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 * The maximum number of pages to writeout in a single bdi flush/kupdate
34 * operation. We do this so we don't hold I_SYNC against an inode for
35 * enormous amounts of time, which would block a userspace task which has
36 * been forced to throttle against that inode. Also, the code reevaluates
37 * the dirty each time it has written this many pages.
39 #define MAX_WRITEBACK_PAGES 1024L
42 * Passed into wb_writeback(), essentially a subset of writeback_control
44 struct wb_writeback_work {
46 struct super_block *sb;
47 unsigned long *older_than_this;
48 enum writeback_sync_modes sync_mode;
49 unsigned int tagged_writepages:1;
50 unsigned int for_kupdate:1;
51 unsigned int range_cyclic:1;
52 unsigned int for_background:1;
54 struct list_head list; /* pending work list */
55 struct completion *done; /* set if the caller waits */
59 * Include the creation of the trace points after defining the
60 * wb_writeback_work structure so that the definition remains local to this
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/writeback.h>
67 * We don't actually have pdflush, but this one is exported though /proc...
69 int nr_pdflush_threads;
72 * writeback_in_progress - determine whether there is writeback in progress
73 * @bdi: the device's backing_dev_info structure.
75 * Determine whether there is writeback waiting to be handled against a
78 int writeback_in_progress(struct backing_dev_info *bdi)
80 return test_bit(BDI_writeback_running, &bdi->state);
83 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
85 struct super_block *sb = inode->i_sb;
87 if (strcmp(sb->s_type->name, "bdev") == 0)
88 return inode->i_mapping->backing_dev_info;
93 static inline struct inode *wb_inode(struct list_head *head)
95 return list_entry(head, struct inode, i_wb_list);
98 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
99 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
102 wake_up_process(bdi->wb.task);
105 * The bdi thread isn't there, wake up the forker thread which
106 * will create and run it.
108 wake_up_process(default_backing_dev_info.wb.task);
112 static void bdi_queue_work(struct backing_dev_info *bdi,
113 struct wb_writeback_work *work)
115 trace_writeback_queue(bdi, work);
117 spin_lock_bh(&bdi->wb_lock);
118 list_add_tail(&work->list, &bdi->work_list);
120 trace_writeback_nothread(bdi, work);
121 bdi_wakeup_flusher(bdi);
122 spin_unlock_bh(&bdi->wb_lock);
126 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
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);
138 trace_writeback_nowork(bdi);
139 wake_up_process(bdi->wb.task);
144 work->sync_mode = WB_SYNC_NONE;
145 work->nr_pages = nr_pages;
146 work->range_cyclic = range_cyclic;
148 bdi_queue_work(bdi, work);
152 * bdi_start_writeback - start writeback
153 * @bdi: the backing device to write from
154 * @nr_pages: the number of pages to write
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)
164 __bdi_start_writeback(bdi, nr_pages, true);
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)
235 * Prevent speculative execution through
236 * spin_unlock(&wb->list_lock);
240 wake_up_bit(&inode->i_state, __I_SYNC);
243 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
245 bool ret = time_after(inode->dirtied_when, t);
248 * For inodes being constantly redirtied, dirtied_when can get stuck.
249 * It _appears_ to be in the future, but is actually in distant past.
250 * This test is necessary to prevent such wrapped-around relative times
251 * from permanently stopping the whole bdi writeback.
253 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
259 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
261 static int move_expired_inodes(struct list_head *delaying_queue,
262 struct list_head *dispatch_queue,
263 unsigned long *older_than_this)
266 struct list_head *pos, *node;
267 struct super_block *sb = NULL;
272 while (!list_empty(delaying_queue)) {
273 inode = wb_inode(delaying_queue->prev);
274 if (older_than_this &&
275 inode_dirtied_after(inode, *older_than_this))
277 if (sb && sb != inode->i_sb)
280 list_move(&inode->i_wb_list, &tmp);
284 /* just one sb in list, splice to dispatch_queue and we're done */
286 list_splice(&tmp, dispatch_queue);
290 /* Move inodes from one superblock together */
291 while (!list_empty(&tmp)) {
292 sb = wb_inode(tmp.prev)->i_sb;
293 list_for_each_prev_safe(pos, node, &tmp) {
294 inode = wb_inode(pos);
295 if (inode->i_sb == sb)
296 list_move(&inode->i_wb_list, dispatch_queue);
304 * Queue all expired dirty inodes for io, eldest first.
