4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/freezer.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
33 * 4MB minimal write chunk size
35 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
38 * Passed into wb_writeback(), essentially a subset of writeback_control
40 struct wb_writeback_work {
42 struct super_block *sb;
43 unsigned long *older_than_this;
44 enum writeback_sync_modes sync_mode;
45 unsigned int tagged_writepages:1;
46 unsigned int for_kupdate:1;
47 unsigned int range_cyclic:1;
48 unsigned int for_background:1;
49 enum wb_reason reason; /* why was writeback initiated? */
51 struct list_head list; /* pending work list */
52 struct completion *done; /* set if the caller waits */
56 * We don't actually have pdflush, but this one is exported though /proc...
58 int nr_pdflush_threads;
61 * writeback_in_progress - determine whether there is writeback in progress
62 * @bdi: the device's backing_dev_info structure.
64 * Determine whether there is writeback waiting to be handled against a
67 int writeback_in_progress(struct backing_dev_info *bdi)
69 return test_bit(BDI_writeback_running, &bdi->state);
71 EXPORT_SYMBOL(writeback_in_progress);
73 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
75 struct super_block *sb = inode->i_sb;
77 if (strcmp(sb->s_type->name, "bdev") == 0)
78 return inode->i_mapping->backing_dev_info;
83 static inline struct inode *wb_inode(struct list_head *head)
85 return list_entry(head, struct inode, i_wb_list);
89 * Include the creation of the trace points after defining the
90 * wb_writeback_work structure and inline functions so that the definition
91 * remains local to this file.
93 #define CREATE_TRACE_POINTS
94 #include <trace/events/writeback.h>
96 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
97 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
100 wake_up_process(bdi->wb.task);
103 * The bdi thread isn't there, wake up the forker thread which
104 * will create and run it.
106 wake_up_process(default_backing_dev_info.wb.task);
110 static void bdi_queue_work(struct backing_dev_info *bdi,
111 struct wb_writeback_work *work)
113 trace_writeback_queue(bdi, work);
115 spin_lock_bh(&bdi->wb_lock);
116 list_add_tail(&work->list, &bdi->work_list);
118 trace_writeback_nothread(bdi, work);
119 bdi_wakeup_flusher(bdi);
120 spin_unlock_bh(&bdi->wb_lock);
124 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
125 bool range_cyclic, enum wb_reason reason)
127 struct wb_writeback_work *work;
130 * This is WB_SYNC_NONE writeback, so if allocation fails just
131 * wakeup the thread for old dirty data writeback
133 work = kzalloc(sizeof(*work), GFP_ATOMIC);
136 trace_writeback_nowork(bdi);
137 wake_up_process(bdi->wb.task);
142 work->sync_mode = WB_SYNC_NONE;
143 work->nr_pages = nr_pages;
144 work->range_cyclic = range_cyclic;
145 work->reason = reason;
147 bdi_queue_work(bdi, work);
151 * bdi_start_writeback - start writeback
152 * @bdi: the backing device to write from
153 * @nr_pages: the number of pages to write
154 * @reason: reason why some writeback work was initiated
157 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
158 * started when this function returns, we make no guarantees on
159 * completion. Caller need not hold sb s_umount semaphore.
162 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
163 enum wb_reason reason)
165 __bdi_start_writeback(bdi, nr_pages, true, reason);
169 * bdi_start_background_writeback - start background writeback
170 * @bdi: the backing device to write from
173 * This makes sure WB_SYNC_NONE background writeback happens. When
174 * this function returns, it is only guaranteed that for given BDI
175 * some IO is happening if we are over background dirty threshold.
176 * Caller need not hold sb s_umount semaphore.
178 void bdi_start_background_writeback(struct backing_dev_info *bdi)
181 * We just wake up the flusher thread. It will perform background
182 * writeback as soon as there is no other work to do.
184 trace_writeback_wake_background(bdi);
185 spin_lock_bh(&bdi->wb_lock);
186 bdi_wakeup_flusher(bdi);
187 spin_unlock_bh(&bdi->wb_lock);
191 * Remove the inode from the writeback list it is on.
193 void inode_wb_list_del(struct inode *inode)
195 struct backing_dev_info *bdi = inode_to_bdi(inode);
197 spin_lock(&bdi->wb.list_lock);
198 list_del_init(&inode->i_wb_list);
199 spin_unlock(&bdi->wb.list_lock);
203 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
204 * furthest end of its superblock's dirty-inode list.
