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/module.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
23 #include <linux/kthread.h>
24 #include <linux/freezer.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/buffer_head.h>
29 #include <linux/tracepoint.h>
33 * Passed into wb_writeback(), essentially a subset of writeback_control
35 struct wb_writeback_work {
37 struct super_block *sb;
38 unsigned long *older_than_this;
39 enum writeback_sync_modes sync_mode;
40 unsigned int tagged_writepages:1;
41 unsigned int for_kupdate:1;
42 unsigned int range_cyclic:1;
43 unsigned int for_background:1;
44 enum wb_reason reason; /* why was writeback initiated? */
46 struct list_head list; /* pending work list */
47 struct completion *done; /* set if the caller waits */
51 * We don't actually have pdflush, but this one is exported though /proc...
53 int nr_pdflush_threads;
56 * writeback_in_progress - determine whether there is writeback in progress
57 * @bdi: the device's backing_dev_info structure.
59 * Determine whether there is writeback waiting to be handled against a
62 int writeback_in_progress(struct backing_dev_info *bdi)
64 return test_bit(BDI_writeback_running, &bdi->state);
66 EXPORT_SYMBOL(writeback_in_progress);
68 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
70 struct super_block *sb = inode->i_sb;
72 if (strcmp(sb->s_type->name, "bdev") == 0)
73 return inode->i_mapping->backing_dev_info;
78 static inline struct inode *wb_inode(struct list_head *head)
80 return list_entry(head, struct inode, i_wb_list);
84 * Include the creation of the trace points after defining the
85 * wb_writeback_work structure and inline functions so that the definition
86 * remains local to this file.
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/writeback.h>
91 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
92 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
95 wake_up_process(bdi->wb.task);
98 * The bdi thread isn't there, wake up the forker thread which
99 * will create and run it.
101 wake_up_process(default_backing_dev_info.wb.task);
105 static void bdi_queue_work(struct backing_dev_info *bdi,
106 struct wb_writeback_work *work)
108 trace_writeback_queue(bdi, work);
110 spin_lock_bh(&bdi->wb_lock);
111 list_add_tail(&work->list, &bdi->work_list);
113 trace_writeback_nothread(bdi, work);
114 bdi_wakeup_flusher(bdi);
115 spin_unlock_bh(&bdi->wb_lock);
119 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
120 bool range_cyclic, enum wb_reason reason)
122 struct wb_writeback_work *work;
125 * This is WB_SYNC_NONE writeback, so if allocation fails just
126 * wakeup the thread for old dirty data writeback
128 work = kzalloc(sizeof(*work), GFP_ATOMIC);
131 trace_writeback_nowork(bdi);
132 wake_up_process(bdi->wb.task);
137 work->sync_mode = WB_SYNC_NONE;
138 work->nr_pages = nr_pages;
139 work->range_cyclic = range_cyclic;
140 work->reason = reason;
142 bdi_queue_work(bdi, work);
146 * bdi_start_writeback - start writeback
147 * @bdi: the backing device to write from
148 * @nr_pages: the number of pages to write
149 * @reason: reason why some writeback work was initiated
152 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
153 * started when this function returns, we make no guarantees on
154 * completion. Caller need not hold sb s_umount semaphore.
157 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
158 enum wb_reason reason)
160 __bdi_start_writeback(bdi, nr_pages, true, reason);
164 * bdi_start_background_writeback - start background writeback
165 * @bdi: the backing device to write from
168 * This makes sure WB_SYNC_NONE background writeback happens. When
169 * this function returns, it is only guaranteed that for given BDI
170 * some IO is happening if we are over background dirty threshold.
171 * Caller need not hold sb s_umount semaphore.
173 void bdi_start_background_writeback(struct backing_dev_info *bdi)
176 * We just wake up the flusher thread. It will perform background
177 * writeback as soon as there is no other work to do.
179 trace_writeback_wake_background(bdi);
180 spin_lock_bh(&bdi->wb_lock);
181 bdi_wakeup_flusher(bdi);
182 spin_unlock_bh(&bdi->wb_lock);
186 * Remove the inode from the writeback list it is on.
188 void inode_wb_list_del(struct inode *inode)
190 struct backing_dev_info *bdi = inode_to_bdi(inode);
192 spin_lock(&bdi->wb.list_lock);
193 list_del_init(&inode->i_wb_list);
194 spin_unlock(&bdi->wb.list_lock);
198 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
199 * furthest end of its superblock's dirty-inode list.