306 * newly dirtied b_dirty b_io b_more_io
307 * =============> gf edc BA
309 * newly dirtied b_dirty b_io b_more_io
310 * =============> g fBAedc
312 * +--> dequeue for IO
314 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
317 assert_spin_locked(&wb->list_lock);
318 list_splice_init(&wb->b_more_io, &wb->b_io);
319 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
320 trace_writeback_queue_io(wb, older_than_this, moved);
323 static int write_inode(struct inode *inode, struct writeback_control *wbc)
325 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
326 return inode->i_sb->s_op->write_inode(inode, wbc);
331 * Wait for writeback on an inode to complete.
333 static void inode_wait_for_writeback(struct inode *inode,
334 struct bdi_writeback *wb)
336 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
337 wait_queue_head_t *wqh;
339 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
340 while (inode->i_state & I_SYNC) {
341 spin_unlock(&inode->i_lock);
342 spin_unlock(&wb->list_lock);
343 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
344 spin_lock(&wb->list_lock);
345 spin_lock(&inode->i_lock);
350 * Write out an inode's dirty pages. Called under wb->list_lock and
351 * inode->i_lock. Either the caller has an active reference on the inode or
352 * the inode has I_WILL_FREE set.
354 * If `wait' is set, wait on the writeout.
356 * The whole writeout design is quite complex and fragile. We want to avoid
357 * starvation of particular inodes when others are being redirtied, prevent
361 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
362 struct writeback_control *wbc)
364 struct address_space *mapping = inode->i_mapping;
365 long nr_to_write = wbc->nr_to_write;
369 assert_spin_locked(&wb->list_lock);
370 assert_spin_locked(&inode->i_lock);
372 if (!atomic_read(&inode->i_count))
373 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
375 WARN_ON(inode->i_state & I_WILL_FREE);
377 if (inode->i_state & I_SYNC) {
379 * If this inode is locked for writeback and we are not doing
380 * writeback-for-data-integrity, move it to b_more_io so that
381 * writeback can proceed with the other inodes on s_io.
383 * We'll have another go at writing back this inode when we
384 * completed a full scan of b_io.
386 if (wbc->sync_mode != WB_SYNC_ALL) {
387 requeue_io(inode, wb);
388 trace_writeback_single_inode_requeue(inode, wbc,
394 * It's a data-integrity sync. We must wait.
396 inode_wait_for_writeback(inode, wb);
399 BUG_ON(inode->i_state & I_SYNC);
401 /* Set I_SYNC, reset I_DIRTY_PAGES */
402 inode->i_state |= I_SYNC;
403 inode->i_state &= ~I_DIRTY_PAGES;
404 spin_unlock(&inode->i_lock);
405 spin_unlock(&wb->list_lock);
407 ret = do_writepages(mapping, wbc);
410 * Make sure to wait on the data before writing out the metadata.
411 * This is important for filesystems that modify metadata on data
414 if (wbc->sync_mode == WB_SYNC_ALL) {
415 int err = filemap_fdatawait(mapping);
421 * Some filesystems may redirty the inode during the writeback
422 * due to delalloc, clear dirty metadata flags right before
425 spin_lock(&inode->i_lock);
426 dirty = inode->i_state & I_DIRTY;
427 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
428 spin_unlock(&inode->i_lock);
429 /* Don't write the inode if only I_DIRTY_PAGES was set */
430 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
431 int err = write_inode(inode, wbc);
436 spin_lock(&wb->list_lock);
437 spin_lock(&inode->i_lock);
438 inode->i_state &= ~I_SYNC;
439 if (!(inode->i_state & I_FREEING)) {
441 * Sync livelock prevention. Each inode is tagged and synced in
442 * one shot. If still dirty, it will be redirty_tail()'ed below.
