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/freezer.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.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 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 * We don't actually have pdflush, but this one is exported though /proc...
58 int nr_pdflush_threads;
61 * writeback_in_progress - determine whether there is writeback in progress
62 * @bdi: the device's backing_dev_info structure.
64 * Determine whether there is writeback waiting to be handled against a
67 int writeback_in_progress(struct backing_dev_info *bdi)
69 return test_bit(BDI_writeback_running, &bdi->state);
72 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
74 struct super_block *sb = inode->i_sb;
76 if (strcmp(sb->s_type->name, "bdev") == 0)
77 return inode->i_mapping->backing_dev_info;
82 static inline struct inode *wb_inode(struct list_head *head)
84 return list_entry(head, struct inode, i_wb_list);
88 * Include the creation of the trace points after defining the
89 * wb_writeback_work structure and inline functions so that the definition
90 * remains local to this file.
92 #define CREATE_TRACE_POINTS
93 #include <trace/events/writeback.h>
95 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
96 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
99 wake_up_process(bdi->wb.task);
102 * The bdi thread isn't there, wake up the forker thread which
103 * will create and run it.
105 wake_up_process(default_backing_dev_info.wb.task);
109 static void bdi_queue_work(struct backing_dev_info *bdi,
110 struct wb_writeback_work *work)
112 trace_writeback_queue(bdi, work);
114 spin_lock_bh(&bdi->wb_lock);
115 list_add_tail(&work->list, &bdi->work_list);
117 trace_writeback_nothread(bdi, work);
118 bdi_wakeup_flusher(bdi);
119 spin_unlock_bh(&bdi->wb_lock);
123 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
124 bool range_cyclic, enum wb_reason reason)
126 struct wb_writeback_work *work;
129 * This is WB_SYNC_NONE writeback, so if allocation fails just
130 * wakeup the thread for old dirty data writeback
132 work = kzalloc(sizeof(*work), GFP_ATOMIC);
135 trace_writeback_nowork(bdi);
136 wake_up_process(bdi->wb.task);
141 work->sync_mode = WB_SYNC_NONE;
142 work->nr_pages = nr_pages;
143 work->range_cyclic = range_cyclic;
144 work->reason = reason;
146 bdi_queue_work(bdi, work);
150 * bdi_start_writeback - start writeback
151 * @bdi: the backing device to write from
152 * @nr_pages: the number of pages to write
153 * @reason: reason why some writeback work was initiated
156 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
157 * started when this function returns, we make no guarantees on
158 * completion. Caller need not hold sb s_umount semaphore.
161 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
162 enum wb_reason reason)
164 __bdi_start_writeback(bdi, nr_pages, true, reason);
168 * bdi_start_background_writeback - start background writeback
169 * @bdi: the backing device to write from
172 * This makes sure WB_SYNC_NONE background writeback happens. When
173 * this function returns, it is only guaranteed that for given BDI
174 * some IO is happening if we are over background dirty threshold.
175 * Caller need not hold sb s_umount semaphore.
177 void bdi_start_background_writeback(struct backing_dev_info *bdi)
180 * We just wake up the flusher thread. It will perform background
181 * writeback as soon as there is no other work to do.
183 trace_writeback_wake_background(bdi);
184 spin_lock_bh(&bdi->wb_lock);
185 bdi_wakeup_flusher(bdi);
186 spin_unlock_bh(&bdi->wb_lock);
190 * Remove the inode from the writeback list it is on.
192 void inode_wb_list_del(struct inode *inode)
194 struct backing_dev_info *bdi = inode_to_bdi(inode);
196 spin_lock(&bdi->wb.list_lock);
197 list_del_init(&inode->i_wb_list);
198 spin_unlock(&bdi->wb.list_lock);
202 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
203 * furthest end of its superblock's dirty-inode list.
205 * Before stamping the inode's ->dirtied_when, we check to see whether it is
206 * already the most-recently-dirtied inode on the b_dirty list. If that is
207 * the case then the inode must have been redirtied while it was being written
208 * out and we don't reset its dirtied_when.
210 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
212 assert_spin_locked(&wb->list_lock);
213 if (!list_empty(&wb->b_dirty)) {
216 tail = wb_inode(wb->b_dirty.next);
217 if (time_before(inode->dirtied_when, tail->dirtied_when))
218 inode->dirtied_when = jiffies;
220 list_move(&inode->i_wb_list, &wb->b_dirty);
224 * requeue inode for re-scanning after bdi->b_io list is exhausted.
