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;
45 struct list_head list; /* pending work list */
46 struct completion *done; /* set if the caller waits */
50 * Include the creation of the trace points after defining the
51 * wb_writeback_work structure so that the definition remains local to this
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/writeback.h>
58 * We don't actually have pdflush, but this one is exported though /proc...
60 int nr_pdflush_threads;
63 * writeback_in_progress - determine whether there is writeback in progress
64 * @bdi: the device's backing_dev_info structure.
66 * Determine whether there is writeback waiting to be handled against a
69 int writeback_in_progress(struct backing_dev_info *bdi)
71 return test_bit(BDI_writeback_running, &bdi->state);
74 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
76 struct super_block *sb = inode->i_sb;
78 if (strcmp(sb->s_type->name, "bdev") == 0)
79 return inode->i_mapping->backing_dev_info;
84 static inline struct inode *wb_inode(struct list_head *head)
86 return list_entry(head, struct inode, i_wb_list);
89 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
90 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
93 wake_up_process(bdi->wb.task);
96 * The bdi thread isn't there, wake up the forker thread which
97 * will create and run it.
99 wake_up_process(default_backing_dev_info.wb.task);
103 static void bdi_queue_work(struct backing_dev_info *bdi,
104 struct wb_writeback_work *work)
106 trace_writeback_queue(bdi, work);
108 spin_lock_bh(&bdi->wb_lock);
109 list_add_tail(&work->list, &bdi->work_list);
111 trace_writeback_nothread(bdi, work);
112 bdi_wakeup_flusher(bdi);
113 spin_unlock_bh(&bdi->wb_lock);
117 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
120 struct wb_writeback_work *work;
123 * This is WB_SYNC_NONE writeback, so if allocation fails just
124 * wakeup the thread for old dirty data writeback
126 work = kzalloc(sizeof(*work), GFP_ATOMIC);
129 trace_writeback_nowork(bdi);
130 wake_up_process(bdi->wb.task);
135 work->sync_mode = WB_SYNC_NONE;
136 work->nr_pages = nr_pages;
137 work->range_cyclic = range_cyclic;
139 bdi_queue_work(bdi, work);
143 * bdi_start_writeback - start writeback
144 * @bdi: the backing device to write from
145 * @nr_pages: the number of pages to write
148 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
149 * started when this function returns, we make no guarantees on
150 * completion. Caller need not hold sb s_umount semaphore.
153 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
155 __bdi_start_writeback(bdi, nr_pages, true);
159 * bdi_start_background_writeback - start background writeback
160 * @bdi: the backing device to write from
163 * This makes sure WB_SYNC_NONE background writeback happens. When
164 * this function returns, it is only guaranteed that for given BDI
165 * some IO is happening if we are over background dirty threshold.
166 * Caller need not hold sb s_umount semaphore.
168 void bdi_start_background_writeback(struct backing_dev_info *bdi)
171 * We just wake up the flusher thread. It will perform background
172 * writeback as soon as there is no other work to do.
174 trace_writeback_wake_background(bdi);
175 spin_lock_bh(&bdi->wb_lock);
176 bdi_wakeup_flusher(bdi);
177 spin_unlock_bh(&bdi->wb_lock);
181 * Remove the inode from the writeback list it is on.
183 void inode_wb_list_del(struct inode *inode)
185 struct backing_dev_info *bdi = inode_to_bdi(inode);
187 spin_lock(&bdi->wb.list_lock);
188 list_del_init(&inode->i_wb_list);
189 spin_unlock(&bdi->wb.list_lock);
193 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
194 * furthest end of its superblock's dirty-inode list.
196 * Before stamping the inode's ->dirtied_when, we check to see whether it is
197 * already the most-recently-dirtied inode on the b_dirty list. If that is
198 * the case then the inode must have been redirtied while it was being written
199 * out and we don't reset its dirtied_when.
201 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
203 assert_spin_locked(&wb->list_lock);
204 if (!list_empty(&wb->b_dirty)) {
207 tail = wb_inode(wb->b_dirty.next);
208 if (time_before(inode->dirtied_when, tail->dirtied_when))
209 inode->dirtied_when = jiffies;
211 list_move(&inode->i_wb_list, &wb->b_dirty);
215 * requeue inode for re-scanning after bdi->b_io list is exhausted.