443 * Update the dirty time to prevent enqueue and sync it again.
445 if ((inode->i_state & I_DIRTY) &&
446 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
447 inode->dirtied_when = jiffies;
449 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
451 * We didn't write back all the pages. nfs_writepages()
452 * sometimes bales out without doing anything.
454 inode->i_state |= I_DIRTY_PAGES;
455 if (wbc->nr_to_write <= 0) {
457 * slice used up: queue for next turn
459 requeue_io(inode, wb);
462 * Writeback blocked by something other than
463 * congestion. Delay the inode for some time to
464 * avoid spinning on the CPU (100% iowait)
465 * retrying writeback of the dirty page/inode
466 * that cannot be performed immediately.
468 redirty_tail(inode, wb);
470 } else if (inode->i_state & I_DIRTY) {
472 * Filesystems can dirty the inode during writeback
473 * operations, such as delayed allocation during
474 * submission or metadata updates after data IO
477 redirty_tail(inode, wb);
480 * The inode is clean. At this point we either have
481 * a reference to the inode or it's on it's way out.
482 * No need to add it back to the LRU.
484 list_del_init(&inode->i_wb_list);
487 inode_sync_complete(inode);
488 trace_writeback_single_inode(inode, wbc, nr_to_write);
493 * For background writeback the caller does not have the sb pinned
494 * before calling writeback. So make sure that we do pin it, so it doesn't
495 * go away while we are writing inodes from it.
497 static bool pin_sb_for_writeback(struct super_block *sb)
500 if (list_empty(&sb->s_instances)) {
501 spin_unlock(&sb_lock);
506 spin_unlock(&sb_lock);
508 if (down_read_trylock(&sb->s_umount)) {
511 up_read(&sb->s_umount);
518 static long writeback_chunk_size(struct wb_writeback_work *work)
523 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
524 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
525 * here avoids calling into writeback_inodes_wb() more than once.
527 * The intended call sequence for WB_SYNC_ALL writeback is:
530 * writeback_sb_inodes() <== called only once
531 * write_cache_pages() <== called once for each inode
532 * (quickly) tag currently dirty pages
533 * (maybe slowly) sync all tagged pages
535 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
538 pages = min(MAX_WRITEBACK_PAGES, work->nr_pages);
544 * Write a portion of b_io inodes which belong to @sb.
546 * If @only_this_sb is true, then find and write all such
547 * inodes. Otherwise write only ones which go sequentially
550 * Return the number of pages and/or inodes written.
552 static long writeback_sb_inodes(struct super_block *sb,
553 struct bdi_writeback *wb,
554 struct wb_writeback_work *work)
556 struct writeback_control wbc = {
557 .sync_mode = work->sync_mode,
558 .tagged_writepages = work->tagged_writepages,
559 .for_kupdate = work->for_kupdate,
560 .for_background = work->for_background,
561 .range_cyclic = work->range_cyclic,
563 .range_end = LLONG_MAX,
565 unsigned long start_time = jiffies;
567 long wrote = 0; /* count both pages and inodes */
569 while (!list_empty(&wb->b_io)) {
570 struct inode *inode = wb_inode(wb->b_io.prev);
572 if (inode->i_sb != sb) {
575 * We only want to write back data for this
576 * superblock, move all inodes not belonging
577 * to it back onto the dirty list.
579 redirty_tail(inode, wb);
584 * The inode belongs to a different superblock.
585 * Bounce back to the caller to unpin this and
586 * pin the next superblock.
592 * Don't bother with new inodes or inodes beeing freed, first
593 * kind does not need peridic writeout yet, and for the latter
594 * kind writeout is handled by the freer.
