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/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
32 * 4MB minimal write chunk size
34 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
37 * Passed into wb_writeback(), essentially a subset of writeback_control
39 struct wb_writeback_work {
41 struct super_block *sb;
43 * Write only inodes dirtied before this time. Don't forget to set
44 * older_than_this_is_set when you set this.
46 unsigned long older_than_this;
47 enum writeback_sync_modes sync_mode;
48 unsigned int tagged_writepages:1;
49 unsigned int for_kupdate:1;
50 unsigned int range_cyclic:1;
51 unsigned int for_background:1;
52 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
53 unsigned int older_than_this_is_set:1;
54 enum wb_reason reason; /* why was writeback initiated? */
56 struct list_head list; /* pending work list */
57 struct completion *done; /* set if the caller waits */
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 (sb_is_blkdev_sb(sb))
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 static void bdi_queue_work(struct backing_dev_info *bdi,
97 struct wb_writeback_work *work)
99 trace_writeback_queue(bdi, work);
101 spin_lock_bh(&bdi->wb_lock);
102 list_add_tail(&work->list, &bdi->work_list);
103 spin_unlock_bh(&bdi->wb_lock);
105 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
109 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
110 bool range_cyclic, enum wb_reason reason)
112 struct wb_writeback_work *work;
115 * This is WB_SYNC_NONE writeback, so if allocation fails just
116 * wakeup the thread for old dirty data writeback
118 work = kzalloc(sizeof(*work), GFP_ATOMIC);
120 trace_writeback_nowork(bdi);
121 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
125 work->sync_mode = WB_SYNC_NONE;
126 work->nr_pages = nr_pages;
127 work->range_cyclic = range_cyclic;
128 work->reason = reason;
130 bdi_queue_work(bdi, work);
134 * bdi_start_writeback - start writeback
135 * @bdi: the backing device to write from
136 * @nr_pages: the number of pages to write
137 * @reason: reason why some writeback work was initiated
140 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
141 * started when this function returns, we make no guarantees on
142 * completion. Caller need not hold sb s_umount semaphore.
145 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
146 enum wb_reason reason)
148 __bdi_start_writeback(bdi, nr_pages, true, reason);
152 * bdi_start_background_writeback - start background writeback
153 * @bdi: the backing device to write from
156 * This makes sure WB_SYNC_NONE background writeback happens. When
157 * this function returns, it is only guaranteed that for given BDI
158 * some IO is happening if we are over background dirty threshold.
159 * Caller need not hold sb s_umount semaphore.
161 void bdi_start_background_writeback(struct backing_dev_info *bdi)
164 * We just wake up the flusher thread. It will perform background
165 * writeback as soon as there is no other work to do.
167 trace_writeback_wake_background(bdi);
168 mod_delayed_work(bdi_wq, &bdi->wb.dwork, 0);
172 * Remove the inode from the writeback list it is on.
174 void inode_wb_list_del(struct inode *inode)
176 struct backing_dev_info *bdi = inode_to_bdi(inode);
178 spin_lock(&bdi->wb.list_lock);
179 list_del_init(&inode->i_wb_list);
180 spin_unlock(&bdi->wb.list_lock);
184 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
185 * furthest end of its superblock's dirty-inode list.
187 * Before stamping the inode's ->dirtied_when, we check to see whether it is
188 * already the most-recently-dirtied inode on the b_dirty list. If that is
189 * the case then the inode must have been redirtied while it was being written
190 * out and we don't reset its dirtied_when.
192 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
194 assert_spin_locked(&wb->list_lock);
195 if (!list_empty(&wb->b_dirty)) {
198 tail = wb_inode(wb->b_dirty.next);
199 if (time_before(inode->dirtied_when, tail->dirtied_when))
200 inode->dirtied_when = jiffies;
202 list_move(&inode->i_wb_list, &wb->b_dirty);
206 * requeue inode for re-scanning after bdi->b_io list is exhausted.
208 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
210 assert_spin_locked(&wb->list_lock);
211 list_move(&inode->i_wb_list, &wb->b_more_io);
214 static void inode_sync_complete(struct inode *inode)
216 inode->i_state &= ~I_SYNC;
217 /* If inode is clean an unused, put it into LRU now... */
218 inode_add_lru(inode);
219 /* Waiters must see I_SYNC cleared before being woken up */
221 wake_up_bit(&inode->i_state, __I_SYNC);
224 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
226 bool ret = time_after(inode->dirtied_when, t);
229 * For inodes being constantly redirtied, dirtied_when can get stuck.
230 * It _appears_ to be in the future, but is actually in distant past.
