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 (sb_is_blkdev_sb(sb))
72 return inode->i_mapping->backing_dev_info;
77 static inline struct inode *wb_inode(struct list_head *head)
79 return list_entry(head, struct inode, i_wb_list);
83 * Include the creation of the trace points after defining the
84 * wb_writeback_work structure and inline functions so that the definition
85 * remains local to this file.
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/writeback.h>
90 static void bdi_queue_work(struct backing_dev_info *bdi,
91 struct wb_writeback_work *work)
93 trace_writeback_queue(bdi, work);
95 spin_lock_bh(&bdi->wb_lock);
96 list_add_tail(&work->list, &bdi->work_list);
97 spin_unlock_bh(&bdi->wb_lock);
99 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
103 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
104 bool range_cyclic, enum wb_reason reason)
106 struct wb_writeback_work *work;
109 * This is WB_SYNC_NONE writeback, so if allocation fails just
110 * wakeup the thread for old dirty data writeback
112 work = kzalloc(sizeof(*work), GFP_ATOMIC);
114 trace_writeback_nowork(bdi);
115 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
119 work->sync_mode = WB_SYNC_NONE;
120 work->nr_pages = nr_pages;
121 work->range_cyclic = range_cyclic;
122 work->reason = reason;
124 bdi_queue_work(bdi, work);
128 * bdi_start_writeback - start writeback
129 * @bdi: the backing device to write from
130 * @nr_pages: the number of pages to write
131 * @reason: reason why some writeback work was initiated
134 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
135 * started when this function returns, we make no guarantees on
136 * completion. Caller need not hold sb s_umount semaphore.
139 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
140 enum wb_reason reason)
142 __bdi_start_writeback(bdi, nr_pages, true, reason);
146 * bdi_start_background_writeback - start background writeback
147 * @bdi: the backing device to write from
150 * This makes sure WB_SYNC_NONE background writeback happens. When
151 * this function returns, it is only guaranteed that for given BDI
152 * some IO is happening if we are over background dirty threshold.
153 * Caller need not hold sb s_umount semaphore.
155 void bdi_start_background_writeback(struct backing_dev_info *bdi)
158 * We just wake up the flusher thread. It will perform background
159 * writeback as soon as there is no other work to do.
161 trace_writeback_wake_background(bdi);
162 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
166 * Remove the inode from the writeback list it is on.
168 void inode_wb_list_del(struct inode *inode)
170 struct backing_dev_info *bdi = inode_to_bdi(inode);
172 spin_lock(&bdi->wb.list_lock);
173 list_del_init(&inode->i_wb_list);
174 spin_unlock(&bdi->wb.list_lock);
178 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
179 * furthest end of its superblock's dirty-inode list.
181 * Before stamping the inode's ->dirtied_when, we check to see whether it is
182 * already the most-recently-dirtied inode on the b_dirty list. If that is
183 * the case then the inode must have been redirtied while it was being written
184 * out and we don't reset its dirtied_when.
186 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
188 assert_spin_locked(&wb->list_lock);
189 if (!list_empty(&wb->b_dirty)) {
192 tail = wb_inode(wb->b_dirty.next);
193 if (time_before(inode->dirtied_when, tail->dirtied_when))
194 inode->dirtied_when = jiffies;
196 list_move(&inode->i_wb_list, &wb->b_dirty);
200 * requeue inode for re-scanning after bdi->b_io list is exhausted.
202 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
204 assert_spin_locked(&wb->list_lock);
205 list_move(&inode->i_wb_list, &wb->b_more_io);
208 static void inode_sync_complete(struct inode *inode)
210 inode->i_state &= ~I_SYNC;
211 /* If inode is clean an unused, put it into LRU now... */
212 inode_add_lru(inode);
213 /* Waiters must see I_SYNC cleared before being woken up */
215 wake_up_bit(&inode->i_state, __I_SYNC);
218 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
220 bool ret = time_after(inode->dirtied_when, t);
223 * For inodes being constantly redirtied, dirtied_when can get stuck.
224 * It _appears_ to be in the future, but is actually in distant past.
225 * This test is necessary to prevent such wrapped-around relative times
226 * from permanently stopping the whole bdi writeback.
228 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
234 * Move expired (dirtied before work->older_than_this) dirty inodes from
235 * @delaying_queue to @dispatch_queue.