596 spin_lock(&inode->i_lock);
597 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
598 spin_unlock(&inode->i_lock);
599 requeue_io(inode, wb);
603 write_chunk = writeback_chunk_size(work);
604 wbc.nr_to_write = write_chunk;
605 wbc.pages_skipped = 0;
607 writeback_single_inode(inode, wb, &wbc);
609 work->nr_pages -= write_chunk - wbc.nr_to_write;
610 wrote += write_chunk - wbc.nr_to_write;
611 if (!(inode->i_state & I_DIRTY))
613 if (wbc.pages_skipped) {
615 * writeback is not making progress due to locked
616 * buffers. Skip this inode for now.
618 redirty_tail(inode, wb);
620 spin_unlock(&inode->i_lock);
621 spin_unlock(&wb->list_lock);
624 spin_lock(&wb->list_lock);
626 * bail out to wb_writeback() often enough to check
627 * background threshold and other termination conditions.
630 if (time_is_before_jiffies(start_time + HZ / 10UL))
632 if (work->nr_pages <= 0)
639 static long __writeback_inodes_wb(struct bdi_writeback *wb,
640 struct wb_writeback_work *work)
642 unsigned long start_time = jiffies;
645 while (!list_empty(&wb->b_io)) {
646 struct inode *inode = wb_inode(wb->b_io.prev);
647 struct super_block *sb = inode->i_sb;
649 if (!pin_sb_for_writeback(sb)) {
650 requeue_io(inode, wb);
653 wrote += writeback_sb_inodes(sb, wb, work);
656 /* refer to the same tests at the end of writeback_sb_inodes */
658 if (time_is_before_jiffies(start_time + HZ / 10UL))
660 if (work->nr_pages <= 0)
664 /* Leave any unwritten inodes on b_io */
668 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages)
670 struct wb_writeback_work work = {
671 .nr_pages = nr_pages,
672 .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 inline bool over_bground_thresh(void)
687 unsigned long background_thresh, dirty_thresh;
689 global_dirty_limits(&background_thresh, &dirty_thresh);
691 return (global_page_state(NR_FILE_DIRTY) +
692 global_page_state(NR_UNSTABLE_NFS) > background_thresh);
696 * Called under wb->list_lock. If there are multiple wb per bdi,
697 * only the flusher working on the first wb should do it.
699 static void wb_update_bandwidth(struct bdi_writeback *wb,
700 unsigned long start_time)
702 __bdi_update_bandwidth(wb->bdi, start_time);
706 * Explicit flushing or periodic writeback of "old" data.
708 * Define "old": the first time one of an inode's pages is dirtied, we mark the
709 * dirtying-time in the inode's address_space. So this periodic writeback code
710 * just walks the superblock inode list, writing back any inodes which are
711 * older than a specific point in time.
713 * Try to run once per dirty_writeback_interval. But if a writeback event
714 * takes longer than a dirty_writeback_interval interval, then leave a
717 * older_than_this takes precedence over nr_to_write. So we'll only write back
718 * all dirty pages if they are all attached to "old" mappings.
720 static long wb_writeback(struct bdi_writeback *wb,
721 struct wb_writeback_work *work)
723 unsigned long wb_start = jiffies;
724 long nr_pages = work->nr_pages;
725 unsigned long oldest_jif;
729 oldest_jif = jiffies;
730 work->older_than_this = &oldest_jif;
732 spin_lock(&wb->list_lock);
735 * Stop writeback when nr_pages has been consumed
737 if (work->nr_pages <= 0)
741 * Background writeout and kupdate-style writeback may
742 * run forever. Stop them if there is other work to do
743 * so that e.g. sync can proceed. They'll be restarted
744 * after the other works are all done.