231 * This test is necessary to prevent such wrapped-around relative times
232 * from permanently stopping the whole bdi writeback.
234 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
240 * Move expired (dirtied before work->older_than_this) dirty inodes from
241 * @delaying_queue to @dispatch_queue.
243 static int move_expired_inodes(struct list_head *delaying_queue,
244 struct list_head *dispatch_queue,
245 struct wb_writeback_work *work)
248 struct list_head *pos, *node;
249 struct super_block *sb = NULL;
254 WARN_ON_ONCE(!work->older_than_this_is_set);
255 while (!list_empty(delaying_queue)) {
256 inode = wb_inode(delaying_queue->prev);
257 if (inode_dirtied_after(inode, work->older_than_this))
259 list_move(&inode->i_wb_list, &tmp);
261 if (sb_is_blkdev_sb(inode->i_sb))
263 if (sb && sb != inode->i_sb)
268 /* just one sb in list, splice to dispatch_queue and we're done */
270 list_splice(&tmp, dispatch_queue);
274 /* Move inodes from one superblock together */
275 while (!list_empty(&tmp)) {
276 sb = wb_inode(tmp.prev)->i_sb;
277 list_for_each_prev_safe(pos, node, &tmp) {
278 inode = wb_inode(pos);
279 if (inode->i_sb == sb)
280 list_move(&inode->i_wb_list, dispatch_queue);
288 * Queue all expired dirty inodes for io, eldest first.
290 * newly dirtied b_dirty b_io b_more_io
291 * =============> gf edc BA
293 * newly dirtied b_dirty b_io b_more_io
294 * =============> g fBAedc
296 * +--> dequeue for IO
298 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
301 assert_spin_locked(&wb->list_lock);
302 list_splice_init(&wb->b_more_io, &wb->b_io);
303 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
304 trace_writeback_queue_io(wb, work, moved);
307 static int write_inode(struct inode *inode, struct writeback_control *wbc)
311 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
312 trace_writeback_write_inode_start(inode, wbc);
313 ret = inode->i_sb->s_op->write_inode(inode, wbc);
314 trace_writeback_write_inode(inode, wbc);
321 * Wait for writeback on an inode to complete. Called with i_lock held.
322 * Caller must make sure inode cannot go away when we drop i_lock.
324 static void __inode_wait_for_writeback(struct inode *inode)
325 __releases(inode->i_lock)
326 __acquires(inode->i_lock)
328 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
329 wait_queue_head_t *wqh;
331 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
332 while (inode->i_state & I_SYNC) {
333 spin_unlock(&inode->i_lock);
334 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
335 spin_lock(&inode->i_lock);
340 * Wait for writeback on an inode to complete. Caller must have inode pinned.
342 void inode_wait_for_writeback(struct inode *inode)
344 spin_lock(&inode->i_lock);
345 __inode_wait_for_writeback(inode);
346 spin_unlock(&inode->i_lock);
350 * Sleep until I_SYNC is cleared. This function must be called with i_lock
351 * held and drops it. It is aimed for callers not holding any inode reference
352 * so once i_lock is dropped, inode can go away.
354 static void inode_sleep_on_writeback(struct inode *inode)
355 __releases(inode->i_lock)
358 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
361 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
362 sleep = inode->i_state & I_SYNC;
363 spin_unlock(&inode->i_lock);
366 finish_wait(wqh, &wait);
370 * Find proper writeback list for the inode depending on its current state and
371 * possibly also change of its state while we were doing writeback. Here we
372 * handle things such as livelock prevention or fairness of writeback among
373 * inodes. This function can be called only by flusher thread - noone else
374 * processes all inodes in writeback lists and requeueing inodes behind flusher
375 * thread's back can have unexpected consequences.
377 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
378 struct writeback_control *wbc)
380 if (inode->i_state & I_FREEING)
384 * Sync livelock prevention. Each inode is tagged and synced in one
385 * shot. If still dirty, it will be redirty_tail()'ed below. Update
386 * the dirty time to prevent enqueue and sync it again.
388 if ((inode->i_state & I_DIRTY) &&
389 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
390 inode->dirtied_when = jiffies;
392 if (wbc->pages_skipped) {
394 * writeback is not making progress due to locked
395 * buffers. Skip this inode for now.
397 redirty_tail(inode, wb);
401 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
403 * We didn't write back all the pages. nfs_writepages()
404 * sometimes bales out without doing anything.