237 static int move_expired_inodes(struct list_head *delaying_queue,
238 struct list_head *dispatch_queue,
239 struct wb_writeback_work *work)
242 struct list_head *pos, *node;
243 struct super_block *sb = NULL;
248 while (!list_empty(delaying_queue)) {
249 inode = wb_inode(delaying_queue->prev);
250 if (work->older_than_this &&
251 inode_dirtied_after(inode, *work->older_than_this))
253 list_move(&inode->i_wb_list, &tmp);
255 if (sb_is_blkdev_sb(inode->i_sb))
257 if (sb && sb != inode->i_sb)
262 /* just one sb in list, splice to dispatch_queue and we're done */
264 list_splice(&tmp, dispatch_queue);
268 /* Move inodes from one superblock together */
269 while (!list_empty(&tmp)) {
270 sb = wb_inode(tmp.prev)->i_sb;
271 list_for_each_prev_safe(pos, node, &tmp) {
272 inode = wb_inode(pos);
273 if (inode->i_sb == sb)
274 list_move(&inode->i_wb_list, dispatch_queue);
282 * Queue all expired dirty inodes for io, eldest first.
284 * newly dirtied b_dirty b_io b_more_io
285 * =============> gf edc BA
287 * newly dirtied b_dirty b_io b_more_io
288 * =============> g fBAedc
290 * +--> dequeue for IO
292 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
295 assert_spin_locked(&wb->list_lock);
296 list_splice_init(&wb->b_more_io, &wb->b_io);
297 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
298 trace_writeback_queue_io(wb, work, moved);
301 static int write_inode(struct inode *inode, struct writeback_control *wbc)
305 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
306 trace_writeback_write_inode_start(inode, wbc);
307 ret = inode->i_sb->s_op->write_inode(inode, wbc);
308 trace_writeback_write_inode(inode, wbc);
315 * Wait for writeback on an inode to complete. Called with i_lock held.
316 * Caller must make sure inode cannot go away when we drop i_lock.
318 static void __inode_wait_for_writeback(struct inode *inode)
319 __releases(inode->i_lock)
320 __acquires(inode->i_lock)
322 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
323 wait_queue_head_t *wqh;
325 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
326 while (inode->i_state & I_SYNC) {
327 spin_unlock(&inode->i_lock);
328 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
329 spin_lock(&inode->i_lock);
334 * Wait for writeback on an inode to complete. Caller must have inode pinned.
336 void inode_wait_for_writeback(struct inode *inode)
338 spin_lock(&inode->i_lock);
339 __inode_wait_for_writeback(inode);
340 spin_unlock(&inode->i_lock);
344 * Sleep until I_SYNC is cleared. This function must be called with i_lock
345 * held and drops it. It is aimed for callers not holding any inode reference
346 * so once i_lock is dropped, inode can go away.
348 static void inode_sleep_on_writeback(struct inode *inode)
349 __releases(inode->i_lock)
352 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
355 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
356 sleep = inode->i_state & I_SYNC;
357 spin_unlock(&inode->i_lock);
360 finish_wait(wqh, &wait);
364 * Find proper writeback list for the inode depending on its current state and
365 * possibly also change of its state while we were doing writeback. Here we
366 * handle things such as livelock prevention or fairness of writeback among
367 * inodes. This function can be called only by flusher thread - noone else
368 * processes all inodes in writeback lists and requeueing inodes behind flusher
369 * thread's back can have unexpected consequences.
371 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
372 struct writeback_control *wbc)
374 if (inode->i_state & I_FREEING)
378 * Sync livelock prevention. Each inode is tagged and synced in one
379 * shot. If still dirty, it will be redirty_tail()'ed below. Update
380 * the dirty time to prevent enqueue and sync it again.
382 if ((inode->i_state & I_DIRTY) &&
383 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
384 inode->dirtied_when = jiffies;
386 if (wbc->pages_skipped) {
388 * writeback is not making progress due to locked
389 * buffers. Skip this inode for now.
391 redirty_tail(inode, wb);
395 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
397 * We didn't write back all the pages. nfs_writepages()
398 * sometimes bales out without doing anything.
400 if (wbc->nr_to_write <= 0) {
401 /* Slice used up. Queue for next turn. */
402 requeue_io(inode, wb);
405 * Writeback blocked by something other than
406 * congestion. Delay the inode for some time to
407 * avoid spinning on the CPU (100% iowait)
408 * retrying writeback of the dirty page/inode
409 * that cannot be performed immediately.