746 if ((work->for_background || work->for_kupdate) &&
747 !list_empty(&wb->bdi->work_list))
751 * For background writeout, stop when we are below the
752 * background dirty threshold
754 if (work->for_background && !over_bground_thresh())
757 if (work->for_kupdate) {
758 oldest_jif = jiffies -
759 msecs_to_jiffies(dirty_expire_interval * 10);
760 work->older_than_this = &oldest_jif;
763 trace_writeback_start(wb->bdi, work);
764 if (list_empty(&wb->b_io))
765 queue_io(wb, work->older_than_this);
767 progress = writeback_sb_inodes(work->sb, wb, work);
769 progress = __writeback_inodes_wb(wb, work);
770 trace_writeback_written(wb->bdi, work);
772 wb_update_bandwidth(wb, wb_start);
775 * Did we write something? Try for more
777 * Dirty inodes are moved to b_io for writeback in batches.
778 * The completion of the current batch does not necessarily
779 * mean the overall work is done. So we keep looping as long
780 * as made some progress on cleaning pages or inodes.
785 * No more inodes for IO, bail
787 if (list_empty(&wb->b_more_io))
790 * Nothing written. Wait for some inode to
791 * become available for writeback. Otherwise
792 * we'll just busyloop.
794 if (!list_empty(&wb->b_more_io)) {
795 trace_writeback_wait(wb->bdi, work);
796 inode = wb_inode(wb->b_more_io.prev);
797 spin_lock(&inode->i_lock);
798 inode_wait_for_writeback(inode, wb);
799 spin_unlock(&inode->i_lock);
802 spin_unlock(&wb->list_lock);
804 return nr_pages - work->nr_pages;
808 * Return the next wb_writeback_work struct that hasn't been processed yet.
810 static struct wb_writeback_work *
811 get_next_work_item(struct backing_dev_info *bdi)
813 struct wb_writeback_work *work = NULL;
815 spin_lock_bh(&bdi->wb_lock);
816 if (!list_empty(&bdi->work_list)) {
817 work = list_entry(bdi->work_list.next,
818 struct wb_writeback_work, list);
819 list_del_init(&work->list);
821 spin_unlock_bh(&bdi->wb_lock);
826 * Add in the number of potentially dirty inodes, because each inode
827 * write can dirty pagecache in the underlying blockdev.
829 static unsigned long get_nr_dirty_pages(void)
831 return global_page_state(NR_FILE_DIRTY) +
832 global_page_state(NR_UNSTABLE_NFS) +
833 get_nr_dirty_inodes();
836 static long wb_check_background_flush(struct bdi_writeback *wb)
838 if (over_bground_thresh()) {
840 struct wb_writeback_work work = {
841 .nr_pages = LONG_MAX,
842 .sync_mode = WB_SYNC_NONE,
847 return wb_writeback(wb, &work);
853 static long wb_check_old_data_flush(struct bdi_writeback *wb)
855 unsigned long expired;
859 * When set to zero, disable periodic writeback
861 if (!dirty_writeback_interval)
864 expired = wb->last_old_flush +
865 msecs_to_jiffies(dirty_writeback_interval * 10);
866 if (time_before(jiffies, expired))
869 wb->last_old_flush = jiffies;
870 nr_pages = get_nr_dirty_pages();
873 struct wb_writeback_work work = {
874 .nr_pages = nr_pages,
875 .sync_mode = WB_SYNC_NONE,
880 return wb_writeback(wb, &work);
887 * Retrieve work items and do the writeback they describe
889 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
891 struct backing_dev_info *bdi = wb->bdi;
892 struct wb_writeback_work *work;
895 set_bit(BDI_writeback_running, &wb->bdi->state);
896 while ((work = get_next_work_item(bdi)) != NULL) {
898 * Override sync mode, in case we must wait for completion
899 * because this thread is exiting now.
902 work->sync_mode = WB_SYNC_ALL;
904 trace_writeback_exec(bdi, work);
906 wrote += wb_writeback(wb, work);
909 * Notify the caller of completion if this is a synchronous
910 * work item, otherwise just free it.
913 complete(work->done);
919 * Check for periodic writeback, kupdated() style
921 wrote += wb_check_old_data_flush(wb);
922 wrote += wb_check_background_flush(wb);
923 clear_bit(BDI_writeback_running, &wb->bdi->state);
929 * Handle writeback of dirty data for the device backed by this bdi. Also
930 * wakes up periodically and does kupdated style flushing.