406 if (wbc->nr_to_write <= 0) {
407 /* Slice used up. Queue for next turn. */
408 requeue_io(inode, wb);
411 * Writeback blocked by something other than
412 * congestion. Delay the inode for some time to
413 * avoid spinning on the CPU (100% iowait)
414 * retrying writeback of the dirty page/inode
415 * that cannot be performed immediately.
417 redirty_tail(inode, wb);
419 } else if (inode->i_state & I_DIRTY) {
421 * Filesystems can dirty the inode during writeback operations,
422 * such as delayed allocation during submission or metadata
423 * updates after data IO completion.
425 redirty_tail(inode, wb);
427 /* The inode is clean. Remove from writeback lists. */
428 list_del_init(&inode->i_wb_list);
433 * Write out an inode and its dirty pages. Do not update the writeback list
434 * linkage. That is left to the caller. The caller is also responsible for
435 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
438 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
440 struct address_space *mapping = inode->i_mapping;
441 long nr_to_write = wbc->nr_to_write;
445 WARN_ON(!(inode->i_state & I_SYNC));
447 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
449 ret = do_writepages(mapping, wbc);
452 * Make sure to wait on the data before writing out the metadata.
453 * This is important for filesystems that modify metadata on data
454 * I/O completion. We don't do it for sync(2) writeback because it has a
455 * separate, external IO completion path and ->sync_fs for guaranteeing
456 * inode metadata is written back correctly.
458 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
459 int err = filemap_fdatawait(mapping);
465 * Some filesystems may redirty the inode during the writeback
466 * due to delalloc, clear dirty metadata flags right before
469 spin_lock(&inode->i_lock);
470 /* Clear I_DIRTY_PAGES if we've written out all dirty pages */
471 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
472 inode->i_state &= ~I_DIRTY_PAGES;
473 dirty = inode->i_state & I_DIRTY;
474 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
475 spin_unlock(&inode->i_lock);
476 /* Don't write the inode if only I_DIRTY_PAGES was set */
477 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
478 int err = write_inode(inode, wbc);
482 trace_writeback_single_inode(inode, wbc, nr_to_write);
487 * Write out an inode's dirty pages. Either the caller has an active reference
488 * on the inode or the inode has I_WILL_FREE set.
490 * This function is designed to be called for writing back one inode which
491 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
492 * and does more profound writeback list handling in writeback_sb_inodes().
495 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
496 struct writeback_control *wbc)
500 spin_lock(&inode->i_lock);
501 if (!atomic_read(&inode->i_count))
502 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
504 WARN_ON(inode->i_state & I_WILL_FREE);
506 if (inode->i_state & I_SYNC) {
507 if (wbc->sync_mode != WB_SYNC_ALL)
510 * It's a data-integrity sync. We must wait. Since callers hold
511 * inode reference or inode has I_WILL_FREE set, it cannot go
514 __inode_wait_for_writeback(inode);
516 WARN_ON(inode->i_state & I_SYNC);
518 * Skip inode if it is clean. We don't want to mess with writeback
519 * lists in this function since flusher thread may be doing for example
520 * sync in parallel and if we move the inode, it could get skipped. So
521 * here we make sure inode is on some writeback list and leave it there
522 * unless we have completely cleaned the inode.
524 if (!(inode->i_state & I_DIRTY))
526 inode->i_state |= I_SYNC;
527 spin_unlock(&inode->i_lock);
529 ret = __writeback_single_inode(inode, wbc);
531 spin_lock(&wb->list_lock);
532 spin_lock(&inode->i_lock);
534 * If inode is clean, remove it from writeback lists. Otherwise don't
535 * touch it. See comment above for explanation.
537 if (!(inode->i_state & I_DIRTY))
538 list_del_init(&inode->i_wb_list);
539 spin_unlock(&wb->list_lock);
540 inode_sync_complete(inode);
542 spin_unlock(&inode->i_lock);
546 static long writeback_chunk_size(struct backing_dev_info *bdi,
547 struct wb_writeback_work *work)
552 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
553 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
554 * here avoids calling into writeback_inodes_wb() more than once.
556 * The intended call sequence for WB_SYNC_ALL writeback is:
559 * writeback_sb_inodes() <== called only once
560 * write_cache_pages() <== called once for each inode
561 * (quickly) tag currently dirty pages
562 * (maybe slowly) sync all tagged pages
564 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
567 pages = min(bdi->avg_write_bandwidth / 2,
568 global_dirty_limit / DIRTY_SCOPE);
569 pages = min(pages, work->nr_pages);
570 pages = round_down(pages + MIN_WRITEBACK_PAGES,
571 MIN_WRITEBACK_PAGES);
578 * Write a portion of b_io inodes which belong to @sb.