226 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
228 assert_spin_locked(&wb->list_lock);
229 list_move(&inode->i_wb_list, &wb->b_more_io);
232 static void inode_sync_complete(struct inode *inode)
234 inode->i_state &= ~I_SYNC;
235 /* Waiters must see I_SYNC cleared before being woken up */
237 wake_up_bit(&inode->i_state, __I_SYNC);
240 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
242 bool ret = time_after(inode->dirtied_when, t);
245 * For inodes being constantly redirtied, dirtied_when can get stuck.
246 * It _appears_ to be in the future, but is actually in distant past.
247 * This test is necessary to prevent such wrapped-around relative times
248 * from permanently stopping the whole bdi writeback.
250 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
256 * Move expired (dirtied after work->older_than_this) dirty inodes from
257 * @delaying_queue to @dispatch_queue.
259 static int move_expired_inodes(struct list_head *delaying_queue,
260 struct list_head *dispatch_queue,
261 struct wb_writeback_work *work)
264 struct list_head *pos, *node;
265 struct super_block *sb = NULL;
270 while (!list_empty(delaying_queue)) {
271 inode = wb_inode(delaying_queue->prev);
272 if (work->older_than_this &&
273 inode_dirtied_after(inode, *work->older_than_this))
275 if (sb && sb != inode->i_sb)
278 list_move(&inode->i_wb_list, &tmp);
282 /* just one sb in list, splice to dispatch_queue and we're done */
284 list_splice(&tmp, dispatch_queue);
288 /* Move inodes from one superblock together */
289 while (!list_empty(&tmp)) {
290 sb = wb_inode(tmp.prev)->i_sb;
291 list_for_each_prev_safe(pos, node, &tmp) {
292 inode = wb_inode(pos);
293 if (inode->i_sb == sb)
294 list_move(&inode->i_wb_list, dispatch_queue);
302 * Queue all expired dirty inodes for io, eldest first.
304 * newly dirtied b_dirty b_io b_more_io
305 * =============> gf edc BA
307 * newly dirtied b_dirty b_io b_more_io
308 * =============> g fBAedc
310 * +--> dequeue for IO
312 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
315 assert_spin_locked(&wb->list_lock);
316 list_splice_init(&wb->b_more_io, &wb->b_io);
317 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
318 trace_writeback_queue_io(wb, work, moved);
321 static int write_inode(struct inode *inode, struct writeback_control *wbc)
323 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
324 return inode->i_sb->s_op->write_inode(inode, wbc);
329 * Wait for writeback on an inode to complete.
331 static void inode_wait_for_writeback(struct inode *inode)
333 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
334 wait_queue_head_t *wqh;
336 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
337 while (inode->i_state & I_SYNC) {
338 spin_unlock(&inode->i_lock);
339 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
340 spin_lock(&inode->i_lock);
345 * Find proper writeback list for the inode depending on its current state and
346 * possibly also change of its state while we were doing writeback. Here we
347 * handle things such as livelock prevention or fairness of writeback among
348 * inodes. This function can be called only by flusher thread - noone else
349 * processes all inodes in writeback lists and requeueing inodes behind flusher
350 * thread's back can have unexpected consequences.
352 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
353 struct writeback_control *wbc)
355 if (inode->i_state & I_FREEING)
359 * Sync livelock prevention. Each inode is tagged and synced in one
360 * shot. If still dirty, it will be redirty_tail()'ed below. Update
361 * the dirty time to prevent enqueue and sync it again.
363 if ((inode->i_state & I_DIRTY) &&
364 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
365 inode->dirtied_when = jiffies;
367 if (wbc->pages_skipped) {
369 * writeback is not making progress due to locked
370 * buffers. Skip this inode for now.
372 redirty_tail(inode, wb);
376 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
378 * We didn't write back all the pages. nfs_writepages()
379 * sometimes bales out without doing anything.
381 if (wbc->nr_to_write <= 0) {
382 /* Slice used up. Queue for next turn. */
383 requeue_io(inode, wb);
386 * Writeback blocked by something other than
387 * congestion. Delay the inode for some time to
388 * avoid spinning on the CPU (100% iowait)
389 * retrying writeback of the dirty page/inode
390 * that cannot be performed immediately.
