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_wakeup_thread(struct backing_dev_info
*bdi
)
92 spin_lock_bh(&bdi
->wb_lock
);
93 if (test_bit(BDI_registered
, &bdi
->state
))
94 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
95 spin_unlock_bh(&bdi
->wb_lock
);
98 static void bdi_queue_work(struct backing_dev_info
*bdi
,
99 struct wb_writeback_work
*work
)
101 trace_writeback_queue(bdi
, work
);
103 spin_lock_bh(&bdi
->wb_lock
);
104 if (!test_bit(BDI_registered
, &bdi
->state
)) {
106 complete(work
->done
);
109 list_add_tail(&work
->list
, &bdi
->work_list
);
110 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
112 spin_unlock_bh(&bdi
->wb_lock
);
116 __bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
117 bool range_cyclic
, enum wb_reason reason
)
119 struct wb_writeback_work
*work
;
122 * This is WB_SYNC_NONE writeback, so if allocation fails just
123 * wakeup the thread for old dirty data writeback
125 work
= kzalloc(sizeof(*work
), GFP_ATOMIC
);
127 trace_writeback_nowork(bdi
);
128 bdi_wakeup_thread(bdi
);
132 work
->sync_mode
= WB_SYNC_NONE
;
133 work
->nr_pages
= nr_pages
;
134 work
->range_cyclic
= range_cyclic
;
135 work
->reason
= reason
;
137 bdi_queue_work(bdi
, work
);
141 * bdi_start_writeback - start writeback
142 * @bdi: the backing device to write from
143 * @nr_pages: the number of pages to write
144 * @reason: reason why some writeback work was initiated
147 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
148 * started when this function returns, we make no guarantees on
149 * completion. Caller need not hold sb s_umount semaphore.
152 void bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
153 enum wb_reason reason
)
155 __bdi_start_writeback(bdi
, nr_pages
, true, reason
);
159 * bdi_start_background_writeback - start background writeback
160 * @bdi: the backing device to write from
163 * This makes sure WB_SYNC_NONE background writeback happens. When
164 * this function returns, it is only guaranteed that for given BDI
165 * some IO is happening if we are over background dirty threshold.
166 * Caller need not hold sb s_umount semaphore.
168 void bdi_start_background_writeback(struct backing_dev_info
*bdi
)
171 * We just wake up the flusher thread. It will perform background
172 * writeback as soon as there is no other work to do.
174 trace_writeback_wake_background(bdi
);
175 bdi_wakeup_thread(bdi
);
179 * Remove the inode from the writeback list it is on.
181 void inode_wb_list_del(struct inode
*inode
)
183 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
185 spin_lock(&bdi
->wb
.list_lock
);
186 list_del_init(&inode
->i_wb_list
);
187 spin_unlock(&bdi
->wb
.list_lock
);
191 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
192 * furthest end of its superblock's dirty-inode list.
194 * Before stamping the inode's ->dirtied_when, we check to see whether it is
195 * already the most-recently-dirtied inode on the b_dirty list. If that is
196 * the case then the inode must have been redirtied while it was being written
197 * out and we don't reset its dirtied_when.
199 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
201 assert_spin_locked(&wb
->list_lock
);
202 if (!list_empty(&wb
->b_dirty
)) {
205 tail
= wb_inode(wb
->b_dirty
.next
);
206 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
207 inode
->dirtied_when
= jiffies
;
209 list_move(&inode
->i_wb_list
, &wb
->b_dirty
);
213 * requeue inode for re-scanning after bdi->b_io list is exhausted.
215 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
217 assert_spin_locked(&wb
->list_lock
);
218 list_move(&inode
->i_wb_list
, &wb
->b_more_io
);
221 static void inode_sync_complete(struct inode
*inode
)
223 inode
->i_state
&= ~I_SYNC
;
224 /* If inode is clean an unused, put it into LRU now... */
225 inode_add_lru(inode
);
226 /* Waiters must see I_SYNC cleared before being woken up */
228 wake_up_bit(&inode
->i_state
, __I_SYNC
);
231 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
233 bool ret
= time_after(inode
->dirtied_when
, t
);
236 * For inodes being constantly redirtied, dirtied_when can get stuck.
237 * It _appears_ to be in the future, but is actually in distant past.
238 * This test is necessary to prevent such wrapped-around relative times
239 * from permanently stopping the whole bdi writeback.
241 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
247 * Move expired (dirtied before work->older_than_this) dirty inodes from
248 * @delaying_queue to @dispatch_queue.