411 redirty_tail(inode, wb);
413 } else if (inode->i_state & I_DIRTY) {
415 * Filesystems can dirty the inode during writeback operations,
416 * such as delayed allocation during submission or metadata
417 * updates after data IO completion.
419 redirty_tail(inode, wb);
421 /* The inode is clean. Remove from writeback lists. */
422 list_del_init(&inode->i_wb_list);
427 * Write out an inode and its dirty pages. Do not update the writeback list
428 * linkage. That is left to the caller. The caller is also responsible for
429 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
432 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
434 struct address_space *mapping = inode->i_mapping;
435 long nr_to_write = wbc->nr_to_write;
439 WARN_ON(!(inode->i_state & I_SYNC));
441 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
443 ret = do_writepages(mapping, wbc);
446 * Make sure to wait on the data before writing out the metadata.
447 * This is important for filesystems that modify metadata on data
450 if (wbc->sync_mode == WB_SYNC_ALL) {
451 int err = filemap_fdatawait(mapping);
457 * Some filesystems may redirty the inode during the writeback
458 * due to delalloc, clear dirty metadata flags right before
461 spin_lock(&inode->i_lock);
462 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
463 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
464 inode->i_state &= ~I_DIRTY_PAGES;
465 dirty = inode->i_state & I_DIRTY;
466 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
467 spin_unlock(&inode->i_lock);
468 /* Don't write the inode if only I_DIRTY_PAGES was set */
469 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
470 int err = write_inode(inode, wbc);
474 trace_writeback_single_inode(inode, wbc, nr_to_write);
479 * Write out an inode's dirty pages. Either the caller has an active reference
480 * on the inode or the inode has I_WILL_FREE set.
482 * This function is designed to be called for writing back one inode which
483 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
484 * and does more profound writeback list handling in writeback_sb_inodes().
487 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
488 struct writeback_control *wbc)
492 spin_lock(&inode->i_lock);
493 if (!atomic_read(&inode->i_count))
494 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
496 WARN_ON(inode->i_state & I_WILL_FREE);
498 if (inode->i_state & I_SYNC) {
499 if (wbc->sync_mode != WB_SYNC_ALL)
502 * It's a data-integrity sync. We must wait. Since callers hold
503 * inode reference or inode has I_WILL_FREE set, it cannot go
506 __inode_wait_for_writeback(inode);
508 WARN_ON(inode->i_state & I_SYNC);
510 * Skip inode if it is clean. We don't want to mess with writeback
511 * lists in this function since flusher thread may be doing for example
512 * sync in parallel and if we move the inode, it could get skipped. So
513 * here we make sure inode is on some writeback list and leave it there
514 * unless we have completely cleaned the inode.
516 if (!(inode->i_state & I_DIRTY))
518 inode->i_state |= I_SYNC;
519 spin_unlock(&inode->i_lock);
521 ret = __writeback_single_inode(inode, wbc);
523 spin_lock(&wb->list_lock);
524 spin_lock(&inode->i_lock);
526 * If inode is clean, remove it from writeback lists. Otherwise don't
527 * touch it. See comment above for explanation.
529 if (!(inode->i_state & I_DIRTY))
530 list_del_init(&inode->i_wb_list);
531 spin_unlock(&wb->list_lock);
532 inode_sync_complete(inode);
534 spin_unlock(&inode->i_lock);
538 static long writeback_chunk_size(struct backing_dev_info *bdi,
539 struct wb_writeback_work *work)
544 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
545 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
546 * here avoids calling into writeback_inodes_wb() more than once.
548 * The intended call sequence for WB_SYNC_ALL writeback is:
551 * writeback_sb_inodes() <== called only once
552 * write_cache_pages() <== called once for each inode
553 * (quickly) tag currently dirty pages
554 * (maybe slowly) sync all tagged pages
556 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
559 pages = min(bdi->avg_write_bandwidth / 2,
560 global_dirty_limit / DIRTY_SCOPE);
561 pages = min(pages, work->nr_pages);
562 pages = round_down(pages + MIN_WRITEBACK_PAGES,
563 MIN_WRITEBACK_PAGES);
570 * Write a portion of b_io inodes which belong to @sb.
572 * Return the number of pages and/or inodes written.