201 * Before stamping the inode's ->dirtied_when, we check to see whether it is
202 * already the most-recently-dirtied inode on the b_dirty list. If that is
203 * the case then the inode must have been redirtied while it was being written
204 * out and we don't reset its dirtied_when.
206 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
208 assert_spin_locked(&wb->list_lock);
209 if (!list_empty(&wb->b_dirty)) {
212 tail = wb_inode(wb->b_dirty.next);
213 if (time_before(inode->dirtied_when, tail->dirtied_when))
214 inode->dirtied_when = jiffies;
216 list_move(&inode->i_wb_list, &wb->b_dirty);
220 * requeue inode for re-scanning after bdi->b_io list is exhausted.
222 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
224 assert_spin_locked(&wb->list_lock);
225 list_move(&inode->i_wb_list, &wb->b_more_io);
228 static void inode_sync_complete(struct inode *inode)
231 * Prevent speculative execution through
232 * spin_unlock(&wb->list_lock);
236 wake_up_bit(&inode->i_state, __I_SYNC);
239 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
241 bool ret = time_after(inode->dirtied_when, t);
244 * For inodes being constantly redirtied, dirtied_when can get stuck.
245 * It _appears_ to be in the future, but is actually in distant past.
246 * This test is necessary to prevent such wrapped-around relative times
247 * from permanently stopping the whole bdi writeback.
249 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
255 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
257 static int move_expired_inodes(struct list_head *delaying_queue,
258 struct list_head *dispatch_queue,
259 struct wb_writeback_work *work)
262 struct list_head *pos, *node;
263 struct super_block *sb = NULL;
268 while (!list_empty(delaying_queue)) {
269 inode = wb_inode(delaying_queue->prev);
270 if (work->older_than_this &&
271 inode_dirtied_after(inode, *work->older_than_this))
273 if (sb && sb != inode->i_sb)
276 list_move(&inode->i_wb_list, &tmp);
280 /* just one sb in list, splice to dispatch_queue and we're done */
282 list_splice(&tmp, dispatch_queue);
286 /* Move inodes from one superblock together */
287 while (!list_empty(&tmp)) {
288 sb = wb_inode(tmp.prev)->i_sb;
289 list_for_each_prev_safe(pos, node, &tmp) {
290 inode = wb_inode(pos);
291 if (inode->i_sb == sb)
292 list_move(&inode->i_wb_list, dispatch_queue);
300 * Queue all expired dirty inodes for io, eldest first.
302 * newly dirtied b_dirty b_io b_more_io
303 * =============> gf edc BA
305 * newly dirtied b_dirty b_io b_more_io
306 * =============> g fBAedc
308 * +--> dequeue for IO
310 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
313 assert_spin_locked(&wb->list_lock);
314 list_splice_init(&wb->b_more_io, &wb->b_io);
315 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
316 trace_writeback_queue_io(wb, work, moved);
319 static int write_inode(struct inode *inode, struct writeback_control *wbc)
321 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
322 return inode->i_sb->s_op->write_inode(inode, wbc);
327 * Wait for writeback on an inode to complete.
329 static void inode_wait_for_writeback(struct inode *inode,
330 struct bdi_writeback *wb)
332 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
333 wait_queue_head_t *wqh;
335 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
336 while (inode->i_state & I_SYNC) {
337 spin_unlock(&inode->i_lock);
338 spin_unlock(&wb->list_lock);
339 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
340 spin_lock(&wb->list_lock);
341 spin_lock(&inode->i_lock);
346 * Write out an inode's dirty pages. Called under wb->list_lock and
347 * inode->i_lock. Either the caller has an active reference on the inode or
348 * the inode has I_WILL_FREE set.
350 * If `wait' is set, wait on the writeout.
352 * The whole writeout design is quite complex and fragile. We want to avoid
353 * starvation of particular inodes when others are being redirtied, prevent
357 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
358 struct writeback_control *wbc)
360 struct address_space *mapping = inode->i_mapping;
361 long nr_to_write = wbc->nr_to_write;
365 assert_spin_locked(&wb->list_lock);
366 assert_spin_locked(&inode->i_lock);
368 if (!atomic_read(&inode->i_count))
369 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
371 WARN_ON(inode->i_state & I_WILL_FREE);
373 if (inode->i_state & I_SYNC) {
375 * If this inode is locked for writeback and we are not doing
376 * writeback-for-data-integrity, move it to b_more_io so that
377 * writeback can proceed with the other inodes on s_io.
