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>
29 #include <linux/device.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 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
50 enum wb_reason reason; /* why was writeback initiated? */
52 struct list_head list; /* pending work list */
53 struct completion *done; /* set if the caller waits */
57 * If an inode is constantly having its pages dirtied, but then the
58 * updates stop dirtytime_expire_interval seconds in the past, it's
59 * possible for the worst case time between when an inode has its
60 * timestamps updated and when they finally get written out to be two
61 * dirtytime_expire_intervals. We set the default to 12 hours (in
62 * seconds), which means most of the time inodes will have their
63 * timestamps written to disk after 12 hours, but in the worst case a
64 * few inodes might not their timestamps updated for 24 hours.
66 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
69 * writeback_in_progress - determine whether there is writeback in progress
70 * @bdi: the device's backing_dev_info structure.
72 * Determine whether there is writeback waiting to be handled against a
75 int writeback_in_progress(struct backing_dev_info *bdi)
77 return test_bit(WB_writeback_running, &bdi->wb.state);
79 EXPORT_SYMBOL(writeback_in_progress);
81 static inline struct inode *wb_inode(struct list_head *head)
83 return list_entry(head, struct inode, i_wb_list);
87 * Include the creation of the trace points after defining the
88 * wb_writeback_work structure and inline functions so that the definition
89 * remains local to this file.
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/writeback.h>
94 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
96 static void wb_wakeup(struct bdi_writeback *wb)
98 spin_lock_bh(&wb->work_lock);
99 if (test_bit(WB_registered, &wb->state))
100 mod_delayed_work(bdi_wq, &wb->dwork, 0);
101 spin_unlock_bh(&wb->work_lock);
104 static void wb_queue_work(struct bdi_writeback *wb,
105 struct wb_writeback_work *work)
107 trace_writeback_queue(wb->bdi, work);
109 spin_lock_bh(&wb->work_lock);
110 if (!test_bit(WB_registered, &wb->state)) {
112 complete(work->done);
115 list_add_tail(&work->list, &wb->work_list);
116 mod_delayed_work(bdi_wq, &wb->dwork, 0);
118 spin_unlock_bh(&wb->work_lock);
121 static void __wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
122 bool range_cyclic, enum wb_reason reason)
124 struct wb_writeback_work *work;
127 * This is WB_SYNC_NONE writeback, so if allocation fails just
128 * wakeup the thread for old dirty data writeback
130 work = kzalloc(sizeof(*work), GFP_ATOMIC);
132 trace_writeback_nowork(wb->bdi);
137 work->sync_mode = WB_SYNC_NONE;
138 work->nr_pages = nr_pages;
139 work->range_cyclic = range_cyclic;
140 work->reason = reason;
142 wb_queue_work(wb, work);
145 #ifdef CONFIG_CGROUP_WRITEBACK
148 * inode_congested - test whether an inode is congested
149 * @inode: inode to test for congestion
150 * @cong_bits: mask of WB_[a]sync_congested bits to test
152 * Tests whether @inode is congested. @cong_bits is the mask of congestion
153 * bits to test and the return value is the mask of set bits.
155 * If cgroup writeback is enabled for @inode, the congestion state is
156 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
157 * associated with @inode is congested; otherwise, the root wb's congestion
160 int inode_congested(struct inode *inode, int cong_bits)
163 struct bdi_writeback *wb = inode_to_wb(inode);
165 return wb_congested(wb, cong_bits);
168 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
170 EXPORT_SYMBOL_GPL(inode_congested);
172 #endif /* CONFIG_CGROUP_WRITEBACK */
175 * bdi_start_writeback - start writeback
176 * @bdi: the backing device to write from
177 * @nr_pages: the number of pages to write
178 * @reason: reason why some writeback work was initiated
181 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
182 * started when this function returns, we make no guarantees on
183 * completion. Caller need not hold sb s_umount semaphore.
186 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
187 enum wb_reason reason)
189 __wb_start_writeback(&bdi->wb, nr_pages, true, reason);
193 * bdi_start_background_writeback - start background writeback
194 * @bdi: the backing device to write from
197 * This makes sure WB_SYNC_NONE background writeback happens. When
198 * this function returns, it is only guaranteed that for given BDI
199 * some IO is happening if we are over background dirty threshold.
200 * Caller need not hold sb s_umount semaphore.
202 void bdi_start_background_writeback(struct backing_dev_info *bdi)
205 * We just wake up the flusher thread. It will perform background
206 * writeback as soon as there is no other work to do.
208 trace_writeback_wake_background(bdi);
213 * Remove the inode from the writeback list it is on.
215 void inode_wb_list_del(struct inode *inode)
217 struct bdi_writeback *wb = inode_to_wb(inode);
219 spin_lock(&wb->list_lock);
220 list_del_init(&inode->i_wb_list);
221 spin_unlock(&wb->list_lock);
225 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
226 * furthest end of its superblock's dirty-inode list.