217 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
219 assert_spin_locked(&wb->list_lock);
220 list_move(&inode->i_wb_list, &wb->b_more_io);
223 static void inode_sync_complete(struct inode *inode)
226 * Prevent speculative execution through
227 * spin_unlock(&wb->list_lock);
231 wake_up_bit(&inode->i_state, __I_SYNC);
234 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
236 bool ret = time_after(inode->dirtied_when, t);
239 * For inodes being constantly redirtied, dirtied_when can get stuck.
240 * It _appears_ to be in the future, but is actually in distant past.
241 * This test is necessary to prevent such wrapped-around relative times
242 * from permanently stopping the whole bdi writeback.
244 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
250 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
252 static int move_expired_inodes(struct list_head *delaying_queue,
253 struct list_head *dispatch_queue,
254 unsigned long *older_than_this)
257 struct list_head *pos, *node;
258 struct super_block *sb = NULL;
263 while (!list_empty(delaying_queue)) {
264 inode = wb_inode(delaying_queue->prev);
265 if (older_than_this &&
266 inode_dirtied_after(inode, *older_than_this))
268 if (sb && sb != inode->i_sb)
271 list_move(&inode->i_wb_list, &tmp);
275 /* just one sb in list, splice to dispatch_queue and we're done */
277 list_splice(&tmp, dispatch_queue);
281 /* Move inodes from one superblock together */
282 while (!list_empty(&tmp)) {
283 sb = wb_inode(tmp.prev)->i_sb;
284 list_for_each_prev_safe(pos, node, &tmp) {
285 inode = wb_inode(pos);
286 if (inode->i_sb == sb)
287 list_move(&inode->i_wb_list, dispatch_queue);
295 * Queue all expired dirty inodes for io, eldest first.
297 * newly dirtied b_dirty b_io b_more_io
298 * =============> gf edc BA
300 * newly dirtied b_dirty b_io b_more_io
301 * =============> g fBAedc
303 * +--> dequeue for IO
305 static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
308 assert_spin_locked(&wb->list_lock);
309 list_splice_init(&wb->b_more_io, &wb->b_io);
310 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
311 trace_writeback_queue_io(wb, older_than_this, moved);
314 static int write_inode(struct inode *inode, struct writeback_control *wbc)
316 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
317 return inode->i_sb->s_op->write_inode(inode, wbc);
322 * Wait for writeback on an inode to complete.
324 static void inode_wait_for_writeback(struct inode *inode,
325 struct bdi_writeback *wb)
327 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
328 wait_queue_head_t *wqh;
330 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
331 while (inode->i_state & I_SYNC) {
332 spin_unlock(&inode->i_lock);
333 spin_unlock(&wb->list_lock);
334 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
335 spin_lock(&wb->list_lock);
336 spin_lock(&inode->i_lock);
341 * Write out an inode's dirty pages. Called under wb->list_lock and
342 * inode->i_lock. Either the caller has an active reference on the inode or
343 * the inode has I_WILL_FREE set.
345 * If `wait' is set, wait on the writeout.
347 * The whole writeout design is quite complex and fragile. We want to avoid
348 * starvation of particular inodes when others are being redirtied, prevent
352 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
353 struct writeback_control *wbc)
355 struct address_space *mapping = inode->i_mapping;
356 long nr_to_write = wbc->nr_to_write;
360 assert_spin_locked(&wb->list_lock);
361 assert_spin_locked(&inode->i_lock);
363 if (!atomic_read(&inode->i_count))
364 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
366 WARN_ON(inode->i_state & I_WILL_FREE);
368 if (inode->i_state & I_SYNC) {
370 * If this inode is locked for writeback and we are not doing
371 * writeback-for-data-integrity, move it to b_more_io so that
372 * writeback can proceed with the other inodes on s_io.
374 * We'll have another go at writing back this inode when we
375 * completed a full scan of b_io.
377 if (wbc->sync_mode != WB_SYNC_ALL) {
378 requeue_io(inode, wb);
379 trace_writeback_single_inode_requeue(inode, wbc,
385 * It's a data-integrity sync. We must wait.