379 * We'll have another go at writing back this inode when we
380 * completed a full scan of b_io.
382 if (wbc->sync_mode != WB_SYNC_ALL) {
383 requeue_io(inode, wb);
384 trace_writeback_single_inode_requeue(inode, wbc,
390 * It's a data-integrity sync. We must wait.
392 inode_wait_for_writeback(inode, wb);
395 BUG_ON(inode->i_state & I_SYNC);
397 /* Set I_SYNC, reset I_DIRTY_PAGES */
398 inode->i_state |= I_SYNC;
399 spin_unlock(&inode->i_lock);
400 spin_unlock(&wb->list_lock);
402 ret = do_writepages(mapping, wbc);
405 * Make sure to wait on the data before writing out the metadata.
406 * This is important for filesystems that modify metadata on data
409 if (wbc->sync_mode == WB_SYNC_ALL) {
410 int err = filemap_fdatawait(mapping);
416 * Some filesystems may redirty the inode during the writeback
417 * due to delalloc, clear dirty metadata flags right before
420 spin_lock(&inode->i_lock);
422 dirty = inode->i_state & I_DIRTY;
423 inode->i_state &= ~I_DIRTY;
426 * Paired with smp_mb() in __mark_inode_dirty(). This allows
427 * __mark_inode_dirty() to test i_state without grabbing i_lock -
428 * either they see the I_DIRTY bits cleared or we see the dirtied
431 * I_DIRTY_PAGES is always cleared together above even if @mapping
432 * still has dirty pages. The flag is reinstated after smp_mb() if
433 * necessary. This guarantees that either __mark_inode_dirty()
434 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
438 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
439 inode->i_state |= I_DIRTY_PAGES;
441 spin_unlock(&inode->i_lock);
443 /* Don't write the inode if only I_DIRTY_PAGES was set */
444 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
445 int err = write_inode(inode, wbc);
450 spin_lock(&wb->list_lock);
451 spin_lock(&inode->i_lock);
452 inode->i_state &= ~I_SYNC;
453 if (!(inode->i_state & I_FREEING)) {
455 * Sync livelock prevention. Each inode is tagged and synced in
456 * one shot. If still dirty, it will be redirty_tail()'ed below.
457 * Update the dirty time to prevent enqueue and sync it again.
459 if ((inode->i_state & I_DIRTY) &&
460 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
461 inode->dirtied_when = jiffies;
463 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
465 * We didn't write back all the pages. nfs_writepages()
466 * sometimes bales out without doing anything.
468 if (wbc->nr_to_write <= 0) {
470 * slice used up: queue for next turn
472 requeue_io(inode, wb);
475 * Writeback blocked by something other than
476 * congestion. Delay the inode for some time to
477 * avoid spinning on the CPU (100% iowait)
478 * retrying writeback of the dirty page/inode
479 * that cannot be performed immediately.
481 redirty_tail(inode, wb);
483 } else if (inode->i_state & I_DIRTY) {
485 * Filesystems can dirty the inode during writeback
486 * operations, such as delayed allocation during
487 * submission or metadata updates after data IO
490 redirty_tail(inode, wb);
493 * The inode is clean. At this point we either have
494 * a reference to the inode or it's on it's way out.
495 * No need to add it back to the LRU.
497 list_del_init(&inode->i_wb_list);
500 inode_sync_complete(inode);
501 trace_writeback_single_inode(inode, wbc, nr_to_write);
505 static long writeback_chunk_size(struct backing_dev_info *bdi,
506 struct wb_writeback_work *work)
511 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
512 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
513 * here avoids calling into writeback_inodes_wb() more than once.
515 * The intended call sequence for WB_SYNC_ALL writeback is:
518 * writeback_sb_inodes() <== called only once
519 * write_cache_pages() <== called once for each inode
520 * (quickly) tag currently dirty pages
521 * (maybe slowly) sync all tagged pages
523 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
526 pages = min(bdi->avg_write_bandwidth / 2,
527 global_dirty_limit / DIRTY_SCOPE);
528 pages = min(pages, work->nr_pages);
529 pages = round_down(pages + MIN_WRITEBACK_PAGES,
530 MIN_WRITEBACK_PAGES);
537 * Write a portion of b_io inodes which belong to @sb.
539 * If @only_this_sb is true, then find and write all such
540 * inodes. Otherwise write only ones which go sequentially
543 * Return the number of pages and/or inodes written.
