more_io
.prev
);
914 spin_lock(&inode
->i_lock
);
915 spin_unlock(&wb
->list_lock
);
916 /* This function drops i_lock... */
917 inode_sleep_on_writeback(inode
);
918 spin_lock(&wb
->list_lock
);
921 spin_unlock(&wb
->list_lock
);
923 return nr_pages
- work
->nr_pages
;
927 * Return the next wb_writeback_work struct that hasn't been processed yet.
929 static struct wb_writeback_work
*
930 get_next_work_item(struct backing_dev_info
*bdi
)
932 struct wb_writeback_work
*work
= NULL
;
934 spin_lock_bh(&bdi
->wb_lock
);
935 if (!list_empty(&bdi
->work_list
)) {
936 work
= list_entry(bdi
->work_list
.next
,
937 struct wb_writeback_work
, list
);
938 list_del_init(&work
->list
);
940 spin_unlock_bh(&bdi
->wb_lock
);
945 * Add in the number of potentially dirty inodes, because each inode
946 * write can dirty pagecache in the underlying blockdev.
948 static unsigned long get_nr_dirty_pages(void)
950 return global_page_state(NR_FILE_DIRTY
) +
951 global_page_state(NR_UNSTABLE_NFS
) +
952 get_nr_dirty_inodes();
955 static long wb_check_background_flush(struct bdi_writeback
*wb
)
957 if (over_bground_thresh(wb
->bdi
)) {
959 struct wb_writeback_work work
= {
960 .nr_pages
= LONG_MAX
,
961 .sync_mode
= WB_SYNC_NONE
,
964 .reason
= WB_REASON_BACKGROUND
,
967 return wb_writeback(wb
, &work
);
973 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
975 unsigned long expired
;
979 * When set to zero, disable periodic writeback
981 if (!dirty_writeback_interval
)
984 expired
= wb
->last_old_flush
+
985 msecs_to_jiffies(dirty_writeback_interval
* 10);
986 if (time_before(jiffies
, expired
))
989 wb
->last_old_flush
= jiffies
;
990 nr_pages
= get_nr_dirty_pages();
993 struct wb_writeback_work work
= {
994 .nr_pages
= nr_pages
,
995 .sync_mode
= WB_SYNC_NONE
,
998 .reason
= WB_REASON_PERIODIC
,
1001 return wb_writeback(wb
, &work
);
1008 * Retrieve work items and do the writeback they describe
1010 static long wb_do_writeback(struct bdi_writeback
*wb
)
1012 struct backing_dev_info
*bdi
= wb
->bdi
;
1013 struct wb_writeback_work
*work
;
1016 set_bit(BDI_writeback_running
, &wb
->bdi
->state
);
1017 while ((work
= get_next_work_item(bdi
)) != NULL
) {
1019 trace_writeback_exec(bdi
, work
);
1021 wrote
+= wb_writeback(wb
, work
);
1024 * Notify the caller of completion if this is a synchronous
1025 * work item, otherwise just free it.
1028 complete(work
->done
);
1034 * Check for periodic writeback, kupdated() style
1036 wrote
+= wb_check_old_data_flush(wb
);
1037 wrote
+= wb_check_background_flush(wb
);
1038 clear_bit(BDI_writeback_running
, &wb
->bdi
->state
);
1044 * Handle writeback of dirty data for the device backed by this bdi. Also
1045 * reschedules periodically and does kupdated style flushing.
1047 void bdi_writeback_workfn(struct work_struct
*work
)
1049 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
1050 struct bdi_writeback
, dwork
);
1051 struct backing_dev_info
*bdi
= wb
->bdi
;
1054 set_worker_desc("flush-%s", dev_name(bdi
->dev
));
1055 current
->flags
|= PF_SWAPWRITE
;
1057 if (likely(!current_is_workqueue_rescuer() ||
1058 !test_bit(BDI_registered
, &bdi
->state
))) {
1060 * The normal path. Keep writing back @bdi until its
1061 * work_list is empty. Note that this path is also taken
1062 * if @bdi is shutting down even when we're running off the
1063 * rescuer as work_list needs to be drained.