574 static long writeback_sb_inodes(struct super_block *sb,
575 struct bdi_writeback *wb,
576 struct wb_writeback_work *work)
578 struct writeback_control wbc = {
579 .sync_mode = work->sync_mode,
580 .tagged_writepages = work->tagged_writepages,
581 .for_kupdate = work->for_kupdate,
582 .for_background = work->for_background,
583 .range_cyclic = work->range_cyclic,
585 .range_end = LLONG_MAX,
587 unsigned long start_time = jiffies;
589 long wrote = 0; /* count both pages and inodes */
591 while (!list_empty(&wb->b_io)) {
592 struct inode *inode = wb_inode(wb->b_io.prev);
594 if (inode->i_sb != sb) {
597 * We only want to write back data for this
598 * superblock, move all inodes not belonging
599 * to it back onto the dirty list.
601 redirty_tail(inode, wb);
606 * The inode belongs to a different superblock.
607 * Bounce back to the caller to unpin this and
608 * pin the next superblock.
614 * Don't bother with new inodes or inodes being freed, first
615 * kind does not need periodic writeout yet, and for the latter
616 * kind writeout is handled by the freer.
618 spin_lock(&inode->i_lock);
619 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
620 spin_unlock(&inode->i_lock);
621 redirty_tail(inode, wb);
624 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
626 * If this inode is locked for writeback and we are not
627 * doing writeback-for-data-integrity, move it to
628 * b_more_io so that writeback can proceed with the
629 * other inodes on s_io.
631 * We'll have another go at writing back this inode
632 * when we completed a full scan of b_io.
634 spin_unlock(&inode->i_lock);
635 requeue_io(inode, wb);
636 trace_writeback_sb_inodes_requeue(inode);
639 spin_unlock(&wb->list_lock);
642 * We already requeued the inode if it had I_SYNC set and we
643 * are doing WB_SYNC_NONE writeback. So this catches only the
646 if (inode->i_state & I_SYNC) {
647 /* Wait for I_SYNC. This function drops i_lock... */
648 inode_sleep_on_writeback(inode);
649 /* Inode may be gone, start again */
650 spin_lock(&wb->list_lock);
653 inode->i_state |= I_SYNC;
654 spin_unlock(&inode->i_lock);
656 write_chunk = writeback_chunk_size(wb->bdi, work);
657 wbc.nr_to_write = write_chunk;
658 wbc.pages_skipped = 0;
661 * We use I_SYNC to pin the inode in memory. While it is set
662 * evict_inode() will wait so the inode cannot be freed.
664 __writeback_single_inode(inode, &wbc);
666 work->nr_pages -= write_chunk - wbc.nr_to_write;
667 wrote += write_chunk - wbc.nr_to_write;
668 spin_lock(&wb->list_lock);
669 spin_lock(&inode->i_lock);
670 if (!(inode->i_state & I_DIRTY))
672 requeue_inode(inode, wb, &wbc);
673 inode_sync_complete(inode);
674 spin_unlock(&inode->i_lock);
675 cond_resched_lock(&wb->list_lock);
677 * bail out to wb_writeback() often enough to check
678 * background threshold and other termination conditions.
681 if (time_is_before_jiffies(start_time + HZ / 10UL))
683 if (work->nr_pages <= 0)
690 static long __writeback_inodes_wb(struct bdi_writeback *wb,
691 struct wb_writeback_work *work)
693 unsigned long start_time = jiffies;
696 while (!list_empty(&wb->b_io)) {
697 struct inode *inode = wb_inode(wb->b_io.prev);
698 struct super_block *sb = inode->i_sb;
700 if (!grab_super_passive(sb)) {
702 * grab_super_passive() may fail consistently due to
703 * s_umount being grabbed by someone else. Don't use
704 * requeue_io() to avoid busy retrying the inode/sb.
706 redirty_tail(inode, wb);
709 wrote += writeback_sb_inodes(sb, wb, work);
712 /* refer to the same tests at the end of writeback_sb_inodes */
714 if (time_is_before_jiffies(start_time + HZ / 10UL))
716 if (work->nr_pages <= 0)
720 /* Leave any unwritten inodes on b_io */
724 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
725 enum wb_reason reason)
727 struct wb_writeback_work work = {
728 .nr_pages = nr_pages,
729 .sync_mode = WB_SYNC_NONE,
734 spin_lock(&wb->list_lock);
735 if (list_empty(&wb->b_io))
737 __writeback_inodes_wb(wb, &work);
738 spin_unlock(&wb->list_lock);
740 return nr_pages - work.nr_pages;
743 static bool over_bground_thresh(struct backing_dev_info *bdi)
745 unsigned long background_thresh, dirty_thresh;
747 global_dirty_limits(&background_thresh, &dirty_thresh);
749 if (global_page_state(NR_FILE_DIRTY) +
750 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
753 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
754 bdi_dirty_limit(bdi, background_thresh))
761 * Called under wb->list_lock. If there are multiple wb per bdi,
762 * only the flusher working on the first wb should do it.