580 * Return the number of pages and/or inodes written.
582 static long writeback_sb_inodes(struct super_block *sb,
583 struct bdi_writeback *wb,
584 struct wb_writeback_work *work)
586 struct writeback_control wbc = {
587 .sync_mode = work->sync_mode,
588 .tagged_writepages = work->tagged_writepages,
589 .for_kupdate = work->for_kupdate,
590 .for_background = work->for_background,
591 .for_sync = work->for_sync,
592 .range_cyclic = work->range_cyclic,
594 .range_end = LLONG_MAX,
596 unsigned long start_time = jiffies;
598 long wrote = 0; /* count both pages and inodes */
600 while (!list_empty(&wb->b_io)) {
601 struct inode *inode = wb_inode(wb->b_io.prev);
603 if (inode->i_sb != sb) {
606 * We only want to write back data for this
607 * superblock, move all inodes not belonging
608 * to it back onto the dirty list.
610 redirty_tail(inode, wb);
615 * The inode belongs to a different superblock.
616 * Bounce back to the caller to unpin this and
617 * pin the next superblock.
623 * Don't bother with new inodes or inodes being freed, first
624 * kind does not need periodic writeout yet, and for the latter
625 * kind writeout is handled by the freer.
627 spin_lock(&inode->i_lock);
628 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
629 spin_unlock(&inode->i_lock);
630 redirty_tail(inode, wb);
633 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
635 * If this inode is locked for writeback and we are not
636 * doing writeback-for-data-integrity, move it to
637 * b_more_io so that writeback can proceed with the
638 * other inodes on s_io.
640 * We'll have another go at writing back this inode
641 * when we completed a full scan of b_io.
643 spin_unlock(&inode->i_lock);
644 requeue_io(inode, wb);
645 trace_writeback_sb_inodes_requeue(inode);
648 spin_unlock(&wb->list_lock);
651 * We already requeued the inode if it had I_SYNC set and we
652 * are doing WB_SYNC_NONE writeback. So this catches only the
655 if (inode->i_state & I_SYNC) {
656 /* Wait for I_SYNC. This function drops i_lock... */
657 inode_sleep_on_writeback(inode);
658 /* Inode may be gone, start again */
659 spin_lock(&wb->list_lock);
662 inode->i_state |= I_SYNC;
663 spin_unlock(&inode->i_lock);
665 write_chunk = writeback_chunk_size(wb->bdi, work);
666 wbc.nr_to_write = write_chunk;
667 wbc.pages_skipped = 0;
670 * We use I_SYNC to pin the inode in memory. While it is set
671 * evict_inode() will wait so the inode cannot be freed.
673 __writeback_single_inode(inode, &wbc);
675 work->nr_pages -= write_chunk - wbc.nr_to_write;
676 wrote += write_chunk - wbc.nr_to_write;
677 spin_lock(&wb->list_lock);
678 spin_lock(&inode->i_lock);
679 if (!(inode->i_state & I_DIRTY))
681 requeue_inode(inode, wb, &wbc);
682 inode_sync_complete(inode);
683 spin_unlock(&inode->i_lock);
684 cond_resched_lock(&wb->list_lock);
686 * bail out to wb_writeback() often enough to check
687 * background threshold and other termination conditions.
690 if (time_is_before_jiffies(start_time + HZ / 10UL))
692 if (work->nr_pages <= 0)
699 static long __writeback_inodes_wb(struct bdi_writeback *wb,
700 struct wb_writeback_work *work)
702 unsigned long start_time = jiffies;
705 while (!list_empty(&wb->b_io)) {
706 struct inode *inode = wb_inode(wb->b_io.prev);
707 struct super_block *sb = inode->i_sb;
709 if (!grab_super_passive(sb)) {
711 * grab_super_passive() may fail consistently due to
712 * s_umount being grabbed by someone else. Don't use
713 * requeue_io() to avoid busy retrying the inode/sb.
715 redirty_tail(inode, wb);
718 wrote += writeback_sb_inodes(sb, wb, work);
721 /* refer to the same tests at the end of writeback_sb_inodes */
723 if (time_is_before_jiffies(start_time + HZ / 10UL))
725 if (work->nr_pages <= 0)
729 /* Leave any unwritten inodes on b_io */
733 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
734 enum wb_reason reason)
736 struct wb_writeback_work work = {
737 .nr_pages = nr_pages,
738 .sync_mode = WB_SYNC_NONE,
741 .older_than_this = jiffies,
742 .older_than_this_is_set = 1,
745 spin_lock(&wb->list_lock);
746 if (list_empty(&wb->b_io))
748 __writeback_inodes_wb(wb, &work);
749 spin_unlock(&wb->list_lock);
751 return nr_pages - work.nr_pages;
754 static bool over_bground_thresh(struct backing_dev_info *bdi)
756 unsigned long background_thresh, dirty_thresh;
758 global_dirty_limits(&background_thresh, &dirty_thresh);
760 if (global_page_state(NR_FILE_DIRTY) +
761 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
764 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
765 bdi_dirty_limit(bdi, background_thresh))
772 * Called under wb->list_lock. If there are multiple wb per bdi,
773 * only the flusher working on the first wb should do it.
