4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
29 #include <linux/device.h>
33 * 4MB minimal write chunk size
35 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
38 * Passed into wb_writeback(), essentially a subset of writeback_control
40 struct wb_writeback_work {
42 struct super_block *sb;
43 unsigned long *older_than_this;
44 enum writeback_sync_modes sync_mode;
45 unsigned int tagged_writepages:1;
46 unsigned int for_kupdate:1;
47 unsigned int range_cyclic:1;
48 unsigned int for_background:1;
49 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
50 enum wb_reason reason; /* why was writeback initiated? */
52 struct list_head list; /* pending work list */
53 struct completion *done; /* set if the caller waits */
57 * writeback_in_progress - determine whether there is writeback in progress
58 * @bdi: the device's backing_dev_info structure.
60 * Determine whether there is writeback waiting to be handled against a
63 int writeback_in_progress(struct backing_dev_info *bdi)
65 return test_bit(BDI_writeback_running, &bdi->state);
67 EXPORT_SYMBOL(writeback_in_progress);
69 struct backing_dev_info *inode_to_bdi(struct inode *inode)
71 struct super_block *sb;
74 return &noop_backing_dev_info;
78 if (sb_is_blkdev_sb(sb))
79 return blk_get_backing_dev_info(I_BDEV(inode));
83 EXPORT_SYMBOL_GPL(inode_to_bdi);
85 static inline struct inode *wb_inode(struct list_head *head)
87 return list_entry(head, struct inode, i_wb_list);
91 * Include the creation of the trace points after defining the
92 * wb_writeback_work structure and inline functions so that the definition
93 * remains local to this file.
95 #define CREATE_TRACE_POINTS
96 #include <trace/events/writeback.h>
98 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
100 static void bdi_wakeup_thread(struct backing_dev_info *bdi)
102 spin_lock_bh(&bdi->wb_lock);
103 if (test_bit(BDI_registered, &bdi->state))
104 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
105 spin_unlock_bh(&bdi->wb_lock);
108 static void bdi_queue_work(struct backing_dev_info *bdi,
109 struct wb_writeback_work *work)
111 trace_writeback_queue(bdi, work);
113 spin_lock_bh(&bdi->wb_lock);
114 if (!test_bit(BDI_registered, &bdi->state)) {
116 complete(work->done);
119 list_add_tail(&work->list, &bdi->work_list);
120 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
122 spin_unlock_bh(&bdi->wb_lock);
126 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
127 bool range_cyclic, enum wb_reason reason)
129 struct wb_writeback_work *work;
132 * This is WB_SYNC_NONE writeback, so if allocation fails just
133 * wakeup the thread for old dirty data writeback
135 work = kzalloc(sizeof(*work), GFP_ATOMIC);
137 trace_writeback_nowork(bdi);
138 bdi_wakeup_thread(bdi);
142 work->sync_mode = WB_SYNC_NONE;
143 work->nr_pages = nr_pages;
144 work->range_cyclic = range_cyclic;
145 work->reason = reason;
147 bdi_queue_work(bdi, work);
151 * bdi_start_writeback - start writeback
152 * @bdi: the backing device to write from
153 * @nr_pages: the number of pages to write
154 * @reason: reason why some writeback work was initiated
157 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
158 * started when this function returns, we make no guarantees on
159 * completion. Caller need not hold sb s_umount semaphore.
162 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
163 enum wb_reason reason)
165 __bdi_start_writeback(bdi, nr_pages, true, reason);
169 * bdi_start_background_writeback - start background writeback
170 * @bdi: the backing device to write from
173 * This makes sure WB_SYNC_NONE background writeback happens. When
174 * this function returns, it is only guaranteed that for given BDI
175 * some IO is happening if we are over background dirty threshold.
176 * Caller need not hold sb s_umount semaphore.
178 void bdi_start_background_writeback(struct backing_dev_info *bdi)
181 * We just wake up the flusher thread. It will perform background
182 * writeback as soon as there is no other work to do.
184 trace_writeback_wake_background(bdi);
185 bdi_wakeup_thread(bdi);
189 * Remove the inode from the writeback list it is on.
191 void inode_wb_list_del(struct inode *inode)
193 struct backing_dev_info *bdi = inode_to_bdi(inode);
195 spin_lock(&bdi->wb.list_lock);
196 list_del_init(&inode->i_wb_list);
197 spin_unlock(&bdi->wb.list_lock);
201 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
202 * furthest end of its superblock's dirty-inode list.
204 * Before stamping the inode's ->dirtied_when, we check to see whether it is
205 * already the most-recently-dirtied inode on the b_dirty list. If that is
206 * the case then the inode must have been redirtied while it was being written
207 * out and we don't reset its dirtied_when.
209 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
211 assert_spin_locked(&wb->list_lock);
212 if (!list_empty(&wb->b_dirty)) {
215 tail = wb_inode(wb->b_dirty.next);
216 if (time_before(inode->dirtied_when, tail->dirtied_when))
217 inode->dirtied_when = jiffies;
219 list_move(&inode->i_wb_list, &wb->b_dirty);
223 * requeue inode for re-scanning after bdi->b_io list is exhausted.