206 * Before stamping the inode's ->dirtied_when, we check to see whether it is
207 * already the most-recently-dirtied inode on the b_dirty list. If that is
208 * the case then the inode must have been redirtied while it was being written
209 * out and we don't reset its dirtied_when.
211 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
213 assert_spin_locked(&wb->list_lock);
214 if (!list_empty(&wb->b_dirty)) {
217 tail = wb_inode(wb->b_dirty.next);
218 if (time_before(inode->dirtied_when, tail->dirtied_when))
219 inode->dirtied_when = jiffies;
221 list_move(&inode->i_wb_list, &wb->b_dirty);
225 * requeue inode for re-scanning after bdi->b_io list is exhausted.
227 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
229 assert_spin_locked(&wb->list_lock);
230 list_move(&inode->i_wb_list, &wb->b_more_io);
233 static void inode_sync_complete(struct inode *inode)
236 * Prevent speculative execution through
237 * spin_unlock(&wb->list_lock);
241 wake_up_bit(&inode->i_state, __I_SYNC);
244 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
246 bool ret = time_after(inode->dirtied_when, t);
249 * For inodes being constantly redirtied, dirtied_when can get stuck.
250 * It _appears_ to be in the future, but is actually in distant past.
251 * This test is necessary to prevent such wrapped-around relative times
252 * from permanently stopping the whole bdi writeback.
254 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
260 * Move expired (dirtied after work->older_than_this) dirty inodes from
261 * @delaying_queue to @dispatch_queue.
263 static int move_expired_inodes(struct list_head *delaying_queue,
264 struct list_head *dispatch_queue,
265 struct wb_writeback_work *work)
268 struct list_head *pos, *node;
269 struct super_block *sb = NULL;
274 while (!list_empty(delaying_queue)) {
275 inode = wb_inode(delaying_queue->prev);
276 if (work->older_than_this &&
277 inode_dirtied_after(inode, *work->older_than_this))
279 if (sb && sb != inode->i_sb)
282 list_move(&inode->i_wb_list, &tmp);
286 /* just one sb in list, splice to dispatch_queue and we're done */
288 list_splice(&tmp, dispatch_queue);
292 /* Move inodes from one superblock together */
293 while (!list_empty(&tmp)) {
294 sb = wb_inode(tmp.prev)->i_sb;
295 list_for_each_prev_safe(pos, node, &tmp) {
296 inode = wb_inode(pos);
297 if (inode->i_sb == sb)
298 list_move(&inode->i_wb_list, dispatch_queue);
306 * Queue all expired dirty inodes for io, eldest first.
308 * newly dirtied b_dirty b_io b_more_io
309 * =============> gf edc BA
311 * newly dirtied b_dirty b_io b_more_io
312 * =============> g fBAedc
314 * +--> dequeue for IO
316 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
319 assert_spin_locked(&wb->list_lock);
320 list_splice_init(&wb->b_more_io, &wb->b_io);
321 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
322 trace_writeback_queue_io(wb, work, moved);
325 static int write_inode(struct inode *inode, struct writeback_control *wbc)
327 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
328 return inode->i_sb->s_op->write_inode(inode, wbc);
333 * Wait for writeback on an inode to complete.
335 static void inode_wait_for_writeback(struct inode *inode,
336 struct bdi_writeback *wb)
338 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
339 wait_queue_head_t *wqh;
341 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
342 while (inode->i_state & I_SYNC) {
343 spin_unlock(&inode->i_lock);
344 spin_unlock(&wb->list_lock);
345 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
346 spin_lock(&wb->list_lock);
347 spin_lock(&inode->i_lock);
352 * Write out an inode's dirty pages. Called under wb->list_lock and
353 * inode->i_lock. Either the caller has an active reference on the inode or
354 * the inode has I_WILL_FREE set.
356 * If `wait' is set, wait on the writeout.