1066 pages_written
= wb_do_writeback(wb
);
1067 trace_writeback_pages_written(pages_written
);
1068 } while (!list_empty(&bdi
->work_list
));
1071 * bdi_wq can't get enough workers and we're running off
1072 * the emergency worker. Don't hog it. Hopefully, 1024 is
1073 * enough for efficient IO.
1075 pages_written
= writeback_inodes_wb(&bdi
->wb
, 1024,
1076 WB_REASON_FORKER_THREAD
);
1077 trace_writeback_pages_written(pages_written
);
1080 if (!list_empty(&bdi
->work_list
))
1081 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
1082 else if (wb_has_dirty_io(wb
) && dirty_writeback_interval
)
1083 bdi_wakeup_thread_delayed(bdi
);
1085 current
->flags
&= ~PF_SWAPWRITE
;
1089 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1092 void wakeup_flusher_threads(long nr_pages
, enum wb_reason reason
)
1094 struct backing_dev_info
*bdi
;
1097 nr_pages
= get_nr_dirty_pages();
1100 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
1101 if (!bdi_has_dirty_io(bdi
))
1103 __bdi_start_writeback(bdi
, nr_pages
, false, reason
);
1108 static noinline
void block_dump___mark_inode_dirty(struct inode
*inode
)
1110 if (inode
->i_ino
|| strcmp(inode
->i_sb
->s_id
, "bdev")) {
1111 struct dentry
*dentry
;
1112 const char *name
= "?";
1114 dentry
= d_find_alias(inode
);
1116 spin_lock(&dentry
->d_lock
);
1117 name
= (const char *) dentry
->d_name
.name
;
1120 "%s(%d): dirtied inode %lu (%s) on %s\n",
1121 current
->comm
, task_pid_nr(current
), inode
->i_ino
,
1122 name
, inode
->i_sb
->s_id
);
1124 spin_unlock(&dentry
->d_lock
);
1131 * __mark_inode_dirty - internal function
1132 * @inode: inode to mark
1133 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1134 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1135 * mark_inode_dirty_sync.
1137 * Put the inode on the super block's dirty list.
1139 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1140 * dirty list only if it is hashed or if it refers to a blockdev.
1141 * If it was not hashed, it will never be added to the dirty list
1142 * even if it is later hashed, as it will have been marked dirty already.
1144 * In short, make sure you hash any inodes _before_ you start marking
1147 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1148 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1149 * the kernel-internal blockdev inode represents the dirtying time of the
1150 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1151 * page->mapping->host, so the page-dirtying time is recorded in the internal
1154 void __mark_inode_dirty(struct inode
*inode
, int flags
)
1156 struct super_block
*sb
= inode
->i_sb
;
1157 struct backing_dev_info
*bdi
= NULL
;
1160 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1161 * dirty the inode itself
1163 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
1164 trace_writeback_dirty_inode_start(inode
, flags
);
1166 if (sb
->s_op
->dirty_inode
)
1167 sb
->s_op
->dirty_inode(inode
, flags
);
1169 trace_writeback_dirty_inode(inode
, flags
);
1173 * Paired with smp_mb() in __writeback_single_inode() for the
1174 * following lockless i_state test. See there for details.
1178 if ((inode
->i_state
& flags
) == flags
)
1181 if (unlikely(block_dump
))
1182 block_dump___mark_inode_dirty(inode
);
1184 spin_lock(&inode
->i_lock
);
1185 if ((inode
->i_state
& flags
) != flags
) {
1186 const int was_dirty
= inode
->i_state
& I_DIRTY
;
1188 inode
->i_state
|= flags
;
1191 * If the inode is being synced, just update its dirty state.
1192 * The unlocker will place the inode on the appropriate
1193 * superblock list, based upon its state.
1195 if (inode
->i_state
& I_SYNC
)
1196 goto out_unlock_inode
;
1199 * Only add valid (hashed) inodes to the superblock's
1200 * dirty list. Add blockdev inodes as well.