250 static int move_expired_inodes(struct list_head
*delaying_queue
,
251 struct list_head
*dispatch_queue
,
252 struct wb_writeback_work
*work
)
255 struct list_head
*pos
, *node
;
256 struct super_block
*sb
= NULL
;
261 while (!list_empty(delaying_queue
)) {
262 inode
= wb_inode(delaying_queue
->prev
);
263 if (work
->older_than_this
&&
264 inode_dirtied_after(inode
, *work
->older_than_this
))
266 if (sb
&& sb
!= inode
->i_sb
)
269 list_move(&inode
->i_wb_list
, &tmp
);
273 /* just one sb in list, splice to dispatch_queue and we're done */
275 list_splice(&tmp
, dispatch_queue
);
279 /* Move inodes from one superblock together */
280 while (!list_empty(&tmp
)) {
281 sb
= wb_inode(tmp
.prev
)->i_sb
;
282 list_for_each_prev_safe(pos
, node
, &tmp
) {
283 inode
= wb_inode(pos
);
284 if (inode
->i_sb
== sb
)
285 list_move(&inode
->i_wb_list
, dispatch_queue
);
293 * Queue all expired dirty inodes for io, eldest first.
295 * newly dirtied b_dirty b_io b_more_io
296 * =============> gf edc BA
298 * newly dirtied b_dirty b_io b_more_io
299 * =============> g fBAedc
301 * +--> dequeue for IO
303 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
)
306 assert_spin_locked(&wb
->list_lock
);
307 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
308 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, work
);
309 trace_writeback_queue_io(wb
, work
, moved
);
312 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
316 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
317 trace_writeback_write_inode_start(inode
, wbc
);
318 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
319 trace_writeback_write_inode(inode
, wbc
);
326 * Wait for writeback on an inode to complete. Called with i_lock held.
327 * Caller must make sure inode cannot go away when we drop i_lock.
329 static void __inode_wait_for_writeback(struct inode
*inode
)
330 __releases(inode
->i_lock
)
331 __acquires(inode
->i_lock
)
333 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
334 wait_queue_head_t
*wqh
;
336 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
337 while (inode
->i_state
& I_SYNC
) {
338 spin_unlock(&inode
->i_lock
);
339 __wait_on_bit(wqh
, &wq
, inode_wait
, TASK_UNINTERRUPTIBLE
);
340 spin_lock(&inode
->i_lock
);
345 * Wait for writeback on an inode to complete. Caller must have inode pinned.
347 void inode_wait_for_writeback(struct inode
*inode
)
349 spin_lock(&inode
->i_lock
);
350 __inode_wait_for_writeback(inode
);
351 spin_unlock(&inode
->i_lock
);
355 * Sleep until I_SYNC is cleared. This function must be called with i_lock
356 * held and drops it. It is aimed for callers not holding any inode reference
357 * so once i_lock is dropped, inode can go away.
359 static void inode_sleep_on_writeback(struct inode
*inode
)
360 __releases(inode
->i_lock
)
363 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
366 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
367 sleep
= inode
->i_state
& I_SYNC
;
368 spin_unlock(&inode
->i_lock
);
371 finish_wait(wqh
, &wait
);
375 * Find proper writeback list for the inode depending on its current state and
376 * possibly also change of its state while we were doing writeback. Here we
377 * handle things such as livelock prevention or fairness of writeback among
378 * inodes. This function can be called only by flusher thread - noone else
379 * processes all inodes in writeback lists and requeueing inodes behind flusher
380 * thread's back can have unexpected consequences.
382 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
383 struct writeback_control
*wbc
)
385 if (inode
->i_state
& I_FREEING
)
389 * Sync livelock prevention. Each inode is tagged and synced in one
390 * shot. If still dirty, it will be redirty_tail()'ed below. Update
391 * the dirty time to prevent enqueue and sync it again.
393 if ((inode
->i_state
& I_DIRTY
) &&
394 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
395 inode
->dirtied_when
= jiffies
;
397 if (wbc
->pages_skipped
) {
399 * writeback is not making progress due to locked
400 * buffers. Skip this inode for now.
402 redirty_tail(inode
, wb
);
406 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
408 * We didn't write back all the pages. nfs_writepages()
409 * sometimes bales out without doing anything.
411 if (wbc
->nr_to_write
<= 0) {
412 /* Slice used up. Queue for next turn. */
413 requeue_io(inode
, wb
);
416 * Writeback blocked by something other than
417 * congestion. Delay the inode for some time to
418 * avoid spinning on the CPU (100% iowait)
419 * retrying writeback of the dirty page/inode
420 * that cannot be performed immediately.
422 redirty_tail(inode
, wb
);
424 } else if (inode
->i_state
& I_DIRTY
) {
426 * Filesystems can dirty the inode during writeback operations,
427 * such as delayed allocation during submission or metadata
428 * updates after data IO completion.