358 * The whole writeout design is quite complex and fragile. We want to avoid
359 * starvation of particular inodes when others are being redirtied, prevent
363 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
364 struct writeback_control *wbc)
366 struct address_space *mapping = inode->i_mapping;
367 long nr_to_write = wbc->nr_to_write;
371 assert_spin_locked(&wb->list_lock);
372 assert_spin_locked(&inode->i_lock);
374 if (!atomic_read(&inode->i_count))
375 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
377 WARN_ON(inode->i_state & I_WILL_FREE);
379 if (inode->i_state & I_SYNC) {
381 * If this inode is locked for writeback and we are not doing
382 * writeback-for-data-integrity, move it to b_more_io so that
383 * writeback can proceed with the other inodes on s_io.
385 * We'll have another go at writing back this inode when we
386 * completed a full scan of b_io.
388 if (wbc->sync_mode != WB_SYNC_ALL) {
389 requeue_io(inode, wb);
390 trace_writeback_single_inode_requeue(inode, wbc,
396 * It's a data-integrity sync. We must wait.
398 inode_wait_for_writeback(inode, wb);
401 BUG_ON(inode->i_state & I_SYNC);
403 /* Set I_SYNC, reset I_DIRTY_PAGES */
404 inode->i_state |= I_SYNC;
405 inode->i_state &= ~I_DIRTY_PAGES;
406 spin_unlock(&inode->i_lock);
407 spin_unlock(&wb->list_lock);
409 ret = do_writepages(mapping, wbc);
412 * Make sure to wait on the data before writing out the metadata.
413 * This is important for filesystems that modify metadata on data
416 if (wbc->sync_mode == WB_SYNC_ALL) {
417 int err = filemap_fdatawait(mapping);
423 * Some filesystems may redirty the inode during the writeback
424 * due to delalloc, clear dirty metadata flags right before
427 spin_lock(&inode->i_lock);
428 dirty = inode->i_state & I_DIRTY;
429 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
430 spin_unlock(&inode->i_lock);
431 /* Don't write the inode if only I_DIRTY_PAGES was set */
432 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
433 int err = write_inode(inode, wbc);
438 spin_lock(&wb->list_lock);
439 spin_lock(&inode->i_lock);
440 inode->i_state &= ~I_SYNC;
441 if (!(inode->i_state & I_FREEING)) {
443 * Sync livelock prevention. Each inode is tagged and synced in
444 * one shot. If still dirty, it will be redirty_tail()'ed below.
445 * Update the dirty time to prevent enqueue and sync it again.
447 if ((inode->i_state & I_DIRTY) &&
448 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
449 inode->dirtied_when = jiffies;
451 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
453 * We didn't write back all the pages. nfs_writepages()
454 * sometimes bales out without doing anything.
456 inode->i_state |= I_DIRTY_PAGES;
457 if (wbc->nr_to_write <= 0) {
459 * slice used up: queue for next turn
461 requeue_io(inode, wb);
464 * Writeback blocked by something other than
465 * congestion. Delay the inode for some time to
466 * avoid spinning on the CPU (100% iowait)
467 * retrying writeback of the dirty page/inode
468 * that cannot be performed immediately.
470 redirty_tail(inode, wb);
472 } else if (inode->i_state & I_DIRTY) {
474 * Filesystems can dirty the inode during writeback
475 * operations, such as delayed allocation during
476 * submission or metadata updates after data IO
479 redirty_tail(inode, wb);
482 * The inode is clean. At this point we either have
483 * a reference to the inode or it's on it's way out.
484 * No need to add it back to the LRU.
486 list_del_init(&inode->i_wb_list);
489 inode_sync_complete(inode);
490 trace_writeback_single_inode(inode, wbc, nr_to_write);
494 static long writeback_chunk_size(struct backing_dev_info *bdi,
495 struct wb_writeback_work *work)
500 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
501 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
502 * here avoids calling into writeback_inodes_wb() more than once.
504 * The intended call sequence for WB_SYNC_ALL writeback is:
507 * writeback_sb_inodes() <== called only once
508 * write_cache_pages() <== called once for each inode
509 * (quickly) tag currently dirty pages
510 * (maybe slowly) sync all tagged pages
512 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
515 pages = min(bdi->avg_write_bandwidth / 2,
516 global_dirty_limit / DIRTY_SCOPE);
517 pages = min(pages, work->nr_pages);
518 pages = round_down(pages + MIN_WRITEBACK_PAGES,
519 MIN_WRITEBACK_PAGES);
526 * Write a portion of b_io inodes which belong to @sb.
528 * If @only_this_sb is true, then find and write all such
529 * inodes. Otherwise write only ones which go sequentially
532 * Return the number of pages and/or inodes written.