775 static void wb_update_bandwidth(struct bdi_writeback *wb,
776 unsigned long start_time)
778 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
782 * Explicit flushing or periodic writeback of "old" data.
784 * Define "old": the first time one of an inode's pages is dirtied, we mark the
785 * dirtying-time in the inode's address_space. So this periodic writeback code
786 * just walks the superblock inode list, writing back any inodes which are
787 * older than a specific point in time.
789 * Try to run once per dirty_writeback_interval. But if a writeback event
790 * takes longer than a dirty_writeback_interval interval, then leave a
793 * older_than_this takes precedence over nr_to_write. So we'll only write back
794 * all dirty pages if they are all attached to "old" mappings.
796 static long wb_writeback(struct bdi_writeback *wb,
797 struct wb_writeback_work *work)
799 unsigned long wb_start = jiffies;
800 long nr_pages = work->nr_pages;
804 if (!work->older_than_this_is_set) {
805 work->older_than_this = jiffies;
806 work->older_than_this_is_set = 1;
809 spin_lock(&wb->list_lock);
812 * Stop writeback when nr_pages has been consumed
814 if (work->nr_pages <= 0)
818 * Background writeout and kupdate-style writeback may
819 * run forever. Stop them if there is other work to do
820 * so that e.g. sync can proceed. They'll be restarted
821 * after the other works are all done.
823 if ((work->for_background || work->for_kupdate) &&
824 !list_empty(&wb->bdi->work_list))
828 * For background writeout, stop when we are below the
829 * background dirty threshold
831 if (work->for_background && !over_bground_thresh(wb->bdi))
835 * Kupdate and background works are special and we want to
836 * include all inodes that need writing. Livelock avoidance is
837 * handled by these works yielding to any other work so we are
840 if (work->for_kupdate) {
841 work->older_than_this = jiffies -
842 msecs_to_jiffies(dirty_expire_interval * 10);
843 } else if (work->for_background)
844 work->older_than_this = jiffies;
846 trace_writeback_start(wb->bdi, work);
847 if (list_empty(&wb->b_io))
850 progress = writeback_sb_inodes(work->sb, wb, work);
852 progress = __writeback_inodes_wb(wb, work);
853 trace_writeback_written(wb->bdi, work);
855 wb_update_bandwidth(wb, wb_start);
858 * Did we write something? Try for more
860 * Dirty inodes are moved to b_io for writeback in batches.
861 * The completion of the current batch does not necessarily
862 * mean the overall work is done. So we keep looping as long
863 * as made some progress on cleaning pages or inodes.
868 * No more inodes for IO, bail
870 if (list_empty(&wb->b_more_io))
873 * Nothing written. Wait for some inode to
874 * become available for writeback. Otherwise
875 * we'll just busyloop.
877 if (!list_empty(&wb->b_more_io)) {
878 trace_writeback_wait(wb->bdi, work);
879 inode = wb_inode(wb->b_more_io.prev);
880 spin_lock(&inode->i_lock);
881 spin_unlock(&wb->list_lock);
882 /* This function drops i_lock... */
883 inode_sleep_on_writeback(inode);
884 spin_lock(&wb->list_lock);
887 spin_unlock(&wb->list_lock);
889 return nr_pages - work->nr_pages;
893 * Return the next wb_writeback_work struct that hasn't been processed yet.
895 static struct wb_writeback_work *
896 get_next_work_item(struct backing_dev_info *bdi)
898 struct wb_writeback_work *work = NULL;
900 spin_lock_bh(&bdi->wb_lock);
901 if (!list_empty(&bdi->work_list)) {
902 work = list_entry(bdi->work_list.next,
903 struct wb_writeback_work, list);
904 list_del_init(&work->list);
906 spin_unlock_bh(&bdi->wb_lock);
911 * Add in the number of potentially dirty inodes, because each inode
912 * write can dirty pagecache in the underlying blockdev.
