4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/freezer.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
33 * 4MB minimal write chunk size
35 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
38 * Passed into wb_writeback(), essentially a subset of writeback_control
40 struct wb_writeback_work {
42 struct super_block *sb;
43 unsigned long *older_than_this;
44 enum writeback_sync_modes sync_mode;
45 unsigned int tagged_writepages:1;
46 unsigned int for_kupdate:1;
47 unsigned int range_cyclic:1;
48 unsigned int for_background:1;
49 enum wb_reason reason; /* why was writeback initiated? */
51 struct list_head list; /* pending work list */
52 struct completion *done; /* set if the caller waits */
56 * We don't actually have pdflush, but this one is exported though /proc...
58 int nr_pdflush_threads;
61 * writeback_in_progress - determine whether there is writeback in progress
62 * @bdi: the device's backing_dev_info structure.
64 * Determine whether there is writeback waiting to be handled against a
67 int writeback_in_progress(struct backing_dev_info *bdi)
69 return test_bit(BDI_writeback_running, &bdi->state);
72 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
74 struct super_block *sb = inode->i_sb;
76 if (strcmp(sb->s_type->name, "bdev") == 0)
77 return inode->i_mapping->backing_dev_info;
82 static inline struct inode *wb_inode(struct list_head *head)
84 return list_entry(head, struct inode, i_wb_list);
88 * Include the creation of the trace points after defining the
89 * wb_writeback_work structure and inline functions so that the definition
90 * remains local to this file.
92 #define CREATE_TRACE_POINTS
93 #include <trace/events/writeback.h>
95 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
96 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
99 wake_up_process(bdi->wb.task);
102 * The bdi thread isn't there, wake up the forker thread which
103 * will create and run it.
105 wake_up_process(default_backing_dev_info.wb.task);
109 static void bdi_queue_work(struct backing_dev_info *bdi,
110 struct wb_writeback_work *work)
112 trace_writeback_queue(bdi, work);
114 spin_lock_bh(&bdi->wb_lock);
115 list_add_tail(&work->list, &bdi->work_list);
117 trace_writeback_nothread(bdi, work);
118 bdi_wakeup_flusher(bdi);
119 spin_unlock_bh(&bdi->wb_lock);
123 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
124 bool range_cyclic, enum wb_reason reason)
126 struct wb_writeback_work *work;
129 * This is WB_SYNC_NONE writeback, so if allocation fails just
130 * wakeup the thread for old dirty data writeback
132 work = kzalloc(sizeof(*work), GFP_ATOMIC);
135 trace_writeback_nowork(bdi);
136 wake_up_process(bdi->wb.task);
141 work->sync_mode = WB_SYNC_NONE;
142 work->nr_pages = nr_pages;
143 work->range_cyclic = range_cyclic;
144 work->reason = reason;
146 bdi_queue_work(bdi, work);
150 * bdi_start_writeback - start writeback
151 * @bdi: the backing device to write from
152 * @nr_pages: the number of pages to write
153 * @reason: reason why some writeback work was initiated
156 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
157 * started when this function returns, we make no guarantees on
158 * completion. Caller need not hold sb s_umount semaphore.
161 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
162 enum wb_reason reason)
164 __bdi_start_writeback(bdi, nr_pages, true, reason);
168 * bdi_start_background_writeback - start background writeback
169 * @bdi: the backing device to write from
172 * This makes sure WB_SYNC_NONE background writeback happens. When
173 * this function returns, it is only guaranteed that for given BDI
174 * some IO is happening if we are over background dirty threshold.
175 * Caller need not hold sb s_umount semaphore.
177 void bdi_start_background_writeback(struct backing_dev_info *bdi)
180 * We just wake up the flusher thread. It will perform background
181 * writeback as soon as there is no other work to do.
183 trace_writeback_wake_background(bdi);
184 spin_lock_bh(&bdi->wb_lock);
185 bdi_wakeup_flusher(bdi);
186 spin_unlock_bh(&bdi->wb_lock);
190 * Remove the inode from the writeback list it is on.
192 void inode_wb_list_del(struct inode *inode)
194 struct backing_dev_info *bdi = inode_to_bdi(inode);
196 spin_lock(&bdi->wb.list_lock);
197 list_del_init(&inode->i_wb_list);
198 spin_unlock(&bdi->wb.list_lock);
202 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
203 * furthest end of its superblock's dirty-inode list.