534 static long writeback_sb_inodes(struct super_block *sb,
535 struct bdi_writeback *wb,
536 struct wb_writeback_work *work)
538 struct writeback_control wbc = {
539 .sync_mode = work->sync_mode,
540 .tagged_writepages = work->tagged_writepages,
541 .for_kupdate = work->for_kupdate,
542 .for_background = work->for_background,
543 .range_cyclic = work->range_cyclic,
545 .range_end = LLONG_MAX,
547 unsigned long start_time = jiffies;
549 long wrote = 0; /* count both pages and inodes */
551 while (!list_empty(&wb->b_io)) {
552 struct inode *inode = wb_inode(wb->b_io.prev);
554 if (inode->i_sb != sb) {
557 * We only want to write back data for this
558 * superblock, move all inodes not belonging
559 * to it back onto the dirty list.
561 redirty_tail(inode, wb);
566 * The inode belongs to a different superblock.
567 * Bounce back to the caller to unpin this and
568 * pin the next superblock.
574 * Don't bother with new inodes or inodes beeing freed, first
575 * kind does not need peridic writeout yet, and for the latter
576 * kind writeout is handled by the freer.
578 spin_lock(&inode->i_lock);
579 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
580 spin_unlock(&inode->i_lock);
581 redirty_tail(inode, wb);
585 write_chunk = writeback_chunk_size(wb->bdi, work);
586 wbc.nr_to_write = write_chunk;
587 wbc.pages_skipped = 0;
589 writeback_single_inode(inode, wb, &wbc);
591 work->nr_pages -= write_chunk - wbc.nr_to_write;
592 wrote += write_chunk - wbc.nr_to_write;
593 if (!(inode->i_state & I_DIRTY))
595 if (wbc.pages_skipped) {
597 * writeback is not making progress due to locked
598 * buffers. Skip this inode for now.
600 redirty_tail(inode, wb);
602 spin_unlock(&inode->i_lock);
603 spin_unlock(&wb->list_lock);
606 spin_lock(&wb->list_lock);
608 * bail out to wb_writeback() often enough to check
609 * background threshold and other termination conditions.
612 if (time_is_before_jiffies(start_time + HZ / 10UL))
614 if (work->nr_pages <= 0)
621 static long __writeback_inodes_wb(struct bdi_writeback *wb,
622 struct wb_writeback_work *work)
624 unsigned long start_time = jiffies;
627 while (!list_empty(&wb->b_io)) {
628 struct inode *inode = wb_inode(wb->b_io.prev);
629 struct super_block *sb = inode->i_sb;
631 if (!grab_super_passive(sb)) {
633 * grab_super_passive() may fail consistently due to
634 * s_umount being grabbed by someone else. Don't use
635 * requeue_io() to avoid busy retrying the inode/sb.
637 redirty_tail(inode, wb);
640 wrote += writeback_sb_inodes(sb, wb, work);
643 /* refer to the same tests at the end of writeback_sb_inodes */
645 if (time_is_before_jiffies(start_time + HZ / 10UL))
647 if (work->nr_pages <= 0)
651 /* Leave any unwritten inodes on b_io */
655 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
656 enum wb_reason reason)
658 struct wb_writeback_work work = {
659 .nr_pages = nr_pages,
660 .sync_mode = WB_SYNC_NONE,
665 spin_lock(&wb->list_lock);
666 if (list_empty(&wb->b_io))
668 __writeback_inodes_wb(wb, &work);
669 spin_unlock(&wb->list_lock);
671 return nr_pages - work.nr_pages;
674 static bool over_bground_thresh(struct backing_dev_info *bdi)
676 unsigned long background_thresh, dirty_thresh;
678 global_dirty_limits(&background_thresh, &dirty_thresh);
680 if (global_page_state(NR_FILE_DIRTY) +
681 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
684 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
685 bdi_dirty_limit(bdi, background_thresh))
692 * Called under wb->list_lock. If there are multiple wb per bdi,
693 * only the flusher working on the first wb should do it.
695 static void wb_update_bandwidth(struct bdi_writeback *wb,
696 unsigned long start_time)
698 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
702 * Explicit flushing or periodic writeback of "old" data.
704 * Define "old": the first time one of an inode's pages is dirtied, we mark the
705 * dirtying-time in the inode's address_space. So this periodic writeback code
706 * just walks the superblock inode list, writing back any inodes which are
707 * older than a specific point in time.