225 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
227 assert_spin_locked(&wb->list_lock);
228 list_move(&inode->i_wb_list, &wb->b_more_io);
231 static void inode_sync_complete(struct inode *inode)
233 inode->i_state &= ~I_SYNC;
234 /* If inode is clean an unused, put it into LRU now... */
235 inode_add_lru(inode);
236 /* Waiters must see I_SYNC cleared before being woken up */
238 wake_up_bit(&inode->i_state, __I_SYNC);
241 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
243 bool ret = time_after(inode->dirtied_when, t);
246 * For inodes being constantly redirtied, dirtied_when can get stuck.
247 * It _appears_ to be in the future, but is actually in distant past.
248 * This test is necessary to prevent such wrapped-around relative times
249 * from permanently stopping the whole bdi writeback.
251 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
256 #define EXPIRE_DIRTY_ATIME 0x0001
259 * Move expired (dirtied before work->older_than_this) dirty inodes from
260 * @delaying_queue to @dispatch_queue.
262 static int move_expired_inodes(struct list_head *delaying_queue,
263 struct list_head *dispatch_queue,
265 struct wb_writeback_work *work)
267 unsigned long *older_than_this = NULL;
268 unsigned long expire_time;
270 struct list_head *pos, *node;
271 struct super_block *sb = NULL;
276 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
277 older_than_this = work->older_than_this;
278 else if ((work->reason == WB_REASON_SYNC) == 0) {
279 expire_time = jiffies - (HZ * 86400);
280 older_than_this = &expire_time;
282 while (!list_empty(delaying_queue)) {
283 inode = wb_inode(delaying_queue->prev);
284 if (older_than_this &&
285 inode_dirtied_after(inode, *older_than_this))
287 list_move(&inode->i_wb_list, &tmp);
289 if (flags & EXPIRE_DIRTY_ATIME)
290 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
291 if (sb_is_blkdev_sb(inode->i_sb))
293 if (sb && sb != inode->i_sb)
298 /* just one sb in list, splice to dispatch_queue and we're done */
300 list_splice(&tmp, dispatch_queue);
304 /* Move inodes from one superblock together */
305 while (!list_empty(&tmp)) {
306 sb = wb_inode(tmp.prev)->i_sb;
307 list_for_each_prev_safe(pos, node, &tmp) {
308 inode = wb_inode(pos);
309 if (inode->i_sb == sb)
310 list_move(&inode->i_wb_list, dispatch_queue);
318 * Queue all expired dirty inodes for io, eldest first.
320 * newly dirtied b_dirty b_io b_more_io
321 * =============> gf edc BA
323 * newly dirtied b_dirty b_io b_more_io
324 * =============> g fBAedc
326 * +--> dequeue for IO
328 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
332 assert_spin_locked(&wb->list_lock);
333 list_splice_init(&wb->b_more_io, &wb->b_io);
334 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
335 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
336 EXPIRE_DIRTY_ATIME, work);
337 trace_writeback_queue_io(wb, work, moved);
340 static int write_inode(struct inode *inode, struct writeback_control *wbc)
344 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
345 trace_writeback_write_inode_start(inode, wbc);
346 ret = inode->i_sb->s_op->write_inode(inode, wbc);
347 trace_writeback_write_inode(inode, wbc);
354 * Wait for writeback on an inode to complete. Called with i_lock held.
355 * Caller must make sure inode cannot go away when we drop i_lock.
357 static void __inode_wait_for_writeback(struct inode *inode)
358 __releases(inode->i_lock)
359 __acquires(inode->i_lock)
361 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
362 wait_queue_head_t *wqh;
364 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
365 while (inode->i_state & I_SYNC) {
366 spin_unlock(&inode->i_lock);
367 __wait_on_bit(wqh, &wq, bit_wait,
368 TASK_UNINTERRUPTIBLE);
369 spin_lock(&inode->i_lock);
374 * Wait for writeback on an inode to complete. Caller must have inode pinned.
376 void inode_wait_for_writeback(struct inode *inode)
378 spin_lock(&inode->i_lock);
379 __inode_wait_for_writeback(inode);
380 spin_unlock(&inode->i_lock);
384 * Sleep until I_SYNC is cleared. This function must be called with i_lock
385 * held and drops it. It is aimed for callers not holding any inode reference
386 * so once i_lock is dropped, inode can go away.
388 static void inode_sleep_on_writeback(struct inode *inode)
389 __releases(inode->i_lock)
392 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
395 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
396 sleep = inode->i_state & I_SYNC;
397 spin_unlock(&inode->i_lock);
400 finish_wait(wqh, &wait);
404 * Find proper writeback list for the inode depending on its current state and
405 * possibly also change of its state while we were doing writeback. Here we
406 * handle things such as livelock prevention or fairness of writeback among
407 * inodes. This function can be called only by flusher thread - noone else
408 * processes all inodes in writeback lists and requeueing inodes behind flusher
409 * thread's back can have unexpected consequences.