764 static void wb_update_bandwidth(struct bdi_writeback *wb,
765 unsigned long start_time)
767 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
771 * Explicit flushing or periodic writeback of "old" data.
773 * Define "old": the first time one of an inode's pages is dirtied, we mark the
774 * dirtying-time in the inode's address_space. So this periodic writeback code
775 * just walks the superblock inode list, writing back any inodes which are
776 * older than a specific point in time.
778 * Try to run once per dirty_writeback_interval. But if a writeback event
779 * takes longer than a dirty_writeback_interval interval, then leave a
782 * older_than_this takes precedence over nr_to_write. So we'll only write back
783 * all dirty pages if they are all attached to "old" mappings.
785 static long wb_writeback(struct bdi_writeback *wb,
786 struct wb_writeback_work *work)
788 unsigned long wb_start = jiffies;
789 long nr_pages = work->nr_pages;
790 unsigned long oldest_jif;
794 oldest_jif = jiffies;
795 work->older_than_this = &oldest_jif;
797 spin_lock(&wb->list_lock);
800 * Stop writeback when nr_pages has been consumed
802 if (work->nr_pages <= 0)
806 * Background writeout and kupdate-style writeback may
807 * run forever. Stop them if there is other work to do
808 * so that e.g. sync can proceed. They'll be restarted
809 * after the other works are all done.
811 if ((work->for_background || work->for_kupdate) &&
812 !list_empty(&wb->bdi->work_list))
816 * For background writeout, stop when we are below the
817 * background dirty threshold
819 if (work->for_background && !over_bground_thresh(wb->bdi))
823 * Kupdate and background works are special and we want to
824 * include all inodes that need writing. Livelock avoidance is
825 * handled by these works yielding to any other work so we are
828 if (work->for_kupdate) {
829 oldest_jif = jiffies -
830 msecs_to_jiffies(dirty_expire_interval * 10);
831 } else if (work->for_background)
832 oldest_jif = jiffies;
834 trace_writeback_start(wb->bdi, work);
835 if (list_empty(&wb->b_io))
838 progress = writeback_sb_inodes(work->sb, wb, work);
840 progress = __writeback_inodes_wb(wb, work);
841 trace_writeback_written(wb->bdi, work);
843 wb_update_bandwidth(wb, wb_start);
846 * Did we write something? Try for more
848 * Dirty inodes are moved to b_io for writeback in batches.
849 * The completion of the current batch does not necessarily
850 * mean the overall work is done. So we keep looping as long
851 * as made some progress on cleaning pages or inodes.
856 * No more inodes for IO, bail
858 if (list_empty(&wb->b_more_io))
861 * Nothing written. Wait for some inode to
862 * become available for writeback. Otherwise
863 * we'll just busyloop.
865 if (!list_empty(&wb->b_more_io)) {
866 trace_writeback_wait(wb->bdi, work);
867 inode = wb_inode(wb->b_more_io.prev);
868 spin_lock(&inode->i_lock);
869 spin_unlock(&wb->list_lock);
870 /* This function drops i_lock... */
871 inode_sleep_on_writeback(inode);
872 spin_lock(&wb->list_lock);
875 spin_unlock(&wb->list_lock);
877 return nr_pages - work->nr_pages;
881 * Return the next wb_writeback_work struct that hasn't been processed yet.
883 static struct wb_writeback_work *
884 get_next_work_item(struct backing_dev_info *bdi)
886 struct wb_writeback_work *work = NULL;
888 spin_lock_bh(&bdi->wb_lock);
889 if (!list_empty(&bdi->work_list)) {
890 work = list_entry(bdi->work_list.next,
891 struct wb_writeback_work, list);
892 list_del_init(&work->list);
894 spin_unlock_bh(&bdi->wb_lock);
899 * Add in the number of potentially dirty inodes, because each inode
900 * write can dirty pagecache in the underlying blockdev.