709 * Try to run once per dirty_writeback_interval. But if a writeback event
710 * takes longer than a dirty_writeback_interval interval, then leave a
713 * older_than_this takes precedence over nr_to_write. So we'll only write back
714 * all dirty pages if they are all attached to "old" mappings.
716 static long wb_writeback(struct bdi_writeback *wb,
717 struct wb_writeback_work *work)
719 unsigned long wb_start = jiffies;
720 long nr_pages = work->nr_pages;
721 unsigned long oldest_jif;
725 oldest_jif = jiffies;
726 work->older_than_this = &oldest_jif;
728 spin_lock(&wb->list_lock);
731 * Stop writeback when nr_pages has been consumed
733 if (work->nr_pages <= 0)
737 * Background writeout and kupdate-style writeback may
738 * run forever. Stop them if there is other work to do
739 * so that e.g. sync can proceed. They'll be restarted
740 * after the other works are all done.
742 if ((work->for_background || work->for_kupdate) &&
743 !list_empty(&wb->bdi->work_list))
747 * For background writeout, stop when we are below the
748 * background dirty threshold
750 if (work->for_background && !over_bground_thresh(wb->bdi))
754 * Kupdate and background works are special and we want to
755 * include all inodes that need writing. Livelock avoidance is
756 * handled by these works yielding to any other work so we are
759 if (work->for_kupdate) {
760 oldest_jif = jiffies -
761 msecs_to_jiffies(dirty_expire_interval * 10);
762 } else if (work->for_background)
763 oldest_jif = jiffies;
765 trace_writeback_start(wb->bdi, work);
766 if (list_empty(&wb->b_io))
769 progress = writeback_sb_inodes(work->sb, wb, work);
771 progress = __writeback_inodes_wb(wb, work);
772 trace_writeback_written(wb->bdi, work);
774 wb_update_bandwidth(wb, wb_start);
777 * Did we write something? Try for more
779 * Dirty inodes are moved to b_io for writeback in batches.
780 * The completion of the current batch does not necessarily
781 * mean the overall work is done. So we keep looping as long
782 * as made some progress on cleaning pages or inodes.
787 * No more inodes for IO, bail
789 if (list_empty(&wb->b_more_io))
792 * Nothing written. Wait for some inode to
793 * become available for writeback. Otherwise
794 * we'll just busyloop.
796 if (!list_empty(&wb->b_more_io)) {
797 trace_writeback_wait(wb->bdi, work);
798 inode = wb_inode(wb->b_more_io.prev);
799 spin_lock(&inode->i_lock);
800 inode_wait_for_writeback(inode, wb);
801 spin_unlock(&inode->i_lock);
804 spin_unlock(&wb->list_lock);
806 return nr_pages - work->nr_pages;
810 * Return the next wb_writeback_work struct that hasn't been processed yet.
812 static struct wb_writeback_work *
813 get_next_work_item(struct backing_dev_info *bdi)
815 struct wb_writeback_work *work = NULL;
817 spin_lock_bh(&bdi->wb_lock);
818 if (!list_empty(&bdi->work_list)) {
819 work = list_entry(bdi->work_list.next,
820 struct wb_writeback_work, list);
821 list_del_init(&work->list);
823 spin_unlock_bh(&bdi->wb_lock);
828 * Add in the number of potentially dirty inodes, because each inode
829 * write can dirty pagecache in the underlying blockdev.
831 static unsigned long get_nr_dirty_pages(void)
833 return global_page_state(NR_FILE_DIRTY) +
834 global_page_state(NR_UNSTABLE_NFS) +
835 get_nr_dirty_inodes();
838 static long wb_check_background_flush(struct bdi_writeback *wb)
840 if (over_bground_thresh(wb->bdi)) {
842 struct wb_writeback_work work = {
843 .nr_pages = LONG_MAX,
844 .sync_mode = WB_SYNC_NONE,
847 .reason = WB_REASON_BACKGROUND,
850 return wb_writeback(wb, &work);
856 static long wb_check_old_data_flush(struct bdi_writeback *wb)
858 unsigned long expired;
862 * When set to zero, disable periodic writeback
864 if (!dirty_writeback_interval)
867 expired = wb->last_old_flush +
868 msecs_to_jiffies(dirty_writeback_interval * 10);
869 if (time_before(jiffies, expired))
872 wb->last_old_flush = jiffies;
873 nr_pages = get_nr_dirty_pages();
876 struct wb_writeback_work work = {
877 .nr_pages = nr_pages,
878 .sync_mode = WB_SYNC_NONE,
881 .reason = WB_REASON_PERIODIC,
884 return wb_writeback(wb, &work);
891 * Retrieve work items and do the writeback they describe
893 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
895 struct backing_dev_info *bdi = wb->bdi;
896 struct wb_writeback_work *work;
899 set_bit(BDI_writeback_running, &wb->bdi->state);
900 while ((work = get_next_work_item(bdi)) != NULL) {
902 * Override sync mode, in case we must wait for completion
903 * because this thread is exiting now.