392 redirty_tail(inode, wb);
394 } else if (inode->i_state & I_DIRTY) {
396 * Filesystems can dirty the inode during writeback operations,
397 * such as delayed allocation during submission or metadata
398 * updates after data IO completion.
400 redirty_tail(inode, wb);
402 /* The inode is clean. Remove from writeback lists. */
403 list_del_init(&inode->i_wb_list);
408 * Write out an inode and its dirty pages. Do not update the writeback list
409 * linkage. That is left to the caller. The caller is also responsible for
410 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
413 __writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
414 struct writeback_control *wbc)
416 struct address_space *mapping = inode->i_mapping;
417 long nr_to_write = wbc->nr_to_write;
421 WARN_ON(!(inode->i_state & I_SYNC));
423 ret = do_writepages(mapping, wbc);
426 * Make sure to wait on the data before writing out the metadata.
427 * This is important for filesystems that modify metadata on data
430 if (wbc->sync_mode == WB_SYNC_ALL) {
431 int err = filemap_fdatawait(mapping);
437 * Some filesystems may redirty the inode during the writeback
438 * due to delalloc, clear dirty metadata flags right before
441 spin_lock(&inode->i_lock);
442 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
443 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
444 inode->i_state &= ~I_DIRTY_PAGES;
445 dirty = inode->i_state & I_DIRTY;
446 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
447 spin_unlock(&inode->i_lock);
448 /* Don't write the inode if only I_DIRTY_PAGES was set */
449 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
450 int err = write_inode(inode, wbc);
454 trace_writeback_single_inode(inode, wbc, nr_to_write);
459 * Write out an inode's dirty pages. Either the caller has an active reference
460 * on the inode or the inode has I_WILL_FREE set.
462 * This function is designed to be called for writing back one inode which
463 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
464 * and does more profound writeback list handling in writeback_sb_inodes().
467 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
468 struct writeback_control *wbc)
472 spin_lock(&inode->i_lock);
473 if (!atomic_read(&inode->i_count))
474 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
476 WARN_ON(inode->i_state & I_WILL_FREE);
478 if (inode->i_state & I_SYNC) {
479 if (wbc->sync_mode != WB_SYNC_ALL)
482 * It's a data-integrity sync. We must wait.
484 inode_wait_for_writeback(inode);
486 WARN_ON(inode->i_state & I_SYNC);
488 * Skip inode if it is clean. We don't want to mess with writeback
489 * lists in this function since flusher thread may be doing for example
490 * sync in parallel and if we move the inode, it could get skipped. So
491 * here we make sure inode is on some writeback list and leave it there
492 * unless we have completely cleaned the inode.
494 if (!(inode->i_state & I_DIRTY))
496 inode->i_state |= I_SYNC;
497 spin_unlock(&inode->i_lock);
499 ret = __writeback_single_inode(inode, wb, wbc);
501 spin_lock(&wb->list_lock);
502 spin_lock(&inode->i_lock);
504 * If inode is clean, remove it from writeback lists. Otherwise don't
505 * touch it. See comment above for explanation.
507 if (!(inode->i_state & I_DIRTY))
508 list_del_init(&inode->i_wb_list);
509 spin_unlock(&wb->list_lock);
510 inode_sync_complete(inode);
512 spin_unlock(&inode->i_lock);
516 static long writeback_chunk_size(struct backing_dev_info *bdi,
517 struct wb_writeback_work *work)
522 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
523 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
524 * here avoids calling into writeback_inodes_wb() more than once.
526 * The intended call sequence for WB_SYNC_ALL writeback is:
529 * writeback_sb_inodes() <== called only once
530 * write_cache_pages() <== called once for each inode
531 * (quickly) tag currently dirty pages
532 * (maybe slowly) sync all tagged pages
534 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
537 pages = min(bdi->avg_write_bandwidth / 2,
538 global_dirty_limit / DIRTY_SCOPE);
539 pages = min(pages, work->nr_pages);
540 pages = round_down(pages + MIN_WRITEBACK_PAGES,
541 MIN_WRITEBACK_PAGES);
548 * Write a portion of b_io inodes which belong to @sb.
550 * If @only_this_sb is true, then find and write all such
551 * inodes. Otherwise write only ones which go sequentially
554 * Return the number of pages and/or inodes written.