387 inode_wait_for_writeback(inode, wb);
390 BUG_ON(inode->i_state & I_SYNC);
392 /* Set I_SYNC, reset I_DIRTY_PAGES */
393 inode->i_state |= I_SYNC;
394 inode->i_state &= ~I_DIRTY_PAGES;
395 spin_unlock(&inode->i_lock);
396 spin_unlock(&wb->list_lock);
398 ret = do_writepages(mapping, wbc);
401 * Make sure to wait on the data before writing out the metadata.
402 * This is important for filesystems that modify metadata on data
405 if (wbc->sync_mode == WB_SYNC_ALL) {
406 int err = filemap_fdatawait(mapping);
412 * Some filesystems may redirty the inode during the writeback
413 * due to delalloc, clear dirty metadata flags right before
416 spin_lock(&inode->i_lock);
417 dirty = inode->i_state & I_DIRTY;
418 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
419 spin_unlock(&inode->i_lock);
420 /* Don't write the inode if only I_DIRTY_PAGES was set */
421 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
422 int err = write_inode(inode, wbc);
427 spin_lock(&wb->list_lock);
428 spin_lock(&inode->i_lock);
429 inode->i_state &= ~I_SYNC;
430 if (!(inode->i_state & I_FREEING)) {
432 * Sync livelock prevention. Each inode is tagged and synced in
433 * one shot. If still dirty, it will be redirty_tail()'ed below.
434 * Update the dirty time to prevent enqueue and sync it again.
436 if ((inode->i_state & I_DIRTY) &&
437 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
438 inode->dirtied_when = jiffies;
440 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
442 * We didn't write back all the pages. nfs_writepages()
443 * sometimes bales out without doing anything.
445 inode->i_state |= I_DIRTY_PAGES;
446 if (wbc->nr_to_write <= 0) {
448 * slice used up: queue for next turn
450 requeue_io(inode, wb);
453 * Writeback blocked by something other than
454 * congestion. Delay the inode for some time to
455 * avoid spinning on the CPU (100% iowait)
456 * retrying writeback of the dirty page/inode
457 * that cannot be performed immediately.
459 redirty_tail(inode, wb);
461 } else if (inode->i_state & I_DIRTY) {
463 * Filesystems can dirty the inode during writeback
464 * operations, such as delayed allocation during
465 * submission or metadata updates after data IO
468 redirty_tail(inode, wb);
471 * The inode is clean. At this point we either have
472 * a reference to the inode or it's on it's way out.
473 * No need to add it back to the LRU.
475 list_del_init(&inode->i_wb_list);
478 inode_sync_complete(inode);
479 trace_writeback_single_inode(inode, wbc, nr_to_write);
484 * For background writeback the caller does not have the sb pinned
485 * before calling writeback. So make sure that we do pin it, so it doesn't
486 * go away while we are writing inodes from it.
488 static bool pin_sb_for_writeback(struct super_block *sb)
491 if (list_empty(&sb->s_instances)) {
492 spin_unlock(&sb_lock);
497 spin_unlock(&sb_lock);
499 if (down_read_trylock(&sb->s_umount)) {
502 up_read(&sb->s_umount);
509 static long writeback_chunk_size(struct backing_dev_info *bdi,
510 struct wb_writeback_work *work)
515 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
516 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
517 * here avoids calling into writeback_inodes_wb() more than once.
519 * The intended call sequence for WB_SYNC_ALL writeback is:
522 * writeback_sb_inodes() <== called only once
523 * write_cache_pages() <== called once for each inode
524 * (quickly) tag currently dirty pages
525 * (maybe slowly) sync all tagged pages
527 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
530 pages = min(bdi->avg_write_bandwidth / 2,
531 global_dirty_limit / DIRTY_SCOPE);
532 pages = min(pages, work->nr_pages);
533 pages = round_down(pages + MIN_WRITEBACK_PAGES,
534 MIN_WRITEBACK_PAGES);
541 * Write a portion of b_io inodes which belong to @sb.
543 * If @only_this_sb is true, then find and write all such
544 * inodes. Otherwise write only ones which go sequentially
547 * Return the number of pages and/or inodes written.