932 int bdi_writeback_thread(void *data)
934 struct bdi_writeback *wb = data;
935 struct backing_dev_info *bdi = wb->bdi;
938 current->flags |= PF_SWAPWRITE;
940 wb->last_active = jiffies;
943 * Our parent may run at a different priority, just set us to normal
945 set_user_nice(current, 0);
947 trace_writeback_thread_start(bdi);
949 while (!kthread_should_stop()) {
951 * Remove own delayed wake-up timer, since we are already awake
952 * and we'll take care of the preriodic write-back.
954 del_timer(&wb->wakeup_timer);
956 pages_written = wb_do_writeback(wb, 0);
958 trace_writeback_pages_written(pages_written);
961 wb->last_active = jiffies;
963 set_current_state(TASK_INTERRUPTIBLE);
964 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
965 __set_current_state(TASK_RUNNING);
969 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
970 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
973 * We have nothing to do, so can go sleep without any
974 * timeout and save power. When a work is queued or
975 * something is made dirty - we will be woken up.
983 /* Flush any work that raced with us exiting */
984 if (!list_empty(&bdi->work_list))
985 wb_do_writeback(wb, 1);
987 trace_writeback_thread_stop(bdi);
993 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
996 void wakeup_flusher_threads(long nr_pages)
998 struct backing_dev_info *bdi;
1001 nr_pages = global_page_state(NR_FILE_DIRTY) +
1002 global_page_state(NR_UNSTABLE_NFS);
1006 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1007 if (!bdi_has_dirty_io(bdi))
1009 __bdi_start_writeback(bdi, nr_pages, false);
1014 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1016 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1017 struct dentry *dentry;
1018 const char *name = "?";
1020 dentry = d_find_alias(inode);
1022 spin_lock(&dentry->d_lock);
1023 name = (const char *) dentry->d_name.name;
1026 "%s(%d): dirtied inode %lu (%s) on %s\n",
1027 current->comm, task_pid_nr(current), inode->i_ino,
1028 name, inode->i_sb->s_id);
1030 spin_unlock(&dentry->d_lock);
1037 * __mark_inode_dirty - internal function
1038 * @inode: inode to mark
1039 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1040 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1041 * mark_inode_dirty_sync.
1043 * Put the inode on the super block's dirty list.
1045 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1046 * dirty list only if it is hashed or if it refers to a blockdev.
1047 * If it was not hashed, it will never be added to the dirty list
1048 * even if it is later hashed, as it will have been marked dirty already.
1050 * In short, make sure you hash any inodes _before_ you start marking
1053 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1054 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1055 * the kernel-internal blockdev inode represents the dirtying time of the
1056 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1057 * page->mapping->host, so the page-dirtying time is recorded in the internal
1060 void __mark_inode_dirty(struct inode *inode, int flags)
1062 struct super_block *sb = inode->i_sb;
1063 struct backing_dev_info *bdi = NULL;
1066 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1067 * dirty the inode itself
1069 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1070 if (sb->s_op->dirty_inode)
1071 sb->s_op->dirty_inode(inode, flags);
1075 * make sure that changes are seen by all cpus before we test i_state
1080 /* avoid the locking if we can */
1081 if ((inode->i_state & flags) == flags)
1084 if (unlikely(block_dump))
1085 block_dump___mark_inode_dirty(inode);
1087 spin_lock(&inode->i_lock);
1088 if ((inode->i_state & flags) != flags) {
1089 const int was_dirty = inode->i_state & I_DIRTY;
1091 inode->i_state |= flags;
1094 * If the inode is being synced, just update its dirty state.
1095 * The unlocker will place the inode on the appropriate
1096 * superblock list, based upon its state.
1098 if (inode->i_state & I_SYNC)
1099 goto out_unlock_inode;
1102 * Only add valid (hashed) inodes to the superblock's
1103 * dirty list. Add blockdev inodes as well.