205 * Before stamping the inode's ->dirtied_when, we check to see whether it is
206 * already the most-recently-dirtied inode on the b_dirty list. If that is
207 * the case then the inode must have been redirtied while it was being written
208 * out and we don't reset its dirtied_when.
210 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
212 assert_spin_locked(&wb->list_lock);
213 if (!list_empty(&wb->b_dirty)) {
216 tail = wb_inode(wb->b_dirty.next);
217 if (time_before(inode->dirtied_when, tail->dirtied_when))
218 inode->dirtied_when = jiffies;
220 list_move(&inode->i_wb_list, &wb->b_dirty);
224 * requeue inode for re-scanning after bdi->b_io list is exhausted.
226 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
228 assert_spin_locked(&wb->list_lock);
229 list_move(&inode->i_wb_list, &wb->b_more_io);
232 static void inode_sync_complete(struct inode *inode)
234 inode->i_state &= ~I_SYNC;
235 /* Waiters must see I_SYNC cleared before being woken up */
237 wake_up_bit(&inode->i_state, __I_SYNC);
240 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
242 bool ret = time_after(inode->dirtied_when, t);
245 * For inodes being constantly redirtied, dirtied_when can get stuck.
246 * It _appears_ to be in the future, but is actually in distant past.
247 * This test is necessary to prevent such wrapped-around relative times
248 * from permanently stopping the whole bdi writeback.
250 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
256 * Move expired (dirtied after work->older_than_this) dirty inodes from
257 * @delaying_queue to @dispatch_queue.
259 static int move_expired_inodes(struct list_head *delaying_queue,
260 struct list_head *dispatch_queue,
261 struct wb_writeback_work *work)
264 struct list_head *pos, *node;
265 struct super_block *sb = NULL;
270 while (!list_empty(delaying_queue)) {
271 inode = wb_inode(delaying_queue->prev);
272 if (work->older_than_this &&
273 inode_dirtied_after(inode, *work->older_than_this))
275 if (sb && sb != inode->i_sb)
278 list_move(&inode->i_wb_list, &tmp);
282 /* just one sb in list, splice to dispatch_queue and we're done */
284 list_splice(&tmp, dispatch_queue);
288 /* Move inodes from one superblock together */
289 while (!list_empty(&tmp)) {
290 sb = wb_inode(tmp.prev)->i_sb;
291 list_for_each_prev_safe(pos, node, &tmp) {
292 inode = wb_inode(pos);
293 if (inode->i_sb == sb)
294 list_move(&inode->i_wb_list, dispatch_queue);
302 * Queue all expired dirty inodes for io, eldest first.
304 * newly dirtied b_dirty b_io b_more_io
305 * =============> gf edc BA
307 * newly dirtied b_dirty b_io b_more_io
308 * =============> g fBAedc
310 * +--> dequeue for IO
312 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
315 assert_spin_locked(&wb->list_lock);
316 list_splice_init(&wb->b_more_io, &wb->b_io);
317 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
318 trace_writeback_queue_io(wb, work, moved);
321 static int write_inode(struct inode *inode, struct writeback_control *wbc)
323 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
324 return inode->i_sb->s_op->write_inode(inode, wbc);
329 * Wait for writeback on an inode to complete.
331 static void inode_wait_for_writeback(struct inode *inode,
332 struct bdi_writeback *wb)
334 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
335 wait_queue_head_t *wqh;
337 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
338 while (inode->i_state & I_SYNC) {
339 spin_unlock(&inode->i_lock);
340 spin_unlock(&wb->list_lock);
341 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
342 spin_lock(&wb->list_lock);
343 spin_lock(&inode->i_lock);
348 * Find proper writeback list for the inode depending on its current state and
349 * possibly also change of its state while we were doing writeback. Here we
350 * handle things such as livelock prevention or fairness of writeback among
351 * inodes. This function can be called only by flusher thread - noone else
352 * processes all inodes in writeback lists and requeueing inodes behind flusher
353 * thread's back can have unexpected consequences.