411 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
412 struct writeback_control *wbc)
414 if (inode->i_state & I_FREEING)
418 * Sync livelock prevention. Each inode is tagged and synced in one
419 * shot. If still dirty, it will be redirty_tail()'ed below. Update
420 * the dirty time to prevent enqueue and sync it again.
422 if ((inode->i_state & I_DIRTY) &&
423 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
424 inode->dirtied_when = jiffies;
426 if (wbc->pages_skipped) {
428 * writeback is not making progress due to locked
429 * buffers. Skip this inode for now.
431 redirty_tail(inode, wb);
435 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
437 * We didn't write back all the pages. nfs_writepages()
438 * sometimes bales out without doing anything.
440 if (wbc->nr_to_write <= 0) {
441 /* Slice used up. Queue for next turn. */
442 requeue_io(inode, wb);
445 * Writeback blocked by something other than
446 * congestion. Delay the inode for some time to
447 * avoid spinning on the CPU (100% iowait)
448 * retrying writeback of the dirty page/inode
449 * that cannot be performed immediately.
451 redirty_tail(inode, wb);
453 } else if (inode->i_state & I_DIRTY) {
455 * Filesystems can dirty the inode during writeback operations,
456 * such as delayed allocation during submission or metadata
457 * updates after data IO completion.
459 redirty_tail(inode, wb);
460 } else if (inode->i_state & I_DIRTY_TIME) {
461 list_move(&inode->i_wb_list, &wb->b_dirty_time);
463 /* The inode is clean. Remove from writeback lists. */
464 list_del_init(&inode->i_wb_list);
469 * Write out an inode and its dirty pages. Do not update the writeback list
470 * linkage. That is left to the caller. The caller is also responsible for
471 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
474 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
476 struct address_space *mapping = inode->i_mapping;
477 long nr_to_write = wbc->nr_to_write;
481 WARN_ON(!(inode->i_state & I_SYNC));
483 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
485 ret = do_writepages(mapping, wbc);
488 * Make sure to wait on the data before writing out the metadata.
489 * This is important for filesystems that modify metadata on data
490 * I/O completion. We don't do it for sync(2) writeback because it has a
491 * separate, external IO completion path and ->sync_fs for guaranteeing
492 * inode metadata is written back correctly.
494 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
495 int err = filemap_fdatawait(mapping);
501 * Some filesystems may redirty the inode during the writeback
502 * due to delalloc, clear dirty metadata flags right before
505 spin_lock(&inode->i_lock);
507 dirty = inode->i_state & I_DIRTY;
508 if (((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) &&
509 (inode->i_state & I_DIRTY_TIME)) ||
510 (inode->i_state & I_DIRTY_TIME_EXPIRED)) {
511 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
512 trace_writeback_lazytime(inode);
514 inode->i_state &= ~dirty;
517 * Paired with smp_mb() in __mark_inode_dirty(). This allows
518 * __mark_inode_dirty() to test i_state without grabbing i_lock -
519 * either they see the I_DIRTY bits cleared or we see the dirtied
522 * I_DIRTY_PAGES is always cleared together above even if @mapping
523 * still has dirty pages. The flag is reinstated after smp_mb() if
524 * necessary. This guarantees that either __mark_inode_dirty()
525 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
529 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
530 inode->i_state |= I_DIRTY_PAGES;
532 spin_unlock(&inode->i_lock);
534 if (dirty & I_DIRTY_TIME)
535 mark_inode_dirty_sync(inode);
536 /* Don't write the inode if only I_DIRTY_PAGES was set */
537 if (dirty & ~I_DIRTY_PAGES) {
538 int err = write_inode(inode, wbc);
542 trace_writeback_single_inode(inode, wbc, nr_to_write);
547 * Write out an inode's dirty pages. Either the caller has an active reference
548 * on the inode or the inode has I_WILL_FREE set.
550 * This function is designed to be called for writing back one inode which
551 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
552 * and does more profound writeback list handling in writeback_sb_inodes().
555 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
556 struct writeback_control *wbc)
560 spin_lock(&inode->i_lock);
561 if (!atomic_read(&inode->i_count))
562 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
564 WARN_ON(inode->i_state & I_WILL_FREE);
566 if (inode->i_state & I_SYNC) {
567 if (wbc->sync_mode != WB_SYNC_ALL)
570 * It's a data-integrity sync. We must wait. Since callers hold
571 * inode reference or inode has I_WILL_FREE set, it cannot go
574 __inode_wait_for_writeback(inode);
576 WARN_ON(inode->i_state & I_SYNC);
578 * Skip inode if it is clean and we have no outstanding writeback in
579 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
580 * function since flusher thread may be doing for example sync in
581 * parallel and if we move the inode, it could get skipped. So here we
582 * make sure inode is on some writeback list and leave it there unless
583 * we have completely cleaned the inode.