556 static long writeback_sb_inodes(struct super_block *sb,
557 struct bdi_writeback *wb,
558 struct wb_writeback_work *work)
560 struct writeback_control wbc = {
561 .sync_mode = work->sync_mode,
562 .tagged_writepages = work->tagged_writepages,
563 .for_kupdate = work->for_kupdate,
564 .for_background = work->for_background,
565 .range_cyclic = work->range_cyclic,
567 .range_end = LLONG_MAX,
569 unsigned long start_time = jiffies;
571 long wrote = 0; /* count both pages and inodes */
573 while (!list_empty(&wb->b_io)) {
574 struct inode *inode = wb_inode(wb->b_io.prev);
576 if (inode->i_sb != sb) {
579 * We only want to write back data for this
580 * superblock, move all inodes not belonging
581 * to it back onto the dirty list.
583 redirty_tail(inode, wb);
588 * The inode belongs to a different superblock.
589 * Bounce back to the caller to unpin this and
590 * pin the next superblock.
596 * Don't bother with new inodes or inodes beeing freed, first
597 * kind does not need peridic writeout yet, and for the latter
598 * kind writeout is handled by the freer.
600 spin_lock(&inode->i_lock);
601 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
602 spin_unlock(&inode->i_lock);
603 redirty_tail(inode, wb);
606 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
608 * If this inode is locked for writeback and we are not
609 * doing writeback-for-data-integrity, move it to
610 * b_more_io so that writeback can proceed with the
611 * other inodes on s_io.
613 * We'll have another go at writing back this inode
614 * when we completed a full scan of b_io.
616 spin_unlock(&inode->i_lock);
617 requeue_io(inode, wb);
618 trace_writeback_sb_inodes_requeue(inode);
621 spin_unlock(&wb->list_lock);
625 * We already requeued the inode if it had I_SYNC set and we
626 * are doing WB_SYNC_NONE writeback. So this catches only the
629 if (inode->i_state & I_SYNC)
630 inode_wait_for_writeback(inode);
631 inode->i_state |= I_SYNC;
632 spin_unlock(&inode->i_lock);
633 write_chunk = writeback_chunk_size(wb->bdi, work);
634 wbc.nr_to_write = write_chunk;
635 wbc.pages_skipped = 0;
637 __writeback_single_inode(inode, wb, &wbc);
639 work->nr_pages -= write_chunk - wbc.nr_to_write;
640 wrote += write_chunk - wbc.nr_to_write;
641 spin_lock(&wb->list_lock);
642 spin_lock(&inode->i_lock);
643 if (!(inode->i_state & I_DIRTY))
645 requeue_inode(inode, wb, &wbc);
646 inode_sync_complete(inode);
647 spin_unlock(&inode->i_lock);
648 spin_unlock(&wb->list_lock);
651 spin_lock(&wb->list_lock);
653 * bail out to wb_writeback() often enough to check
654 * background threshold and other termination conditions.
657 if (time_is_before_jiffies(start_time + HZ / 10UL))
659 if (work->nr_pages <= 0)
666 static long __writeback_inodes_wb(struct bdi_writeback *wb,
667 struct wb_writeback_work *work)
669 unsigned long start_time = jiffies;
672 while (!list_empty(&wb->b_io)) {
673 struct inode *inode = wb_inode(wb->b_io.prev);
674 struct super_block *sb = inode->i_sb;
676 if (!grab_super_passive(sb)) {
678 * grab_super_passive() may fail consistently due to
679 * s_umount being grabbed by someone else. Don't use
680 * requeue_io() to avoid busy retrying the inode/sb.
682 redirty_tail(inode, wb);
685 wrote += writeback_sb_inodes(sb, wb, work);
688 /* refer to the same tests at the end of writeback_sb_inodes */
690 if (time_is_before_jiffies(start_time + HZ / 10UL))
692 if (work->nr_pages <= 0)
696 /* Leave any unwritten inodes on b_io */
700 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
701 enum wb_reason reason)
703 struct wb_writeback_work work = {
704 .nr_pages = nr_pages,
705 .sync_mode = WB_SYNC_NONE,
710 spin_lock(&wb->list_lock);
711 if (list_empty(&wb->b_io))
713 __writeback_inodes_wb(wb, &work);
714 spin_unlock(&wb->list_lock);
716 return nr_pages - work.nr_pages;
719 static bool over_bground_thresh(struct backing_dev_info *bdi)
721 unsigned long background_thresh, dirty_thresh;
723 global_dirty_limits(&background_thresh, &dirty_thresh);
725 if (global_page_state(NR_FILE_DIRTY) +
726 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
729 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
730 bdi_dirty_limit(bdi, background_thresh))
737 * Called under wb->list_lock. If there are multiple wb per bdi,
738 * only the flusher working on the first wb should do it.