545 static long writeback_sb_inodes(struct super_block *sb,
546 struct bdi_writeback *wb,
547 struct wb_writeback_work *work)
549 struct writeback_control wbc = {
550 .sync_mode = work->sync_mode,
551 .tagged_writepages = work->tagged_writepages,
552 .for_kupdate = work->for_kupdate,
553 .for_background = work->for_background,
554 .range_cyclic = work->range_cyclic,
556 .range_end = LLONG_MAX,
558 unsigned long start_time = jiffies;
560 long wrote = 0; /* count both pages and inodes */
562 while (!list_empty(&wb->b_io)) {
563 struct inode *inode = wb_inode(wb->b_io.prev);
565 if (inode->i_sb != sb) {
568 * We only want to write back data for this
569 * superblock, move all inodes not belonging
570 * to it back onto the dirty list.
572 redirty_tail(inode, wb);
577 * The inode belongs to a different superblock.
578 * Bounce back to the caller to unpin this and
579 * pin the next superblock.
585 * Don't bother with new inodes or inodes beeing freed, first
586 * kind does not need peridic writeout yet, and for the latter
587 * kind writeout is handled by the freer.
589 spin_lock(&inode->i_lock);
590 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
591 spin_unlock(&inode->i_lock);
592 redirty_tail(inode, wb);
596 write_chunk = writeback_chunk_size(wb->bdi, work);
597 wbc.nr_to_write = write_chunk;
598 wbc.pages_skipped = 0;
600 writeback_single_inode(inode, wb, &wbc);
602 work->nr_pages -= write_chunk - wbc.nr_to_write;
603 wrote += write_chunk - wbc.nr_to_write;
604 if (!(inode->i_state & I_DIRTY))
606 if (wbc.pages_skipped) {
608 * writeback is not making progress due to locked
609 * buffers. Skip this inode for now.
611 redirty_tail(inode, wb);
613 spin_unlock(&inode->i_lock);
614 spin_unlock(&wb->list_lock);
617 spin_lock(&wb->list_lock);
619 * bail out to wb_writeback() often enough to check
620 * background threshold and other termination conditions.
623 if (time_is_before_jiffies(start_time + HZ / 10UL))
625 if (work->nr_pages <= 0)
632 static long __writeback_inodes_wb(struct bdi_writeback *wb,
633 struct wb_writeback_work *work)
635 unsigned long start_time = jiffies;
638 while (!list_empty(&wb->b_io)) {
639 struct inode *inode = wb_inode(wb->b_io.prev);
640 struct super_block *sb = inode->i_sb;
642 if (!grab_super_passive(sb)) {
644 * grab_super_passive() may fail consistently due to
645 * s_umount being grabbed by someone else. Don't use
646 * requeue_io() to avoid busy retrying the inode/sb.
648 redirty_tail(inode, wb);
651 wrote += writeback_sb_inodes(sb, wb, work);
654 /* refer to the same tests at the end of writeback_sb_inodes */
656 if (time_is_before_jiffies(start_time + HZ / 10UL))
658 if (work->nr_pages <= 0)
662 /* Leave any unwritten inodes on b_io */
666 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
667 enum wb_reason reason)
669 struct wb_writeback_work work = {
670 .nr_pages = nr_pages,
671 .sync_mode = WB_SYNC_NONE,
676 spin_lock(&wb->list_lock);
677 if (list_empty(&wb->b_io))
679 __writeback_inodes_wb(wb, &work);
680 spin_unlock(&wb->list_lock);
682 return nr_pages - work.nr_pages;
685 static bool over_bground_thresh(struct backing_dev_info *bdi)
687 unsigned long background_thresh, dirty_thresh;
689 global_dirty_limits(&background_thresh, &dirty_thresh);
691 if (global_page_state(NR_FILE_DIRTY) +
692 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
695 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
696 bdi_dirty_limit(bdi, background_thresh))
703 * Called under wb->list_lock. If there are multiple wb per bdi,
704 * only the flusher working on the first wb should do it.
706 static void wb_update_bandwidth(struct bdi_writeback *wb,
707 unsigned long start_time)
709 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
713 * Explicit flushing or periodic writeback of "old" data.
715 * Define "old": the first time one of an inode's pages is dirtied, we mark the
716 * dirtying-time in the inode's address_space. So this periodic writeback code
717 * just walks the superblock inode list, writing back any inodes which are
718 * older than a specific point in time.