1202 if (!S_ISBLK(inode
->i_mode
)) {
1203 if (inode_unhashed(inode
))
1204 goto out_unlock_inode
;
1206 if (inode
->i_state
& I_FREEING
)
1207 goto out_unlock_inode
;
1210 * If the inode was already on b_dirty/b_io/b_more_io, don't
1211 * reposition it (that would break b_dirty time-ordering).
1214 bool wakeup_bdi
= false;
1215 bdi
= inode_to_bdi(inode
);
1217 spin_unlock(&inode
->i_lock
);
1218 spin_lock(&bdi
->wb
.list_lock
);
1219 if (bdi_cap_writeback_dirty(bdi
)) {
1220 WARN(!test_bit(BDI_registered
, &bdi
->state
),
1221 "bdi-%s not registered\n", bdi
->name
);
1224 * If this is the first dirty inode for this
1225 * bdi, we have to wake-up the corresponding
1226 * bdi thread to make sure background
1227 * write-back happens later.
1229 if (!wb_has_dirty_io(&bdi
->wb
))
1233 inode
->dirtied_when
= jiffies
;
1234 list_move(&inode
->i_wb_list
, &bdi
->wb
.b_dirty
);
1235 spin_unlock(&bdi
->wb
.list_lock
);
1238 bdi_wakeup_thread_delayed(bdi
);
1243 spin_unlock(&inode
->i_lock
);
1246 EXPORT_SYMBOL(__mark_inode_dirty
);
1248 static void wait_sb_inodes(struct super_block
*sb
)
1250 struct inode
*inode
, *old_inode
= NULL
;
1253 * We need to be protected against the filesystem going from
1254 * r/o to r/w or vice versa.
1256 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1258 spin_lock(&inode_sb_list_lock
);
1261 * Data integrity sync. Must wait for all pages under writeback,
1262 * because there may have been pages dirtied before our sync
1263 * call, but which had writeout started before we write it out.
1264 * In which case, the inode may not be on the dirty list, but
1265 * we still have to wait for that writeout.
1267 list_for_each_entry(inode
, &sb
->s_inodes
, i_sb_list
) {
1268 struct address_space
*mapping
= inode
->i_mapping
;
1270 spin_lock(&inode
->i_lock
);
1271 if ((inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) ||
1272 (mapping
->nrpages
== 0)) {
1273 spin_unlock(&inode
->i_lock
);
1277 spin_unlock(&inode
->i_lock
);
1278 spin_unlock(&inode_sb_list_lock
);
1281 * We hold a reference to 'inode' so it couldn't have been
1282 * removed from s_inodes list while we dropped the
1283 * inode_sb_list_lock. We cannot iput the inode now as we can
1284 * be holding the last reference and we cannot iput it under
1285 * inode_sb_list_lock. So we keep the reference and iput it
1291 filemap_fdatawait(mapping
);
1295 spin_lock(&inode_sb_list_lock
);
1297 spin_unlock(&inode_sb_list_lock
);
1302 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1303 * @sb: the superblock
1304 * @nr: the number of pages to write
1305 * @reason: reason why some writeback work initiated
1307 * Start writeback on some inodes on this super_block. No guarantees are made
1308 * on how many (if any) will be written, and this function does not wait
1309 * for IO completion of submitted IO.
1311 void writeback_inodes_sb_nr(struct super_block
*sb
,
1313 enum wb_reason reason
)
1315 DECLARE_COMPLETION_ONSTACK(done
);
1316 struct wb_writeback_work work
= {
1318 .sync_mode
= WB_SYNC_NONE
,
1319 .tagged_writepages
= 1,
1325 if (sb
->s_bdi
== &noop_backing_dev_info
)
1327 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1328 bdi_queue_work(sb
->s_bdi
, &work
);
1329 wait_for_completion(&done
);
1331 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
1334 * writeback_inodes_sb - writeback dirty inodes from given super_block
1335 * @sb: the superblock
1336 * @reason: reason why some writeback work was initiated
1338 * Start writeback on some inodes on this super_block. No guarantees are made
1339 * on how many (if any) will be written, and this function does not wait
1340 * for IO completion of submitted IO.
