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 unsigned int for_sync
:1; /* sync(2) WB_SYNC_ALL writeback */
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 (sb_is_blkdev_sb(sb
))
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 static void bdi_queue_work(struct backing_dev_info
*bdi
,
92 struct wb_writeback_work
*work
)
94 trace_writeback_queue(bdi
, work
);
96 spin_lock_bh(&bdi
->wb_lock
);
97 list_add_tail(&work
->list
, &bdi
->work_list
);
98 spin_unlock_bh(&bdi
->wb_lock
);
100 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
104 __bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
105 bool range_cyclic
, enum wb_reason reason
)
107 struct wb_writeback_work
*work
;
110 * This is WB_SYNC_NONE writeback, so if allocation fails just
111 * wakeup the thread for old dirty data writeback
113 work
= kzalloc(sizeof(*work
), GFP_ATOMIC
);
115 trace_writeback_nowork(bdi
);
116 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
120 work
->sync_mode
= WB_SYNC_NONE
;
121 work
->nr_pages
= nr_pages
;
122 work
->range_cyclic
= range_cyclic
;
123 work
->reason
= reason
;
125 bdi_queue_work(bdi
, work
);
129 * bdi_start_writeback - start writeback
130 * @bdi: the backing device to write from
131 * @nr_pages: the number of pages to write
132 * @reason: reason why some writeback work was initiated
135 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
136 * started when this function returns, we make no guarantees on
137 * completion. Caller need not hold sb s_umount semaphore.
140 void bdi_start_writeback(struct backing_dev_info
*bdi
, long nr_pages
,
141 enum wb_reason reason
)
143 __bdi_start_writeback(bdi
, nr_pages
, true, reason
);
147 * bdi_start_background_writeback - start background writeback
148 * @bdi: the backing device to write from
151 * This makes sure WB_SYNC_NONE background writeback happens. When
152 * this function returns, it is only guaranteed that for given BDI
153 * some IO is happening if we are over background dirty threshold.
154 * Caller need not hold sb s_umount semaphore.
156 void bdi_start_background_writeback(struct backing_dev_info
*bdi
)
159 * We just wake up the flusher thread. It will perform background
160 * writeback as soon as there is no other work to do.
162 trace_writeback_wake_background(bdi
);
163 mod_delayed_work(bdi_wq
, &bdi
->wb
.dwork
, 0);
167 * Remove the inode from the writeback list it is on.
169 void inode_wb_list_del(struct inode
*inode
)
171 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
173 spin_lock(&bdi
->wb
.list_lock
);
174 list_del_init(&inode
->i_wb_list
);
175 spin_unlock(&bdi
->wb
.list_lock
);
179 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
180 * furthest end of its superblock's dirty-inode list.
182 * Before stamping the inode's ->dirtied_when, we check to see whether it is
183 * already the most-recently-dirtied inode on the b_dirty list. If that is
184 * the case then the inode must have been redirtied while it was being written
185 * out and we don't reset its dirtied_when.
187 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
189 assert_spin_locked(&wb
->list_lock
);
190 if (!list_empty(&wb
->b_dirty
)) {
193 tail
= wb_inode(wb
->b_dirty
.next
);
194 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
195 inode
->dirtied_when
= jiffies
;
197 list_move(&inode
->i_wb_list
, &wb
->b_dirty
);
201 * requeue inode for re-scanning after bdi->b_io list is exhausted.
203 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
205 assert_spin_locked(&wb
->list_lock
);
206 list_move(&inode
->i_wb_list
, &wb
->b_more_io
);
209 static void inode_sync_complete(struct inode
*inode
)
211 inode
->i_state
&= ~I_SYNC
;
212 /* If inode is clean an unused, put it into LRU now... */
213 inode_add_lru(inode
);
214 /* Waiters must see I_SYNC cleared before being woken up */
216 wake_up_bit(&inode
->i_state
, __I_SYNC
);
219 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
221 bool ret
= time_after(inode
->dirtied_when
, t
);
224 * For inodes being constantly redirtied, dirtied_when can get stuck.
225 * It _appears_ to be in the future, but is actually in distant past.
226 * This test is necessary to prevent such wrapped-around relative times
227 * from permanently stopping the whole bdi writeback.
229 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
235 * Move expired (dirtied before work->older_than_this) dirty inodes from
236 * @delaying_queue to @dispatch_queue.