720 * Try to run once per dirty_writeback_interval. But if a writeback event
721 * takes longer than a dirty_writeback_interval interval, then leave a
724 * older_than_this takes precedence over nr_to_write. So we'll only write back
725 * all dirty pages if they are all attached to "old" mappings.
727 static long wb_writeback(struct bdi_writeback *wb,
728 struct wb_writeback_work *work)
730 unsigned long wb_start = jiffies;
731 long nr_pages = work->nr_pages;
732 unsigned long oldest_jif;
736 oldest_jif = jiffies;
737 work->older_than_this = &oldest_jif;
739 spin_lock(&wb->list_lock);
742 * Stop writeback when nr_pages has been consumed
744 if (work->nr_pages <= 0)
748 * Background writeout and kupdate-style writeback may
749 * run forever. Stop them if there is other work to do
750 * so that e.g. sync can proceed. They'll be restarted
751 * after the other works are all done.
753 if ((work->for_background || work->for_kupdate) &&
754 !list_empty(&wb->bdi->work_list))
758 * For background writeout, stop when we are below the
759 * background dirty threshold
761 if (work->for_background && !over_bground_thresh(wb->bdi))
764 if (work->for_kupdate) {
765 oldest_jif = jiffies -
766 msecs_to_jiffies(dirty_expire_interval * 10);
767 work->older_than_this = &oldest_jif;
770 trace_writeback_start(wb->bdi, work);
771 if (list_empty(&wb->b_io))
774 progress = writeback_sb_inodes(work->sb, wb, work);
776 progress = __writeback_inodes_wb(wb, work);
777 trace_writeback_written(wb->bdi, work);
779 wb_update_bandwidth(wb, wb_start);
782 * Did we write something? Try for more
784 * Dirty inodes are moved to b_io for writeback in batches.
785 * The completion of the current batch does not necessarily
786 * mean the overall work is done. So we keep looping as long
787 * as made some progress on cleaning pages or inodes.
792 * No more inodes for IO, bail
794 if (list_empty(&wb->b_more_io))
797 * Nothing written. Wait for some inode to
798 * become available for writeback. Otherwise
799 * we'll just busyloop.
801 if (!list_empty(&wb->b_more_io)) {
802 trace_writeback_wait(wb->bdi, work);
803 inode = wb_inode(wb->b_more_io.prev);
804 spin_lock(&inode->i_lock);
805 inode_wait_for_writeback(inode, wb);
806 spin_unlock(&inode->i_lock);
809 spin_unlock(&wb->list_lock);
811 return nr_pages - work->nr_pages;
815 * Return the next wb_writeback_work struct that hasn't been processed yet.
817 static struct wb_writeback_work *
818 get_next_work_item(struct backing_dev_info *bdi)
820 struct wb_writeback_work *work = NULL;
822 spin_lock_bh(&bdi->wb_lock);
823 if (!list_empty(&bdi->work_list)) {
824 work = list_entry(bdi->work_list.next,
825 struct wb_writeback_work, list);
826 list_del_init(&work->list);
828 spin_unlock_bh(&bdi->wb_lock);
833 * Add in the number of potentially dirty inodes, because each inode
834 * write can dirty pagecache in the underlying blockdev.
836 static unsigned long get_nr_dirty_pages(void)
838 return global_page_state(NR_FILE_DIRTY) +
839 global_page_state(NR_UNSTABLE_NFS) +
840 get_nr_dirty_inodes();
843 static long wb_check_background_flush(struct bdi_writeback *wb)
845 if (over_bground_thresh(wb->bdi)) {
847 struct wb_writeback_work work = {
848 .nr_pages = LONG_MAX,
849 .sync_mode = WB_SYNC_NONE,
852 .reason = WB_REASON_BACKGROUND,
855 return wb_writeback(wb, &work);
861 static long wb_check_old_data_flush(struct bdi_writeback *wb)
863 unsigned long expired;
867 * When set to zero, disable periodic writeback
869 if (!dirty_writeback_interval)
872 expired = wb->last_old_flush +
873 msecs_to_jiffies(dirty_writeback_interval * 10);
874 if (time_before(jiffies, expired))
877 wb->last_old_flush = jiffies;
878 nr_pages = get_nr_dirty_pages();
881 struct wb_writeback_work work = {
882 .nr_pages = nr_pages,
883 .sync_mode = WB_SYNC_NONE,
886 .reason = WB_REASON_PERIODIC,
889 return wb_writeback(wb, &work);
896 * Retrieve work items and do the writeback they describe
898 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
900 struct backing_dev_info *bdi = wb->bdi;
901 struct wb_writeback_work *work;
904 set_bit(BDI_writeback_running, &wb->bdi->state);
905 while ((work = get_next_work_item(bdi)) != NULL) {
907 * Override sync mode, in case we must wait for completion
908 * because this thread is exiting now.