740 static void wb_update_bandwidth(struct bdi_writeback *wb,
741 unsigned long start_time)
743 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
747 * Explicit flushing or periodic writeback of "old" data.
749 * Define "old": the first time one of an inode's pages is dirtied, we mark the
750 * dirtying-time in the inode's address_space. So this periodic writeback code
751 * just walks the superblock inode list, writing back any inodes which are
752 * older than a specific point in time.
754 * Try to run once per dirty_writeback_interval. But if a writeback event
755 * takes longer than a dirty_writeback_interval interval, then leave a
758 * older_than_this takes precedence over nr_to_write. So we'll only write back
759 * all dirty pages if they are all attached to "old" mappings.
761 static long wb_writeback(struct bdi_writeback *wb,
762 struct wb_writeback_work *work)
764 unsigned long wb_start = jiffies;
765 long nr_pages = work->nr_pages;
766 unsigned long oldest_jif;
770 oldest_jif = jiffies;
771 work->older_than_this = &oldest_jif;
773 spin_lock(&wb->list_lock);
776 * Stop writeback when nr_pages has been consumed
778 if (work->nr_pages <= 0)
782 * Background writeout and kupdate-style writeback may
783 * run forever. Stop them if there is other work to do
784 * so that e.g. sync can proceed. They'll be restarted
785 * after the other works are all done.
787 if ((work->for_background || work->for_kupdate) &&
788 !list_empty(&wb->bdi->work_list))
792 * For background writeout, stop when we are below the
793 * background dirty threshold
795 if (work->for_background && !over_bground_thresh(wb->bdi))
799 * Kupdate and background works are special and we want to
800 * include all inodes that need writing. Livelock avoidance is
801 * handled by these works yielding to any other work so we are
804 if (work->for_kupdate) {
805 oldest_jif = jiffies -
806 msecs_to_jiffies(dirty_expire_interval * 10);
807 } else if (work->for_background)
808 oldest_jif = jiffies;
810 trace_writeback_start(wb->bdi, work);
811 if (list_empty(&wb->b_io))
814 progress = writeback_sb_inodes(work->sb, wb, work);
816 progress = __writeback_inodes_wb(wb, work);
817 trace_writeback_written(wb->bdi, work);
819 wb_update_bandwidth(wb, wb_start);
822 * Did we write something? Try for more
824 * Dirty inodes are moved to b_io for writeback in batches.
825 * The completion of the current batch does not necessarily
826 * mean the overall work is done. So we keep looping as long
827 * as made some progress on cleaning pages or inodes.
832 * No more inodes for IO, bail
834 if (list_empty(&wb->b_more_io))
837 * Nothing written. Wait for some inode to
838 * become available for writeback. Otherwise
839 * we'll just busyloop.
841 if (!list_empty(&wb->b_more_io)) {
842 trace_writeback_wait(wb->bdi, work);
843 inode = wb_inode(wb->b_more_io.prev);
844 spin_lock(&inode->i_lock);
845 spin_unlock(&wb->list_lock);
846 inode_wait_for_writeback(inode);
847 spin_unlock(&inode->i_lock);
848 spin_lock(&wb->list_lock);
851 spin_unlock(&wb->list_lock);
853 return nr_pages - work->nr_pages;
857 * Return the next wb_writeback_work struct that hasn't been processed yet.
859 static struct wb_writeback_work *
860 get_next_work_item(struct backing_dev_info *bdi)
862 struct wb_writeback_work *work = NULL;
864 spin_lock_bh(&bdi->wb_lock);
865 if (!list_empty(&bdi->work_list)) {
866 work = list_entry(bdi->work_list.next,
867 struct wb_writeback_work, list);
868 list_del_init(&work->list);
870 spin_unlock_bh(&bdi->wb_lock);
875 * Add in the number of potentially dirty inodes, because each inode
876 * write can dirty pagecache in the underlying blockdev.