238 static int move_expired_inodes(struct list_head
*delaying_queue
,
239 struct list_head
*dispatch_queue
,
240 struct wb_writeback_work
*work
)
243 struct list_head
*pos
, *node
;
244 struct super_block
*sb
= NULL
;
249 while (!list_empty(delaying_queue
)) {
250 inode
= wb_inode(delaying_queue
->prev
);
251 if (work
->older_than_this
&&
252 inode_dirtied_after(inode
, *work
->older_than_this
))
254 list_move(&inode
->i_wb_list
, &tmp
);
256 if (sb_is_blkdev_sb(inode
->i_sb
))
258 if (sb
&& sb
!= inode
->i_sb
)
263 /* just one sb in list, splice to dispatch_queue and we're done */
265 list_splice(&tmp
, dispatch_queue
);
269 /* Move inodes from one superblock together */
270 while (!list_empty(&tmp
)) {
271 sb
= wb_inode(tmp
.prev
)->i_sb
;
272 list_for_each_prev_safe(pos
, node
, &tmp
) {
273 inode
= wb_inode(pos
);
274 if (inode
->i_sb
== sb
)
275 list_move(&inode
->i_wb_list
, dispatch_queue
);
283 * Queue all expired dirty inodes for io, eldest first.
285 * newly dirtied b_dirty b_io b_more_io
286 * =============> gf edc BA
288 * newly dirtied b_dirty b_io b_more_io
289 * =============> g fBAedc
291 * +--> dequeue for IO
293 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
)
296 assert_spin_locked(&wb
->list_lock
);
297 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
298 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, work
);
299 trace_writeback_queue_io(wb
, work
, moved
);
302 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
306 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
307 trace_writeback_write_inode_start(inode
, wbc
);
308 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
309 trace_writeback_write_inode(inode
, wbc
);
316 * Wait for writeback on an inode to complete. Called with i_lock held.
317 * Caller must make sure inode cannot go away when we drop i_lock.
319 static void __inode_wait_for_writeback(struct inode
*inode
)
320 __releases(inode
->i_lock
)
321 __acquires(inode
->i_lock
)
323 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
324 wait_queue_head_t
*wqh
;
326 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
327 while (inode
->i_state
& I_SYNC
) {
328 spin_unlock(&inode
->i_lock
);
329 __wait_on_bit(wqh
, &wq
, inode_wait
, TASK_UNINTERRUPTIBLE
);
330 spin_lock(&inode
->i_lock
);
335 * Wait for writeback on an inode to complete. Caller must have inode pinned.
337 void inode_wait_for_writeback(struct inode
*inode
)
339 spin_lock(&inode
->i_lock
);
340 __inode_wait_for_writeback(inode
);
341 spin_unlock(&inode
->i_lock
);
345 * Sleep until I_SYNC is cleared. This function must be called with i_lock
346 * held and drops it. It is aimed for callers not holding any inode reference
347 * so once i_lock is dropped, inode can go away.
349 static void inode_sleep_on_writeback(struct inode
*inode
)
350 __releases(inode
->i_lock
)
353 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
356 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
357 sleep
= inode
->i_state
& I_SYNC
;
358 spin_unlock(&inode
->i_lock
);
361 finish_wait(wqh
, &wait
);
365 * Find proper writeback list for the inode depending on its current state and
366 * possibly also change of its state while we were doing writeback. Here we
367 * handle things such as livelock prevention or fairness of writeback among
368 * inodes. This function can be called only by flusher thread - noone else
369 * processes all inodes in writeback lists and requeueing inodes behind flusher
370 * thread's back can have unexpected consequences.
372 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
373 struct writeback_control
*wbc
)
375 if (inode
->i_state
& I_FREEING
)
379 * Sync livelock prevention. Each inode is tagged and synced in one
380 * shot. If still dirty, it will be redirty_tail()'ed below. Update
381 * the dirty time to prevent enqueue and sync it again.
383 if ((inode
->i_state
& I_DIRTY
) &&
384 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
385 inode
->dirtied_when
= jiffies
;
387 if (wbc
->pages_skipped
) {
389 * writeback is not making progress due to locked
390 * buffers. Skip this inode for now.
392 redirty_tail(inode
, wb
);
396 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
398 * We didn't write back all the pages. nfs_writepages()
399 * sometimes bales out without doing anything.
401 if (wbc
->nr_to_write
<= 0) {
402 /* Slice used up. Queue for next turn. */
403 requeue_io(inode
, wb
);
406 * Writeback blocked by something other than
407 * congestion. Delay the inode for some time to
408 * avoid spinning on the CPU (100% iowait)
409 * retrying writeback of the dirty page/inode
410 * that cannot be performed immediately.
412 redirty_tail(inode
, wb
);
414 } else if (inode
->i_state
& I_DIRTY
) {
416 * Filesystems can dirty the inode during writeback operations,
417 * such as delayed allocation during submission or metadata
418 * updates after data IO completion.