902 static unsigned long get_nr_dirty_pages(void)
904 return global_page_state(NR_FILE_DIRTY) +
905 global_page_state(NR_UNSTABLE_NFS) +
906 get_nr_dirty_inodes();
909 static long wb_check_background_flush(struct bdi_writeback *wb)
911 if (over_bground_thresh(wb->bdi)) {
913 struct wb_writeback_work work = {
914 .nr_pages = LONG_MAX,
915 .sync_mode = WB_SYNC_NONE,
918 .reason = WB_REASON_BACKGROUND,
921 return wb_writeback(wb, &work);
927 static long wb_check_old_data_flush(struct bdi_writeback *wb)
929 unsigned long expired;
933 * When set to zero, disable periodic writeback
935 if (!dirty_writeback_interval)
938 expired = wb->last_old_flush +
939 msecs_to_jiffies(dirty_writeback_interval * 10);
940 if (time_before(jiffies, expired))
943 wb->last_old_flush = jiffies;
944 nr_pages = get_nr_dirty_pages();
947 struct wb_writeback_work work = {
948 .nr_pages = nr_pages,
949 .sync_mode = WB_SYNC_NONE,
952 .reason = WB_REASON_PERIODIC,
955 return wb_writeback(wb, &work);
962 * Retrieve work items and do the writeback they describe
964 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
966 struct backing_dev_info *bdi = wb->bdi;
967 struct wb_writeback_work *work;
970 set_bit(BDI_writeback_running, &wb->bdi->state);
971 while ((work = get_next_work_item(bdi)) != NULL) {
973 * Override sync mode, in case we must wait for completion
974 * because this thread is exiting now.
977 work->sync_mode = WB_SYNC_ALL;
979 trace_writeback_exec(bdi, work);
981 wrote += wb_writeback(wb, work);
984 * Notify the caller of completion if this is a synchronous
985 * work item, otherwise just free it.
988 complete(work->done);
994 * Check for periodic writeback, kupdated() style
996 wrote += wb_check_old_data_flush(wb);
997 wrote += wb_check_background_flush(wb);
998 clear_bit(BDI_writeback_running, &wb->bdi->state);
1004 * Handle writeback of dirty data for the device backed by this bdi. Also
1005 * reschedules periodically and does kupdated style flushing.
1007 void bdi_writeback_workfn(struct work_struct *work)
1009 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1010 struct bdi_writeback, dwork);
1011 struct backing_dev_info *bdi = wb->bdi;
1014 set_worker_desc("flush-%s", dev_name(bdi->dev));
1015 current->flags |= PF_SWAPWRITE;
1017 if (likely(!current_is_workqueue_rescuer() ||
1018 list_empty(&bdi->bdi_list))) {
1020 * The normal path. Keep writing back @bdi until its
1021 * work_list is empty. Note that this path is also taken
1022 * if @bdi is shutting down even when we're running off the
1023 * rescuer as work_list needs to be drained.
1026 pages_written = wb_do_writeback(wb, 0);
1027 trace_writeback_pages_written(pages_written);
1028 } while (!list_empty(&bdi->work_list));
1031 * bdi_wq can't get enough workers and we're running off
1032 * the emergency worker. Don't hog it. Hopefully, 1024 is
1033 * enough for efficient IO.
1035 pages_written = writeback_inodes_wb(&bdi->wb, 1024,
1036 WB_REASON_FORKER_THREAD);
1037 trace_writeback_pages_written(pages_written);
1040 if (!list_empty(&bdi->work_list) ||
1041 (wb_has_dirty_io(wb) && dirty_writeback_interval))
1042 queue_delayed_work(bdi_wq, &wb->dwork,
1043 msecs_to_jiffies(dirty_writeback_interval * 10));
1045 current->flags &= ~PF_SWAPWRITE;
1049 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1052 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1054 struct backing_dev_info *bdi;
1057 nr_pages = global_page_state(NR_FILE_DIRTY) +
1058 global_page_state(NR_UNSTABLE_NFS);
1062 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1063 if (!bdi_has_dirty_io(bdi))
1065 __bdi_start_writeback(bdi, nr_pages, false, reason);
1070 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1072 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1073 struct dentry *dentry;
1074 const char *name = "?";
1076 dentry = d_find_alias(inode);
1078 spin_lock(&dentry->d_lock);
1079 name = (const char *) dentry->d_name.name;
1082 "%s(%d): dirtied inode %lu (%s) on %s\n",
1083 current->comm, task_pid_nr(current), inode->i_ino,
1084 name, inode->i_sb->s_id);
1086 spin_unlock(&dentry->d_lock);
1093 * __mark_inode_dirty - internal function
1094 * @inode: inode to mark
1095 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1096 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1097 * mark_inode_dirty_sync.
1099 * Put the inode on the super block's dirty list.