878 static unsigned long get_nr_dirty_pages(void)
880 return global_page_state(NR_FILE_DIRTY) +
881 global_page_state(NR_UNSTABLE_NFS) +
882 get_nr_dirty_inodes();
885 static long wb_check_background_flush(struct bdi_writeback *wb)
887 if (over_bground_thresh(wb->bdi)) {
889 struct wb_writeback_work work = {
890 .nr_pages = LONG_MAX,
891 .sync_mode = WB_SYNC_NONE,
894 .reason = WB_REASON_BACKGROUND,
897 return wb_writeback(wb, &work);
903 static long wb_check_old_data_flush(struct bdi_writeback *wb)
905 unsigned long expired;
909 * When set to zero, disable periodic writeback
911 if (!dirty_writeback_interval)
914 expired = wb->last_old_flush +
915 msecs_to_jiffies(dirty_writeback_interval * 10);
916 if (time_before(jiffies, expired))
919 wb->last_old_flush = jiffies;
920 nr_pages = get_nr_dirty_pages();
923 struct wb_writeback_work work = {
924 .nr_pages = nr_pages,
925 .sync_mode = WB_SYNC_NONE,
928 .reason = WB_REASON_PERIODIC,
931 return wb_writeback(wb, &work);
938 * Retrieve work items and do the writeback they describe
940 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
942 struct backing_dev_info *bdi = wb->bdi;
943 struct wb_writeback_work *work;
946 set_bit(BDI_writeback_running, &wb->bdi->state);
947 while ((work = get_next_work_item(bdi)) != NULL) {
949 * Override sync mode, in case we must wait for completion
950 * because this thread is exiting now.
953 work->sync_mode = WB_SYNC_ALL;
955 trace_writeback_exec(bdi, work);
957 wrote += wb_writeback(wb, work);
960 * Notify the caller of completion if this is a synchronous
961 * work item, otherwise just free it.
964 complete(work->done);
970 * Check for periodic writeback, kupdated() style
972 wrote += wb_check_old_data_flush(wb);
973 wrote += wb_check_background_flush(wb);
974 clear_bit(BDI_writeback_running, &wb->bdi->state);
980 * Handle writeback of dirty data for the device backed by this bdi. Also
981 * wakes up periodically and does kupdated style flushing.
983 int bdi_writeback_thread(void *data)
985 struct bdi_writeback *wb = data;
986 struct backing_dev_info *bdi = wb->bdi;
989 current->flags |= PF_SWAPWRITE;
991 wb->last_active = jiffies;
994 * Our parent may run at a different priority, just set us to normal
996 set_user_nice(current, 0);
998 trace_writeback_thread_start(bdi);
1000 while (!kthread_freezable_should_stop(NULL)) {
1002 * Remove own delayed wake-up timer, since we are already awake
1003 * and we'll take care of the preriodic write-back.
1005 del_timer(&wb->wakeup_timer);
1007 pages_written = wb_do_writeback(wb, 0);
1009 trace_writeback_pages_written(pages_written);
1012 wb->last_active = jiffies;
1014 set_current_state(TASK_INTERRUPTIBLE);
1015 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
1016 __set_current_state(TASK_RUNNING);
1020 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1021 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
1024 * We have nothing to do, so can go sleep without any
1025 * timeout and save power. When a work is queued or
1026 * something is made dirty - we will be woken up.
1032 /* Flush any work that raced with us exiting */
1033 if (!list_empty(&bdi->work_list))
1034 wb_do_writeback(wb, 1);
1036 trace_writeback_thread_stop(bdi);
1042 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1045 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1047 struct backing_dev_info *bdi;
1050 nr_pages = global_page_state(NR_FILE_DIRTY) +
1051 global_page_state(NR_UNSTABLE_NFS);
1055 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1056 if (!bdi_has_dirty_io(bdi))
1058 __bdi_start_writeback(bdi, nr_pages, false, reason);
1063 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1065 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1066 struct dentry *dentry;
1067 const char *name = "?";
1069 dentry = d_find_alias(inode);
1071 spin_lock(&dentry->d_lock);
1072 name = (const char *) dentry->d_name.name;
1075 "%s(%d): dirtied inode %lu (%s) on %s\n",
1076 current->comm, task_pid_nr(current), inode->i_ino,
1077 name, inode->i_sb->s_id);
1079 spin_unlock(&dentry->d_lock);
1086 * __mark_inode_dirty - internal function
1087 * @inode: inode to mark
1088 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1089 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1090 * mark_inode_dirty_sync.
1092 * Put the inode on the super block's dirty list.
1094 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1095 * dirty list only if it is hashed or if it refers to a blockdev.