914 static unsigned long get_nr_dirty_pages(void)
916 return global_page_state(NR_FILE_DIRTY) +
917 global_page_state(NR_UNSTABLE_NFS) +
918 get_nr_dirty_inodes();
921 static long wb_check_background_flush(struct bdi_writeback *wb)
923 if (over_bground_thresh(wb->bdi)) {
925 struct wb_writeback_work work = {
926 .nr_pages = LONG_MAX,
927 .sync_mode = WB_SYNC_NONE,
930 .reason = WB_REASON_BACKGROUND,
933 return wb_writeback(wb, &work);
939 static long wb_check_old_data_flush(struct bdi_writeback *wb)
941 unsigned long expired;
945 * When set to zero, disable periodic writeback
947 if (!dirty_writeback_interval)
950 expired = wb->last_old_flush +
951 msecs_to_jiffies(dirty_writeback_interval * 10);
952 if (time_before(jiffies, expired))
955 wb->last_old_flush = jiffies;
956 nr_pages = get_nr_dirty_pages();
959 struct wb_writeback_work work = {
960 .nr_pages = nr_pages,
961 .sync_mode = WB_SYNC_NONE,
964 .reason = WB_REASON_PERIODIC,
967 return wb_writeback(wb, &work);
974 * Retrieve work items and do the writeback they describe
976 static long wb_do_writeback(struct bdi_writeback *wb)
978 struct backing_dev_info *bdi = wb->bdi;
979 struct wb_writeback_work *work;
982 set_bit(BDI_writeback_running, &wb->bdi->state);
983 while ((work = get_next_work_item(bdi)) != NULL) {
985 trace_writeback_exec(bdi, work);
987 wrote += wb_writeback(wb, work);
990 * Notify the caller of completion if this is a synchronous
991 * work item, otherwise just free it.
994 complete(work->done);
1000 * Check for periodic writeback, kupdated() style
1002 wrote += wb_check_old_data_flush(wb);
1003 wrote += wb_check_background_flush(wb);
1004 clear_bit(BDI_writeback_running, &wb->bdi->state);
1010 * Handle writeback of dirty data for the device backed by this bdi. Also
1011 * reschedules periodically and does kupdated style flushing.
1013 void bdi_writeback_workfn(struct work_struct *work)
1015 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1016 struct bdi_writeback, dwork);
1017 struct backing_dev_info *bdi = wb->bdi;
1020 set_worker_desc("flush-%s", dev_name(bdi->dev));
1021 current->flags |= PF_SWAPWRITE;
1023 if (likely(!current_is_workqueue_rescuer() ||
1024 list_empty(&bdi->bdi_list))) {
1026 * The normal path. Keep writing back @bdi until its
1027 * work_list is empty. Note that this path is also taken
1028 * if @bdi is shutting down even when we're running off the
1029 * rescuer as work_list needs to be drained.
1032 pages_written = wb_do_writeback(wb);
1033 trace_writeback_pages_written(pages_written);
1034 } while (!list_empty(&bdi->work_list));
1037 * bdi_wq can't get enough workers and we're running off
1038 * the emergency worker. Don't hog it. Hopefully, 1024 is
1039 * enough for efficient IO.
1041 pages_written = writeback_inodes_wb(&bdi->wb, 1024,
1042 WB_REASON_FORKER_THREAD);
1043 trace_writeback_pages_written(pages_written);
1046 if (!list_empty(&bdi->work_list) ||
1047 (wb_has_dirty_io(wb) && dirty_writeback_interval))
1048 queue_delayed_work(bdi_wq, &wb->dwork,
1049 msecs_to_jiffies(dirty_writeback_interval * 10));
1051 current->flags &= ~PF_SWAPWRITE;
1055 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1058 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1060 struct backing_dev_info *bdi;
1063 nr_pages = get_nr_dirty_pages();
1066 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1067 if (!bdi_has_dirty_io(bdi))
1069 __bdi_start_writeback(bdi, nr_pages, false, reason);
1074 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1076 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1077 struct dentry *dentry;
1078 const char *name = "?";
1080 dentry = d_find_alias(inode);
1082 spin_lock(&dentry->d_lock);
1083 name = (const char *) dentry->d_name.name;
1086 "%s(%d): dirtied inode %lu (%s) on %s\n",
1087 current->comm, task_pid_nr(current), inode->i_ino,
1088 name, inode->i_sb->s_id);
1090 spin_unlock(&dentry->d_lock);
1097 * __mark_inode_dirty - internal function
1098 * @inode: inode to mark
1099 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1100 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1101 * mark_inode_dirty_sync.