585 if (!(inode->i_state & I_DIRTY_ALL) &&
586 (wbc->sync_mode != WB_SYNC_ALL ||
587 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
589 inode->i_state |= I_SYNC;
590 spin_unlock(&inode->i_lock);
592 ret = __writeback_single_inode(inode, wbc);
594 spin_lock(&wb->list_lock);
595 spin_lock(&inode->i_lock);
597 * If inode is clean, remove it from writeback lists. Otherwise don't
598 * touch it. See comment above for explanation.
600 if (!(inode->i_state & I_DIRTY_ALL))
601 list_del_init(&inode->i_wb_list);
602 spin_unlock(&wb->list_lock);
603 inode_sync_complete(inode);
605 spin_unlock(&inode->i_lock);
609 static long writeback_chunk_size(struct backing_dev_info *bdi,
610 struct wb_writeback_work *work)
615 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
616 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
617 * here avoids calling into writeback_inodes_wb() more than once.
619 * The intended call sequence for WB_SYNC_ALL writeback is:
622 * writeback_sb_inodes() <== called only once
623 * write_cache_pages() <== called once for each inode
624 * (quickly) tag currently dirty pages
625 * (maybe slowly) sync all tagged pages
627 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
630 pages = min(bdi->avg_write_bandwidth / 2,
631 global_dirty_limit / DIRTY_SCOPE);
632 pages = min(pages, work->nr_pages);
633 pages = round_down(pages + MIN_WRITEBACK_PAGES,
634 MIN_WRITEBACK_PAGES);
641 * Write a portion of b_io inodes which belong to @sb.
643 * Return the number of pages and/or inodes written.
645 static long writeback_sb_inodes(struct super_block *sb,
646 struct bdi_writeback *wb,
647 struct wb_writeback_work *work)
649 struct writeback_control wbc = {
650 .sync_mode = work->sync_mode,
651 .tagged_writepages = work->tagged_writepages,
652 .for_kupdate = work->for_kupdate,
653 .for_background = work->for_background,
654 .for_sync = work->for_sync,
655 .range_cyclic = work->range_cyclic,
657 .range_end = LLONG_MAX,
659 unsigned long start_time = jiffies;
661 long wrote = 0; /* count both pages and inodes */
663 while (!list_empty(&wb->b_io)) {
664 struct inode *inode = wb_inode(wb->b_io.prev);
666 if (inode->i_sb != sb) {
669 * We only want to write back data for this
670 * superblock, move all inodes not belonging
671 * to it back onto the dirty list.
673 redirty_tail(inode, wb);
678 * The inode belongs to a different superblock.
679 * Bounce back to the caller to unpin this and
680 * pin the next superblock.
686 * Don't bother with new inodes or inodes being freed, first
687 * kind does not need periodic writeout yet, and for the latter
688 * kind writeout is handled by the freer.
690 spin_lock(&inode->i_lock);
691 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
692 spin_unlock(&inode->i_lock);
693 redirty_tail(inode, wb);
696 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
698 * If this inode is locked for writeback and we are not
699 * doing writeback-for-data-integrity, move it to
700 * b_more_io so that writeback can proceed with the
701 * other inodes on s_io.
703 * We'll have another go at writing back this inode
704 * when we completed a full scan of b_io.
706 spin_unlock(&inode->i_lock);
707 requeue_io(inode, wb);
708 trace_writeback_sb_inodes_requeue(inode);
711 spin_unlock(&wb->list_lock);
714 * We already requeued the inode if it had I_SYNC set and we
715 * are doing WB_SYNC_NONE writeback. So this catches only the
718 if (inode->i_state & I_SYNC) {
719 /* Wait for I_SYNC. This function drops i_lock... */
720 inode_sleep_on_writeback(inode);
721 /* Inode may be gone, start again */
722 spin_lock(&wb->list_lock);
725 inode->i_state |= I_SYNC;
726 spin_unlock(&inode->i_lock);
728 write_chunk = writeback_chunk_size(wb->bdi, work);
729 wbc.nr_to_write = write_chunk;
730 wbc.pages_skipped = 0;
733 * We use I_SYNC to pin the inode in memory. While it is set
734 * evict_inode() will wait so the inode cannot be freed.
736 __writeback_single_inode(inode, &wbc);
738 work->nr_pages -= write_chunk - wbc.nr_to_write;
739 wrote += write_chunk - wbc.nr_to_write;
740 spin_lock(&wb->list_lock);
741 spin_lock(&inode->i_lock);
742 if (!(inode->i_state & I_DIRTY_ALL))
744 requeue_inode(inode, wb, &wbc);
745 inode_sync_complete(inode);
746 spin_unlock(&inode->i_lock);
747 cond_resched_lock(&wb->list_lock);
749 * bail out to wb_writeback() often enough to check
750 * background threshold and other termination conditions.