906 work->sync_mode = WB_SYNC_ALL;
908 trace_writeback_exec(bdi, work);
910 wrote += wb_writeback(wb, work);
913 * Notify the caller of completion if this is a synchronous
914 * work item, otherwise just free it.
917 complete(work->done);
923 * Check for periodic writeback, kupdated() style
925 wrote += wb_check_old_data_flush(wb);
926 wrote += wb_check_background_flush(wb);
927 clear_bit(BDI_writeback_running, &wb->bdi->state);
933 * Handle writeback of dirty data for the device backed by this bdi. Also
934 * wakes up periodically and does kupdated style flushing.
936 int bdi_writeback_thread(void *data)
938 struct bdi_writeback *wb = data;
939 struct backing_dev_info *bdi = wb->bdi;
942 current->flags |= PF_SWAPWRITE;
944 wb->last_active = jiffies;
947 * Our parent may run at a different priority, just set us to normal
949 set_user_nice(current, 0);
951 trace_writeback_thread_start(bdi);
953 while (!kthread_freezable_should_stop(NULL)) {
955 * Remove own delayed wake-up timer, since we are already awake
956 * and we'll take care of the preriodic write-back.
958 del_timer(&wb->wakeup_timer);
960 pages_written = wb_do_writeback(wb, 0);
962 trace_writeback_pages_written(pages_written);
965 wb->last_active = jiffies;
967 set_current_state(TASK_INTERRUPTIBLE);
968 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
969 __set_current_state(TASK_RUNNING);
973 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
974 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
977 * We have nothing to do, so can go sleep without any
978 * timeout and save power. When a work is queued or
979 * something is made dirty - we will be woken up.
985 /* Flush any work that raced with us exiting */
986 if (!list_empty(&bdi->work_list))
987 wb_do_writeback(wb, 1);
989 trace_writeback_thread_stop(bdi);
995 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
998 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1000 struct backing_dev_info *bdi;
1003 nr_pages = global_page_state(NR_FILE_DIRTY) +
1004 global_page_state(NR_UNSTABLE_NFS);
1008 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1009 if (!bdi_has_dirty_io(bdi))
1011 __bdi_start_writeback(bdi, nr_pages, false, reason);
1016 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1018 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1019 struct dentry *dentry;
1020 const char *name = "?";
1022 dentry = d_find_alias(inode);
1024 spin_lock(&dentry->d_lock);
1025 name = (const char *) dentry->d_name.name;
1028 "%s(%d): dirtied inode %lu (%s) on %s\n",
1029 current->comm, task_pid_nr(current), inode->i_ino,
1030 name, inode->i_sb->s_id);
1032 spin_unlock(&dentry->d_lock);
1039 * __mark_inode_dirty - internal function
1040 * @inode: inode to mark
1041 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1042 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1043 * mark_inode_dirty_sync.
1045 * Put the inode on the super block's dirty list.
1047 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1048 * dirty list only if it is hashed or if it refers to a blockdev.
1049 * If it was not hashed, it will never be added to the dirty list
1050 * even if it is later hashed, as it will have been marked dirty already.