430 redirty_tail(inode
, wb
);
432 /* The inode is clean. Remove from writeback lists. */
433 list_del_init(&inode
->i_wb_list
);
438 * Write out an inode and its dirty pages. Do not update the writeback list
439 * linkage. That is left to the caller. The caller is also responsible for
440 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
443 __writeback_single_inode(struct inode
*inode
, 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 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
454 ret
= do_writepages(mapping
, wbc
);
457 * Make sure to wait on the data before writing out the metadata.
458 * This is important for filesystems that modify metadata on data
461 if (wbc
->sync_mode
== WB_SYNC_ALL
) {
462 int err
= filemap_fdatawait(mapping
);
468 * Some filesystems may redirty the inode during the writeback
469 * due to delalloc, clear dirty metadata flags right before
472 spin_lock(&inode
->i_lock
);
473 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
474 if (!mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
475 inode
->i_state
&= ~I_DIRTY_PAGES
;
476 dirty
= inode
->i_state
& I_DIRTY
;
477 inode
->i_state
&= ~(I_DIRTY_SYNC
| I_DIRTY_DATASYNC
);
478 spin_unlock(&inode
->i_lock
);
479 /* Don't write the inode if only I_DIRTY_PAGES was set */
480 if (dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
481 int err
= write_inode(inode
, wbc
);
485 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
490 * Write out an inode's dirty pages. Either the caller has an active reference
491 * on the inode or the inode has I_WILL_FREE set.
493 * This function is designed to be called for writing back one inode which
494 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
495 * and does more profound writeback list handling in writeback_sb_inodes().
498 writeback_single_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
499 struct writeback_control
*wbc
)
503 spin_lock(&inode
->i_lock
);
504 if (!atomic_read(&inode
->i_count
))
505 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
507 WARN_ON(inode
->i_state
& I_WILL_FREE
);
509 if (inode
->i_state
& I_SYNC
) {
510 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
513 * It's a data-integrity sync. We must wait. Since callers hold
514 * inode reference or inode has I_WILL_FREE set, it cannot go
517 __inode_wait_for_writeback(inode
);
519 WARN_ON(inode
->i_state
& I_SYNC
);
521 * Skip inode if it is clean and we have no outstanding writeback in
522 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
523 * function since flusher thread may be doing for example sync in
524 * parallel and if we move the inode, it could get skipped. So here we
525 * make sure inode is on some writeback list and leave it there unless
526 * we have completely cleaned the inode.
528 if (!(inode
->i_state
& I_DIRTY
) &&
529 (wbc
->sync_mode
!= WB_SYNC_ALL
||
530 !mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
)))
532 inode
->i_state
|= I_SYNC
;
533 spin_unlock(&inode
->i_lock
);
535 ret
= __writeback_single_inode(inode
, wbc
);
537 spin_lock(&wb
->list_lock
);
538 spin_lock(&inode
->i_lock
);
540 * If inode is clean, remove it from writeback lists. Otherwise don't
541 * touch it. See comment above for explanation.
543 if (!(inode
->i_state
& I_DIRTY
))
544 list_del_init(&inode
->i_wb_list
);
545 spin_unlock(&wb
->list_lock
);
546 inode_sync_complete(inode
);
548 spin_unlock(&inode
->i_lock
);
552 static long writeback_chunk_size(struct backing_dev_info
*bdi
,
553 struct wb_writeback_work
*work
)
558 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
559 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
560 * here avoids calling into writeback_inodes_wb() more than once.
562 * The intended call sequence for WB_SYNC_ALL writeback is:
565 * writeback_sb_inodes() <== called only once
566 * write_cache_pages() <== called once for each inode
567 * (quickly) tag currently dirty pages
568 * (maybe slowly) sync all tagged pages
570 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
573 pages
= min(bdi
->avg_write_bandwidth
/ 2,
574 global_dirty_limit
/ DIRTY_SCOPE
);
575 pages
= min(pages
, work
->nr_pages
);
576 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
577 MIN_WRITEBACK_PAGES
);
584 * Write a portion of b_io inodes which belong to @sb.
586 * Return the number of pages and/or inodes written.
588 static long writeback_sb_inodes(struct super_block
*sb
,
589 struct bdi_writeback
*wb
,
590 struct wb_writeback_work
*work
)
592 struct writeback_control wbc
= {
593 .sync_mode
= work
->sync_mode
,
594 .tagged_writepages
= work
->tagged_writepages
,
595 .for_kupdate
= work
->for_kupdate
,
596 .for_background
= work
->for_background
,
597 .range_cyclic
= work
->range_cyclic
,
599 .range_end
= LLONG_MAX
,
601 unsigned long start_time
= jiffies
;
603 long wrote
= 0; /* count both pages and inodes */
605 while (!list_empty(&wb
->b_io
)) {
606 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
608 if (inode
->i_sb
!= sb
) {
611 * We only want to write back data for this
612 * superblock, move all inodes not belonging
613 * to it back onto the dirty list.