753 if (time_is_before_jiffies(start_time + HZ / 10UL))
755 if (work->nr_pages <= 0)
762 static long __writeback_inodes_wb(struct bdi_writeback *wb,
763 struct wb_writeback_work *work)
765 unsigned long start_time = jiffies;
768 while (!list_empty(&wb->b_io)) {
769 struct inode *inode = wb_inode(wb->b_io.prev);
770 struct super_block *sb = inode->i_sb;
772 if (!trylock_super(sb)) {
774 * trylock_super() may fail consistently due to
775 * s_umount being grabbed by someone else. Don't use
776 * requeue_io() to avoid busy retrying the inode/sb.
778 redirty_tail(inode, wb);
781 wrote += writeback_sb_inodes(sb, wb, work);
782 up_read(&sb->s_umount);
784 /* refer to the same tests at the end of writeback_sb_inodes */
786 if (time_is_before_jiffies(start_time + HZ / 10UL))
788 if (work->nr_pages <= 0)
792 /* Leave any unwritten inodes on b_io */
796 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
797 enum wb_reason reason)
799 struct wb_writeback_work work = {
800 .nr_pages = nr_pages,
801 .sync_mode = WB_SYNC_NONE,
806 spin_lock(&wb->list_lock);
807 if (list_empty(&wb->b_io))
809 __writeback_inodes_wb(wb, &work);
810 spin_unlock(&wb->list_lock);
812 return nr_pages - work.nr_pages;
815 static bool over_bground_thresh(struct backing_dev_info *bdi)
817 unsigned long background_thresh, dirty_thresh;
819 global_dirty_limits(&background_thresh, &dirty_thresh);
821 if (global_page_state(NR_FILE_DIRTY) +
822 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
825 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
826 bdi_dirty_limit(bdi, background_thresh))
833 * Called under wb->list_lock. If there are multiple wb per bdi,
834 * only the flusher working on the first wb should do it.
836 static void wb_update_bandwidth(struct bdi_writeback *wb,
837 unsigned long start_time)
839 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
843 * Explicit flushing or periodic writeback of "old" data.
845 * Define "old": the first time one of an inode's pages is dirtied, we mark the
846 * dirtying-time in the inode's address_space. So this periodic writeback code
847 * just walks the superblock inode list, writing back any inodes which are
848 * older than a specific point in time.
850 * Try to run once per dirty_writeback_interval. But if a writeback event
851 * takes longer than a dirty_writeback_interval interval, then leave a
854 * older_than_this takes precedence over nr_to_write. So we'll only write back
855 * all dirty pages if they are all attached to "old" mappings.
857 static long wb_writeback(struct bdi_writeback *wb,
858 struct wb_writeback_work *work)
860 unsigned long wb_start = jiffies;
861 long nr_pages = work->nr_pages;
862 unsigned long oldest_jif;
866 oldest_jif = jiffies;
867 work->older_than_this = &oldest_jif;
869 spin_lock(&wb->list_lock);
872 * Stop writeback when nr_pages has been consumed
874 if (work->nr_pages <= 0)
878 * Background writeout and kupdate-style writeback may
879 * run forever. Stop them if there is other work to do
880 * so that e.g. sync can proceed. They'll be restarted
881 * after the other works are all done.
883 if ((work->for_background || work->for_kupdate) &&
884 !list_empty(&wb->bdi->work_list))
888 * For background writeout, stop when we are below the
889 * background dirty threshold
891 if (work->for_background && !over_bground_thresh(wb->bdi))
895 * Kupdate and background works are special and we want to
896 * include all inodes that need writing. Livelock avoidance is
897 * handled by these works yielding to any other work so we are
900 if (work->for_kupdate) {
901 oldest_jif = jiffies -
902 msecs_to_jiffies(dirty_expire_interval * 10);
903 } else if (work->for_background)
904 oldest_jif = jiffies;
906 trace_writeback_start(wb->bdi, work);
907 if (list_empty(&wb->b_io))
910 progress = writeback_sb_inodes(work->sb, wb, work);
912 progress = __writeback_inodes_wb(wb, work);
913 trace_writeback_written(wb->bdi, work);
915 wb_update_bandwidth(wb, wb_start);
918 * Did we write something? Try for more
920 * Dirty inodes are moved to b_io for writeback in batches.
921 * The completion of the current batch does not necessarily
922 * mean the overall work is done. So we keep looping as long
923 * as made some progress on cleaning pages or inodes.
928 * No more inodes for IO, bail
930 if (list_empty(&wb->b_more_io))
933 * Nothing written. Wait for some inode to
934 * become available for writeback. Otherwise
935 * we'll just busyloop.
937 if (!list_empty(&wb->b_more_io)) {
938 trace_writeback_wait(wb->bdi, work);
939 inode = wb_inode(wb->b_more_io.prev);
940 spin_lock(&inode->i_lock);
941 spin_unlock(&wb->list_lock);
942 /* This function drops i_lock... */
943 inode_sleep_on_writeback(inode);
944 spin_lock(&wb->list_lock);
947 spin_unlock(&wb->list_lock);
949 return nr_pages - work->nr_pages;
953 * Return the next wb_writeback_work struct that hasn't been processed yet.