1052 * In short, make sure you hash any inodes _before_ you start marking
1055 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1056 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1057 * the kernel-internal blockdev inode represents the dirtying time of the
1058 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1059 * page->mapping->host, so the page-dirtying time is recorded in the internal
1062 void __mark_inode_dirty(struct inode *inode, int flags)
1064 struct super_block *sb = inode->i_sb;
1065 struct backing_dev_info *bdi = NULL;
1068 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1069 * dirty the inode itself
1071 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1072 if (sb->s_op->dirty_inode)
1073 sb->s_op->dirty_inode(inode, flags);
1077 * make sure that changes are seen by all cpus before we test i_state
1082 /* avoid the locking if we can */
1083 if ((inode->i_state & flags) == flags)
1086 if (unlikely(block_dump > 1))
1087 block_dump___mark_inode_dirty(inode);
1089 spin_lock(&inode->i_lock);
1090 if ((inode->i_state & flags) != flags) {
1091 const int was_dirty = inode->i_state & I_DIRTY;
1093 inode->i_state |= flags;
1096 * If the inode is being synced, just update its dirty state.
1097 * The unlocker will place the inode on the appropriate
1098 * superblock list, based upon its state.
1100 if (inode->i_state & I_SYNC)
1101 goto out_unlock_inode;
1104 * Only add valid (hashed) inodes to the superblock's
1105 * dirty list. Add blockdev inodes as well.
1107 if (!S_ISBLK(inode->i_mode)) {
1108 if (inode_unhashed(inode))
1109 goto out_unlock_inode;
1111 if (inode->i_state & I_FREEING)
1112 goto out_unlock_inode;
1115 * If the inode was already on b_dirty/b_io/b_more_io, don't
1116 * reposition it (that would break b_dirty time-ordering).
1119 bool wakeup_bdi = false;
1120 bdi = inode_to_bdi(inode);
1122 if (bdi_cap_writeback_dirty(bdi)) {
1123 WARN(!test_bit(BDI_registered, &bdi->state),
1124 "bdi-%s not registered\n", bdi->name);
1127 * If this is the first dirty inode for this
1128 * bdi, we have to wake-up the corresponding
1129 * bdi thread to make sure background
1130 * write-back happens later.
1132 if (!wb_has_dirty_io(&bdi->wb))
1136 spin_unlock(&inode->i_lock);
1137 spin_lock(&bdi->wb.list_lock);
1138 inode->dirtied_when = jiffies;
1139 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1140 spin_unlock(&bdi->wb.list_lock);
1143 bdi_wakeup_thread_delayed(bdi);
1148 spin_unlock(&inode->i_lock);
1151 EXPORT_SYMBOL(__mark_inode_dirty);
1153 static void wait_sb_inodes(struct super_block *sb)
1155 struct inode *inode, *old_inode = NULL;
1158 * We need to be protected against the filesystem going from
1159 * r/o to r/w or vice versa.
1161 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1163 spin_lock(&inode_sb_list_lock);
1166 * Data integrity sync. Must wait for all pages under writeback,
1167 * because there may have been pages dirtied before our sync
1168 * call, but which had writeout started before we write it out.
1169 * In which case, the inode may not be on the dirty list, but
1170 * we still have to wait for that writeout.
1172 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1173 struct address_space *mapping = inode->i_mapping;
1175 spin_lock(&inode->i_lock);
1176 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1177 (mapping->nrpages == 0)) {
1178 spin_unlock(&inode->i_lock);
1182 spin_unlock(&inode->i_lock);
1183 spin_unlock(&inode_sb_list_lock);
1186 * We hold a reference to 'inode' so it couldn't have been
1187 * removed from s_inodes list while we dropped the
1188 * inode_sb_list_lock. We cannot iput the inode now as we can
1189 * be holding the last reference and we cannot iput it under
1190 * inode_sb_list_lock. So we keep the reference and iput it
1196 filemap_fdatawait(mapping);
1200 spin_lock(&inode_sb_list_lock);
1202 spin_unlock(&inode_sb_list_lock);
1207 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1208 * @sb: the superblock
1209 * @nr: the number of pages to write
1210 * @reason: reason why some writeback work initiated
1212 * Start writeback on some inodes on this super_block. No guarantees are made
1213 * on how many (if any) will be written, and this function does not wait
1214 * for IO completion of submitted IO.
1216 void writeback_inodes_sb_nr(struct super_block *sb,
1218 enum wb_reason reason)
1220 DECLARE_COMPLETION_ONSTACK(done);
1221 struct wb_writeback_work work = {
1223 .sync_mode = WB_SYNC_NONE,
1224 .tagged_writepages = 1,
1230 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1231 bdi_queue_work(sb->s_bdi, &work);
1232 wait_for_completion(&done);
1234 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1237 * writeback_inodes_sb - writeback dirty inodes from given super_block
1238 * @sb: the superblock
1239 * @reason: reason why some writeback work was initiated
1241 * Start writeback on some inodes on this super_block. No guarantees are made
1242 * on how many (if any) will be written, and this function does not wait
1243 * for IO completion of submitted IO.