615 redirty_tail(inode
, wb
);
620 * The inode belongs to a different superblock.
621 * Bounce back to the caller to unpin this and
622 * pin the next superblock.
628 * Don't bother with new inodes or inodes being freed, first
629 * kind does not need periodic writeout yet, and for the latter
630 * kind writeout is handled by the freer.
632 spin_lock(&inode
->i_lock
);
633 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
634 spin_unlock(&inode
->i_lock
);
635 redirty_tail(inode
, wb
);
638 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
640 * If this inode is locked for writeback and we are not
641 * doing writeback-for-data-integrity, move it to
642 * b_more_io so that writeback can proceed with the
643 * other inodes on s_io.
645 * We'll have another go at writing back this inode
646 * when we completed a full scan of b_io.
648 spin_unlock(&inode
->i_lock
);
649 requeue_io(inode
, wb
);
650 trace_writeback_sb_inodes_requeue(inode
);
653 spin_unlock(&wb
->list_lock
);
656 * We already requeued the inode if it had I_SYNC set and we
657 * are doing WB_SYNC_NONE writeback. So this catches only the
660 if (inode
->i_state
& I_SYNC
) {
661 /* Wait for I_SYNC. This function drops i_lock... */
662 inode_sleep_on_writeback(inode
);
663 /* Inode may be gone, start again */
664 spin_lock(&wb
->list_lock
);
667 inode
->i_state
|= I_SYNC
;
668 spin_unlock(&inode
->i_lock
);
670 write_chunk
= writeback_chunk_size(wb
->bdi
, work
);
671 wbc
.nr_to_write
= write_chunk
;
672 wbc
.pages_skipped
= 0;
675 * We use I_SYNC to pin the inode in memory. While it is set
676 * evict_inode() will wait so the inode cannot be freed.
678 __writeback_single_inode(inode
, &wbc
);
680 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
681 wrote
+= write_chunk
- wbc
.nr_to_write
;
682 spin_lock(&wb
->list_lock
);
683 spin_lock(&inode
->i_lock
);
684 if (!(inode
->i_state
& I_DIRTY
))
686 requeue_inode(inode
, wb
, &wbc
);
687 inode_sync_complete(inode
);
688 spin_unlock(&inode
->i_lock
);
689 cond_resched_lock(&wb
->list_lock
);
691 * bail out to wb_writeback() often enough to check
692 * background threshold and other termination conditions.
695 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
697 if (work
->nr_pages
<= 0)
704 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
705 struct wb_writeback_work
*work
)
707 unsigned long start_time
= jiffies
;
710 while (!list_empty(&wb
->b_io
)) {
711 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
712 struct super_block
*sb
= inode
->i_sb
;
714 if (!grab_super_passive(sb
)) {
716 * grab_super_passive() may fail consistently due to
717 * s_umount being grabbed by someone else. Don't use
718 * requeue_io() to avoid busy retrying the inode/sb.
720 redirty_tail(inode
, wb
);
723 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
726 /* refer to the same tests at the end of writeback_sb_inodes */
728 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
730 if (work
->nr_pages
<= 0)
734 /* Leave any unwritten inodes on b_io */
738 long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
739 enum wb_reason reason
)
741 struct wb_writeback_work work
= {
742 .nr_pages
= nr_pages
,
743 .sync_mode
= WB_SYNC_NONE
,
748 spin_lock(&wb
->list_lock
);
749 if (list_empty(&wb
->b_io
))
751 __writeback_inodes_wb(wb
, &work
);
752 spin_unlock(&wb
->list_lock
);
754 return nr_pages
- work
.nr_pages
;
757 static bool over_bground_thresh(struct backing_dev_info
*bdi
)
759 unsigned long background_thresh
, dirty_thresh
;
761 global_dirty_limits(&background_thresh
, &dirty_thresh
);
763 if (global_page_state(NR_FILE_DIRTY
) +
764 global_page_state(NR_UNSTABLE_NFS
) > background_thresh
)
767 if (bdi_stat(bdi
, BDI_RECLAIMABLE
) >
768 bdi_dirty_limit(bdi
, background_thresh
))
775 * Called under wb->list_lock. If there are multiple wb per bdi,
776 * only the flusher working on the first wb should do it.
778 static void wb_update_bandwidth(struct bdi_writeback
*wb
,
779 unsigned long start_time
)
781 __bdi_update_bandwidth(wb
->bdi
, 0, 0, 0, 0, 0, start_time
);
785 * Explicit flushing or periodic writeback of "old" data.