228 * Before stamping the inode's ->dirtied_when, we check to see whether it is
229 * already the most-recently-dirtied inode on the b_dirty list. If that is
230 * the case then the inode must have been redirtied while it was being written
231 * out and we don't reset its dirtied_when.
233 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
235 assert_spin_locked(&wb->list_lock);
236 if (!list_empty(&wb->b_dirty)) {
239 tail = wb_inode(wb->b_dirty.next);
240 if (time_before(inode->dirtied_when, tail->dirtied_when))
241 inode->dirtied_when = jiffies;
243 list_move(&inode->i_wb_list, &wb->b_dirty);
247 * requeue inode for re-scanning after bdi->b_io list is exhausted.
249 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
251 assert_spin_locked(&wb->list_lock);
252 list_move(&inode->i_wb_list, &wb->b_more_io);
255 static void inode_sync_complete(struct inode *inode)
257 inode->i_state &= ~I_SYNC;
258 /* If inode is clean an unused, put it into LRU now... */
259 inode_add_lru(inode);
260 /* Waiters must see I_SYNC cleared before being woken up */
262 wake_up_bit(&inode->i_state, __I_SYNC);
265 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
267 bool ret = time_after(inode->dirtied_when, t);
270 * For inodes being constantly redirtied, dirtied_when can get stuck.
271 * It _appears_ to be in the future, but is actually in distant past.
272 * This test is necessary to prevent such wrapped-around relative times
273 * from permanently stopping the whole bdi writeback.
275 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
280 #define EXPIRE_DIRTY_ATIME 0x0001
283 * Move expired (dirtied before work->older_than_this) dirty inodes from
284 * @delaying_queue to @dispatch_queue.
286 static int move_expired_inodes(struct list_head *delaying_queue,
287 struct list_head *dispatch_queue,
289 struct wb_writeback_work *work)
291 unsigned long *older_than_this = NULL;
292 unsigned long expire_time;
294 struct list_head *pos, *node;
295 struct super_block *sb = NULL;
300 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
301 older_than_this = work->older_than_this;
302 else if (!work->for_sync) {
303 expire_time = jiffies - (dirtytime_expire_interval * HZ);
304 older_than_this = &expire_time;
306 while (!list_empty(delaying_queue)) {
307 inode = wb_inode(delaying_queue->prev);
308 if (older_than_this &&
309 inode_dirtied_after(inode, *older_than_this))
311 list_move(&inode->i_wb_list, &tmp);
313 if (flags & EXPIRE_DIRTY_ATIME)
314 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
315 if (sb_is_blkdev_sb(inode->i_sb))
317 if (sb && sb != inode->i_sb)
322 /* just one sb in list, splice to dispatch_queue and we're done */
324 list_splice(&tmp, dispatch_queue);
328 /* Move inodes from one superblock together */
329 while (!list_empty(&tmp)) {
330 sb = wb_inode(tmp.prev)->i_sb;
331 list_for_each_prev_safe(pos, node, &tmp) {
332 inode = wb_inode(pos);
333 if (inode->i_sb == sb)
334 list_move(&inode->i_wb_list, dispatch_queue);
342 * Queue all expired dirty inodes for io, eldest first.
344 * newly dirtied b_dirty b_io b_more_io
345 * =============> gf edc BA
347 * newly dirtied b_dirty b_io b_more_io
348 * =============> g fBAedc
350 * +--> dequeue for IO
352 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
356 assert_spin_locked(&wb->list_lock);
357 list_splice_init(&wb->b_more_io, &wb->b_io);
358 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
359 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
360 EXPIRE_DIRTY_ATIME, work);
361 trace_writeback_queue_io(wb, work, moved);
364 static int write_inode(struct inode *inode, struct writeback_control *wbc)
368 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
369 trace_writeback_write_inode_start(inode, wbc);
370 ret = inode->i_sb->s_op->write_inode(inode, wbc);
371 trace_writeback_write_inode(inode, wbc);
378 * Wait for writeback on an inode to complete. Called with i_lock held.
379 * Caller must make sure inode cannot go away when we drop i_lock.
381 static void __inode_wait_for_writeback(struct inode *inode)
382 __releases(inode->i_lock)
383 __acquires(inode->i_lock)
385 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
386 wait_queue_head_t *wqh;
388 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
389 while (inode->i_state & I_SYNC) {
390 spin_unlock(&inode->i_lock);
391 __wait_on_bit(wqh, &wq, bit_wait,
392 TASK_UNINTERRUPTIBLE);
393 spin_lock(&inode->i_lock);
398 * Wait for writeback on an inode to complete. Caller must have inode pinned.
400 void inode_wait_for_writeback(struct inode *inode)
402 spin_lock(&inode->i_lock);
403 __inode_wait_for_writeback(inode);
404 spin_unlock(&inode->i_lock);
408 * Sleep until I_SYNC is cleared. This function must be called with i_lock
409 * held and drops it. It is aimed for callers not holding any inode reference
410 * so once i_lock is dropped, inode can go away.