355 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
356 struct writeback_control *wbc)
358 if (inode->i_state & I_FREEING)
362 * Sync livelock prevention. Each inode is tagged and synced in one
363 * shot. If still dirty, it will be redirty_tail()'ed below. Update
364 * the dirty time to prevent enqueue and sync it again.
366 if ((inode->i_state & I_DIRTY) &&
367 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
368 inode->dirtied_when = jiffies;
370 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
372 * We didn't write back all the pages. nfs_writepages()
373 * sometimes bales out without doing anything.
375 if (wbc->nr_to_write <= 0) {
376 /* Slice used up. Queue for next turn. */
377 requeue_io(inode, wb);
380 * Writeback blocked by something other than
381 * congestion. Delay the inode for some time to
382 * avoid spinning on the CPU (100% iowait)
383 * retrying writeback of the dirty page/inode
384 * that cannot be performed immediately.
386 redirty_tail(inode, wb);
388 } else if (inode->i_state & I_DIRTY) {
390 * Filesystems can dirty the inode during writeback operations,
391 * such as delayed allocation during submission or metadata
392 * updates after data IO completion.
394 redirty_tail(inode, wb);
396 /* The inode is clean. Remove from writeback lists. */
397 list_del_init(&inode->i_wb_list);
402 * Write out an inode's dirty pages. Called under wb->list_lock and
403 * inode->i_lock. Either the caller has an active reference on the inode or
404 * the inode has I_WILL_FREE set.
406 * If `wait' is set, wait on the writeout.
408 * The whole writeout design is quite complex and fragile. We want to avoid
409 * starvation of particular inodes when others are being redirtied, prevent
413 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
414 struct writeback_control *wbc)
416 struct address_space *mapping = inode->i_mapping;
417 long nr_to_write = wbc->nr_to_write;
421 assert_spin_locked(&wb->list_lock);
422 assert_spin_locked(&inode->i_lock);
424 if (!atomic_read(&inode->i_count))
425 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
427 WARN_ON(inode->i_state & I_WILL_FREE);
429 if (inode->i_state & I_SYNC) {
430 if (wbc->sync_mode != WB_SYNC_ALL)
433 * It's a data-integrity sync. We must wait.
435 inode_wait_for_writeback(inode, wb);
438 BUG_ON(inode->i_state & I_SYNC);
440 /* Set I_SYNC, reset I_DIRTY_PAGES */
441 inode->i_state |= I_SYNC;
442 spin_unlock(&inode->i_lock);
443 spin_unlock(&wb->list_lock);
445 ret = do_writepages(mapping, wbc);
448 * Make sure to wait on the data before writing out the metadata.
449 * This is important for filesystems that modify metadata on data
452 if (wbc->sync_mode == WB_SYNC_ALL) {
453 int err = filemap_fdatawait(mapping);
459 * Some filesystems may redirty the inode during the writeback
460 * due to delalloc, clear dirty metadata flags right before
463 spin_lock(&inode->i_lock);
464 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
465 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
466 inode->i_state &= ~I_DIRTY_PAGES;
467 dirty = inode->i_state & I_DIRTY;
468 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
469 spin_unlock(&inode->i_lock);
470 /* Don't write the inode if only I_DIRTY_PAGES was set */
471 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
472 int err = write_inode(inode, wbc);
477 spin_lock(&wb->list_lock);
478 spin_lock(&inode->i_lock);
479 requeue_inode(inode, wb, wbc);
480 inode_sync_complete(inode);
481 trace_writeback_single_inode(inode, wbc, nr_to_write);
485 static long writeback_chunk_size(struct backing_dev_info *bdi,
486 struct wb_writeback_work *work)
491 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
492 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
493 * here avoids calling into writeback_inodes_wb() more than once.
495 * The intended call sequence for WB_SYNC_ALL writeback is:
498 * writeback_sb_inodes() <== called only once
499 * write_cache_pages() <== called once for each inode
500 * (quickly) tag currently dirty pages
501 * (maybe slowly) sync all tagged pages
503 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
506 pages = min(bdi->avg_write_bandwidth / 2,
507 global_dirty_limit / DIRTY_SCOPE);
508 pages = min(pages, work->nr_pages);
509 pages = round_down(pages + MIN_WRITEBACK_PAGES,
510 MIN_WRITEBACK_PAGES);
517 * Write a portion of b_io inodes which belong to @sb.