955 static struct wb_writeback_work *
956 get_next_work_item(struct backing_dev_info *bdi)
958 struct wb_writeback_work *work = NULL;
960 spin_lock_bh(&bdi->wb_lock);
961 if (!list_empty(&bdi->work_list)) {
962 work = list_entry(bdi->work_list.next,
963 struct wb_writeback_work, list);
964 list_del_init(&work->list);
966 spin_unlock_bh(&bdi->wb_lock);
971 * Add in the number of potentially dirty inodes, because each inode
972 * write can dirty pagecache in the underlying blockdev.
974 static unsigned long get_nr_dirty_pages(void)
976 return global_page_state(NR_FILE_DIRTY) +
977 global_page_state(NR_UNSTABLE_NFS) +
978 get_nr_dirty_inodes();
981 static long wb_check_background_flush(struct bdi_writeback *wb)
983 if (over_bground_thresh(wb->bdi)) {
985 struct wb_writeback_work work = {
986 .nr_pages = LONG_MAX,
987 .sync_mode = WB_SYNC_NONE,
990 .reason = WB_REASON_BACKGROUND,
993 return wb_writeback(wb, &work);
999 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1001 unsigned long expired;
1005 * When set to zero, disable periodic writeback
1007 if (!dirty_writeback_interval)
1010 expired = wb->last_old_flush +
1011 msecs_to_jiffies(dirty_writeback_interval * 10);
1012 if (time_before(jiffies, expired))
1015 wb->last_old_flush = jiffies;
1016 nr_pages = get_nr_dirty_pages();
1019 struct wb_writeback_work work = {
1020 .nr_pages = nr_pages,
1021 .sync_mode = WB_SYNC_NONE,
1024 .reason = WB_REASON_PERIODIC,
1027 return wb_writeback(wb, &work);
1034 * Retrieve work items and do the writeback they describe
1036 static long wb_do_writeback(struct bdi_writeback *wb)
1038 struct backing_dev_info *bdi = wb->bdi;
1039 struct wb_writeback_work *work;
1042 set_bit(BDI_writeback_running, &wb->bdi->state);
1043 while ((work = get_next_work_item(bdi)) != NULL) {
1045 trace_writeback_exec(bdi, work);
1047 wrote += wb_writeback(wb, work);
1050 * Notify the caller of completion if this is a synchronous
1051 * work item, otherwise just free it.
1054 complete(work->done);
1060 * Check for periodic writeback, kupdated() style
1062 wrote += wb_check_old_data_flush(wb);
1063 wrote += wb_check_background_flush(wb);
1064 clear_bit(BDI_writeback_running, &wb->bdi->state);
1070 * Handle writeback of dirty data for the device backed by this bdi. Also
1071 * reschedules periodically and does kupdated style flushing.
1073 void bdi_writeback_workfn(struct work_struct *work)
1075 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1076 struct bdi_writeback, dwork);
1077 struct backing_dev_info *bdi = wb->bdi;
1080 set_worker_desc("flush-%s", dev_name(bdi->dev));
1081 current->flags |= PF_SWAPWRITE;
1083 if (likely(!current_is_workqueue_rescuer() ||
1084 !test_bit(BDI_registered, &bdi->state))) {
1086 * The normal path. Keep writing back @bdi until its
1087 * work_list is empty. Note that this path is also taken
1088 * if @bdi is shutting down even when we're running off the
1089 * rescuer as work_list needs to be drained.
1092 pages_written = wb_do_writeback(wb);
1093 trace_writeback_pages_written(pages_written);
1094 } while (!list_empty(&bdi->work_list));
1097 * bdi_wq can't get enough workers and we're running off
1098 * the emergency worker. Don't hog it. Hopefully, 1024 is
1099 * enough for efficient IO.
1101 pages_written = writeback_inodes_wb(&bdi->wb, 1024,
1102 WB_REASON_FORKER_THREAD);
1103 trace_writeback_pages_written(pages_written);
1106 if (!list_empty(&bdi->work_list))
1107 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1108 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1109 bdi_wakeup_thread_delayed(bdi);
1111 current->flags &= ~PF_SWAPWRITE;
1115 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1118 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1120 struct backing_dev_info *bdi;
1123 nr_pages = get_nr_dirty_pages();
1126 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1127 if (!bdi_has_dirty_io(bdi))
1129 __bdi_start_writeback(bdi, nr_pages, false, reason);
1134 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1136 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1137 struct dentry *dentry;
1138 const char *name = "?";
1140 dentry = d_find_alias(inode);
1142 spin_lock(&dentry->d_lock);
1143 name = (const char *) dentry->d_name.name;
1146 "%s(%d): dirtied inode %lu (%s) on %s\n",
1147 current->comm, task_pid_nr(current), inode->i_ino,
1148 name, inode->i_sb->s_id);
1150 spin_unlock(&dentry->d_lock);
1157 * __mark_inode_dirty - internal function
1158 * @inode: inode to mark
1159 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1160 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1161 * mark_inode_dirty_sync.