1096 * If it was not hashed, it will never be added to the dirty list
1097 * even if it is later hashed, as it will have been marked dirty already.
1099 * In short, make sure you hash any inodes _before_ you start marking
1102 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1103 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1104 * the kernel-internal blockdev inode represents the dirtying time of the
1105 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1106 * page->mapping->host, so the page-dirtying time is recorded in the internal
1109 void __mark_inode_dirty(struct inode *inode, int flags)
1111 struct super_block *sb = inode->i_sb;
1112 struct backing_dev_info *bdi = NULL;
1115 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1116 * dirty the inode itself
1118 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1119 if (sb->s_op->dirty_inode)
1120 sb->s_op->dirty_inode(inode, flags);
1124 * make sure that changes are seen by all cpus before we test i_state
1129 /* avoid the locking if we can */
1130 if ((inode->i_state & flags) == flags)
1133 if (unlikely(block_dump))
1134 block_dump___mark_inode_dirty(inode);
1136 spin_lock(&inode->i_lock);
1137 if ((inode->i_state & flags) != flags) {
1138 const int was_dirty = inode->i_state & I_DIRTY;
1140 inode->i_state |= flags;
1143 * If the inode is being synced, just update its dirty state.
1144 * The unlocker will place the inode on the appropriate
1145 * superblock list, based upon its state.
1147 if (inode->i_state & I_SYNC)
1148 goto out_unlock_inode;
1151 * Only add valid (hashed) inodes to the superblock's
1152 * dirty list. Add blockdev inodes as well.
1154 if (!S_ISBLK(inode->i_mode)) {
1155 if (inode_unhashed(inode))
1156 goto out_unlock_inode;
1158 if (inode->i_state & I_FREEING)
1159 goto out_unlock_inode;
1162 * If the inode was already on b_dirty/b_io/b_more_io, don't
1163 * reposition it (that would break b_dirty time-ordering).
1166 bool wakeup_bdi = false;
1167 bdi = inode_to_bdi(inode);
1169 if (bdi_cap_writeback_dirty(bdi)) {
1170 WARN(!test_bit(BDI_registered, &bdi->state),
1171 "bdi-%s not registered\n", bdi->name);
1174 * If this is the first dirty inode for this
1175 * bdi, we have to wake-up the corresponding
1176 * bdi thread to make sure background
1177 * write-back happens later.
1179 if (!wb_has_dirty_io(&bdi->wb))
1183 spin_unlock(&inode->i_lock);
1184 spin_lock(&bdi->wb.list_lock);
1185 inode->dirtied_when = jiffies;
1186 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1187 spin_unlock(&bdi->wb.list_lock);
1190 bdi_wakeup_thread_delayed(bdi);
1195 spin_unlock(&inode->i_lock);
1198 EXPORT_SYMBOL(__mark_inode_dirty);
1200 static void wait_sb_inodes(struct super_block *sb)
1202 struct inode *inode, *old_inode = NULL;
1205 * We need to be protected against the filesystem going from
1206 * r/o to r/w or vice versa.
1208 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1210 spin_lock(&inode_sb_list_lock);
1213 * Data integrity sync. Must wait for all pages under writeback,
1214 * because there may have been pages dirtied before our sync
1215 * call, but which had writeout started before we write it out.
1216 * In which case, the inode may not be on the dirty list, but
1217 * we still have to wait for that writeout.
1219 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1220 struct address_space *mapping = inode->i_mapping;
1222 spin_lock(&inode->i_lock);
1223 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1224 (mapping->nrpages == 0)) {
1225 spin_unlock(&inode->i_lock);
1229 spin_unlock(&inode->i_lock);
1230 spin_unlock(&inode_sb_list_lock);
1233 * We hold a reference to 'inode' so it couldn't have been
1234 * removed from s_inodes list while we dropped the
1235 * inode_sb_list_lock. We cannot iput the inode now as we can
1236 * be holding the last reference and we cannot iput it under
1237 * inode_sb_list_lock. So we keep the reference and iput it
1243 filemap_fdatawait(mapping);
1247 spin_lock(&inode_sb_list_lock);
1249 spin_unlock(&inode_sb_list_lock);
1254 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1255 * @sb: the superblock
1256 * @nr: the number of pages to write
1257 * @reason: reason why some writeback work initiated
1259 * Start writeback on some inodes on this super_block. No guarantees are made
1260 * on how many (if any) will be written, and this function does not wait
1261 * for IO completion of submitted IO.