787 * Define "old": the first time one of an inode's pages is dirtied, we mark the
788 * dirtying-time in the inode's address_space. So this periodic writeback code
789 * just walks the superblock inode list, writing back any inodes which are
790 * older than a specific point in time.
792 * Try to run once per dirty_writeback_interval. But if a writeback event
793 * takes longer than a dirty_writeback_interval interval, then leave a
796 * older_than_this takes precedence over nr_to_write. So we'll only write back
797 * all dirty pages if they are all attached to "old" mappings.
799 static long wb_writeback(struct bdi_writeback
*wb
,
800 struct wb_writeback_work
*work
)
802 unsigned long wb_start
= jiffies
;
803 long nr_pages
= work
->nr_pages
;
804 unsigned long oldest_jif
;
808 oldest_jif
= jiffies
;
809 work
->older_than_this
= &oldest_jif
;
811 spin_lock(&wb
->list_lock
);
814 * Stop writeback when nr_pages has been consumed
816 if (work
->nr_pages
<= 0)
820 * Background writeout and kupdate-style writeback may
821 * run forever. Stop them if there is other work to do
822 * so that e.g. sync can proceed. They'll be restarted
823 * after the other works are all done.
825 if ((work
->for_background
|| work
->for_kupdate
) &&
826 !list_empty(&wb
->bdi
->work_list
))
830 * For background writeout, stop when we are below the
831 * background dirty threshold
833 if (work
->for_background
&& !over_bground_thresh(wb
->bdi
))
837 * Kupdate and background works are special and we want to
838 * include all inodes that need writing. Livelock avoidance is
839 * handled by these works yielding to any other work so we are
842 if (work
->for_kupdate
) {
843 oldest_jif
= jiffies
-
844 msecs_to_jiffies(dirty_expire_interval
* 10);
845 } else if (work
->for_background
)
846 oldest_jif
= jiffies
;
848 trace_writeback_start(wb
->bdi
, work
);
849 if (list_empty(&wb
->b_io
))
852 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
854 progress
= __writeback_inodes_wb(wb
, work
);
855 trace_writeback_written(wb
->bdi
, work
);
857 wb_update_bandwidth(wb
, wb_start
);
860 * Did we write something? Try for more
862 * Dirty inodes are moved to b_io for writeback in batches.
863 * The completion of the current batch does not necessarily
864 * mean the overall work is done. So we keep looping as long
865 * as made some progress on cleaning pages or inodes.
870 * No more inodes for IO, bail
872 if (list_empty(&wb
->b_more_io
))
875 * Nothing written. Wait for some inode to
876 * become available for writeback. Otherwise
877 * we'll just busyloop.
879 if (!list_empty(&wb
->b_more_io
)) {
880 trace_writeback_wait(wb
->bdi
, work
);
881 inode
= wb_inode(wb
->b_more_io
.prev
);
882 spin_lock(&inode
->i_lock
);
883 spin_unlock(&wb
->list_lock
);
884 /* This function drops i_lock... */
885 inode_sleep_on_writeback(inode
);
886 spin_lock(&wb
->list_lock
);
889 spin_unlock(&wb
->list_lock
);
891 return nr_pages
- work
->nr_pages
;
895 * Return the next wb_writeback_work struct that hasn't been processed yet.
897 static struct wb_writeback_work
*
898 get_next_work_item(struct backing_dev_info
*bdi
)
900 struct wb_writeback_work
*work
= NULL
;
902 spin_lock_bh(&bdi
->wb_lock
);
903 if (!list_empty(&bdi
->work_list
)) {
904 work
= list_entry(bdi
->work_list
.next
,
905 struct wb_writeback_work
, list
);
906 list_del_init(&work
->list
);
908 spin_unlock_bh(&bdi
->wb_lock
);
913 * Add in the number of potentially dirty inodes, because each inode
914 * write can dirty pagecache in the underlying blockdev.
916 static unsigned long get_nr_dirty_pages(void)
918 return global_page_state(NR_FILE_DIRTY
) +
919 global_page_state(NR_UNSTABLE_NFS
) +
920 get_nr_dirty_inodes();
923 static long wb_check_background_flush(struct bdi_writeback
*wb
)
925 if (over_bground_thresh(wb
->bdi
)) {
927 struct wb_writeback_work work
= {
928 .nr_pages
= LONG_MAX
,
929 .sync_mode
= WB_SYNC_NONE
,
932 .reason
= WB_REASON_BACKGROUND
,
935 return wb_writeback(wb
, &work
);
941 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
943 unsigned long expired
;
947 * When set to zero, disable periodic writeback
949 if (!dirty_writeback_interval
)
952 expired
= wb
->last_old_flush
+
953 msecs_to_jiffies(dirty_writeback_interval
* 10);
954 if (time_before(jiffies
, expired
))
957 wb
->last_old_flush
= jiffies
;
958 nr_pages
= get_nr_dirty_pages();
961 struct wb_writeback_work work
= {
962 .nr_pages
= nr_pages
,
963 .sync_mode
= WB_SYNC_NONE
,
966 .reason
= WB_REASON_PERIODIC
,
969 return wb_writeback(wb
, &work
);
976 * Retrieve work items and do the writeback they describe
978 long wb_do_writeback(struct bdi_writeback
*wb
, int force_wait
)
980 struct backing_dev_info
*bdi
= wb
->bdi
;
981 struct wb_writeback_work
*work
;
984 set_bit(BDI_writeback_running
, &wb
->bdi
->state
);
985 while ((work
= get_next_work_item(bdi
)) != NULL
) {
987 * Override sync mode, in case we must wait for completion
988 * because this thread is exiting now.