412 static void inode_sleep_on_writeback(struct inode *inode)
413 __releases(inode->i_lock)
416 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
419 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
420 sleep = inode->i_state & I_SYNC;
421 spin_unlock(&inode->i_lock);
424 finish_wait(wqh, &wait);
428 * Find proper writeback list for the inode depending on its current state and
429 * possibly also change of its state while we were doing writeback. Here we
430 * handle things such as livelock prevention or fairness of writeback among
431 * inodes. This function can be called only by flusher thread - noone else
432 * processes all inodes in writeback lists and requeueing inodes behind flusher
433 * thread's back can have unexpected consequences.
435 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
436 struct writeback_control *wbc)
438 if (inode->i_state & I_FREEING)
442 * Sync livelock prevention. Each inode is tagged and synced in one
443 * shot. If still dirty, it will be redirty_tail()'ed below. Update
444 * the dirty time to prevent enqueue and sync it again.
446 if ((inode->i_state & I_DIRTY) &&
447 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
448 inode->dirtied_when = jiffies;
450 if (wbc->pages_skipped) {
452 * writeback is not making progress due to locked
453 * buffers. Skip this inode for now.
455 redirty_tail(inode, wb);
459 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
461 * We didn't write back all the pages. nfs_writepages()
462 * sometimes bales out without doing anything.
464 if (wbc->nr_to_write <= 0) {
465 /* Slice used up. Queue for next turn. */
466 requeue_io(inode, wb);
469 * Writeback blocked by something other than
470 * congestion. Delay the inode for some time to
471 * avoid spinning on the CPU (100% iowait)
472 * retrying writeback of the dirty page/inode
473 * that cannot be performed immediately.
475 redirty_tail(inode, wb);
477 } else if (inode->i_state & I_DIRTY) {
479 * Filesystems can dirty the inode during writeback operations,
480 * such as delayed allocation during submission or metadata
481 * updates after data IO completion.
483 redirty_tail(inode, wb);
484 } else if (inode->i_state & I_DIRTY_TIME) {
485 inode->dirtied_when = jiffies;
486 list_move(&inode->i_wb_list, &wb->b_dirty_time);
488 /* The inode is clean. Remove from writeback lists. */
489 list_del_init(&inode->i_wb_list);
494 * Write out an inode and its dirty pages. Do not update the writeback list
495 * linkage. That is left to the caller. The caller is also responsible for
496 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
499 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
501 struct address_space *mapping = inode->i_mapping;
502 long nr_to_write = wbc->nr_to_write;
506 WARN_ON(!(inode->i_state & I_SYNC));
508 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
510 ret = do_writepages(mapping, wbc);
513 * Make sure to wait on the data before writing out the metadata.
514 * This is important for filesystems that modify metadata on data
515 * I/O completion. We don't do it for sync(2) writeback because it has a
516 * separate, external IO completion path and ->sync_fs for guaranteeing
517 * inode metadata is written back correctly.
519 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
520 int err = filemap_fdatawait(mapping);
526 * Some filesystems may redirty the inode during the writeback
527 * due to delalloc, clear dirty metadata flags right before
530 spin_lock(&inode->i_lock);
532 dirty = inode->i_state & I_DIRTY;
533 if (inode->i_state & I_DIRTY_TIME) {
534 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
535 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
536 unlikely(time_after(jiffies,
537 (inode->dirtied_time_when +
538 dirtytime_expire_interval * HZ)))) {
539 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
540 trace_writeback_lazytime(inode);
543 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
544 inode->i_state &= ~dirty;
547 * Paired with smp_mb() in __mark_inode_dirty(). This allows
548 * __mark_inode_dirty() to test i_state without grabbing i_lock -
549 * either they see the I_DIRTY bits cleared or we see the dirtied
552 * I_DIRTY_PAGES is always cleared together above even if @mapping
553 * still has dirty pages. The flag is reinstated after smp_mb() if
554 * necessary. This guarantees that either __mark_inode_dirty()
555 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
559 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
560 inode->i_state |= I_DIRTY_PAGES;
562 spin_unlock(&inode->i_lock);
564 if (dirty & I_DIRTY_TIME)
565 mark_inode_dirty_sync(inode);
566 /* Don't write the inode if only I_DIRTY_PAGES was set */
567 if (dirty & ~I_DIRTY_PAGES) {
568 int err = write_inode(inode, wbc);
572 trace_writeback_single_inode(inode, wbc, nr_to_write);
577 * Write out an inode's dirty pages. Either the caller has an active reference
578 * on the inode or the inode has I_WILL_FREE set.
580 * This function is designed to be called for writing back one inode which
581 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
582 * and does more profound writeback list handling in writeback_sb_inodes().
585 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
586 struct writeback_control *wbc)
590 spin_lock(&inode->i_lock);
591 if (!atomic_read(&inode->i_count))
592 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
594 WARN_ON(inode->i_state & I_WILL_FREE);
596 if (inode->i_state & I_SYNC) {
597 if (wbc->sync_mode != WB_SYNC_ALL)
600 * It's a data-integrity sync. We must wait. Since callers hold
601 * inode reference or inode has I_WILL_FREE set, it cannot go
604 __inode_wait_for_writeback(inode);
606 WARN_ON(inode->i_state & I_SYNC);
608 * Skip inode if it is clean and we have no outstanding writeback in
609 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
610 * function since flusher thread may be doing for example sync in
611 * parallel and if we move the inode, it could get skipped. So here we
612 * make sure inode is on some writeback list and leave it there unless
613 * we have completely cleaned the inode.