420 redirty_tail(inode
, wb
);
422 /* The inode is clean. Remove from writeback lists. */
423 list_del_init(&inode
->i_wb_list
);
428 * Write out an inode and its dirty pages. Do not update the writeback list
429 * linkage. That is left to the caller. The caller is also responsible for
430 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
433 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
435 struct address_space
*mapping
= inode
->i_mapping
;
436 long nr_to_write
= wbc
->nr_to_write
;
440 WARN_ON(!(inode
->i_state
& I_SYNC
));
442 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
444 ret
= do_writepages(mapping
, wbc
);
447 * Make sure to wait on the data before writing out the metadata.
448 * This is important for filesystems that modify metadata on data
449 * I/O completion. We don't do it for sync(2) writeback because it has a
450 * separate, external IO completion path and ->sync_fs for guaranteeing
451 * inode metadata is written back correctly.
453 if (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
) {
454 int err
= filemap_fdatawait(mapping
);
460 * Some filesystems may redirty the inode during the writeback
461 * due to delalloc, clear dirty metadata flags right before
464 spin_lock(&inode
->i_lock
);
465 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
466 if (!mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
467 inode
->i_state
&= ~I_DIRTY_PAGES
;
468 dirty
= inode
->i_state
& I_DIRTY
;
469 inode
->i_state
&= ~(I_DIRTY_SYNC
| I_DIRTY_DATASYNC
);
470 spin_unlock(&inode
->i_lock
);
471 /* Don't write the inode if only I_DIRTY_PAGES was set */
472 if (dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
473 int err
= write_inode(inode
, wbc
);
477 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
482 * Write out an inode's dirty pages. Either the caller has an active reference
483 * on the inode or the inode has I_WILL_FREE set.
485 * This function is designed to be called for writing back one inode which
486 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
487 * and does more profound writeback list handling in writeback_sb_inodes().
490 writeback_single_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
491 struct writeback_control
*wbc
)
495 spin_lock(&inode
->i_lock
);
496 if (!atomic_read(&inode
->i_count
))
497 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
499 WARN_ON(inode
->i_state
& I_WILL_FREE
);
501 if (inode
->i_state
& I_SYNC
) {
502 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
505 * It's a data-integrity sync. We must wait. Since callers hold
506 * inode reference or inode has I_WILL_FREE set, it cannot go
509 __inode_wait_for_writeback(inode
);
511 WARN_ON(inode
->i_state
& I_SYNC
);
513 * Skip inode if it is clean. We don't want to mess with writeback
514 * lists in this function since flusher thread may be doing for example
515 * sync in parallel and if we move the inode, it could get skipped. So
516 * here we make sure inode is on some writeback list and leave it there
517 * unless we have completely cleaned the inode.
519 if (!(inode
->i_state
& I_DIRTY
))
521 inode
->i_state
|= I_SYNC
;
522 spin_unlock(&inode
->i_lock
);
524 ret
= __writeback_single_inode(inode
, wbc
);
526 spin_lock(&wb
->list_lock
);
527 spin_lock(&inode
->i_lock
);
529 * If inode is clean, remove it from writeback lists. Otherwise don't
530 * touch it. See comment above for explanation.
532 if (!(inode
->i_state
& I_DIRTY
))
533 list_del_init(&inode
->i_wb_list
);
534 spin_unlock(&wb
->list_lock
);
535 inode_sync_complete(inode
);
537 spin_unlock(&inode
->i_lock
);
541 static long writeback_chunk_size(struct backing_dev_info
*bdi
,
542 struct wb_writeback_work
*work
)
547 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
548 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
549 * here avoids calling into writeback_inodes_wb() more than once.
551 * The intended call sequence for WB_SYNC_ALL writeback is:
554 * writeback_sb_inodes() <== called only once
555 * write_cache_pages() <== called once for each inode
556 * (quickly) tag currently dirty pages
557 * (maybe slowly) sync all tagged pages
559 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
562 pages
= min(bdi
->avg_write_bandwidth
/ 2,
563 global_dirty_limit
/ DIRTY_SCOPE
);
564 pages
= min(pages
, work
->nr_pages
);
565 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
566 MIN_WRITEBACK_PAGES
);
573 * Write a portion of b_io inodes which belong to @sb.
575 * Return the number of pages and/or inodes written.
577 static long writeback_sb_inodes(struct super_block
*sb
,
578 struct bdi_writeback
*wb
,
579 struct wb_writeback_work
*work
)
581 struct writeback_control wbc
= {
582 .sync_mode
= work
->sync_mode
,
583 .tagged_writepages
= work
->tagged_writepages
,
584 .for_kupdate
= work
->for_kupdate
,
585 .for_background
= work
->for_background
,
586 .for_sync
= work
->for_sync
,
587 .range_cyclic
= work
->range_cyclic
,
589 .range_end
= LLONG_MAX
,
591 unsigned long start_time
= jiffies
;
593 long wrote
= 0; /* count both pages and inodes */
595 while (!list_empty(&wb
->b_io
)) {
596 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
598 if (inode
->i_sb
!= sb
) {
601 * We only want to write back data for this
602 * superblock, move all inodes not belonging
603 * to it back onto the dirty list.