1101 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1102 * dirty list only if it is hashed or if it refers to a blockdev.
1103 * If it was not hashed, it will never be added to the dirty list
1104 * even if it is later hashed, as it will have been marked dirty already.
1106 * In short, make sure you hash any inodes _before_ you start marking
1109 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1110 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1111 * the kernel-internal blockdev inode represents the dirtying time of the
1112 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1113 * page->mapping->host, so the page-dirtying time is recorded in the internal
1116 void __mark_inode_dirty(struct inode *inode, int flags)
1118 struct super_block *sb = inode->i_sb;
1119 struct backing_dev_info *bdi = NULL;
1122 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1123 * dirty the inode itself
1125 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1126 trace_writeback_dirty_inode_start(inode, flags);
1128 if (sb->s_op->dirty_inode)
1129 sb->s_op->dirty_inode(inode, flags);
1131 trace_writeback_dirty_inode(inode, flags);
1135 * make sure that changes are seen by all cpus before we test i_state
1140 /* avoid the locking if we can */
1141 if ((inode->i_state & flags) == flags)
1144 if (unlikely(block_dump))
1145 block_dump___mark_inode_dirty(inode);
1147 spin_lock(&inode->i_lock);
1148 if ((inode->i_state & flags) != flags) {
1149 const int was_dirty = inode->i_state & I_DIRTY;
1151 inode->i_state |= flags;
1154 * If the inode is being synced, just update its dirty state.
1155 * The unlocker will place the inode on the appropriate
1156 * superblock list, based upon its state.
1158 if (inode->i_state & I_SYNC)
1159 goto out_unlock_inode;
1162 * Only add valid (hashed) inodes to the superblock's
1163 * dirty list. Add blockdev inodes as well.
1165 if (!S_ISBLK(inode->i_mode)) {
1166 if (inode_unhashed(inode))
1167 goto out_unlock_inode;
1169 if (inode->i_state & I_FREEING)
1170 goto out_unlock_inode;
1173 * If the inode was already on b_dirty/b_io/b_more_io, don't
1174 * reposition it (that would break b_dirty time-ordering).
1177 bool wakeup_bdi = false;
1178 bdi = inode_to_bdi(inode);
1180 if (bdi_cap_writeback_dirty(bdi)) {
1181 WARN(!test_bit(BDI_registered, &bdi->state),
1182 "bdi-%s not registered\n", bdi->name);
1185 * If this is the first dirty inode for this
1186 * bdi, we have to wake-up the corresponding
1187 * bdi thread to make sure background
1188 * write-back happens later.
1190 if (!wb_has_dirty_io(&bdi->wb))
1194 spin_unlock(&inode->i_lock);
1195 spin_lock(&bdi->wb.list_lock);
1196 inode->dirtied_when = jiffies;
1197 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1198 spin_unlock(&bdi->wb.list_lock);
1201 bdi_wakeup_thread_delayed(bdi);
1206 spin_unlock(&inode->i_lock);
1209 EXPORT_SYMBOL(__mark_inode_dirty);
1211 static void wait_sb_inodes(struct super_block *sb)
1213 struct inode *inode, *old_inode = NULL;
1216 * We need to be protected against the filesystem going from
1217 * r/o to r/w or vice versa.
1219 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1221 spin_lock(&inode_sb_list_lock);
1224 * Data integrity sync. Must wait for all pages under writeback,
1225 * because there may have been pages dirtied before our sync
1226 * call, but which had writeout started before we write it out.
1227 * In which case, the inode may not be on the dirty list, but
1228 * we still have to wait for that writeout.
1230 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1231 struct address_space *mapping = inode->i_mapping;
1233 spin_lock(&inode->i_lock);
1234 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1235 (mapping->nrpages == 0)) {
1236 spin_unlock(&inode->i_lock);
1240 spin_unlock(&inode->i_lock);
1241 spin_unlock(&inode_sb_list_lock);
1244 * We hold a reference to 'inode' so it couldn't have been
1245 * removed from s_inodes list while we dropped the
1246 * inode_sb_list_lock. We cannot iput the inode now as we can
1247 * be holding the last reference and we cannot iput it under
1248 * inode_sb_list_lock. So we keep the reference and iput it
1254 filemap_fdatawait(mapping);
1258 spin_lock(&inode_sb_list_lock);
1260 spin_unlock(&inode_sb_list_lock);
1265 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1266 * @sb: the superblock
1267 * @nr: the number of pages to write
1268 * @reason: reason why some writeback work initiated
1270 * Start writeback on some inodes on this super_block. No guarantees are made
1271 * on how many (if any) will be written, and this function does not wait
1272 * for IO completion of submitted IO.