549 static long writeback_sb_inodes(struct super_block *sb,
550 struct bdi_writeback *wb,
551 struct wb_writeback_work *work)
553 struct writeback_control wbc = {
554 .sync_mode = work->sync_mode,
555 .tagged_writepages = work->tagged_writepages,
556 .for_kupdate = work->for_kupdate,
557 .for_background = work->for_background,
558 .range_cyclic = work->range_cyclic,
560 .range_end = LLONG_MAX,
562 unsigned long start_time = jiffies;
564 long wrote = 0; /* count both pages and inodes */
566 while (!list_empty(&wb->b_io)) {
567 struct inode *inode = wb_inode(wb->b_io.prev);
569 if (inode->i_sb != sb) {
572 * We only want to write back data for this
573 * superblock, move all inodes not belonging
574 * to it back onto the dirty list.
576 redirty_tail(inode, wb);
581 * The inode belongs to a different superblock.
582 * Bounce back to the caller to unpin this and
583 * pin the next superblock.
589 * Don't bother with new inodes or inodes beeing freed, first
590 * kind does not need peridic writeout yet, and for the latter
591 * kind writeout is handled by the freer.
593 spin_lock(&inode->i_lock);
594 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
595 spin_unlock(&inode->i_lock);
596 redirty_tail(inode, wb);
600 write_chunk = writeback_chunk_size(wb->bdi, work);
601 wbc.nr_to_write = write_chunk;
602 wbc.pages_skipped = 0;
604 writeback_single_inode(inode, wb, &wbc);
606 work->nr_pages -= write_chunk - wbc.nr_to_write;
607 wrote += write_chunk - wbc.nr_to_write;
608 if (!(inode->i_state & I_DIRTY))
610 if (wbc.pages_skipped) {
612 * writeback is not making progress due to locked
613 * buffers. Skip this inode for now.
615 redirty_tail(inode, wb);
617 spin_unlock(&inode->i_lock);
618 spin_unlock(&wb->list_lock);
621 spin_lock(&wb->list_lock);
623 * bail out to wb_writeback() often enough to check
624 * background threshold and other termination conditions.
627 if (time_is_before_jiffies(start_time + HZ / 10UL))
629 if (work->nr_pages <= 0)
636 static long __writeback_inodes_wb(struct bdi_writeback *wb,
637 struct wb_writeback_work *work)
639 unsigned long start_time = jiffies;
642 while (!list_empty(&wb->b_io)) {
643 struct inode *inode = wb_inode(wb->b_io.prev);
644 struct super_block *sb = inode->i_sb;
646 if (!pin_sb_for_writeback(sb)) {
647 requeue_io(inode, wb);
650 wrote += writeback_sb_inodes(sb, wb, work);
653 /* refer to the same tests at the end of writeback_sb_inodes */
655 if (time_is_before_jiffies(start_time + HZ / 10UL))
657 if (work->nr_pages <= 0)
661 /* Leave any unwritten inodes on b_io */
665 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages)
667 struct wb_writeback_work work = {
668 .nr_pages = nr_pages,
669 .sync_mode = WB_SYNC_NONE,
673 spin_lock(&wb->list_lock);
674 if (list_empty(&wb->b_io))
676 __writeback_inodes_wb(wb, &work);
677 spin_unlock(&wb->list_lock);
679 return nr_pages - work.nr_pages;
682 static inline bool over_bground_thresh(void)
684 unsigned long background_thresh, dirty_thresh;
686 global_dirty_limits(&background_thresh, &dirty_thresh);
688 return (global_page_state(NR_FILE_DIRTY) +
689 global_page_state(NR_UNSTABLE_NFS) > background_thresh);
693 * Called under wb->list_lock. If there are multiple wb per bdi,
694 * only the flusher working on the first wb should do it.
696 static void wb_update_bandwidth(struct bdi_writeback *wb,
697 unsigned long start_time)
699 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, start_time);
703 * Explicit flushing or periodic writeback of "old" data.
705 * Define "old": the first time one of an inode's pages is dirtied, we mark the
706 * dirtying-time in the inode's address_space. So this periodic writeback code
707 * just walks the superblock inode list, writing back any inodes which are
708 * older than a specific point in time.
710 * Try to run once per dirty_writeback_interval. But if a writeback event
711 * takes longer than a dirty_writeback_interval interval, then leave a
714 * older_than_this takes precedence over nr_to_write. So we'll only write back
715 * all dirty pages if they are all attached to "old" mappings.