991 work
->sync_mode
= WB_SYNC_ALL
;
993 trace_writeback_exec(bdi
, work
);
995 wrote
+= wb_writeback(wb
, work
);
998 * Notify the caller of completion if this is a synchronous
999 * work item, otherwise just free it.
1002 complete(work
->done
);
1008 * Check for periodic writeback, kupdated() style
1010 wrote
+= wb_check_old_data_flush(wb
);
1011 wrote
+= wb_check_background_flush(wb
);
1012 clear_bit(BDI_writeback_running
, &wb
->bdi
->state
);
1018 * Handle writeback of dirty data for the device backed by this bdi. Also
1019 * reschedules periodically and does kupdated style flushing.
1021 void bdi_writeback_workfn(struct work_struct
*work
)
1023 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
1024 struct bdi_writeback
, dwork
);
1025 struct backing_dev_info
*bdi
= wb
->bdi
;
1028 set_worker_desc("flush-%s", dev_name(bdi
->dev
));
1029 current
->flags
|= PF_SWAPWRITE
;
1031 if (likely(!current_is_workqueue_rescuer() ||
1032 !test_bit(BDI_registered
, &bdi
->state
))) {
1034 * The normal path. Keep writing back @bdi until its
1035 * work_list is empty. Note that this path is also taken
1036 * if @bdi is shutting down even when we're running off the
1037 * rescuer as work_list needs to be drained.
1040 pages_written
= wb_do_writeback(wb
, 0);
1041 trace_writeback_pages_written(pages_written
);
1042 } while (!list_empty(&bdi
->work_list
));
1045 * bdi_wq can't get enough workers and we're running off
1046 * the emergency worker. Don't hog it. Hopefully, 1024 is
1047 * enough for efficient IO.
1049 pages_written
= writeback_inodes_wb(&bdi
->wb
, 1024,
1050 WB_REASON_FORKER_THREAD
);
1051 trace_writeback_pages_written(pages_written
);
1054 if (!list_empty(&bdi
->work_list
))
1055 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
1056 else if (wb_has_dirty_io(wb
) && dirty_writeback_interval
)
1057 bdi_wakeup_thread_delayed(bdi
);
1059 current
->flags
&= ~PF_SWAPWRITE
;
1063 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1066 void wakeup_flusher_threads(long nr_pages
, enum wb_reason reason
)
1068 struct backing_dev_info
*bdi
;
1071 nr_pages
= global_page_state(NR_FILE_DIRTY
) +
1072 global_page_state(NR_UNSTABLE_NFS
);
1076 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
1077 if (!bdi_has_dirty_io(bdi
))
1079 __bdi_start_writeback(bdi
, nr_pages
, false, reason
);
1084 static noinline
void block_dump___mark_inode_dirty(struct inode
*inode
)
1086 if (inode
->i_ino
|| strcmp(inode
->i_sb
->s_id
, "bdev")) {
1087 struct dentry
*dentry
;
1088 const char *name
= "?";
1090 dentry
= d_find_alias(inode
);
1092 spin_lock(&dentry
->d_lock
);
1093 name
= (const char *) dentry
->d_name
.name
;
1096 "%s(%d): dirtied inode %lu (%s) on %s\n",
1097 current
->comm
, task_pid_nr(current
), inode
->i_ino
,
1098 name
, inode
->i_sb
->s_id
);
1100 spin_unlock(&dentry
->d_lock
);
1107 * __mark_inode_dirty - internal function
1108 * @inode: inode to mark
1109 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1110 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1111 * mark_inode_dirty_sync.
1113 * Put the inode on the super block's dirty list.
1115 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1116 * dirty list only if it is hashed or if it refers to a blockdev.
1117 * If it was not hashed, it will never be added to the dirty list
1118 * even if it is later hashed, as it will have been marked dirty already.