1103 * Put the inode on the super block's dirty list.
1105 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1106 * dirty list only if it is hashed or if it refers to a blockdev.
1107 * If it was not hashed, it will never be added to the dirty list
1108 * even if it is later hashed, as it will have been marked dirty already.
1110 * In short, make sure you hash any inodes _before_ you start marking
1113 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1114 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1115 * the kernel-internal blockdev inode represents the dirtying time of the
1116 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1117 * page->mapping->host, so the page-dirtying time is recorded in the internal
1120 void __mark_inode_dirty(struct inode *inode, int flags)
1122 struct super_block *sb = inode->i_sb;
1123 struct backing_dev_info *bdi = NULL;
1126 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1127 * dirty the inode itself
1129 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1130 trace_writeback_dirty_inode_start(inode, flags);
1132 if (sb->s_op->dirty_inode)
1133 sb->s_op->dirty_inode(inode, flags);
1135 trace_writeback_dirty_inode(inode, flags);
1139 * make sure that changes are seen by all cpus before we test i_state
1144 /* avoid the locking if we can */
1145 if ((inode->i_state & flags) == flags)
1148 if (unlikely(block_dump))
1149 block_dump___mark_inode_dirty(inode);
1151 spin_lock(&inode->i_lock);
1152 if ((inode->i_state & flags) != flags) {
1153 const int was_dirty = inode->i_state & I_DIRTY;
1155 inode->i_state |= flags;
1158 * If the inode is being synced, just update its dirty state.
1159 * The unlocker will place the inode on the appropriate
1160 * superblock list, based upon its state.
1162 if (inode->i_state & I_SYNC)
1163 goto out_unlock_inode;
1166 * Only add valid (hashed) inodes to the superblock's
1167 * dirty list. Add blockdev inodes as well.
1169 if (!S_ISBLK(inode->i_mode)) {
1170 if (inode_unhashed(inode))
1171 goto out_unlock_inode;
1173 if (inode->i_state & I_FREEING)
1174 goto out_unlock_inode;
1177 * If the inode was already on b_dirty/b_io/b_more_io, don't
1178 * reposition it (that would break b_dirty time-ordering).
1181 bool wakeup_bdi = false;
1182 bdi = inode_to_bdi(inode);
1184 spin_unlock(&inode->i_lock);
1185 spin_lock(&bdi->wb.list_lock);
1186 if (bdi_cap_writeback_dirty(bdi)) {
1187 WARN(!test_bit(BDI_registered, &bdi->state),
1188 "bdi-%s not registered\n", bdi->name);
1191 * If this is the first dirty inode for this
1192 * bdi, we have to wake-up the corresponding
1193 * bdi thread to make sure background
1194 * write-back happens later.
1196 if (!wb_has_dirty_io(&bdi->wb))
1200 inode->dirtied_when = jiffies;
1201 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1202 spin_unlock(&bdi->wb.list_lock);
1205 bdi_wakeup_thread_delayed(bdi);
1210 spin_unlock(&inode->i_lock);
1213 EXPORT_SYMBOL(__mark_inode_dirty);
1215 static void wait_sb_inodes(struct super_block *sb)
1217 struct inode *inode, *old_inode = NULL;
1220 * We need to be protected against the filesystem going from
1221 * r/o to r/w or vice versa.
1223 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1225 spin_lock(&inode_sb_list_lock);
1228 * Data integrity sync. Must wait for all pages under writeback,
1229 * because there may have been pages dirtied before our sync
1230 * call, but which had writeout started before we write it out.
1231 * In which case, the inode may not be on the dirty list, but
1232 * we still have to wait for that writeout.
1234 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1235 struct address_space *mapping = inode->i_mapping;
1237 spin_lock(&inode->i_lock);
1238 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1239 (mapping->nrpages == 0)) {
1240 spin_unlock(&inode->i_lock);
1244 spin_unlock(&inode->i_lock);
1245 spin_unlock(&inode_sb_list_lock);
1248 * We hold a reference to 'inode' so it couldn't have been
1249 * removed from s_inodes list while we dropped the
1250 * inode_sb_list_lock. We cannot iput the inode now as we can
1251 * be holding the last reference and we cannot iput it under
1252 * inode_sb_list_lock. So we keep the reference and iput it
1258 filemap_fdatawait(mapping);
1262 spin_lock(&inode_sb_list_lock);
1264 spin_unlock(&inode_sb_list_lock);
1269 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1270 * @sb: the superblock
1271 * @nr: the number of pages to write
1272 * @reason: reason why some writeback work initiated
1274 * Start writeback on some inodes on this super_block. No guarantees are made
1275 * on how many (if any) will be written, and this function does not wait
1276 * for IO completion of submitted IO.