1245 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1247 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1249 EXPORT_SYMBOL(writeback_inodes_sb);
1252 * writeback_inodes_sb_if_idle - start writeback if none underway
1253 * @sb: the superblock
1254 * @reason: reason why some writeback work was initiated
1256 * Invoke writeback_inodes_sb if no writeback is currently underway.
1257 * Returns 1 if writeback was started, 0 if not.
1259 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1261 if (!writeback_in_progress(sb->s_bdi)) {
1262 down_read(&sb->s_umount);
1263 writeback_inodes_sb(sb, reason);
1264 up_read(&sb->s_umount);
1269 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1272 * writeback_inodes_sb_nr_if_idle - start writeback if none underway
1273 * @sb: the superblock
1274 * @nr: the number of pages to write
1275 * @reason: reason why some writeback work was initiated
1277 * Invoke writeback_inodes_sb if no writeback is currently underway.
1278 * Returns 1 if writeback was started, 0 if not.
1280 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1282 enum wb_reason reason)
1284 if (!writeback_in_progress(sb->s_bdi)) {
1285 down_read(&sb->s_umount);
1286 writeback_inodes_sb_nr(sb, nr, reason);
1287 up_read(&sb->s_umount);
1292 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1295 * sync_inodes_sb - sync sb inode pages
1296 * @sb: the superblock
1298 * This function writes and waits on any dirty inode belonging to this
1301 void sync_inodes_sb(struct super_block *sb)
1303 DECLARE_COMPLETION_ONSTACK(done);
1304 struct wb_writeback_work work = {
1306 .sync_mode = WB_SYNC_ALL,
1307 .nr_pages = LONG_MAX,
1310 .reason = WB_REASON_SYNC,
1313 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1315 bdi_queue_work(sb->s_bdi, &work);
1316 wait_for_completion(&done);
1320 EXPORT_SYMBOL(sync_inodes_sb);
1323 * write_inode_now - write an inode to disk
1324 * @inode: inode to write to disk
1325 * @sync: whether the write should be synchronous or not
1327 * This function commits an inode to disk immediately if it is dirty. This is
1328 * primarily needed by knfsd.
1330 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1332 int write_inode_now(struct inode *inode, int sync)
1334 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1336 struct writeback_control wbc = {
1337 .nr_to_write = LONG_MAX,
1338 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1340 .range_end = LLONG_MAX,
1343 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1344 wbc.nr_to_write = 0;
1347 spin_lock(&wb->list_lock);
1348 spin_lock(&inode->i_lock);
1349 ret = writeback_single_inode(inode, wb, &wbc);
1350 spin_unlock(&inode->i_lock);
1351 spin_unlock(&wb->list_lock);
1354 EXPORT_SYMBOL(write_inode_now);
1357 * sync_inode - write an inode and its pages to disk.
1358 * @inode: the inode to sync
1359 * @wbc: controls the writeback mode
1361 * sync_inode() will write an inode and its pages to disk. It will also
1362 * correctly update the inode on its superblock's dirty inode lists and will
1363 * update inode->i_state.
1365 * The caller must have a ref on the inode.
1367 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1369 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1372 spin_lock(&wb->list_lock);
1373 spin_lock(&inode->i_lock);
1374 ret = writeback_single_inode(inode, wb, wbc);
1375 spin_unlock(&inode->i_lock);
1376 spin_unlock(&wb->list_lock);
1379 EXPORT_SYMBOL(sync_inode);
1382 * sync_inode_metadata - write an inode to disk
1383 * @inode: the inode to sync
1384 * @wait: wait for I/O to complete.
1386 * Write an inode to disk and adjust its dirty state after completion.
1388 * Note: only writes the actual inode, no associated data or other metadata.
1390 int sync_inode_metadata(struct inode *inode, int wait)
1392 struct writeback_control wbc = {
1393 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1394 .nr_to_write = 0, /* metadata-only */
1397 return sync_inode(inode, &wbc);
1399 EXPORT_SYMBOL(sync_inode_metadata);