519 * If @only_this_sb is true, then find and write all such
520 * inodes. Otherwise write only ones which go sequentially
523 * Return the number of pages and/or inodes written.
525 static long writeback_sb_inodes(struct super_block *sb,
526 struct bdi_writeback *wb,
527 struct wb_writeback_work *work)
529 struct writeback_control wbc = {
530 .sync_mode = work->sync_mode,
531 .tagged_writepages = work->tagged_writepages,
532 .for_kupdate = work->for_kupdate,
533 .for_background = work->for_background,
534 .range_cyclic = work->range_cyclic,
536 .range_end = LLONG_MAX,
538 unsigned long start_time = jiffies;
540 long wrote = 0; /* count both pages and inodes */
542 while (!list_empty(&wb->b_io)) {
543 struct inode *inode = wb_inode(wb->b_io.prev);
545 if (inode->i_sb != sb) {
548 * We only want to write back data for this
549 * superblock, move all inodes not belonging
550 * to it back onto the dirty list.
552 redirty_tail(inode, wb);
557 * The inode belongs to a different superblock.
558 * Bounce back to the caller to unpin this and
559 * pin the next superblock.
565 * Don't bother with new inodes or inodes beeing freed, first
566 * kind does not need peridic writeout yet, and for the latter
567 * kind writeout is handled by the freer.
569 spin_lock(&inode->i_lock);
570 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
571 spin_unlock(&inode->i_lock);
572 redirty_tail(inode, wb);
575 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
577 * If this inode is locked for writeback and we are not
578 * doing writeback-for-data-integrity, move it to
579 * b_more_io so that writeback can proceed with the
580 * other inodes on s_io.
582 * We'll have another go at writing back this inode
583 * when we completed a full scan of b_io.
585 spin_unlock(&inode->i_lock);
586 requeue_io(inode, wb);
587 trace_writeback_sb_inodes_requeue(inode);
591 write_chunk = writeback_chunk_size(wb->bdi, work);
592 wbc.nr_to_write = write_chunk;
593 wbc.pages_skipped = 0;
595 writeback_single_inode(inode, wb, &wbc);
597 work->nr_pages -= write_chunk - wbc.nr_to_write;
598 wrote += write_chunk - wbc.nr_to_write;
599 if (!(inode->i_state & I_DIRTY))
601 if (wbc.pages_skipped) {
603 * writeback is not making progress due to locked
604 * buffers. Skip this inode for now.
606 redirty_tail(inode, wb);
608 spin_unlock(&inode->i_lock);
609 spin_unlock(&wb->list_lock);
612 spin_lock(&wb->list_lock);
614 * bail out to wb_writeback() often enough to check
615 * background threshold and other termination conditions.
618 if (time_is_before_jiffies(start_time + HZ / 10UL))
620 if (work->nr_pages <= 0)
627 static long __writeback_inodes_wb(struct bdi_writeback *wb,
628 struct wb_writeback_work *work)
630 unsigned long start_time = jiffies;
633 while (!list_empty(&wb->b_io)) {
634 struct inode *inode = wb_inode(wb->b_io.prev);
635 struct super_block *sb = inode->i_sb;
637 if (!grab_super_passive(sb)) {
639 * grab_super_passive() may fail consistently due to
640 * s_umount being grabbed by someone else. Don't use
641 * requeue_io() to avoid busy retrying the inode/sb.
643 redirty_tail(inode, wb);
646 wrote += writeback_sb_inodes(sb, wb, work);
649 /* refer to the same tests at the end of writeback_sb_inodes */
651 if (time_is_before_jiffies(start_time + HZ / 10UL))
653 if (work->nr_pages <= 0)
657 /* Leave any unwritten inodes on b_io */
661 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
662 enum wb_reason reason)
664 struct wb_writeback_work work = {
665 .nr_pages = nr_pages,
666 .sync_mode = WB_SYNC_NONE,
671 spin_lock(&wb->list_lock);
672 if (list_empty(&wb->b_io))
674 __writeback_inodes_wb(wb, &work);
675 spin_unlock(&wb->list_lock);
677 return nr_pages - work.nr_pages;
680 static bool over_bground_thresh(struct backing_dev_info *bdi)
682 unsigned long background_thresh, dirty_thresh;
684 global_dirty_limits(&background_thresh, &dirty_thresh);
686 if (global_page_state(NR_FILE_DIRTY) +
687 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
690 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
691 bdi_dirty_limit(bdi, background_thresh))
698 * Called under wb->list_lock. If there are multiple wb per bdi,
699 * only the flusher working on the first wb should do it.