605 redirty_tail(inode
, wb
);
610 * The inode belongs to a different superblock.
611 * Bounce back to the caller to unpin this and
612 * pin the next superblock.
618 * Don't bother with new inodes or inodes being freed, first
619 * kind does not need periodic writeout yet, and for the latter
620 * kind writeout is handled by the freer.
622 spin_lock(&inode
->i_lock
);
623 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
624 spin_unlock(&inode
->i_lock
);
625 redirty_tail(inode
, wb
);
628 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
630 * If this inode is locked for writeback and we are not
631 * doing writeback-for-data-integrity, move it to
632 * b_more_io so that writeback can proceed with the
633 * other inodes on s_io.
635 * We'll have another go at writing back this inode
636 * when we completed a full scan of b_io.
638 spin_unlock(&inode
->i_lock
);
639 requeue_io(inode
, wb
);
640 trace_writeback_sb_inodes_requeue(inode
);
643 spin_unlock(&wb
->list_lock
);
646 * We already requeued the inode if it had I_SYNC set and we
647 * are doing WB_SYNC_NONE writeback. So this catches only the
650 if (inode
->i_state
& I_SYNC
) {
651 /* Wait for I_SYNC. This function drops i_lock... */
652 inode_sleep_on_writeback(inode
);
653 /* Inode may be gone, start again */
654 spin_lock(&wb
->list_lock
);
657 inode
->i_state
|= I_SYNC
;
658 spin_unlock(&inode
->i_lock
);
660 write_chunk
= writeback_chunk_size(wb
->bdi
, work
);
661 wbc
.nr_to_write
= write_chunk
;
662 wbc
.pages_skipped
= 0;
665 * We use I_SYNC to pin the inode in memory. While it is set
666 * evict_inode() will wait so the inode cannot be freed.
668 __writeback_single_inode(inode
, &wbc
);
670 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
671 wrote
+= write_chunk
- wbc
.nr_to_write
;
672 spin_lock(&wb
->list_lock
);
673 spin_lock(&inode
->i_lock
);
674 if (!(inode
->i_state
& I_DIRTY
))
676 requeue_inode(inode
, wb
, &wbc
);
677 inode_sync_complete(inode
);
678 spin_unlock(&inode
->i_lock
);
679 cond_resched_lock(&wb
->list_lock
);
681 * bail out to wb_writeback() often enough to check
682 * background threshold and other termination conditions.
685 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
687 if (work
->nr_pages
<= 0)
694 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
695 struct wb_writeback_work
*work
)
697 unsigned long start_time
= jiffies
;
700 while (!list_empty(&wb
->b_io
)) {
701 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
702 struct super_block
*sb
= inode
->i_sb
;
704 if (!grab_super_passive(sb
)) {
706 * grab_super_passive() may fail consistently due to
707 * s_umount being grabbed by someone else. Don't use
708 * requeue_io() to avoid busy retrying the inode/sb.
710 redirty_tail(inode
, wb
);
713 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
716 /* refer to the same tests at the end of writeback_sb_inodes */
718 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
720 if (work
->nr_pages
<= 0)
724 /* Leave any unwritten inodes on b_io */
728 static long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
729 enum wb_reason reason
)
731 struct wb_writeback_work work
= {
732 .nr_pages
= nr_pages
,
733 .sync_mode
= WB_SYNC_NONE
,
738 spin_lock(&wb
->list_lock
);
739 if (list_empty(&wb
->b_io
))
741 __writeback_inodes_wb(wb
, &work
);
742 spin_unlock(&wb
->list_lock
);
744 return nr_pages
- work
.nr_pages
;
747 static bool over_bground_thresh(struct backing_dev_info
*bdi
)
749 unsigned long background_thresh
, dirty_thresh
;
751 global_dirty_limits(&background_thresh
, &dirty_thresh
);
753 if (global_page_state(NR_FILE_DIRTY
) +
754 global_page_state(NR_UNSTABLE_NFS
) > background_thresh
)
757 if (bdi_stat(bdi
, BDI_RECLAIMABLE
) >
758 bdi_dirty_limit(bdi
, background_thresh
))
765 * Called under wb->list_lock. If there are multiple wb per bdi,
766 * only the flusher working on the first wb should do it.
768 static void wb_update_bandwidth(struct bdi_writeback
*wb
,
769 unsigned long start_time
)
771 __bdi_update_bandwidth(wb
->bdi
, 0, 0, 0, 0, 0, start_time
);
775 * Explicit flushing or periodic writeback of "old" data.