1120 * In short, make sure you hash any inodes _before_ you start marking
1123 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1124 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1125 * the kernel-internal blockdev inode represents the dirtying time of the
1126 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1127 * page->mapping->host, so the page-dirtying time is recorded in the internal
1130 void __mark_inode_dirty(struct inode
*inode
, int flags
)
1132 struct super_block
*sb
= inode
->i_sb
;
1133 struct backing_dev_info
*bdi
= NULL
;
1136 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1137 * dirty the inode itself
1139 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
1140 trace_writeback_dirty_inode_start(inode
, flags
);
1142 if (sb
->s_op
->dirty_inode
)
1143 sb
->s_op
->dirty_inode(inode
, flags
);
1145 trace_writeback_dirty_inode(inode
, flags
);
1149 * make sure that changes are seen by all cpus before we test i_state
1154 /* avoid the locking if we can */
1155 if ((inode
->i_state
& flags
) == flags
)
1158 if (unlikely(block_dump
> 1))
1159 block_dump___mark_inode_dirty(inode
);
1161 spin_lock(&inode
->i_lock
);
1162 if ((inode
->i_state
& flags
) != flags
) {
1163 const int was_dirty
= inode
->i_state
& I_DIRTY
;
1165 inode
->i_state
|= flags
;
1168 * If the inode is being synced, just update its dirty state.
1169 * The unlocker will place the inode on the appropriate
1170 * superblock list, based upon its state.
1172 if (inode
->i_state
& I_SYNC
)
1173 goto out_unlock_inode
;
1176 * Only add valid (hashed) inodes to the superblock's
1177 * dirty list. Add blockdev inodes as well.
1179 if (!S_ISBLK(inode
->i_mode
)) {
1180 if (inode_unhashed(inode
))
1181 goto out_unlock_inode
;
1183 if (inode
->i_state
& I_FREEING
)
1184 goto out_unlock_inode
;
1187 * If the inode was already on b_dirty/b_io/b_more_io, don't
1188 * reposition it (that would break b_dirty time-ordering).
1191 bool wakeup_bdi
= false;
1192 bdi
= inode_to_bdi(inode
);
1194 if (bdi_cap_writeback_dirty(bdi
)) {
1195 WARN(!test_bit(BDI_registered
, &bdi
->state
),
1196 "bdi-%s not registered\n", bdi
->name
);
1199 * If this is the first dirty inode for this
1200 * bdi, we have to wake-up the corresponding
1201 * bdi thread to make sure background
1202 * write-back happens later.
1204 if (!wb_has_dirty_io(&bdi
->wb
))
1208 spin_unlock(&inode
->i_lock
);
1209 spin_lock(&bdi
->wb
.list_lock
);
1210 inode
->dirtied_when
= jiffies
;
1211 list_move(&inode
->i_wb_list
, &bdi
->wb
.b_dirty
);
1212 spin_unlock(&bdi
->wb
.list_lock
);
1215 bdi_wakeup_thread_delayed(bdi
);
1220 spin_unlock(&inode
->i_lock
);
1223 EXPORT_SYMBOL(__mark_inode_dirty
);
1225 static void wait_sb_inodes(struct super_block
*sb
)
1227 struct inode
*inode
, *old_inode
= NULL
;
1230 * We need to be protected against the filesystem going from
1231 * r/o to r/w or vice versa.
1233 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1235 spin_lock(&inode_sb_list_lock
);
1238 * Data integrity sync. Must wait for all pages under writeback,
1239 * because there may have been pages dirtied before our sync
1240 * call, but which had writeout started before we write it out.
1241 * In which case, the inode may not be on the dirty list, but
1242 * we still have to wait for that writeout.
1244 list_for_each_entry(inode
, &sb
->s_inodes
, i_sb_list
) {
1245 struct address_space
*mapping
= inode
->i_mapping
;
1247 spin_lock(&inode
->i_lock
);
1248 if ((inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) ||
1249 (mapping
->nrpages
== 0)) {
1250 spin_unlock(&inode
->i_lock
);
1254 spin_unlock(&inode
->i_lock
);
1255 spin_unlock(&inode_sb_list_lock
);
1258 * We hold a reference to 'inode' so it couldn't have been
1259 * removed from s_inodes list while we dropped the
1260 * inode_sb_list_lock. We cannot iput the inode now as we can
1261 * be holding the last reference and we cannot iput it under
1262 * inode_sb_list_lock. So we keep the reference and iput it
1268 filemap_fdatawait(mapping
);
1272 spin_lock(&inode_sb_list_lock
);
1274 spin_unlock(&inode_sb_list_lock
);
1279 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1280 * @sb: the superblock
1281 * @nr: the number of pages to write
1282 * @reason: reason why some writeback work initiated
1284 * Start writeback on some inodes on this super_block. No guarantees are made
1285 * on how many (if any) will be written, and this function does not wait
1286 * for IO completion of submitted IO.