1163 * Put the inode on the super block's dirty list.
1165 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1166 * dirty list only if it is hashed or if it refers to a blockdev.
1167 * If it was not hashed, it will never be added to the dirty list
1168 * even if it is later hashed, as it will have been marked dirty already.
1170 * In short, make sure you hash any inodes _before_ you start marking
1173 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1174 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1175 * the kernel-internal blockdev inode represents the dirtying time of the
1176 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1177 * page->mapping->host, so the page-dirtying time is recorded in the internal
1180 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
1181 void __mark_inode_dirty(struct inode *inode, int flags)
1183 struct super_block *sb = inode->i_sb;
1184 struct backing_dev_info *bdi = NULL;
1187 trace_writeback_mark_inode_dirty(inode, flags);
1190 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1191 * dirty the inode itself
1193 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
1194 trace_writeback_dirty_inode_start(inode, flags);
1196 if (sb->s_op->dirty_inode)
1197 sb->s_op->dirty_inode(inode, flags);
1199 trace_writeback_dirty_inode(inode, flags);
1201 if (flags & I_DIRTY_INODE)
1202 flags &= ~I_DIRTY_TIME;
1203 dirtytime = flags & I_DIRTY_TIME;
1206 * Paired with smp_mb() in __writeback_single_inode() for the
1207 * following lockless i_state test. See there for details.
1211 if (((inode->i_state & flags) == flags) ||
1212 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
1215 if (unlikely(block_dump))
1216 block_dump___mark_inode_dirty(inode);
1218 spin_lock(&inode->i_lock);
1219 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
1220 goto out_unlock_inode;
1221 if ((inode->i_state & flags) != flags) {
1222 const int was_dirty = inode->i_state & I_DIRTY;
1224 if (flags & I_DIRTY_INODE)
1225 inode->i_state &= ~I_DIRTY_TIME;
1226 inode->i_state |= flags;
1229 * If the inode is being synced, just update its dirty state.
1230 * The unlocker will place the inode on the appropriate
1231 * superblock list, based upon its state.
1233 if (inode->i_state & I_SYNC)
1234 goto out_unlock_inode;
1237 * Only add valid (hashed) inodes to the superblock's
1238 * dirty list. Add blockdev inodes as well.
1240 if (!S_ISBLK(inode->i_mode)) {
1241 if (inode_unhashed(inode))
1242 goto out_unlock_inode;
1244 if (inode->i_state & I_FREEING)
1245 goto out_unlock_inode;
1248 * If the inode was already on b_dirty/b_io/b_more_io, don't
1249 * reposition it (that would break b_dirty time-ordering).
1252 bool wakeup_bdi = false;
1253 bdi = inode_to_bdi(inode);
1255 spin_unlock(&inode->i_lock);
1256 spin_lock(&bdi->wb.list_lock);
1257 if (bdi_cap_writeback_dirty(bdi)) {
1258 WARN(!test_bit(BDI_registered, &bdi->state),
1259 "bdi-%s not registered\n", bdi->name);
1262 * If this is the first dirty inode for this
1263 * bdi, we have to wake-up the corresponding
1264 * bdi thread to make sure background
1265 * write-back happens later.
1267 if (!wb_has_dirty_io(&bdi->wb))
1271 inode->dirtied_when = jiffies;
1272 list_move(&inode->i_wb_list, dirtytime ?
1273 &bdi->wb.b_dirty_time : &bdi->wb.b_dirty);
1274 spin_unlock(&bdi->wb.list_lock);
1275 trace_writeback_dirty_inode_enqueue(inode);
1278 bdi_wakeup_thread_delayed(bdi);
1283 spin_unlock(&inode->i_lock);
1286 EXPORT_SYMBOL(__mark_inode_dirty);
1288 static void wait_sb_inodes(struct super_block *sb)
1290 struct inode *inode, *old_inode = NULL;
1293 * We need to be protected against the filesystem going from
1294 * r/o to r/w or vice versa.
1296 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1298 spin_lock(&inode_sb_list_lock);
1301 * Data integrity sync. Must wait for all pages under writeback,
1302 * because there may have been pages dirtied before our sync
1303 * call, but which had writeout started before we write it out.
1304 * In which case, the inode may not be on the dirty list, but
1305 * we still have to wait for that writeout.