717 static long wb_writeback(struct bdi_writeback *wb,
718 struct wb_writeback_work *work)
720 unsigned long wb_start = jiffies;
721 long nr_pages = work->nr_pages;
722 unsigned long oldest_jif;
726 oldest_jif = jiffies;
727 work->older_than_this = &oldest_jif;
729 spin_lock(&wb->list_lock);
732 * Stop writeback when nr_pages has been consumed
734 if (work->nr_pages <= 0)
738 * Background writeout and kupdate-style writeback may
739 * run forever. Stop them if there is other work to do
740 * so that e.g. sync can proceed. They'll be restarted
741 * after the other works are all done.
743 if ((work->for_background || work->for_kupdate) &&
744 !list_empty(&wb->bdi->work_list))
748 * For background writeout, stop when we are below the
749 * background dirty threshold
751 if (work->for_background && !over_bground_thresh())
754 if (work->for_kupdate) {
755 oldest_jif = jiffies -
756 msecs_to_jiffies(dirty_expire_interval * 10);
757 work->older_than_this = &oldest_jif;
760 trace_writeback_start(wb->bdi, work);
761 if (list_empty(&wb->b_io))
762 queue_io(wb, work->older_than_this);
764 progress = writeback_sb_inodes(work->sb, wb, work);
766 progress = __writeback_inodes_wb(wb, work);
767 trace_writeback_written(wb->bdi, work);
769 wb_update_bandwidth(wb, wb_start);
772 * Did we write something? Try for more
774 * Dirty inodes are moved to b_io for writeback in batches.
775 * The completion of the current batch does not necessarily
776 * mean the overall work is done. So we keep looping as long
777 * as made some progress on cleaning pages or inodes.
782 * No more inodes for IO, bail
784 if (list_empty(&wb->b_more_io))
787 * Nothing written. Wait for some inode to
788 * become available for writeback. Otherwise
789 * we'll just busyloop.
791 if (!list_empty(&wb->b_more_io)) {
792 trace_writeback_wait(wb->bdi, work);
793 inode = wb_inode(wb->b_more_io.prev);
794 spin_lock(&inode->i_lock);
795 inode_wait_for_writeback(inode, wb);
796 spin_unlock(&inode->i_lock);
799 spin_unlock(&wb->list_lock);
801 return nr_pages - work->nr_pages;
805 * Return the next wb_writeback_work struct that hasn't been processed yet.
807 static struct wb_writeback_work *
808 get_next_work_item(struct backing_dev_info *bdi)
810 struct wb_writeback_work *work = NULL;
812 spin_lock_bh(&bdi->wb_lock);
813 if (!list_empty(&bdi->work_list)) {
814 work = list_entry(bdi->work_list.next,
815 struct wb_writeback_work, list);
816 list_del_init(&work->list);
818 spin_unlock_bh(&bdi->wb_lock);
823 * Add in the number of potentially dirty inodes, because each inode
824 * write can dirty pagecache in the underlying blockdev.
826 static unsigned long get_nr_dirty_pages(void)
828 return global_page_state(NR_FILE_DIRTY) +
829 global_page_state(NR_UNSTABLE_NFS) +
830 get_nr_dirty_inodes();
833 static long wb_check_background_flush(struct bdi_writeback *wb)
835 if (over_bground_thresh()) {
837 struct wb_writeback_work work = {
838 .nr_pages = LONG_MAX,
839 .sync_mode = WB_SYNC_NONE,
844 return wb_writeback(wb, &work);
850 static long wb_check_old_data_flush(struct bdi_writeback *wb)
852 unsigned long expired;
856 * When set to zero, disable periodic writeback
858 if (!dirty_writeback_interval)
861 expired = wb->last_old_flush +
862 msecs_to_jiffies(dirty_writeback_interval * 10);
863 if (time_before(jiffies, expired))
866 wb->last_old_flush = jiffies;
867 nr_pages = get_nr_dirty_pages();
870 struct wb_writeback_work work = {
871 .nr_pages = nr_pages,
872 .sync_mode = WB_SYNC_NONE,
877 return wb_writeback(wb, &work);
884 * Retrieve work items and do the writeback they describe
886 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
888 struct backing_dev_info *bdi = wb->bdi;
889 struct wb_writeback_work *work;
892 set_bit(BDI_writeback_running, &wb->bdi->state);
893 while ((work = get_next_work_item(bdi)) != NULL) {
895 * Override sync mode, in case we must wait for completion
896 * because this thread is exiting now.