701 static void wb_update_bandwidth(struct bdi_writeback *wb,
702 unsigned long start_time)
704 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
708 * Explicit flushing or periodic writeback of "old" data.
710 * Define "old": the first time one of an inode's pages is dirtied, we mark the
711 * dirtying-time in the inode's address_space. So this periodic writeback code
712 * just walks the superblock inode list, writing back any inodes which are
713 * older than a specific point in time.
715 * Try to run once per dirty_writeback_interval. But if a writeback event
716 * takes longer than a dirty_writeback_interval interval, then leave a
719 * older_than_this takes precedence over nr_to_write. So we'll only write back
720 * all dirty pages if they are all attached to "old" mappings.
722 static long wb_writeback(struct bdi_writeback *wb,
723 struct wb_writeback_work *work)
725 unsigned long wb_start = jiffies;
726 long nr_pages = work->nr_pages;
727 unsigned long oldest_jif;
731 oldest_jif = jiffies;
732 work->older_than_this = &oldest_jif;
734 spin_lock(&wb->list_lock);
737 * Stop writeback when nr_pages has been consumed
739 if (work->nr_pages <= 0)
743 * Background writeout and kupdate-style writeback may
744 * run forever. Stop them if there is other work to do
745 * so that e.g. sync can proceed. They'll be restarted
746 * after the other works are all done.
748 if ((work->for_background || work->for_kupdate) &&
749 !list_empty(&wb->bdi->work_list))
753 * For background writeout, stop when we are below the
754 * background dirty threshold
756 if (work->for_background && !over_bground_thresh(wb->bdi))
760 * Kupdate and background works are special and we want to
761 * include all inodes that need writing. Livelock avoidance is
762 * handled by these works yielding to any other work so we are
765 if (work->for_kupdate) {
766 oldest_jif = jiffies -
767 msecs_to_jiffies(dirty_expire_interval * 10);
768 } else if (work->for_background)
769 oldest_jif = jiffies;
771 trace_writeback_start(wb->bdi, work);
772 if (list_empty(&wb->b_io))
775 progress = writeback_sb_inodes(work->sb, wb, work);
777 progress = __writeback_inodes_wb(wb, work);
778 trace_writeback_written(wb->bdi, work);
780 wb_update_bandwidth(wb, wb_start);
783 * Did we write something? Try for more
785 * Dirty inodes are moved to b_io for writeback in batches.
786 * The completion of the current batch does not necessarily
787 * mean the overall work is done. So we keep looping as long
788 * as made some progress on cleaning pages or inodes.
793 * No more inodes for IO, bail
795 if (list_empty(&wb->b_more_io))
798 * Nothing written. Wait for some inode to
799 * become available for writeback. Otherwise
800 * we'll just busyloop.
802 if (!list_empty(&wb->b_more_io)) {
803 trace_writeback_wait(wb->bdi, work);
804 inode = wb_inode(wb->b_more_io.prev);
805 spin_lock(&inode->i_lock);
806 inode_wait_for_writeback(inode, wb);
807 spin_unlock(&inode->i_lock);
810 spin_unlock(&wb->list_lock);
812 return nr_pages - work->nr_pages;
816 * Return the next wb_writeback_work struct that hasn't been processed yet.
818 static struct wb_writeback_work *
819 get_next_work_item(struct backing_dev_info *bdi)
821 struct wb_writeback_work *work = NULL;
823 spin_lock_bh(&bdi->wb_lock);
824 if (!list_empty(&bdi->work_list)) {
825 work = list_entry(bdi->work_list.next,
826 struct wb_writeback_work, list);
827 list_del_init(&work->list);
829 spin_unlock_bh(&bdi->wb_lock);
834 * Add in the number of potentially dirty inodes, because each inode
835 * write can dirty pagecache in the underlying blockdev.