1274 void writeback_inodes_sb_nr(struct super_block *sb,
1276 enum wb_reason reason)
1278 DECLARE_COMPLETION_ONSTACK(done);
1279 struct wb_writeback_work work = {
1281 .sync_mode = WB_SYNC_NONE,
1282 .tagged_writepages = 1,
1288 if (sb->s_bdi == &noop_backing_dev_info)
1290 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1291 bdi_queue_work(sb->s_bdi, &work);
1292 wait_for_completion(&done);
1294 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1297 * writeback_inodes_sb - writeback dirty inodes from given super_block
1298 * @sb: the superblock
1299 * @reason: reason why some writeback work was initiated
1301 * Start writeback on some inodes on this super_block. No guarantees are made
1302 * on how many (if any) will be written, and this function does not wait
1303 * for IO completion of submitted IO.
1305 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1307 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1309 EXPORT_SYMBOL(writeback_inodes_sb);
1312 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1313 * @sb: the superblock
1314 * @nr: the number of pages to write
1315 * @reason: the reason of writeback
1317 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1318 * Returns 1 if writeback was started, 0 if not.
1320 int try_to_writeback_inodes_sb_nr(struct super_block *sb,
1322 enum wb_reason reason)
1324 if (writeback_in_progress(sb->s_bdi))
1327 if (!down_read_trylock(&sb->s_umount))
1330 writeback_inodes_sb_nr(sb, nr, reason);
1331 up_read(&sb->s_umount);
1334 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
1337 * try_to_writeback_inodes_sb - try to start writeback if none underway
1338 * @sb: the superblock
1339 * @reason: reason why some writeback work was initiated
1341 * Implement by try_to_writeback_inodes_sb_nr()
1342 * Returns 1 if writeback was started, 0 if not.
1344 int try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1346 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1348 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
1351 * sync_inodes_sb - sync sb inode pages
1352 * @sb: the superblock
1354 * This function writes and waits on any dirty inode belonging to this
1357 void sync_inodes_sb(struct super_block *sb)
1359 DECLARE_COMPLETION_ONSTACK(done);
1360 struct wb_writeback_work work = {
1362 .sync_mode = WB_SYNC_ALL,
1363 .nr_pages = LONG_MAX,
1366 .reason = WB_REASON_SYNC,
1369 /* Nothing to do? */
1370 if (sb->s_bdi == &noop_backing_dev_info)
1372 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1374 bdi_queue_work(sb->s_bdi, &work);
1375 wait_for_completion(&done);
1379 EXPORT_SYMBOL(sync_inodes_sb);
1382 * write_inode_now - write an inode to disk
1383 * @inode: inode to write to disk
1384 * @sync: whether the write should be synchronous or not
1386 * This function commits an inode to disk immediately if it is dirty. This is
1387 * primarily needed by knfsd.
1389 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1391 int write_inode_now(struct inode *inode, int sync)
1393 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1394 struct writeback_control wbc = {
1395 .nr_to_write = LONG_MAX,
1396 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1398 .range_end = LLONG_MAX,
1401 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1402 wbc.nr_to_write = 0;
1405 return writeback_single_inode(inode, wb, &wbc);
1407 EXPORT_SYMBOL(write_inode_now);
1410 * sync_inode - write an inode and its pages to disk.
1411 * @inode: the inode to sync
1412 * @wbc: controls the writeback mode
1414 * sync_inode() will write an inode and its pages to disk. It will also
1415 * correctly update the inode on its superblock's dirty inode lists and will
1416 * update inode->i_state.
1418 * The caller must have a ref on the inode.
1420 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1422 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
1424 EXPORT_SYMBOL(sync_inode);
1427 * sync_inode_metadata - write an inode to disk
1428 * @inode: the inode to sync
1429 * @wait: wait for I/O to complete.
1431 * Write an inode to disk and adjust its dirty state after completion.
1433 * Note: only writes the actual inode, no associated data or other metadata.
1435 int sync_inode_metadata(struct inode *inode, int wait)
1437 struct writeback_control wbc = {
1438 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1439 .nr_to_write = 0, /* metadata-only */
1442 return sync_inode(inode, &wbc);
1444 EXPORT_SYMBOL(sync_inode_metadata);