615 if (!(inode->i_state & I_DIRTY_ALL) &&
616 (wbc->sync_mode != WB_SYNC_ALL ||
617 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
619 inode->i_state |= I_SYNC;
620 spin_unlock(&inode->i_lock);
622 ret = __writeback_single_inode(inode, wbc);
624 spin_lock(&wb->list_lock);
625 spin_lock(&inode->i_lock);
627 * If inode is clean, remove it from writeback lists. Otherwise don't
628 * touch it. See comment above for explanation.
630 if (!(inode->i_state & I_DIRTY_ALL))
631 list_del_init(&inode->i_wb_list);
632 spin_unlock(&wb->list_lock);
633 inode_sync_complete(inode);
635 spin_unlock(&inode->i_lock);
639 static long writeback_chunk_size(struct bdi_writeback *wb,
640 struct wb_writeback_work *work)
645 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
646 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
647 * here avoids calling into writeback_inodes_wb() more than once.
649 * The intended call sequence for WB_SYNC_ALL writeback is:
652 * writeback_sb_inodes() <== called only once
653 * write_cache_pages() <== called once for each inode
654 * (quickly) tag currently dirty pages
655 * (maybe slowly) sync all tagged pages
657 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
660 pages = min(wb->avg_write_bandwidth / 2,
661 global_dirty_limit / DIRTY_SCOPE);
662 pages = min(pages, work->nr_pages);
663 pages = round_down(pages + MIN_WRITEBACK_PAGES,
664 MIN_WRITEBACK_PAGES);
671 * Write a portion of b_io inodes which belong to @sb.
673 * Return the number of pages and/or inodes written.
675 static long writeback_sb_inodes(struct super_block *sb,
676 struct bdi_writeback *wb,
677 struct wb_writeback_work *work)
679 struct writeback_control wbc = {
680 .sync_mode = work->sync_mode,
681 .tagged_writepages = work->tagged_writepages,
682 .for_kupdate = work->for_kupdate,
683 .for_background = work->for_background,
684 .for_sync = work->for_sync,
685 .range_cyclic = work->range_cyclic,
687 .range_end = LLONG_MAX,
689 unsigned long start_time = jiffies;
691 long wrote = 0; /* count both pages and inodes */
693 while (!list_empty(&wb->b_io)) {
694 struct inode *inode = wb_inode(wb->b_io.prev);
696 if (inode->i_sb != sb) {
699 * We only want to write back data for this
700 * superblock, move all inodes not belonging
701 * to it back onto the dirty list.
703 redirty_tail(inode, wb);
708 * The inode belongs to a different superblock.
709 * Bounce back to the caller to unpin this and
710 * pin the next superblock.
716 * Don't bother with new inodes or inodes being freed, first
717 * kind does not need periodic writeout yet, and for the latter
718 * kind writeout is handled by the freer.
720 spin_lock(&inode->i_lock);
721 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
722 spin_unlock(&inode->i_lock);
723 redirty_tail(inode, wb);
726 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
728 * If this inode is locked for writeback and we are not
729 * doing writeback-for-data-integrity, move it to
730 * b_more_io so that writeback can proceed with the
731 * other inodes on s_io.
733 * We'll have another go at writing back this inode
734 * when we completed a full scan of b_io.
736 spin_unlock(&inode->i_lock);
737 requeue_io(inode, wb);
738 trace_writeback_sb_inodes_requeue(inode);
741 spin_unlock(&wb->list_lock);
744 * We already requeued the inode if it had I_SYNC set and we
745 * are doing WB_SYNC_NONE writeback. So this catches only the
748 if (inode->i_state & I_SYNC) {
749 /* Wait for I_SYNC. This function drops i_lock... */
750 inode_sleep_on_writeback(inode);
751 /* Inode may be gone, start again */
752 spin_lock(&wb->list_lock);
755 inode->i_state |= I_SYNC;
756 spin_unlock(&inode->i_lock);
758 write_chunk = writeback_chunk_size(wb, work);
759 wbc.nr_to_write = write_chunk;
760 wbc.pages_skipped = 0;
763 * We use I_SYNC to pin the inode in memory. While it is set
764 * evict_inode() will wait so the inode cannot be freed.
766 __writeback_single_inode(inode, &wbc);
768 work->nr_pages -= write_chunk - wbc.nr_to_write;
769 wrote += write_chunk - wbc.nr_to_write;
770 spin_lock(&wb->list_lock);
771 spin_lock(&inode->i_lock);
772 if (!(inode->i_state & I_DIRTY_ALL))
774 requeue_inode(inode, wb, &wbc);
775 inode_sync_complete(inode);
776 spin_unlock(&inode->i_lock);
777 cond_resched_lock(&wb->list_lock);
779 * bail out to wb_writeback() often enough to check
780 * background threshold and other termination conditions.