777 * Define "old": the first time one of an inode's pages is dirtied, we mark the
778 * dirtying-time in the inode's address_space. So this periodic writeback code
779 * just walks the superblock inode list, writing back any inodes which are
780 * older than a specific point in time.
782 * Try to run once per dirty_writeback_interval. But if a writeback event
783 * takes longer than a dirty_writeback_interval interval, then leave a
786 * older_than_this takes precedence over nr_to_write. So we'll only write back
787 * all dirty pages if they are all attached to "old" mappings.
789 static long wb_writeback(struct bdi_writeback
*wb
,
790 struct wb_writeback_work
*work
)
792 unsigned long wb_start
= jiffies
;
793 long nr_pages
= work
->nr_pages
;
794 unsigned long oldest_jif
;
798 oldest_jif
= jiffies
;
799 work
->older_than_this
= &oldest_jif
;
801 spin_lock(&wb
->list_lock
);
804 * Stop writeback when nr_pages has been consumed
806 if (work
->nr_pages
<= 0)
810 * Background writeout and kupdate-style writeback may
811 * run forever. Stop them if there is other work to do
812 * so that e.g. sync can proceed. They'll be restarted
813 * after the other works are all done.
815 if ((work
->for_background
|| work
->for_kupdate
) &&
816 !list_empty(&wb
->bdi
->work_list
))
820 * For background writeout, stop when we are below the
821 * background dirty threshold
823 if (work
->for_background
&& !over_bground_thresh(wb
->bdi
))
827 * Kupdate and background works are special and we want to
828 * include all inodes that need writing. Livelock avoidance is
829 * handled by these works yielding to any other work so we are
832 if (work
->for_kupdate
) {
833 oldest_jif
= jiffies
-
834 msecs_to_jiffies(dirty_expire_interval
* 10);
835 } else if (work
->for_background
)
836 oldest_jif
= jiffies
;
838 trace_writeback_start(wb
->bdi
, work
);
839 if (list_empty(&wb
->b_io
))
842 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
844 progress
= __writeback_inodes_wb(wb
, work
);
845 trace_writeback_written(wb
->bdi
, work
);
847 wb_update_bandwidth(wb
, wb_start
);
850 * Did we write something? Try for more
852 * Dirty inodes are moved to b_io for writeback in batches.
853 * The completion of the current batch does not necessarily
854 * mean the overall work is done. So we keep looping as long
855 * as made some progress on cleaning pages or inodes.
860 * No more inodes for IO, bail
862 if (list_empty(&wb
->b_more_io
))
865 * Nothing written. Wait for some inode to
866 * become available for writeback. Otherwise
867 * we'll just busyloop.
869 if (!list_empty(&wb
->b_more_io
)) {
870 trace_writeback_wait(wb
->bdi
, work
);
871 inode
= wb_inode(wb
->b_more_io
.prev
);
872 spin_lock(&inode
->i_lock
);
873 spin_unlock(&wb
->list_lock
);
874 /* This function drops i_lock... */
875 inode_sleep_on_writeback(inode
);
876 spin_lock(&wb
->list_lock
);
879 spin_unlock(&wb
->list_lock
);
881 return nr_pages
- work
->nr_pages
;
885 * Return the next wb_writeback_work struct that hasn't been processed yet.
887 static struct wb_writeback_work
*
888 get_next_work_item(struct backing_dev_info
*bdi
)
890 struct wb_writeback_work
*work
= NULL
;
892 spin_lock_bh(&bdi
->wb_lock
);
893 if (!list_empty(&bdi
->work_list
)) {
894 work
= list_entry(bdi
->work_list
.next
,
895 struct wb_writeback_work
, list
);
896 list_del_init(&work
->list
);
898 spin_unlock_bh(&bdi
->wb_lock
);
903 * Add in the number of potentially dirty inodes, because each inode
904 * write can dirty pagecache in the underlying blockdev.