1105 if (!S_ISBLK(inode->i_mode)) {
1106 if (inode_unhashed(inode))
1107 goto out_unlock_inode;
1109 if (inode->i_state & I_FREEING)
1110 goto out_unlock_inode;
1113 * If the inode was already on b_dirty/b_io/b_more_io, don't
1114 * reposition it (that would break b_dirty time-ordering).
1117 bool wakeup_bdi = false;
1118 bdi = inode_to_bdi(inode);
1120 if (bdi_cap_writeback_dirty(bdi)) {
1121 WARN(!test_bit(BDI_registered, &bdi->state),
1122 "bdi-%s not registered\n", bdi->name);
1125 * If this is the first dirty inode for this
1126 * bdi, we have to wake-up the corresponding
1127 * bdi thread to make sure background
1128 * write-back happens later.
1130 if (!wb_has_dirty_io(&bdi->wb))
1134 spin_unlock(&inode->i_lock);
1135 spin_lock(&bdi->wb.list_lock);
1136 inode->dirtied_when = jiffies;
1137 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1138 spin_unlock(&bdi->wb.list_lock);
1141 bdi_wakeup_thread_delayed(bdi);
1146 spin_unlock(&inode->i_lock);
1149 EXPORT_SYMBOL(__mark_inode_dirty);
1152 * Write out a superblock's list of dirty inodes. A wait will be performed
1153 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1155 * If older_than_this is non-NULL, then only write out inodes which
1156 * had their first dirtying at a time earlier than *older_than_this.
1158 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1159 * This function assumes that the blockdev superblock's inodes are backed by
1160 * a variety of queues, so all inodes are searched. For other superblocks,
1161 * assume that all inodes are backed by the same queue.
1163 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1164 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1165 * on the writer throttling path, and we get decent balancing between many
1166 * throttled threads: we don't want them all piling up on inode_sync_wait.
1168 static void wait_sb_inodes(struct super_block *sb)
1170 struct inode *inode, *old_inode = NULL;
1173 * We need to be protected against the filesystem going from
1174 * r/o to r/w or vice versa.
1176 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1178 spin_lock(&inode_sb_list_lock);
1181 * Data integrity sync. Must wait for all pages under writeback,
1182 * because there may have been pages dirtied before our sync
1183 * call, but which had writeout started before we write it out.
1184 * In which case, the inode may not be on the dirty list, but
1185 * we still have to wait for that writeout.
1187 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1188 struct address_space *mapping = inode->i_mapping;
1190 spin_lock(&inode->i_lock);
1191 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1192 (mapping->nrpages == 0)) {
1193 spin_unlock(&inode->i_lock);
1197 spin_unlock(&inode->i_lock);
1198 spin_unlock(&inode_sb_list_lock);
1201 * We hold a reference to 'inode' so it couldn't have been
1202 * removed from s_inodes list while we dropped the
1203 * inode_sb_list_lock. We cannot iput the inode now as we can
1204 * be holding the last reference and we cannot iput it under
1205 * inode_sb_list_lock. So we keep the reference and iput it
1211 filemap_fdatawait(mapping);
1215 spin_lock(&inode_sb_list_lock);
1217 spin_unlock(&inode_sb_list_lock);
1222 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1223 * @sb: the superblock
1224 * @nr: the number of pages to write
1226 * Start writeback on some inodes on this super_block. No guarantees are made
1227 * on how many (if any) will be written, and this function does not wait
1228 * for IO completion of submitted IO.
1230 void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
1232 DECLARE_COMPLETION_ONSTACK(done);
1233 struct wb_writeback_work work = {
1235 .sync_mode = WB_SYNC_NONE,
1236 .tagged_writepages = 1,
1241 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1242 bdi_queue_work(sb->s_bdi, &work);
1243 wait_for_completion(&done);
1245 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1248 * writeback_inodes_sb - writeback dirty inodes from given super_block
1249 * @sb: the superblock
1251 * Start writeback on some inodes on this super_block. No guarantees are made
1252 * on how many (if any) will be written, and this function does not wait
1253 * for IO completion of submitted IO.