783 if (time_is_before_jiffies(start_time + HZ / 10UL))
785 if (work->nr_pages <= 0)
792 static long __writeback_inodes_wb(struct bdi_writeback *wb,
793 struct wb_writeback_work *work)
795 unsigned long start_time = jiffies;
798 while (!list_empty(&wb->b_io)) {
799 struct inode *inode = wb_inode(wb->b_io.prev);
800 struct super_block *sb = inode->i_sb;
802 if (!trylock_super(sb)) {
804 * trylock_super() may fail consistently due to
805 * s_umount being grabbed by someone else. Don't use
806 * requeue_io() to avoid busy retrying the inode/sb.
808 redirty_tail(inode, wb);
811 wrote += writeback_sb_inodes(sb, wb, work);
812 up_read(&sb->s_umount);
814 /* refer to the same tests at the end of writeback_sb_inodes */
816 if (time_is_before_jiffies(start_time + HZ / 10UL))
818 if (work->nr_pages <= 0)
822 /* Leave any unwritten inodes on b_io */
826 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
827 enum wb_reason reason)
829 struct wb_writeback_work work = {
830 .nr_pages = nr_pages,
831 .sync_mode = WB_SYNC_NONE,
836 spin_lock(&wb->list_lock);
837 if (list_empty(&wb->b_io))
839 __writeback_inodes_wb(wb, &work);
840 spin_unlock(&wb->list_lock);
842 return nr_pages - work.nr_pages;
845 static bool over_bground_thresh(struct bdi_writeback *wb)
847 unsigned long background_thresh, dirty_thresh;
849 global_dirty_limits(&background_thresh, &dirty_thresh);
851 if (global_page_state(NR_FILE_DIRTY) +
852 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
855 if (wb_stat(wb, WB_RECLAIMABLE) > wb_dirty_limit(wb, background_thresh))
862 * Called under wb->list_lock. If there are multiple wb per bdi,
863 * only the flusher working on the first wb should do it.
865 static void wb_update_bandwidth(struct bdi_writeback *wb,
866 unsigned long start_time)
868 __wb_update_bandwidth(wb, 0, 0, 0, 0, 0, start_time);
872 * Explicit flushing or periodic writeback of "old" data.
874 * Define "old": the first time one of an inode's pages is dirtied, we mark the
875 * dirtying-time in the inode's address_space. So this periodic writeback code
876 * just walks the superblock inode list, writing back any inodes which are
877 * older than a specific point in time.
879 * Try to run once per dirty_writeback_interval. But if a writeback event
880 * takes longer than a dirty_writeback_interval interval, then leave a
883 * older_than_this takes precedence over nr_to_write. So we'll only write back
884 * all dirty pages if they are all attached to "old" mappings.
886 static long wb_writeback(struct bdi_writeback *wb,
887 struct wb_writeback_work *work)
889 unsigned long wb_start = jiffies;
890 long nr_pages = work->nr_pages;
891 unsigned long oldest_jif;
895 oldest_jif = jiffies;
896 work->older_than_this = &oldest_jif;
898 spin_lock(&wb->list_lock);
901 * Stop writeback when nr_pages has been consumed
903 if (work->nr_pages <= 0)
907 * Background writeout and kupdate-style writeback may
908 * run forever. Stop them if there is other work to do
909 * so that e.g. sync can proceed. They'll be restarted
910 * after the other works are all done.
912 if ((work->for_background || work->for_kupdate) &&
913 !list_empty(&wb->work_list))
917 * For background writeout, stop when we are below the
918 * background dirty threshold
920 if (work->for_background && !over_bground_thresh(wb))
924 * Kupdate and background works are special and we want to
925 * include all inodes that need writing. Livelock avoidance is
926 * handled by these works yielding to any other work so we are
929 if (work->for_kupdate) {
930 oldest_jif = jiffies -
931 msecs_to_jiffies(dirty_expire_interval * 10);
932 } else if (work->for_background)
933 oldest_jif = jiffies;
935 trace_writeback_start(wb->bdi, work);
936 if (list_empty(&wb->b_io))
939 progress = writeback_sb_inodes(work->sb, wb, work);
941 progress = __writeback_inodes_wb(wb, work);
942 trace_writeback_written(wb->bdi, work);
944 wb_update_bandwidth(wb, wb_start);
947 * Did we write something? Try for more
949 * Dirty inodes are moved to b_io for writeback in batches.
950 * The completion of the current batch does not necessarily
951 * mean the overall work is done. So we keep looping as long
952 * as made some progress on cleaning pages or inodes.
957 * No more inodes for IO, bail
959 if (list_empty(&wb->b_more_io))
962 * Nothing written. Wait for some inode to
963 * become available for writeback. Otherwise
964 * we'll just busyloop.