899 work->sync_mode = WB_SYNC_ALL;
901 trace_writeback_exec(bdi, work);
903 wrote += wb_writeback(wb, work);
906 * Notify the caller of completion if this is a synchronous
907 * work item, otherwise just free it.
910 complete(work->done);
916 * Check for periodic writeback, kupdated() style
918 wrote += wb_check_old_data_flush(wb);
919 wrote += wb_check_background_flush(wb);
920 clear_bit(BDI_writeback_running, &wb->bdi->state);
926 * Handle writeback of dirty data for the device backed by this bdi. Also
927 * wakes up periodically and does kupdated style flushing.
929 int bdi_writeback_thread(void *data)
931 struct bdi_writeback *wb = data;
932 struct backing_dev_info *bdi = wb->bdi;
935 current->flags |= PF_SWAPWRITE;
937 wb->last_active = jiffies;
940 * Our parent may run at a different priority, just set us to normal
942 set_user_nice(current, 0);
944 trace_writeback_thread_start(bdi);
946 while (!kthread_should_stop()) {
948 * Remove own delayed wake-up timer, since we are already awake
949 * and we'll take care of the preriodic write-back.
951 del_timer(&wb->wakeup_timer);
953 pages_written = wb_do_writeback(wb, 0);
955 trace_writeback_pages_written(pages_written);
958 wb->last_active = jiffies;
960 set_current_state(TASK_INTERRUPTIBLE);
961 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
962 __set_current_state(TASK_RUNNING);
966 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
967 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
970 * We have nothing to do, so can go sleep without any
971 * timeout and save power. When a work is queued or
972 * something is made dirty - we will be woken up.
980 /* Flush any work that raced with us exiting */
981 if (!list_empty(&bdi->work_list))
982 wb_do_writeback(wb, 1);
984 trace_writeback_thread_stop(bdi);
990 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
993 void wakeup_flusher_threads(long nr_pages)
995 struct backing_dev_info *bdi;
998 nr_pages = global_page_state(NR_FILE_DIRTY) +
999 global_page_state(NR_UNSTABLE_NFS);
1003 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1004 if (!bdi_has_dirty_io(bdi))
1006 __bdi_start_writeback(bdi, nr_pages, false);
1011 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1013 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1014 struct dentry *dentry;
1015 const char *name = "?";
1017 dentry = d_find_alias(inode);
1019 spin_lock(&dentry->d_lock);
1020 name = (const char *) dentry->d_name.name;
1023 "%s(%d): dirtied inode %lu (%s) on %s\n",
1024 current->comm, task_pid_nr(current), inode->i_ino,
1025 name, inode->i_sb->s_id);
1027 spin_unlock(&dentry->d_lock);
1034 * __mark_inode_dirty - internal function
1035 * @inode: inode to mark
1036 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1037 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1038 * mark_inode_dirty_sync.
1040 * Put the inode on the super block's dirty list.
1042 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1043 * dirty list only if it is hashed or if it refers to a blockdev.
1044 * If it was not hashed, it will never be added to the dirty list
1045 * even if it is later hashed, as it will have been marked dirty already.
1047 * In short, make sure you hash any inodes _before_ you start marking
1050 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1051 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1052 * the kernel-internal blockdev inode represents the dirtying time of the
1053 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1054 * page->mapping->host, so the page-dirtying time is recorded in the internal
1057 void __mark_inode_dirty(struct inode *inode, int flags)
1059 struct super_block *sb = inode->i_sb;
1060 struct backing_dev_info *bdi = NULL;
1063 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1064 * dirty the inode itself
1066 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1067 if (sb->s_op->dirty_inode)
1068 sb->s_op->dirty_inode(inode, flags);
1072 * make sure that changes are seen by all cpus before we test i_state
1077 /* avoid the locking if we can */
1078 if ((inode->i_state & flags) == flags)
1081 if (unlikely(block_dump))
1082 block_dump___mark_inode_dirty(inode);
1084 spin_lock(&inode->i_lock);
1085 if ((inode->i_state & flags) != flags) {
1086 const int was_dirty = inode->i_state & I_DIRTY;
1088 inode->i_state |= flags;
1091 * If the inode is being synced, just update its dirty state.