911 work->sync_mode = WB_SYNC_ALL;
913 trace_writeback_exec(bdi, work);
915 wrote += wb_writeback(wb, work);
918 * Notify the caller of completion if this is a synchronous
919 * work item, otherwise just free it.
922 complete(work->done);
928 * Check for periodic writeback, kupdated() style
930 wrote += wb_check_old_data_flush(wb);
931 wrote += wb_check_background_flush(wb);
932 clear_bit(BDI_writeback_running, &wb->bdi->state);
938 * Handle writeback of dirty data for the device backed by this bdi. Also
939 * wakes up periodically and does kupdated style flushing.
941 int bdi_writeback_thread(void *data)
943 struct bdi_writeback *wb = data;
944 struct backing_dev_info *bdi = wb->bdi;
947 current->flags |= PF_SWAPWRITE;
949 wb->last_active = jiffies;
952 * Our parent may run at a different priority, just set us to normal
954 set_user_nice(current, 0);
956 trace_writeback_thread_start(bdi);
958 while (!kthread_should_stop()) {
960 * Remove own delayed wake-up timer, since we are already awake
961 * and we'll take care of the preriodic write-back.
963 del_timer(&wb->wakeup_timer);
965 pages_written = wb_do_writeback(wb, 0);
967 trace_writeback_pages_written(pages_written);
970 wb->last_active = jiffies;
972 set_current_state(TASK_INTERRUPTIBLE);
973 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
974 __set_current_state(TASK_RUNNING);
978 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
979 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
982 * We have nothing to do, so can go sleep without any
983 * timeout and save power. When a work is queued or
984 * something is made dirty - we will be woken up.
992 /* Flush any work that raced with us exiting */
993 if (!list_empty(&bdi->work_list))
994 wb_do_writeback(wb, 1);
996 trace_writeback_thread_stop(bdi);
1002 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1005 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1007 struct backing_dev_info *bdi;
1010 nr_pages = global_page_state(NR_FILE_DIRTY) +
1011 global_page_state(NR_UNSTABLE_NFS);
1015 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1016 if (!bdi_has_dirty_io(bdi))
1018 __bdi_start_writeback(bdi, nr_pages, false, reason);
1023 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1025 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1026 struct dentry *dentry;
1027 const char *name = "?";
1029 dentry = d_find_alias(inode);
1031 spin_lock(&dentry->d_lock);
1032 name = (const char *) dentry->d_name.name;
1035 "%s(%d): dirtied inode %lu (%s) on %s\n",
1036 current->comm, task_pid_nr(current), inode->i_ino,
1037 name, inode->i_sb->s_id);
1039 spin_unlock(&dentry->d_lock);
1046 * __mark_inode_dirty - internal function
1047 * @inode: inode to mark
1048 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1049 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1050 * mark_inode_dirty_sync.
1052 * Put the inode on the super block's dirty list.
1054 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1055 * dirty list only if it is hashed or if it refers to a blockdev.
1056 * If it was not hashed, it will never be added to the dirty list
1057 * even if it is later hashed, as it will have been marked dirty already.
1059 * In short, make sure you hash any inodes _before_ you start marking
1062 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1063 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1064 * the kernel-internal blockdev inode represents the dirtying time of the
1065 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1066 * page->mapping->host, so the page-dirtying time is recorded in the internal
1069 void __mark_inode_dirty(struct inode *inode, int flags)
1071 struct super_block *sb = inode->i_sb;
1072 struct backing_dev_info *bdi = NULL;
1075 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1076 * dirty the inode itself
1078 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1079 if (sb->s_op->dirty_inode)
1080 sb->s_op->dirty_inode(inode, flags);
1084 * Paired with smp_mb() in __writeback_single_inode() for the
1085 * following lockless i_state test. See there for details.