966 if (!list_empty(&wb->b_more_io)) {
967 trace_writeback_wait(wb->bdi, work);
968 inode = wb_inode(wb->b_more_io.prev);
969 spin_lock(&inode->i_lock);
970 spin_unlock(&wb->list_lock);
971 /* This function drops i_lock... */
972 inode_sleep_on_writeback(inode);
973 spin_lock(&wb->list_lock);
976 spin_unlock(&wb->list_lock);
978 return nr_pages - work->nr_pages;
982 * Return the next wb_writeback_work struct that hasn't been processed yet.
984 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
986 struct wb_writeback_work *work = NULL;
988 spin_lock_bh(&wb->work_lock);
989 if (!list_empty(&wb->work_list)) {
990 work = list_entry(wb->work_list.next,
991 struct wb_writeback_work, list);
992 list_del_init(&work->list);
994 spin_unlock_bh(&wb->work_lock);
999 * Add in the number of potentially dirty inodes, because each inode
1000 * write can dirty pagecache in the underlying blockdev.
1002 static unsigned long get_nr_dirty_pages(void)
1004 return global_page_state(NR_FILE_DIRTY) +
1005 global_page_state(NR_UNSTABLE_NFS) +
1006 get_nr_dirty_inodes();
1009 static long wb_check_background_flush(struct bdi_writeback *wb)
1011 if (over_bground_thresh(wb)) {
1013 struct wb_writeback_work work = {
1014 .nr_pages = LONG_MAX,
1015 .sync_mode = WB_SYNC_NONE,
1016 .for_background = 1,
1018 .reason = WB_REASON_BACKGROUND,
1021 return wb_writeback(wb, &work);
1027 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1029 unsigned long expired;
1033 * When set to zero, disable periodic writeback
1035 if (!dirty_writeback_interval)
1038 expired = wb->last_old_flush +
1039 msecs_to_jiffies(dirty_writeback_interval * 10);
1040 if (time_before(jiffies, expired))
1043 wb->last_old_flush = jiffies;
1044 nr_pages = get_nr_dirty_pages();
1047 struct wb_writeback_work work = {
1048 .nr_pages = nr_pages,
1049 .sync_mode = WB_SYNC_NONE,
1052 .reason = WB_REASON_PERIODIC,
1055 return wb_writeback(wb, &work);
1062 * Retrieve work items and do the writeback they describe
1064 static long wb_do_writeback(struct bdi_writeback *wb)
1066 struct wb_writeback_work *work;
1069 set_bit(WB_writeback_running, &wb->state);
1070 while ((work = get_next_work_item(wb)) != NULL) {
1072 trace_writeback_exec(wb->bdi, work);
1074 wrote += wb_writeback(wb, work);
1077 * Notify the caller of completion if this is a synchronous
1078 * work item, otherwise just free it.
1081 complete(work->done);
1087 * Check for periodic writeback, kupdated() style
1089 wrote += wb_check_old_data_flush(wb);
1090 wrote += wb_check_background_flush(wb);
1091 clear_bit(WB_writeback_running, &wb->state);
1097 * Handle writeback of dirty data for the device backed by this bdi. Also
1098 * reschedules periodically and does kupdated style flushing.
1100 void wb_workfn(struct work_struct *work)
1102 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1103 struct bdi_writeback, dwork);
1106 set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1107 current->flags |= PF_SWAPWRITE;
1109 if (likely(!current_is_workqueue_rescuer() ||
1110 !test_bit(WB_registered, &wb->state))) {
1112 * The normal path. Keep writing back @wb until its
1113 * work_list is empty. Note that this path is also taken
1114 * if @wb is shutting down even when we're running off the
1115 * rescuer as work_list needs to be drained.
1118 pages_written = wb_do_writeback(wb);
1119 trace_writeback_pages_written(pages_written);
1120 } while (!list_empty(&wb->work_list));
1123 * bdi_wq can't get enough workers and we're running off
1124 * the emergency worker. Don't hog it. Hopefully, 1024 is
1125 * enough for efficient IO.
1127 pages_written = writeback_inodes_wb(wb, 1024,
1128 WB_REASON_FORKER_THREAD);
1129 trace_writeback_pages_written(pages_written);
1132 if (!list_empty(&wb->work_list))
1133 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1134 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1135 wb_wakeup_delayed(wb);
1137 current->flags &= ~PF_SWAPWRITE;
1141 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1144 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1146 struct backing_dev_info *bdi;
1149 nr_pages = get_nr_dirty_pages();
1152 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1153 if (!bdi_has_dirty_io(bdi))
1155 __wb_start_writeback(&bdi->wb, nr_pages, false, reason);
1161 * Wake up bdi's periodically to make sure dirtytime inodes gets
1162 * written back periodically. We deliberately do *not* check the
1163 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1164 * kernel to be constantly waking up once there are any dirtytime
1165 * inodes on the system. So instead we define a separate delayed work
1166 * function which gets called much more rarely. (By default, only
1167 * once every 12 hours.)