1263 void writeback_inodes_sb_nr(struct super_block *sb,
1265 enum wb_reason reason)
1267 DECLARE_COMPLETION_ONSTACK(done);
1268 struct wb_writeback_work work = {
1270 .sync_mode = WB_SYNC_NONE,
1271 .tagged_writepages = 1,
1277 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1278 bdi_queue_work(sb->s_bdi, &work);
1279 wait_for_completion(&done);
1281 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1284 * writeback_inodes_sb - writeback dirty inodes from given super_block
1285 * @sb: the superblock
1286 * @reason: reason why some writeback work was initiated
1288 * Start writeback on some inodes on this super_block. No guarantees are made
1289 * on how many (if any) will be written, and this function does not wait
1290 * for IO completion of submitted IO.
1292 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1294 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1296 EXPORT_SYMBOL(writeback_inodes_sb);
1299 * writeback_inodes_sb_if_idle - start writeback if none underway
1300 * @sb: the superblock
1301 * @reason: reason why some writeback work was initiated
1303 * Invoke writeback_inodes_sb if no writeback is currently underway.
1304 * Returns 1 if writeback was started, 0 if not.
1306 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1308 if (!writeback_in_progress(sb->s_bdi)) {
1309 down_read(&sb->s_umount);
1310 writeback_inodes_sb(sb, reason);
1311 up_read(&sb->s_umount);
1316 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1319 * writeback_inodes_sb_nr_if_idle - start writeback if none underway
1320 * @sb: the superblock
1321 * @nr: the number of pages to write
1322 * @reason: reason why some writeback work was initiated
1324 * Invoke writeback_inodes_sb if no writeback is currently underway.
1325 * Returns 1 if writeback was started, 0 if not.
1327 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1329 enum wb_reason reason)
1331 if (!writeback_in_progress(sb->s_bdi)) {
1332 down_read(&sb->s_umount);
1333 writeback_inodes_sb_nr(sb, nr, reason);
1334 up_read(&sb->s_umount);
1339 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1342 * sync_inodes_sb - sync sb inode pages
1343 * @sb: the superblock
1345 * This function writes and waits on any dirty inode belonging to this
1348 void sync_inodes_sb(struct super_block *sb)
1350 DECLARE_COMPLETION_ONSTACK(done);
1351 struct wb_writeback_work work = {
1353 .sync_mode = WB_SYNC_ALL,
1354 .nr_pages = LONG_MAX,
1357 .reason = WB_REASON_SYNC,
1360 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1362 bdi_queue_work(sb->s_bdi, &work);
1363 wait_for_completion(&done);
1367 EXPORT_SYMBOL(sync_inodes_sb);
1370 * write_inode_now - write an inode to disk
1371 * @inode: inode to write to disk
1372 * @sync: whether the write should be synchronous or not
1374 * This function commits an inode to disk immediately if it is dirty. This is
1375 * primarily needed by knfsd.
1377 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1379 int write_inode_now(struct inode *inode, int sync)
1381 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1382 struct writeback_control wbc = {
1383 .nr_to_write = LONG_MAX,
1384 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1386 .range_end = LLONG_MAX,
1389 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1390 wbc.nr_to_write = 0;
1393 return writeback_single_inode(inode, wb, &wbc);
1395 EXPORT_SYMBOL(write_inode_now);
1398 * sync_inode - write an inode and its pages to disk.
1399 * @inode: the inode to sync
1400 * @wbc: controls the writeback mode
1402 * sync_inode() will write an inode and its pages to disk. It will also
1403 * correctly update the inode on its superblock's dirty inode lists and will
1404 * update inode->i_state.
1406 * The caller must have a ref on the inode.
1408 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1410 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
1412 EXPORT_SYMBOL(sync_inode);
1415 * sync_inode_metadata - write an inode to disk
1416 * @inode: the inode to sync
1417 * @wait: wait for I/O to complete.
1419 * Write an inode to disk and adjust its dirty state after completion.
1421 * Note: only writes the actual inode, no associated data or other metadata.
1423 int sync_inode_metadata(struct inode *inode, int wait)
1425 struct writeback_control wbc = {
1426 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1427 .nr_to_write = 0, /* metadata-only */
1430 return sync_inode(inode, &wbc);
1432 EXPORT_SYMBOL(sync_inode_metadata);