1092 * The unlocker will place the inode on the appropriate
1093 * superblock list, based upon its state.
1095 if (inode->i_state & I_SYNC)
1096 goto out_unlock_inode;
1099 * Only add valid (hashed) inodes to the superblock's
1100 * dirty list. Add blockdev inodes as well.
1102 if (!S_ISBLK(inode->i_mode)) {
1103 if (inode_unhashed(inode))
1104 goto out_unlock_inode;
1106 if (inode->i_state & I_FREEING)
1107 goto out_unlock_inode;
1110 * If the inode was already on b_dirty/b_io/b_more_io, don't
1111 * reposition it (that would break b_dirty time-ordering).
1114 bool wakeup_bdi = false;
1115 bdi = inode_to_bdi(inode);
1117 if (bdi_cap_writeback_dirty(bdi)) {
1118 WARN(!test_bit(BDI_registered, &bdi->state),
1119 "bdi-%s not registered\n", bdi->name);
1122 * If this is the first dirty inode for this
1123 * bdi, we have to wake-up the corresponding
1124 * bdi thread to make sure background
1125 * write-back happens later.
1127 if (!wb_has_dirty_io(&bdi->wb))
1131 spin_unlock(&inode->i_lock);
1132 spin_lock(&bdi->wb.list_lock);
1133 inode->dirtied_when = jiffies;
1134 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1135 spin_unlock(&bdi->wb.list_lock);
1138 bdi_wakeup_thread_delayed(bdi);
1143 spin_unlock(&inode->i_lock);
1146 EXPORT_SYMBOL(__mark_inode_dirty);
1149 * Write out a superblock's list of dirty inodes. A wait will be performed
1150 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1152 * If older_than_this is non-NULL, then only write out inodes which
1153 * had their first dirtying at a time earlier than *older_than_this.
1155 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1156 * This function assumes that the blockdev superblock's inodes are backed by
1157 * a variety of queues, so all inodes are searched. For other superblocks,
1158 * assume that all inodes are backed by the same queue.
1160 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1161 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1162 * on the writer throttling path, and we get decent balancing between many
1163 * throttled threads: we don't want them all piling up on inode_sync_wait.
1165 static void wait_sb_inodes(struct super_block *sb)
1167 struct inode *inode, *old_inode = NULL;
1170 * We need to be protected against the filesystem going from
1171 * r/o to r/w or vice versa.
1173 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1175 spin_lock(&inode_sb_list_lock);
1178 * Data integrity sync. Must wait for all pages under writeback,
1179 * because there may have been pages dirtied before our sync
1180 * call, but which had writeout started before we write it out.
1181 * In which case, the inode may not be on the dirty list, but
1182 * we still have to wait for that writeout.
1184 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1185 struct address_space *mapping = inode->i_mapping;
1187 spin_lock(&inode->i_lock);
1188 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1189 (mapping->nrpages == 0)) {
1190 spin_unlock(&inode->i_lock);
1194 spin_unlock(&inode->i_lock);
1195 spin_unlock(&inode_sb_list_lock);
1198 * We hold a reference to 'inode' so it couldn't have been
1199 * removed from s_inodes list while we dropped the
1200 * inode_sb_list_lock. We cannot iput the inode now as we can
1201 * be holding the last reference and we cannot iput it under
1202 * inode_sb_list_lock. So we keep the reference and iput it
1208 filemap_fdatawait(mapping);
1212 spin_lock(&inode_sb_list_lock);
1214 spin_unlock(&inode_sb_list_lock);
1219 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1220 * @sb: the superblock
1221 * @nr: the number of pages to write
1223 * Start writeback on some inodes on this super_block. No guarantees are made
1224 * on how many (if any) will be written, and this function does not wait
1225 * for IO completion of submitted IO.
1227 void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
1229 DECLARE_COMPLETION_ONSTACK(done);
1230 struct wb_writeback_work work = {
1232 .sync_mode = WB_SYNC_NONE,
1233 .tagged_writepages = 1,
1238 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1239 bdi_queue_work(sb->s_bdi, &work);
1240 wait_for_completion(&done);
1242 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1245 * writeback_inodes_sb - writeback dirty inodes from given super_block
1246 * @sb: the superblock
1248 * Start writeback on some inodes on this super_block. No guarantees are made
1249 * on how many (if any) will be written, and this function does not wait
1250 * for IO completion of submitted IO.