1089 if ((inode->i_state & flags) == flags)
1092 if (unlikely(block_dump))
1093 block_dump___mark_inode_dirty(inode);
1095 spin_lock(&inode->i_lock);
1096 if ((inode->i_state & flags) != flags) {
1097 const int was_dirty = inode->i_state & I_DIRTY;
1099 inode->i_state |= flags;
1102 * If the inode is being synced, just update its dirty state.
1103 * The unlocker will place the inode on the appropriate
1104 * superblock list, based upon its state.
1106 if (inode->i_state & I_SYNC)
1107 goto out_unlock_inode;
1110 * Only add valid (hashed) inodes to the superblock's
1111 * dirty list. Add blockdev inodes as well.
1113 if (!S_ISBLK(inode->i_mode)) {
1114 if (inode_unhashed(inode))
1115 goto out_unlock_inode;
1117 if (inode->i_state & I_FREEING)
1118 goto out_unlock_inode;
1121 * If the inode was already on b_dirty/b_io/b_more_io, don't
1122 * reposition it (that would break b_dirty time-ordering).
1125 bool wakeup_bdi = false;
1126 bdi = inode_to_bdi(inode);
1128 if (bdi_cap_writeback_dirty(bdi)) {
1129 WARN(!test_bit(BDI_registered, &bdi->state),
1130 "bdi-%s not registered\n", bdi->name);
1133 * If this is the first dirty inode for this
1134 * bdi, we have to wake-up the corresponding
1135 * bdi thread to make sure background
1136 * write-back happens later.
1138 if (!wb_has_dirty_io(&bdi->wb))
1142 spin_unlock(&inode->i_lock);
1143 spin_lock(&bdi->wb.list_lock);
1144 inode->dirtied_when = jiffies;
1145 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1146 spin_unlock(&bdi->wb.list_lock);
1149 bdi_wakeup_thread_delayed(bdi);
1154 spin_unlock(&inode->i_lock);
1157 EXPORT_SYMBOL(__mark_inode_dirty);
1160 * Write out a superblock's list of dirty inodes. A wait will be performed
1161 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1163 * If older_than_this is non-NULL, then only write out inodes which
1164 * had their first dirtying at a time earlier than *older_than_this.
1166 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1167 * This function assumes that the blockdev superblock's inodes are backed by
1168 * a variety of queues, so all inodes are searched. For other superblocks,
1169 * assume that all inodes are backed by the same queue.
1171 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1172 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1173 * on the writer throttling path, and we get decent balancing between many
1174 * throttled threads: we don't want them all piling up on inode_sync_wait.
1176 static void wait_sb_inodes(struct super_block *sb)
1178 struct inode *inode, *old_inode = NULL;
1181 * We need to be protected against the filesystem going from
1182 * r/o to r/w or vice versa.
1184 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1186 spin_lock(&inode_sb_list_lock);
1189 * Data integrity sync. Must wait for all pages under writeback,
1190 * because there may have been pages dirtied before our sync
1191 * call, but which had writeout started before we write it out.
1192 * In which case, the inode may not be on the dirty list, but
1193 * we still have to wait for that writeout.
1195 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1196 struct address_space *mapping = inode->i_mapping;
1198 spin_lock(&inode->i_lock);
1199 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1200 (mapping->nrpages == 0)) {
1201 spin_unlock(&inode->i_lock);
1205 spin_unlock(&inode->i_lock);
1206 spin_unlock(&inode_sb_list_lock);
1209 * We hold a reference to 'inode' so it couldn't have been
1210 * removed from s_inodes list while we dropped the
1211 * inode_sb_list_lock. We cannot iput the inode now as we can
1212 * be holding the last reference and we cannot iput it under
1213 * inode_sb_list_lock. So we keep the reference and iput it
1219 filemap_fdatawait(mapping);
1223 spin_lock(&inode_sb_list_lock);
1225 spin_unlock(&inode_sb_list_lock);
1230 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1231 * @sb: the superblock
1232 * @nr: the number of pages to write
1233 * @reason: reason why some writeback work initiated
1235 * Start writeback on some inodes on this super_block. No guarantees are made
1236 * on how many (if any) will be written, and this function does not wait
1237 * for IO completion of submitted IO.