1278 void writeback_inodes_sb_nr(struct super_block *sb,
1280 enum wb_reason reason)
1282 DECLARE_COMPLETION_ONSTACK(done);
1283 struct wb_writeback_work work = {
1285 .sync_mode = WB_SYNC_NONE,
1286 .tagged_writepages = 1,
1292 if (sb->s_bdi == &noop_backing_dev_info)
1294 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1295 bdi_queue_work(sb->s_bdi, &work);
1296 wait_for_completion(&done);
1298 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1301 * writeback_inodes_sb - writeback dirty inodes from given super_block
1302 * @sb: the superblock
1303 * @reason: reason why some writeback work was initiated
1305 * Start writeback on some inodes on this super_block. No guarantees are made
1306 * on how many (if any) will be written, and this function does not wait
1307 * for IO completion of submitted IO.
1309 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1311 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1313 EXPORT_SYMBOL(writeback_inodes_sb);
1316 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1317 * @sb: the superblock
1318 * @nr: the number of pages to write
1319 * @reason: the reason of writeback
1321 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1322 * Returns 1 if writeback was started, 0 if not.
1324 int try_to_writeback_inodes_sb_nr(struct super_block *sb,
1326 enum wb_reason reason)
1328 if (writeback_in_progress(sb->s_bdi))
1331 if (!down_read_trylock(&sb->s_umount))
1334 writeback_inodes_sb_nr(sb, nr, reason);
1335 up_read(&sb->s_umount);
1338 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
1341 * try_to_writeback_inodes_sb - try to start writeback if none underway
1342 * @sb: the superblock
1343 * @reason: reason why some writeback work was initiated
1345 * Implement by try_to_writeback_inodes_sb_nr()
1346 * Returns 1 if writeback was started, 0 if not.
1348 int try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1350 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1352 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
1355 * sync_inodes_sb - sync sb inode pages
1356 * @sb: the superblock
1357 * @older_than_this: timestamp
1359 * This function writes and waits on any dirty inode belonging to this
1360 * superblock that has been dirtied before given timestamp.
1362 void sync_inodes_sb(struct super_block *sb, unsigned long older_than_this)
1364 DECLARE_COMPLETION_ONSTACK(done);
1365 struct wb_writeback_work work = {
1367 .sync_mode = WB_SYNC_ALL,
1368 .nr_pages = LONG_MAX,
1369 .older_than_this = older_than_this,
1370 .older_than_this_is_set = 1,
1373 .reason = WB_REASON_SYNC,
1377 /* Nothing to do? */
1378 if (sb->s_bdi == &noop_backing_dev_info)
1380 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1382 bdi_queue_work(sb->s_bdi, &work);
1383 wait_for_completion(&done);
1387 EXPORT_SYMBOL(sync_inodes_sb);
1390 * write_inode_now - write an inode to disk
1391 * @inode: inode to write to disk
1392 * @sync: whether the write should be synchronous or not
1394 * This function commits an inode to disk immediately if it is dirty. This is
1395 * primarily needed by knfsd.
1397 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1399 int write_inode_now(struct inode *inode, int sync)
1401 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1402 struct writeback_control wbc = {
1403 .nr_to_write = LONG_MAX,
1404 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1406 .range_end = LLONG_MAX,
1409 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1410 wbc.nr_to_write = 0;
1413 return writeback_single_inode(inode, wb, &wbc);
1415 EXPORT_SYMBOL(write_inode_now);
1418 * sync_inode - write an inode and its pages to disk.
1419 * @inode: the inode to sync
1420 * @wbc: controls the writeback mode
1422 * sync_inode() will write an inode and its pages to disk. It will also
1423 * correctly update the inode on its superblock's dirty inode lists and will
1424 * update inode->i_state.
1426 * The caller must have a ref on the inode.
1428 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1430 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
1432 EXPORT_SYMBOL(sync_inode);
1435 * sync_inode_metadata - write an inode to disk
1436 * @inode: the inode to sync
1437 * @wait: wait for I/O to complete.
1439 * Write an inode to disk and adjust its dirty state after completion.
1441 * Note: only writes the actual inode, no associated data or other metadata.
1443 int sync_inode_metadata(struct inode *inode, int wait)
1445 struct writeback_control wbc = {
1446 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1447 .nr_to_write = 0, /* metadata-only */
1450 return sync_inode(inode, &wbc);
1452 EXPORT_SYMBOL(sync_inode_metadata);