1307 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1308 struct address_space *mapping = inode->i_mapping;
1310 spin_lock(&inode->i_lock);
1311 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1312 (mapping->nrpages == 0)) {
1313 spin_unlock(&inode->i_lock);
1317 spin_unlock(&inode->i_lock);
1318 spin_unlock(&inode_sb_list_lock);
1321 * We hold a reference to 'inode' so it couldn't have been
1322 * removed from s_inodes list while we dropped the
1323 * inode_sb_list_lock. We cannot iput the inode now as we can
1324 * be holding the last reference and we cannot iput it under
1325 * inode_sb_list_lock. So we keep the reference and iput it
1331 filemap_fdatawait(mapping);
1335 spin_lock(&inode_sb_list_lock);
1337 spin_unlock(&inode_sb_list_lock);
1342 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1343 * @sb: the superblock
1344 * @nr: the number of pages to write
1345 * @reason: reason why some writeback work initiated
1347 * Start writeback on some inodes on this super_block. No guarantees are made
1348 * on how many (if any) will be written, and this function does not wait
1349 * for IO completion of submitted IO.
1351 void writeback_inodes_sb_nr(struct super_block *sb,
1353 enum wb_reason reason)
1355 DECLARE_COMPLETION_ONSTACK(done);
1356 struct wb_writeback_work work = {
1358 .sync_mode = WB_SYNC_NONE,
1359 .tagged_writepages = 1,
1365 if (sb->s_bdi == &noop_backing_dev_info)
1367 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1368 bdi_queue_work(sb->s_bdi, &work);
1369 wait_for_completion(&done);
1371 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1374 * writeback_inodes_sb - writeback dirty inodes from given super_block
1375 * @sb: the superblock
1376 * @reason: reason why some writeback work was initiated
1378 * Start writeback on some inodes on this super_block. No guarantees are made
1379 * on how many (if any) will be written, and this function does not wait
1380 * for IO completion of submitted IO.
1382 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1384 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1386 EXPORT_SYMBOL(writeback_inodes_sb);
1389 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1390 * @sb: the superblock
1391 * @nr: the number of pages to write
1392 * @reason: the reason of writeback
1394 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1395 * Returns 1 if writeback was started, 0 if not.
1397 int try_to_writeback_inodes_sb_nr(struct super_block *sb,
1399 enum wb_reason reason)
1401 if (writeback_in_progress(sb->s_bdi))
1404 if (!down_read_trylock(&sb->s_umount))
1407 writeback_inodes_sb_nr(sb, nr, reason);
1408 up_read(&sb->s_umount);
1411 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
1414 * try_to_writeback_inodes_sb - try to start writeback if none underway
1415 * @sb: the superblock
1416 * @reason: reason why some writeback work was initiated
1418 * Implement by try_to_writeback_inodes_sb_nr()
1419 * Returns 1 if writeback was started, 0 if not.
1421 int try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1423 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1425 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
1428 * sync_inodes_sb - sync sb inode pages
1429 * @sb: the superblock
1431 * This function writes and waits on any dirty inode belonging to this
1434 void sync_inodes_sb(struct super_block *sb)
1436 DECLARE_COMPLETION_ONSTACK(done);
1437 struct wb_writeback_work work = {
1439 .sync_mode = WB_SYNC_ALL,
1440 .nr_pages = LONG_MAX,
1443 .reason = WB_REASON_SYNC,
1447 /* Nothing to do? */
1448 if (sb->s_bdi == &noop_backing_dev_info)
1450 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1452 bdi_queue_work(sb->s_bdi, &work);
1453 wait_for_completion(&done);
1457 EXPORT_SYMBOL(sync_inodes_sb);
1460 * write_inode_now - write an inode to disk
1461 * @inode: inode to write to disk
1462 * @sync: whether the write should be synchronous or not
1464 * This function commits an inode to disk immediately if it is dirty. This is
1465 * primarily needed by knfsd.
1467 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1469 int write_inode_now(struct inode *inode, int sync)
1471 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1472 struct writeback_control wbc = {
1473 .nr_to_write = LONG_MAX,
1474 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1476 .range_end = LLONG_MAX,
1479 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1480 wbc.nr_to_write = 0;
1483 return writeback_single_inode(inode, wb, &wbc);
1485 EXPORT_SYMBOL(write_inode_now);
1488 * sync_inode - write an inode and its pages to disk.
1489 * @inode: the inode to sync
1490 * @wbc: controls the writeback mode
1492 * sync_inode() will write an inode and its pages to disk. It will also
1493 * correctly update the inode on its superblock's dirty inode lists and will
1494 * update inode->i_state.
1496 * The caller must have a ref on the inode.
1498 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1500 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
1502 EXPORT_SYMBOL(sync_inode);
1505 * sync_inode_metadata - write an inode to disk
1506 * @inode: the inode to sync
1507 * @wait: wait for I/O to complete.
1509 * Write an inode to disk and adjust its dirty state after completion.
1511 * Note: only writes the actual inode, no associated data or other metadata.
1513 int sync_inode_metadata(struct inode *inode, int wait)
1515 struct writeback_control wbc = {
1516 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1517 .nr_to_write = 0, /* metadata-only */
1520 return sync_inode(inode, &wbc);
1522 EXPORT_SYMBOL(sync_inode_metadata);