906 static unsigned long get_nr_dirty_pages(void)
908 return global_page_state(NR_FILE_DIRTY
) +
909 global_page_state(NR_UNSTABLE_NFS
) +
910 get_nr_dirty_inodes();
913 static long wb_check_background_flush(struct bdi_writeback
*wb
)
915 if (over_bground_thresh(wb
->bdi
)) {
917 struct wb_writeback_work work
= {
918 .nr_pages
= LONG_MAX
,
919 .sync_mode
= WB_SYNC_NONE
,
922 .reason
= WB_REASON_BACKGROUND
,
925 return wb_writeback(wb
, &work
);
931 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
933 unsigned long expired
;
937 * When set to zero, disable periodic writeback
939 if (!dirty_writeback_interval
)
942 expired
= wb
->last_old_flush
+
943 msecs_to_jiffies(dirty_writeback_interval
* 10);
944 if (time_before(jiffies
, expired
))
947 wb
->last_old_flush
= jiffies
;
948 nr_pages
= get_nr_dirty_pages();
951 struct wb_writeback_work work
= {
952 .nr_pages
= nr_pages
,
953 .sync_mode
= WB_SYNC_NONE
,
956 .reason
= WB_REASON_PERIODIC
,
959 return wb_writeback(wb
, &work
);
966 * Retrieve work items and do the writeback they describe
968 static long wb_do_writeback(struct bdi_writeback
*wb
)
970 struct backing_dev_info
*bdi
= wb
->bdi
;
971 struct wb_writeback_work
*work
;
974 set_bit(BDI_writeback_running
, &wb
->bdi
->state
);
975 while ((work
= get_next_work_item(bdi
)) != NULL
) {
977 trace_writeback_exec(bdi
, work
);
979 wrote
+= wb_writeback(wb
, work
);
982 * Notify the caller of completion if this is a synchronous
983 * work item, otherwise just free it.
986 complete(work
->done
);
992 * Check for periodic writeback, kupdated() style
994 wrote
+= wb_check_old_data_flush(wb
);
995 wrote
+= wb_check_background_flush(wb
);
996 clear_bit(BDI_writeback_running
, &wb
->bdi
->state
);
1002 * Handle writeback of dirty data for the device backed by this bdi. Also
1003 * reschedules periodically and does kupdated style flushing.
1005 void bdi_writeback_workfn(struct work_struct
*work
)
1007 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
1008 struct bdi_writeback
, dwork
);
1009 struct backing_dev_info
*bdi
= wb
->bdi
;
1012 set_worker_desc("flush-%s", dev_name(bdi
->dev
));
1013 current
->flags
|= PF_SWAPWRITE
;
1015 if (likely(!current_is_workqueue_rescuer() ||
1016 list_empty(&bdi
->bdi_list
))) {
1018 * The normal path. Keep writing back @bdi until its
1019 * work_list is empty. Note that this path is also taken
1020 * if @bdi is shutting down even when we're running off the
1021 * rescuer as work_list needs to be drained.
1024 pages_written
= wb_do_writeback(wb
);
1025 trace_writeback_pages_written(pages_written
);
1026 } while (!list_empty(&bdi
->work_list
));
1029 * bdi_wq can't get enough workers and we're running off
1030 * the emergency worker. Don't hog it. Hopefully, 1024 is
1031 * enough for efficient IO.
1033 pages_written
= writeback_inodes_wb(&bdi
->wb
, 1024,
1034 WB_REASON_FORKER_THREAD
);
1035 trace_writeback_pages_written(pages_written
);
1038 if (!list_empty(&bdi
->work_list
) ||
1039 (wb_has_dirty_io(wb
) && dirty_writeback_interval
))
1040 queue_delayed_work(bdi_wq
, &wb
->dwork
,
1041 msecs_to_jiffies(dirty_writeback_interval
* 10));
1043 current
->flags
&= ~PF_SWAPWRITE
;
1047 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1050 void wakeup_flusher_threads(long nr_pages
, enum wb_reason reason
)
1052 struct backing_dev_info
*bdi
;
1055 nr_pages
= get_nr_dirty_pages();
1058 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
1059 if (!bdi_has_dirty_io(bdi
))
1061 __bdi_start_writeback(bdi
, nr_pages
, false, reason
);
1066 static noinline
void block_dump___mark_inode_dirty(struct inode
*inode
)
1068 if (inode
->i_ino
|| strcmp(inode
->i_sb
->s_id
, "bdev")) {
1069 struct dentry
*dentry
;
1070 const char *name
= "?";
1072 dentry
= d_find_alias(inode
);
1074 spin_lock(&dentry
->d_lock
);
1075 name
= (const char *) dentry
->d_name
.name
;
1078 "%s(%d): dirtied inode %lu (%s) on %s\n",
1079 current
->comm
, task_pid_nr(current
), inode
->i_ino
,
1080 name
, inode
->i_sb
->s_id
);
1082 spin_unlock(&dentry
->d_lock
);
1089 * __mark_inode_dirty - internal function
1090 * @inode: inode to mark
1091 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1092 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1093 * mark_inode_dirty_sync.
1095 * Put the inode on the super block's dirty list.
1097 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1098 * dirty list only if it is hashed or if it refers to a blockdev.
1099 * If it was not hashed, it will never be added to the dirty list
1100 * even if it is later hashed, as it will have been marked dirty already.