837 static unsigned long get_nr_dirty_pages(void)
839 return global_page_state(NR_FILE_DIRTY) +
840 global_page_state(NR_UNSTABLE_NFS) +
841 get_nr_dirty_inodes();
844 static long wb_check_background_flush(struct bdi_writeback *wb)
846 if (over_bground_thresh(wb->bdi)) {
848 struct wb_writeback_work work = {
849 .nr_pages = LONG_MAX,
850 .sync_mode = WB_SYNC_NONE,
853 .reason = WB_REASON_BACKGROUND,
856 return wb_writeback(wb, &work);
862 static long wb_check_old_data_flush(struct bdi_writeback *wb)
864 unsigned long expired;
868 * When set to zero, disable periodic writeback
870 if (!dirty_writeback_interval)
873 expired = wb->last_old_flush +
874 msecs_to_jiffies(dirty_writeback_interval * 10);
875 if (time_before(jiffies, expired))
878 wb->last_old_flush = jiffies;
879 nr_pages = get_nr_dirty_pages();
882 struct wb_writeback_work work = {
883 .nr_pages = nr_pages,
884 .sync_mode = WB_SYNC_NONE,
887 .reason = WB_REASON_PERIODIC,
890 return wb_writeback(wb, &work);
897 * Retrieve work items and do the writeback they describe
899 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
901 struct backing_dev_info *bdi = wb->bdi;
902 struct wb_writeback_work *work;
905 set_bit(BDI_writeback_running, &wb->bdi->state);
906 while ((work = get_next_work_item(bdi)) != NULL) {
908 * Override sync mode, in case we must wait for completion
909 * because this thread is exiting now.
912 work->sync_mode = WB_SYNC_ALL;
914 trace_writeback_exec(bdi, work);
916 wrote += wb_writeback(wb, work);
919 * Notify the caller of completion if this is a synchronous
920 * work item, otherwise just free it.
923 complete(work->done);
929 * Check for periodic writeback, kupdated() style
931 wrote += wb_check_old_data_flush(wb);
932 wrote += wb_check_background_flush(wb);
933 clear_bit(BDI_writeback_running, &wb->bdi->state);
939 * Handle writeback of dirty data for the device backed by this bdi. Also
940 * wakes up periodically and does kupdated style flushing.
942 int bdi_writeback_thread(void *data)
944 struct bdi_writeback *wb = data;
945 struct backing_dev_info *bdi = wb->bdi;
948 current->flags |= PF_SWAPWRITE;
950 wb->last_active = jiffies;
953 * Our parent may run at a different priority, just set us to normal
955 set_user_nice(current, 0);
957 trace_writeback_thread_start(bdi);
959 while (!kthread_freezable_should_stop(NULL)) {
961 * Remove own delayed wake-up timer, since we are already awake
962 * and we'll take care of the preriodic write-back.
964 del_timer(&wb->wakeup_timer);
966 pages_written = wb_do_writeback(wb, 0);
968 trace_writeback_pages_written(pages_written);
971 wb->last_active = jiffies;
973 set_current_state(TASK_INTERRUPTIBLE);
974 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
975 __set_current_state(TASK_RUNNING);
979 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
980 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
983 * We have nothing to do, so can go sleep without any
984 * timeout and save power. When a work is queued or
985 * something is made dirty - we will be woken up.
991 /* Flush any work that raced with us exiting */
992 if (!list_empty(&bdi->work_list))
993 wb_do_writeback(wb, 1);
995 trace_writeback_thread_stop(bdi);
1001 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1004 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1006 struct backing_dev_info *bdi;
1009 nr_pages = global_page_state(NR_FILE_DIRTY) +
1010 global_page_state(NR_UNSTABLE_NFS);
1014 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1015 if (!bdi_has_dirty_io(bdi))
1017 __bdi_start_writeback(bdi, nr_pages, false, reason);
1022 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1024 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1025 struct dentry *dentry;
1026 const char *name = "?";
1028 dentry = d_find_alias(inode);
1030 spin_lock(&dentry->d_lock);
1031 name = (const char *) dentry->d_name.name;
1034 "%s(%d): dirtied inode %lu (%s) on %s\n",
1035 current->comm, task_pid_nr(current), inode->i_ino,
1036 name, inode->i_sb->s_id);
1038 spin_unlock(&dentry->d_lock);
1045 * __mark_inode_dirty - internal function
1046 * @inode: inode to mark
1047 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1048 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1049 * mark_inode_dirty_sync.