1255 void writeback_inodes_sb(struct super_block *sb)
1257 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages());
1259 EXPORT_SYMBOL(writeback_inodes_sb);
1262 * writeback_inodes_sb_if_idle - start writeback if none underway
1263 * @sb: the superblock
1265 * Invoke writeback_inodes_sb if no writeback is currently underway.
1266 * Returns 1 if writeback was started, 0 if not.
1268 int writeback_inodes_sb_if_idle(struct super_block *sb)
1270 if (!writeback_in_progress(sb->s_bdi)) {
1271 down_read(&sb->s_umount);
1272 writeback_inodes_sb(sb);
1273 up_read(&sb->s_umount);
1278 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1281 * writeback_inodes_sb_if_idle - start writeback if none underway
1282 * @sb: the superblock
1283 * @nr: the number of pages to write
1285 * Invoke writeback_inodes_sb if no writeback is currently underway.
1286 * Returns 1 if writeback was started, 0 if not.
1288 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1291 if (!writeback_in_progress(sb->s_bdi)) {
1292 down_read(&sb->s_umount);
1293 writeback_inodes_sb_nr(sb, nr);
1294 up_read(&sb->s_umount);
1299 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1302 * sync_inodes_sb - sync sb inode pages
1303 * @sb: the superblock
1305 * This function writes and waits on any dirty inode belonging to this
1308 void sync_inodes_sb(struct super_block *sb)
1310 DECLARE_COMPLETION_ONSTACK(done);
1311 struct wb_writeback_work work = {
1313 .sync_mode = WB_SYNC_ALL,
1314 .nr_pages = LONG_MAX,
1319 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1321 bdi_queue_work(sb->s_bdi, &work);
1322 wait_for_completion(&done);
1326 EXPORT_SYMBOL(sync_inodes_sb);
1329 * write_inode_now - write an inode to disk
1330 * @inode: inode to write to disk
1331 * @sync: whether the write should be synchronous or not
1333 * This function commits an inode to disk immediately if it is dirty. This is
1334 * primarily needed by knfsd.
1336 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1338 int write_inode_now(struct inode *inode, int sync)
1340 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1342 struct writeback_control wbc = {
1343 .nr_to_write = LONG_MAX,
1344 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1346 .range_end = LLONG_MAX,
1349 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1350 wbc.nr_to_write = 0;
1353 spin_lock(&wb->list_lock);
1354 spin_lock(&inode->i_lock);
1355 ret = writeback_single_inode(inode, wb, &wbc);
1356 spin_unlock(&inode->i_lock);
1357 spin_unlock(&wb->list_lock);
1359 inode_sync_wait(inode);
1362 EXPORT_SYMBOL(write_inode_now);
1365 * sync_inode - write an inode and its pages to disk.
1366 * @inode: the inode to sync
1367 * @wbc: controls the writeback mode
1369 * sync_inode() will write an inode and its pages to disk. It will also
1370 * correctly update the inode on its superblock's dirty inode lists and will
1371 * update inode->i_state.
1373 * The caller must have a ref on the inode.
1375 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1377 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1380 spin_lock(&wb->list_lock);
1381 spin_lock(&inode->i_lock);
1382 ret = writeback_single_inode(inode, wb, wbc);
1383 spin_unlock(&inode->i_lock);
1384 spin_unlock(&wb->list_lock);
1387 EXPORT_SYMBOL(sync_inode);
1390 * sync_inode_metadata - write an inode to disk
1391 * @inode: the inode to sync
1392 * @wait: wait for I/O to complete.
1394 * Write an inode to disk and adjust its dirty state after completion.
1396 * Note: only writes the actual inode, no associated data or other metadata.
1398 int sync_inode_metadata(struct inode *inode, int wait)
1400 struct writeback_control wbc = {
1401 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1402 .nr_to_write = 0, /* metadata-only */
1405 return sync_inode(inode, &wbc);
1407 EXPORT_SYMBOL(sync_inode_metadata);