1239 void writeback_inodes_sb_nr(struct super_block *sb,
1241 enum wb_reason reason)
1243 DECLARE_COMPLETION_ONSTACK(done);
1244 struct wb_writeback_work work = {
1246 .sync_mode = WB_SYNC_NONE,
1247 .tagged_writepages = 1,
1253 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1254 bdi_queue_work(sb->s_bdi, &work);
1255 wait_for_completion(&done);
1257 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1260 * writeback_inodes_sb - writeback dirty inodes from given super_block
1261 * @sb: the superblock
1262 * @reason: reason why some writeback work was initiated
1264 * Start writeback on some inodes on this super_block. No guarantees are made
1265 * on how many (if any) will be written, and this function does not wait
1266 * for IO completion of submitted IO.
1268 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1270 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1272 EXPORT_SYMBOL(writeback_inodes_sb);
1275 * writeback_inodes_sb_if_idle - start writeback if none underway
1276 * @sb: the superblock
1277 * @reason: reason why some writeback work was initiated
1279 * Invoke writeback_inodes_sb if no writeback is currently underway.
1280 * Returns 1 if writeback was started, 0 if not.
1282 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1284 if (!writeback_in_progress(sb->s_bdi)) {
1285 down_read(&sb->s_umount);
1286 writeback_inodes_sb(sb, reason);
1287 up_read(&sb->s_umount);
1292 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1295 * writeback_inodes_sb_if_idle - start writeback if none underway
1296 * @sb: the superblock
1297 * @nr: the number of pages to write
1298 * @reason: reason why some writeback work was initiated
1300 * Invoke writeback_inodes_sb if no writeback is currently underway.
1301 * Returns 1 if writeback was started, 0 if not.
1303 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1305 enum wb_reason reason)
1307 if (!writeback_in_progress(sb->s_bdi)) {
1308 down_read(&sb->s_umount);
1309 writeback_inodes_sb_nr(sb, nr, reason);
1310 up_read(&sb->s_umount);
1315 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1318 * sync_inodes_sb - sync sb inode pages
1319 * @sb: the superblock
1321 * This function writes and waits on any dirty inode belonging to this
1324 void sync_inodes_sb(struct super_block *sb)
1326 DECLARE_COMPLETION_ONSTACK(done);
1327 struct wb_writeback_work work = {
1329 .sync_mode = WB_SYNC_ALL,
1330 .nr_pages = LONG_MAX,
1333 .reason = WB_REASON_SYNC,
1336 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1338 bdi_queue_work(sb->s_bdi, &work);
1339 wait_for_completion(&done);
1343 EXPORT_SYMBOL(sync_inodes_sb);
1346 * write_inode_now - write an inode to disk
1347 * @inode: inode to write to disk
1348 * @sync: whether the write should be synchronous or not
1350 * This function commits an inode to disk immediately if it is dirty. This is
1351 * primarily needed by knfsd.
1353 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1355 int write_inode_now(struct inode *inode, int sync)
1357 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1359 struct writeback_control wbc = {
1360 .nr_to_write = LONG_MAX,
1361 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1363 .range_end = LLONG_MAX,
1366 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1367 wbc.nr_to_write = 0;
1370 spin_lock(&wb->list_lock);
1371 spin_lock(&inode->i_lock);
1372 ret = writeback_single_inode(inode, wb, &wbc);
1373 spin_unlock(&inode->i_lock);
1374 spin_unlock(&wb->list_lock);
1376 inode_sync_wait(inode);
1379 EXPORT_SYMBOL(write_inode_now);
1382 * sync_inode - write an inode and its pages to disk.
1383 * @inode: the inode to sync
1384 * @wbc: controls the writeback mode
1386 * sync_inode() will write an inode and its pages to disk. It will also
1387 * correctly update the inode on its superblock's dirty inode lists and will
1388 * update inode->i_state.
1390 * The caller must have a ref on the inode.
1392 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1394 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1397 spin_lock(&wb->list_lock);
1398 spin_lock(&inode->i_lock);
1399 ret = writeback_single_inode(inode, wb, wbc);
1400 spin_unlock(&inode->i_lock);
1401 spin_unlock(&wb->list_lock);
1404 EXPORT_SYMBOL(sync_inode);
1407 * sync_inode_metadata - write an inode to disk
1408 * @inode: the inode to sync
1409 * @wait: wait for I/O to complete.
1411 * Write an inode to disk and adjust its dirty state after completion.
1413 * Note: only writes the actual inode, no associated data or other metadata.
1415 int sync_inode_metadata(struct inode *inode, int wait)
1417 struct writeback_control wbc = {
1418 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1419 .nr_to_write = 0, /* metadata-only */
1422 return sync_inode(inode, &wbc);
1424 EXPORT_SYMBOL(sync_inode_metadata);