1169 * If there is any other write activity going on in the file system,
1170 * this function won't be necessary. But if the only thing that has
1171 * happened on the file system is a dirtytime inode caused by an atime
1172 * update, we need this infrastructure below to make sure that inode
1173 * eventually gets pushed out to disk.
1175 static void wakeup_dirtytime_writeback(struct work_struct *w);
1176 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1178 static void wakeup_dirtytime_writeback(struct work_struct *w)
1180 struct backing_dev_info *bdi;
1183 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1184 if (list_empty(&bdi->wb.b_dirty_time))
1186 wb_wakeup(&bdi->wb);
1189 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1192 static int __init start_dirtytime_writeback(void)
1194 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1197 __initcall(start_dirtytime_writeback);
1199 int dirtytime_interval_handler(struct ctl_table *table, int write,
1200 void __user *buffer, size_t *lenp, loff_t *ppos)
1204 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1205 if (ret == 0 && write)
1206 mod_delayed_work(system_wq, &dirtytime_work, 0);
1210 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1212 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1213 struct dentry *dentry;
1214 const char *name = "?";
1216 dentry = d_find_alias(inode);
1218 spin_lock(&dentry->d_lock);
1219 name = (const char *) dentry->d_name.name;
1222 "%s(%d): dirtied inode %lu (%s) on %s\n",
1223 current->comm, task_pid_nr(current), inode->i_ino,
1224 name, inode->i_sb->s_id);
1226 spin_unlock(&dentry->d_lock);
1233 * __mark_inode_dirty - internal function
1234 * @inode: inode to mark
1235 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1236 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1237 * mark_inode_dirty_sync.
1239 * Put the inode on the super block's dirty list.
1241 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1242 * dirty list only if it is hashed or if it refers to a blockdev.
1243 * If it was not hashed, it will never be added to the dirty list
1244 * even if it is later hashed, as it will have been marked dirty already.
1246 * In short, make sure you hash any inodes _before_ you start marking
1249 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1250 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1251 * the kernel-internal blockdev inode represents the dirtying time of the
1252 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1253 * page->mapping->host, so the page-dirtying time is recorded in the internal
1256 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
1257 void __mark_inode_dirty(struct inode *inode, int flags)
1259 struct super_block *sb = inode->i_sb;
1260 struct backing_dev_info *bdi = NULL;
1263 trace_writeback_mark_inode_dirty(inode, flags);
1266 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1267 * dirty the inode itself
1269 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
1270 trace_writeback_dirty_inode_start(inode, flags);
1272 if (sb->s_op->dirty_inode)
1273 sb->s_op->dirty_inode(inode, flags);
1275 trace_writeback_dirty_inode(inode, flags);
1277 if (flags & I_DIRTY_INODE)
1278 flags &= ~I_DIRTY_TIME;
1279 dirtytime = flags & I_DIRTY_TIME;
1282 * Paired with smp_mb() in __writeback_single_inode() for the
1283 * following lockless i_state test. See there for details.
1287 if (((inode->i_state & flags) == flags) ||
1288 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
1291 if (unlikely(block_dump))
1292 block_dump___mark_inode_dirty(inode);
1294 spin_lock(&inode->i_lock);
1295 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
1296 goto out_unlock_inode;
1297 if ((inode->i_state & flags) != flags) {
1298 const int was_dirty = inode->i_state & I_DIRTY;
1300 inode_attach_wb(inode, NULL);
1302 if (flags & I_DIRTY_INODE)
1303 inode->i_state &= ~I_DIRTY_TIME;
1304 inode->i_state |= flags;
1307 * If the inode is being synced, just update its dirty state.
1308 * The unlocker will place the inode on the appropriate
1309 * superblock list, based upon its state.
1311 if (inode->i_state & I_SYNC)
1312 goto out_unlock_inode;
1315 * Only add valid (hashed) inodes to the superblock's
1316 * dirty list. Add blockdev inodes as well.
1318 if (!S_ISBLK(inode->i_mode)) {
1319 if (inode_unhashed(inode))
1320 goto out_unlock_inode;
1322 if (inode->i_state & I_FREEING)
1323 goto out_unlock_inode;
1326 * If the inode was already on b_dirty/b_io/b_more_io, don't
1327 * reposition it (that would break b_dirty time-ordering).
1330 bool wakeup_bdi = false;
1331 bdi = inode_to_bdi(inode);
1333 spin_unlock(&inode->i_lock);
1334 spin_lock(&bdi->wb.list_lock);
1335 if (bdi_cap_writeback_dirty(bdi)) {
1336 WARN(!test_bit(WB_registered, &bdi->wb.state),
1337 "bdi-%s not registered\n", bdi->name);
1340 * If this is the first dirty inode for this
1341 * bdi, we have to wake-up the corresponding
1342 * bdi thread to make sure background
1343 * write-back happens later.