1342 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1344 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1346 EXPORT_SYMBOL(writeback_inodes_sb
);
1349 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1350 * @sb: the superblock
1351 * @nr: the number of pages to write
1352 * @reason: the reason of writeback
1354 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1355 * Returns 1 if writeback was started, 0 if not.
1357 int try_to_writeback_inodes_sb_nr(struct super_block
*sb
,
1359 enum wb_reason reason
)
1361 if (writeback_in_progress(sb
->s_bdi
))
1364 if (!down_read_trylock(&sb
->s_umount
))
1367 writeback_inodes_sb_nr(sb
, nr
, reason
);
1368 up_read(&sb
->s_umount
);
1371 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr
);
1374 * try_to_writeback_inodes_sb - try to start writeback if none underway
1375 * @sb: the superblock
1376 * @reason: reason why some writeback work was initiated
1378 * Implement by try_to_writeback_inodes_sb_nr()
1379 * Returns 1 if writeback was started, 0 if not.
1381 int try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1383 return try_to_writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1385 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
1388 * sync_inodes_sb - sync sb inode pages
1389 * @sb: the superblock
1391 * This function writes and waits on any dirty inode belonging to this
1394 void sync_inodes_sb(struct super_block
*sb
)
1396 DECLARE_COMPLETION_ONSTACK(done
);
1397 struct wb_writeback_work work
= {
1399 .sync_mode
= WB_SYNC_ALL
,
1400 .nr_pages
= LONG_MAX
,
1403 .reason
= WB_REASON_SYNC
,
1407 /* Nothing to do? */
1408 if (sb
->s_bdi
== &noop_backing_dev_info
)
1410 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1412 bdi_queue_work(sb
->s_bdi
, &work
);
1413 wait_for_completion(&done
);
1417 EXPORT_SYMBOL(sync_inodes_sb
);
1420 * write_inode_now - write an inode to disk
1421 * @inode: inode to write to disk
1422 * @sync: whether the write should be synchronous or not
1424 * This function commits an inode to disk immediately if it is dirty. This is
1425 * primarily needed by knfsd.
1427 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1429 int write_inode_now(struct inode
*inode
, int sync
)
1431 struct bdi_writeback
*wb
= &inode_to_bdi(inode
)->wb
;
1432 struct writeback_control wbc
= {
1433 .nr_to_write
= LONG_MAX
,
1434 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1436 .range_end
= LLONG_MAX
,
1439 if (!mapping_cap_writeback_dirty(inode
->i_mapping
))
1440 wbc
.nr_to_write
= 0;
1443 return writeback_single_inode(inode
, wb
, &wbc
);
1445 EXPORT_SYMBOL(write_inode_now
);
1448 * sync_inode - write an inode and its pages to disk.
1449 * @inode: the inode to sync
1450 * @wbc: controls the writeback mode
1452 * sync_inode() will write an inode and its pages to disk. It will also
1453 * correctly update the inode on its superblock's dirty inode lists and will
1454 * update inode->i_state.
1456 * The caller must have a ref on the inode.
1458 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1460 return writeback_single_inode(inode
, &inode_to_bdi(inode
)->wb
, wbc
);
1462 EXPORT_SYMBOL(sync_inode
);
1465 * sync_inode_metadata - write an inode to disk
1466 * @inode: the inode to sync
1467 * @wait: wait for I/O to complete.
1469 * Write an inode to disk and adjust its dirty state after completion.
1471 * Note: only writes the actual inode, no associated data or other metadata.
1473 int sync_inode_metadata(struct inode
*inode
, int wait
)
1475 struct writeback_control wbc
= {
1476 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1477 .nr_to_write
= 0, /* metadata-only */
1480 return sync_inode(inode
, &wbc
);
1482 EXPORT_SYMBOL(sync_inode_metadata
);