1288 void writeback_inodes_sb_nr(struct super_block
*sb
,
1290 enum wb_reason reason
)
1292 DECLARE_COMPLETION_ONSTACK(done
);
1293 struct wb_writeback_work work
= {
1295 .sync_mode
= WB_SYNC_NONE
,
1296 .tagged_writepages
= 1,
1302 if (sb
->s_bdi
== &noop_backing_dev_info
)
1304 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1305 bdi_queue_work(sb
->s_bdi
, &work
);
1306 wait_for_completion(&done
);
1308 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
1311 * writeback_inodes_sb - writeback dirty inodes from given super_block
1312 * @sb: the superblock
1313 * @reason: reason why some writeback work was initiated
1315 * Start writeback on some inodes on this super_block. No guarantees are made
1316 * on how many (if any) will be written, and this function does not wait
1317 * for IO completion of submitted IO.
1319 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1321 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1323 EXPORT_SYMBOL(writeback_inodes_sb
);
1326 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1327 * @sb: the superblock
1328 * @nr: the number of pages to write
1329 * @reason: the reason of writeback
1331 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1332 * Returns 1 if writeback was started, 0 if not.
1334 int try_to_writeback_inodes_sb_nr(struct super_block
*sb
,
1336 enum wb_reason reason
)
1338 if (writeback_in_progress(sb
->s_bdi
))
1341 if (!down_read_trylock(&sb
->s_umount
))
1344 writeback_inodes_sb_nr(sb
, nr
, reason
);
1345 up_read(&sb
->s_umount
);
1348 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr
);
1351 * try_to_writeback_inodes_sb - try to start writeback if none underway
1352 * @sb: the superblock
1353 * @reason: reason why some writeback work was initiated
1355 * Implement by try_to_writeback_inodes_sb_nr()
1356 * Returns 1 if writeback was started, 0 if not.
1358 int try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1360 return try_to_writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1362 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
1365 * sync_inodes_sb - sync sb inode pages
1366 * @sb: the superblock
1368 * This function writes and waits on any dirty inode belonging to this
1371 void sync_inodes_sb(struct super_block
*sb
)
1373 DECLARE_COMPLETION_ONSTACK(done
);
1374 struct wb_writeback_work work
= {
1376 .sync_mode
= WB_SYNC_ALL
,
1377 .nr_pages
= LONG_MAX
,
1380 .reason
= WB_REASON_SYNC
,
1383 /* Nothing to do? */
1384 if (sb
->s_bdi
== &noop_backing_dev_info
)
1386 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1388 bdi_queue_work(sb
->s_bdi
, &work
);
1389 wait_for_completion(&done
);
1393 EXPORT_SYMBOL(sync_inodes_sb
);
1396 * write_inode_now - write an inode to disk
1397 * @inode: inode to write to disk
1398 * @sync: whether the write should be synchronous or not
1400 * This function commits an inode to disk immediately if it is dirty. This is
1401 * primarily needed by knfsd.
1403 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1405 int write_inode_now(struct inode
*inode
, int sync
)
1407 struct bdi_writeback
*wb
= &inode_to_bdi(inode
)->wb
;
1408 struct writeback_control wbc
= {
1409 .nr_to_write
= LONG_MAX
,
1410 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1412 .range_end
= LLONG_MAX
,
1415 if (!mapping_cap_writeback_dirty(inode
->i_mapping
))
1416 wbc
.nr_to_write
= 0;
1419 return writeback_single_inode(inode
, wb
, &wbc
);
1421 EXPORT_SYMBOL(write_inode_now
);
1424 * sync_inode - write an inode and its pages to disk.
1425 * @inode: the inode to sync
1426 * @wbc: controls the writeback mode
1428 * sync_inode() will write an inode and its pages to disk. It will also
1429 * correctly update the inode on its superblock's dirty inode lists and will
1430 * update inode->i_state.
1432 * The caller must have a ref on the inode.
1434 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1436 return writeback_single_inode(inode
, &inode_to_bdi(inode
)->wb
, wbc
);
1438 EXPORT_SYMBOL(sync_inode
);
1441 * sync_inode_metadata - write an inode to disk
1442 * @inode: the inode to sync
1443 * @wait: wait for I/O to complete.
1445 * Write an inode to disk and adjust its dirty state after completion.
1447 * Note: only writes the actual inode, no associated data or other metadata.
1449 int sync_inode_metadata(struct inode
*inode
, int wait
)
1451 struct writeback_control wbc
= {
1452 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1453 .nr_to_write
= 0, /* metadata-only */
1456 return sync_inode(inode
, &wbc
);
1458 EXPORT_SYMBOL(sync_inode_metadata
);