1051 * Put the inode on the super block's dirty list.
1053 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1054 * dirty list only if it is hashed or if it refers to a blockdev.
1055 * If it was not hashed, it will never be added to the dirty list
1056 * even if it is later hashed, as it will have been marked dirty already.
1058 * In short, make sure you hash any inodes _before_ you start marking
1061 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1062 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1063 * the kernel-internal blockdev inode represents the dirtying time of the
1064 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1065 * page->mapping->host, so the page-dirtying time is recorded in the internal
1068 void __mark_inode_dirty(struct inode *inode, int flags)
1070 struct super_block *sb = inode->i_sb;
1071 struct backing_dev_info *bdi = NULL;
1074 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1075 * dirty the inode itself
1077 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1078 if (sb->s_op->dirty_inode)
1079 sb->s_op->dirty_inode(inode, flags);
1083 * make sure that changes are seen by all cpus before we test i_state
1088 /* avoid the locking if we can */
1089 if ((inode->i_state & flags) == flags)
1092 if (unlikely(block_dump))
1093 block_dump___mark_inode_dirty(inode);
1095 spin_lock(&inode->i_lock);
1096 if ((inode->i_state & flags) != flags) {
1097 const int was_dirty = inode->i_state & I_DIRTY;
1099 inode->i_state |= flags;
1102 * If the inode is being synced, just update its dirty state.
1103 * The unlocker will place the inode on the appropriate
1104 * superblock list, based upon its state.
1106 if (inode->i_state & I_SYNC)
1107 goto out_unlock_inode;
1110 * Only add valid (hashed) inodes to the superblock's
1111 * dirty list. Add blockdev inodes as well.
1113 if (!S_ISBLK(inode->i_mode)) {
1114 if (inode_unhashed(inode))
1115 goto out_unlock_inode;
1117 if (inode->i_state & I_FREEING)
1118 goto out_unlock_inode;
1121 * If the inode was already on b_dirty/b_io/b_more_io, don't
1122 * reposition it (that would break b_dirty time-ordering).
1125 bool wakeup_bdi = false;
1126 bdi = inode_to_bdi(inode);
1128 if (bdi_cap_writeback_dirty(bdi)) {
1129 WARN(!test_bit(BDI_registered, &bdi->state),
1130 "bdi-%s not registered\n", bdi->name);
1133 * If this is the first dirty inode for this
1134 * bdi, we have to wake-up the corresponding
1135 * bdi thread to make sure background
1136 * write-back happens later.
1138 if (!wb_has_dirty_io(&bdi->wb))
1142 spin_unlock(&inode->i_lock);
1143 spin_lock(&bdi->wb.list_lock);
1144 inode->dirtied_when = jiffies;
1145 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1146 spin_unlock(&bdi->wb.list_lock);
1149 bdi_wakeup_thread_delayed(bdi);
1154 spin_unlock(&inode->i_lock);
1157 EXPORT_SYMBOL(__mark_inode_dirty);
1159 static void wait_sb_inodes(struct super_block *sb)
1161 struct inode *inode, *old_inode = NULL;
1164 * We need to be protected against the filesystem going from
1165 * r/o to r/w or vice versa.
1167 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1169 spin_lock(&inode_sb_list_lock);
1172 * Data integrity sync. Must wait for all pages under writeback,
1173 * because there may have been pages dirtied before our sync
1174 * call, but which had writeout started before we write it out.
1175 * In which case, the inode may not be on the dirty list, but
1176 * we still have to wait for that writeout.
1178 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1179 struct address_space *mapping = inode->i_mapping;
1181 spin_lock(&inode->i_lock);
1182 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1183 (mapping->nrpages == 0)) {
1184 spin_unlock(&inode->i_lock);
1188 spin_unlock(&inode->i_lock);
1189 spin_unlock(&inode_sb_list_lock);
1192 * We hold a reference to 'inode' so it couldn't have been
1193 * removed from s_inodes list while we dropped the
1194 * inode_sb_list_lock. We cannot iput the inode now as we can
1195 * be holding the last reference and we cannot iput it under
1196 * inode_sb_list_lock. So we keep the reference and iput it
1202 filemap_fdatawait(mapping);
1206 spin_lock(&inode_sb_list_lock);
1208 spin_unlock(&inode_sb_list_lock);
1213 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1214 * @sb: the superblock
1215 * @nr: the number of pages to write
1216 * @reason: reason why some writeback work initiated
1218 * Start writeback on some inodes on this super_block. No guarantees are made
1219 * on how many (if any) will be written, and this function does not wait
1220 * for IO completion of submitted IO.