1102 * In short, make sure you hash any inodes _before_ you start marking
1105 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1106 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1107 * the kernel-internal blockdev inode represents the dirtying time of the
1108 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1109 * page->mapping->host, so the page-dirtying time is recorded in the internal
1112 void __mark_inode_dirty(struct inode
*inode
, int flags
)
1114 struct super_block
*sb
= inode
->i_sb
;
1115 struct backing_dev_info
*bdi
= NULL
;
1118 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1119 * dirty the inode itself
1121 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
1122 trace_writeback_dirty_inode_start(inode
, flags
);
1124 if (sb
->s_op
->dirty_inode
)
1125 sb
->s_op
->dirty_inode(inode
, flags
);
1127 trace_writeback_dirty_inode(inode
, flags
);
1131 * make sure that changes are seen by all cpus before we test i_state
1136 /* avoid the locking if we can */
1137 if ((inode
->i_state
& flags
) == flags
)
1140 if (unlikely(block_dump
))
1141 block_dump___mark_inode_dirty(inode
);
1143 spin_lock(&inode
->i_lock
);
1144 if ((inode
->i_state
& flags
) != flags
) {
1145 const int was_dirty
= inode
->i_state
& I_DIRTY
;
1147 inode
->i_state
|= flags
;
1150 * If the inode is being synced, just update its dirty state.
1151 * The unlocker will place the inode on the appropriate
1152 * superblock list, based upon its state.
1154 if (inode
->i_state
& I_SYNC
)
1155 goto out_unlock_inode
;
1158 * Only add valid (hashed) inodes to the superblock's
1159 * dirty list. Add blockdev inodes as well.
1161 if (!S_ISBLK(inode
->i_mode
)) {
1162 if (inode_unhashed(inode
))
1163 goto out_unlock_inode
;
1165 if (inode
->i_state
& I_FREEING
)
1166 goto out_unlock_inode
;
1169 * If the inode was already on b_dirty/b_io/b_more_io, don't
1170 * reposition it (that would break b_dirty time-ordering).
1173 bool wakeup_bdi
= false;
1174 bdi
= inode_to_bdi(inode
);
1176 spin_unlock(&inode
->i_lock
);
1177 spin_lock(&bdi
->wb
.list_lock
);
1178 if (bdi_cap_writeback_dirty(bdi
)) {
1179 WARN(!test_bit(BDI_registered
, &bdi
->state
),
1180 "bdi-%s not registered\n", bdi
->name
);
1183 * If this is the first dirty inode for this
1184 * bdi, we have to wake-up the corresponding
1185 * bdi thread to make sure background
1186 * write-back happens later.
1188 if (!wb_has_dirty_io(&bdi
->wb
))
1192 inode
->dirtied_when
= jiffies
;
1193 list_move(&inode
->i_wb_list
, &bdi
->wb
.b_dirty
);
1194 spin_unlock(&bdi
->wb
.list_lock
);
1197 bdi_wakeup_thread_delayed(bdi
);
1202 spin_unlock(&inode
->i_lock
);
1205 EXPORT_SYMBOL(__mark_inode_dirty
);
1207 static void wait_sb_inodes(struct super_block
*sb
)
1209 struct inode
*inode
, *old_inode
= NULL
;
1212 * We need to be protected against the filesystem going from
1213 * r/o to r/w or vice versa.
1215 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1217 spin_lock(&inode_sb_list_lock
);
1220 * Data integrity sync. Must wait for all pages under writeback,
1221 * because there may have been pages dirtied before our sync
1222 * call, but which had writeout started before we write it out.
1223 * In which case, the inode may not be on the dirty list, but
1224 * we still have to wait for that writeout.
1226 list_for_each_entry(inode
, &sb
->s_inodes
, i_sb_list
) {
1227 struct address_space
*mapping
= inode
->i_mapping
;
1229 spin_lock(&inode
->i_lock
);
1230 if ((inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) ||
1231 (mapping
->nrpages
== 0)) {
1232 spin_unlock(&inode
->i_lock
);
1236 spin_unlock(&inode
->i_lock
);
1237 spin_unlock(&inode_sb_list_lock
);
1240 * We hold a reference to 'inode' so it couldn't have been
1241 * removed from s_inodes list while we dropped the
1242 * inode_sb_list_lock. We cannot iput the inode now as we can
1243 * be holding the last reference and we cannot iput it under
1244 * inode_sb_list_lock. So we keep the reference and iput it
1250 filemap_fdatawait(mapping
);
1254 spin_lock(&inode_sb_list_lock
);
1256 spin_unlock(&inode_sb_list_lock
);
1261 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1262 * @sb: the superblock
1263 * @nr: the number of pages to write
1264 * @reason: reason why some writeback work initiated
1266 * Start writeback on some inodes on this super_block. No guarantees are made
1267 * on how many (if any) will be written, and this function does not wait
1268 * for IO completion of submitted IO.