1252 void writeback_inodes_sb(struct super_block *sb)
1254 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages());
1256 EXPORT_SYMBOL(writeback_inodes_sb);
1259 * writeback_inodes_sb_if_idle - start writeback if none underway
1260 * @sb: the superblock
1262 * Invoke writeback_inodes_sb if no writeback is currently underway.
1263 * Returns 1 if writeback was started, 0 if not.
1265 int writeback_inodes_sb_if_idle(struct super_block *sb)
1267 if (!writeback_in_progress(sb->s_bdi)) {
1268 down_read(&sb->s_umount);
1269 writeback_inodes_sb(sb);
1270 up_read(&sb->s_umount);
1275 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1278 * writeback_inodes_sb_if_idle - start writeback if none underway
1279 * @sb: the superblock
1280 * @nr: the number of pages to write
1282 * Invoke writeback_inodes_sb if no writeback is currently underway.
1283 * Returns 1 if writeback was started, 0 if not.
1285 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1288 if (!writeback_in_progress(sb->s_bdi)) {
1289 down_read(&sb->s_umount);
1290 writeback_inodes_sb_nr(sb, nr);
1291 up_read(&sb->s_umount);
1296 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1299 * sync_inodes_sb - sync sb inode pages
1300 * @sb: the superblock
1302 * This function writes and waits on any dirty inode belonging to this
1305 void sync_inodes_sb(struct super_block *sb)
1307 DECLARE_COMPLETION_ONSTACK(done);
1308 struct wb_writeback_work work = {
1310 .sync_mode = WB_SYNC_ALL,
1311 .nr_pages = LONG_MAX,
1316 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1318 bdi_queue_work(sb->s_bdi, &work);
1319 wait_for_completion(&done);
1323 EXPORT_SYMBOL(sync_inodes_sb);
1326 * write_inode_now - write an inode to disk
1327 * @inode: inode to write to disk
1328 * @sync: whether the write should be synchronous or not
1330 * This function commits an inode to disk immediately if it is dirty. This is
1331 * primarily needed by knfsd.
1333 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1335 int write_inode_now(struct inode *inode, int sync)
1337 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1339 struct writeback_control wbc = {
1340 .nr_to_write = LONG_MAX,
1341 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1343 .range_end = LLONG_MAX,
1346 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1347 wbc.nr_to_write = 0;
1350 spin_lock(&wb->list_lock);
1351 spin_lock(&inode->i_lock);
1352 ret = writeback_single_inode(inode, wb, &wbc);
1353 spin_unlock(&inode->i_lock);
1354 spin_unlock(&wb->list_lock);
1356 inode_sync_wait(inode);
1359 EXPORT_SYMBOL(write_inode_now);
1362 * sync_inode - write an inode and its pages to disk.
1363 * @inode: the inode to sync
1364 * @wbc: controls the writeback mode
1366 * sync_inode() will write an inode and its pages to disk. It will also
1367 * correctly update the inode on its superblock's dirty inode lists and will
1368 * update inode->i_state.
1370 * The caller must have a ref on the inode.
1372 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1374 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1377 spin_lock(&wb->list_lock);
1378 spin_lock(&inode->i_lock);
1379 ret = writeback_single_inode(inode, wb, wbc);
1380 spin_unlock(&inode->i_lock);
1381 spin_unlock(&wb->list_lock);
1384 EXPORT_SYMBOL(sync_inode);
1387 * sync_inode_metadata - write an inode to disk
1388 * @inode: the inode to sync
1389 * @wait: wait for I/O to complete.
1391 * Write an inode to disk and adjust its dirty state after completion.
1393 * Note: only writes the actual inode, no associated data or other metadata.
1395 int sync_inode_metadata(struct inode *inode, int wait)
1397 struct writeback_control wbc = {
1398 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1399 .nr_to_write = 0, /* metadata-only */
1402 return sync_inode(inode, &wbc);
1404 EXPORT_SYMBOL(sync_inode_metadata);