1222 void writeback_inodes_sb_nr(struct super_block *sb,
1224 enum wb_reason reason)
1226 DECLARE_COMPLETION_ONSTACK(done);
1227 struct wb_writeback_work work = {
1229 .sync_mode = WB_SYNC_NONE,
1230 .tagged_writepages = 1,
1236 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1237 bdi_queue_work(sb->s_bdi, &work);
1238 wait_for_completion(&done);
1240 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1243 * writeback_inodes_sb - writeback dirty inodes from given super_block
1244 * @sb: the superblock
1245 * @reason: reason why some writeback work was initiated
1247 * Start writeback on some inodes on this super_block. No guarantees are made
1248 * on how many (if any) will be written, and this function does not wait
1249 * for IO completion of submitted IO.
1251 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1253 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1255 EXPORT_SYMBOL(writeback_inodes_sb);
1258 * writeback_inodes_sb_if_idle - start writeback if none underway
1259 * @sb: the superblock
1260 * @reason: reason why some writeback work was initiated
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, enum wb_reason reason)
1267 if (!writeback_in_progress(sb->s_bdi)) {
1268 down_read(&sb->s_umount);
1269 writeback_inodes_sb(sb, reason);
1270 up_read(&sb->s_umount);
1275 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1278 * writeback_inodes_sb_nr_if_idle - start writeback if none underway
1279 * @sb: the superblock
1280 * @nr: the number of pages to write
1281 * @reason: reason why some writeback work was initiated
1283 * Invoke writeback_inodes_sb if no writeback is currently underway.
1284 * Returns 1 if writeback was started, 0 if not.
1286 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1288 enum wb_reason reason)
1290 if (!writeback_in_progress(sb->s_bdi)) {
1291 down_read(&sb->s_umount);
1292 writeback_inodes_sb_nr(sb, nr, reason);
1293 up_read(&sb->s_umount);
1298 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1301 * sync_inodes_sb - sync sb inode pages
1302 * @sb: the superblock
1304 * This function writes and waits on any dirty inode belonging to this
1307 void sync_inodes_sb(struct super_block *sb)
1309 DECLARE_COMPLETION_ONSTACK(done);
1310 struct wb_writeback_work work = {
1312 .sync_mode = WB_SYNC_ALL,
1313 .nr_pages = LONG_MAX,
1316 .reason = WB_REASON_SYNC,
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);
1360 EXPORT_SYMBOL(write_inode_now);
1363 * sync_inode - write an inode and its pages to disk.
1364 * @inode: the inode to sync
1365 * @wbc: controls the writeback mode
1367 * sync_inode() will write an inode and its pages to disk. It will also
1368 * correctly update the inode on its superblock's dirty inode lists and will
1369 * update inode->i_state.
1371 * The caller must have a ref on the inode.
1373 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1375 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1378 spin_lock(&wb->list_lock);
1379 spin_lock(&inode->i_lock);
1380 ret = writeback_single_inode(inode, wb, wbc);
1381 spin_unlock(&inode->i_lock);
1382 spin_unlock(&wb->list_lock);
1385 EXPORT_SYMBOL(sync_inode);
1388 * sync_inode_metadata - write an inode to disk
1389 * @inode: the inode to sync
1390 * @wait: wait for I/O to complete.
1392 * Write an inode to disk and adjust its dirty state after completion.
1394 * Note: only writes the actual inode, no associated data or other metadata.
1396 int sync_inode_metadata(struct inode *inode, int wait)
1398 struct writeback_control wbc = {
1399 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1400 .nr_to_write = 0, /* metadata-only */
1403 return sync_inode(inode, &wbc);
1405 EXPORT_SYMBOL(sync_inode_metadata);