1270 void writeback_inodes_sb_nr(struct super_block
*sb
,
1272 enum wb_reason reason
)
1274 DECLARE_COMPLETION_ONSTACK(done
);
1275 struct wb_writeback_work work
= {
1277 .sync_mode
= WB_SYNC_NONE
,
1278 .tagged_writepages
= 1,
1284 if (sb
->s_bdi
== &noop_backing_dev_info
)
1286 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1287 bdi_queue_work(sb
->s_bdi
, &work
);
1288 wait_for_completion(&done
);
1290 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
1293 * writeback_inodes_sb - writeback dirty inodes from given super_block
1294 * @sb: the superblock
1295 * @reason: reason why some writeback work was initiated
1297 * Start writeback on some inodes on this super_block. No guarantees are made
1298 * on how many (if any) will be written, and this function does not wait
1299 * for IO completion of submitted IO.
1301 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1303 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1305 EXPORT_SYMBOL(writeback_inodes_sb
);
1308 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1309 * @sb: the superblock
1310 * @nr: the number of pages to write
1311 * @reason: the reason of writeback
1313 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1314 * Returns 1 if writeback was started, 0 if not.
1316 int try_to_writeback_inodes_sb_nr(struct super_block
*sb
,
1318 enum wb_reason reason
)
1320 if (writeback_in_progress(sb
->s_bdi
))
1323 if (!down_read_trylock(&sb
->s_umount
))
1326 writeback_inodes_sb_nr(sb
, nr
, reason
);
1327 up_read(&sb
->s_umount
);
1330 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr
);
1333 * try_to_writeback_inodes_sb - try to start writeback if none underway
1334 * @sb: the superblock
1335 * @reason: reason why some writeback work was initiated
1337 * Implement by try_to_writeback_inodes_sb_nr()
1338 * Returns 1 if writeback was started, 0 if not.
1340 int try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
1342 return try_to_writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
1344 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
1347 * sync_inodes_sb - sync sb inode pages
1348 * @sb: the superblock
1350 * This function writes and waits on any dirty inode belonging to this
1353 void sync_inodes_sb(struct super_block
*sb
)
1355 DECLARE_COMPLETION_ONSTACK(done
);
1356 struct wb_writeback_work work
= {
1358 .sync_mode
= WB_SYNC_ALL
,
1359 .nr_pages
= LONG_MAX
,
1362 .reason
= WB_REASON_SYNC
,
1366 /* Nothing to do? */
1367 if (sb
->s_bdi
== &noop_backing_dev_info
)
1369 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1371 bdi_queue_work(sb
->s_bdi
, &work
);
1372 wait_for_completion(&done
);
1376 EXPORT_SYMBOL(sync_inodes_sb
);
1379 * write_inode_now - write an inode to disk
1380 * @inode: inode to write to disk
1381 * @sync: whether the write should be synchronous or not
1383 * This function commits an inode to disk immediately if it is dirty. This is
1384 * primarily needed by knfsd.
1386 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1388 int write_inode_now(struct inode
*inode
, int sync
)
1390 struct bdi_writeback
*wb
= &inode_to_bdi(inode
)->wb
;
1391 struct writeback_control wbc
= {
1392 .nr_to_write
= LONG_MAX
,
1393 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1395 .range_end
= LLONG_MAX
,
1398 if (!mapping_cap_writeback_dirty(inode
->i_mapping
))
1399 wbc
.nr_to_write
= 0;
1402 return writeback_single_inode(inode
, wb
, &wbc
);
1404 EXPORT_SYMBOL(write_inode_now
);
1407 * sync_inode - write an inode and its pages to disk.
1408 * @inode: the inode to sync
1409 * @wbc: controls the writeback mode
1411 * sync_inode() will write an inode and its pages to disk. It will also
1412 * correctly update the inode on its superblock's dirty inode lists and will
1413 * update inode->i_state.
1415 * The caller must have a ref on the inode.
1417 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1419 return writeback_single_inode(inode
, &inode_to_bdi(inode
)->wb
, wbc
);
1421 EXPORT_SYMBOL(sync_inode
);
1424 * sync_inode_metadata - write an inode to disk
1425 * @inode: the inode to sync
1426 * @wait: wait for I/O to complete.
1428 * Write an inode to disk and adjust its dirty state after completion.
1430 * Note: only writes the actual inode, no associated data or other metadata.
1432 int sync_inode_metadata(struct inode
*inode
, int wait
)
1434 struct writeback_control wbc
= {
1435 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
1436 .nr_to_write
= 0, /* metadata-only */
1439 return sync_inode(inode
, &wbc
);
1441 EXPORT_SYMBOL(sync_inode_metadata
);