1345 if (!wb_has_dirty_io(&bdi->wb))
1349 inode->dirtied_when = jiffies;
1351 inode->dirtied_time_when = jiffies;
1352 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
1353 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1355 list_move(&inode->i_wb_list,
1356 &bdi->wb.b_dirty_time);
1357 spin_unlock(&bdi->wb.list_lock);
1358 trace_writeback_dirty_inode_enqueue(inode);
1361 wb_wakeup_delayed(&bdi->wb);
1366 spin_unlock(&inode->i_lock);
1369 EXPORT_SYMBOL(__mark_inode_dirty);
1371 static void wait_sb_inodes(struct super_block *sb)
1373 struct inode *inode, *old_inode = NULL;
1376 * We need to be protected against the filesystem going from
1377 * r/o to r/w or vice versa.
1379 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1381 spin_lock(&inode_sb_list_lock);
1384 * Data integrity sync. Must wait for all pages under writeback,
1385 * because there may have been pages dirtied before our sync
1386 * call, but which had writeout started before we write it out.
1387 * In which case, the inode may not be on the dirty list, but
1388 * we still have to wait for that writeout.
1390 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1391 struct address_space *mapping = inode->i_mapping;
1393 spin_lock(&inode->i_lock);
1394 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1395 (mapping->nrpages == 0)) {
1396 spin_unlock(&inode->i_lock);
1400 spin_unlock(&inode->i_lock);
1401 spin_unlock(&inode_sb_list_lock);
1404 * We hold a reference to 'inode' so it couldn't have been
1405 * removed from s_inodes list while we dropped the
1406 * inode_sb_list_lock. We cannot iput the inode now as we can
1407 * be holding the last reference and we cannot iput it under
1408 * inode_sb_list_lock. So we keep the reference and iput it
1414 filemap_fdatawait(mapping);
1418 spin_lock(&inode_sb_list_lock);
1420 spin_unlock(&inode_sb_list_lock);
1425 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1426 * @sb: the superblock
1427 * @nr: the number of pages to write
1428 * @reason: reason why some writeback work initiated
1430 * Start writeback on some inodes on this super_block. No guarantees are made
1431 * on how many (if any) will be written, and this function does not wait
1432 * for IO completion of submitted IO.
1434 void writeback_inodes_sb_nr(struct super_block *sb,
1436 enum wb_reason reason)
1438 DECLARE_COMPLETION_ONSTACK(done);
1439 struct wb_writeback_work work = {
1441 .sync_mode = WB_SYNC_NONE,
1442 .tagged_writepages = 1,
1448 if (sb->s_bdi == &noop_backing_dev_info)
1450 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1451 wb_queue_work(&sb->s_bdi->wb, &work);
1452 wait_for_completion(&done);
1454 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1457 * writeback_inodes_sb - writeback dirty inodes from given super_block
1458 * @sb: the superblock
1459 * @reason: reason why some writeback work was initiated
1461 * Start writeback on some inodes on this super_block. No guarantees are made
1462 * on how many (if any) will be written, and this function does not wait
1463 * for IO completion of submitted IO.
1465 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1467 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1469 EXPORT_SYMBOL(writeback_inodes_sb);
1472 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
1473 * @sb: the superblock
1474 * @nr: the number of pages to write
1475 * @reason: the reason of writeback
1477 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
1478 * Returns 1 if writeback was started, 0 if not.
1480 int try_to_writeback_inodes_sb_nr(struct super_block *sb,
1482 enum wb_reason reason)
1484 if (writeback_in_progress(sb->s_bdi))
1487 if (!down_read_trylock(&sb->s_umount))
1490 writeback_inodes_sb_nr(sb, nr, reason);
1491 up_read(&sb->s_umount);
1494 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
1497 * try_to_writeback_inodes_sb - try to start writeback if none underway
1498 * @sb: the superblock
1499 * @reason: reason why some writeback work was initiated
1501 * Implement by try_to_writeback_inodes_sb_nr()
1502 * Returns 1 if writeback was started, 0 if not.
1504 int try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1506 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1508 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
1511 * sync_inodes_sb - sync sb inode pages
1512 * @sb: the superblock
1514 * This function writes and waits on any dirty inode belonging to this
1517 void sync_inodes_sb(struct super_block *sb)
1519 DECLARE_COMPLETION_ONSTACK(done);
1520 struct wb_writeback_work work = {
1522 .sync_mode = WB_SYNC_ALL,
1523 .nr_pages = LONG_MAX,
1526 .reason = WB_REASON_SYNC,
1530 /* Nothing to do? */
1531 if (sb->s_bdi == &noop_backing_dev_info)
1533 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1535 wb_queue_work(&sb->s_bdi->wb, &work);
1536 wait_for_completion(&done);
1540 EXPORT_SYMBOL(sync_inodes_sb);
1543 * write_inode_now - write an inode to disk
1544 * @inode: inode to write to disk
1545 * @sync: whether the write should be synchronous or not
1547 * This function commits an inode to disk immediately if it is dirty. This is
1548 * primarily needed by knfsd.
1550 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1552 int write_inode_now(struct inode *inode, int sync)
1554 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1555 struct writeback_control wbc = {
1556 .nr_to_write = LONG_MAX,
1557 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1559 .range_end = LLONG_MAX,
1562 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1563 wbc.nr_to_write = 0;
1566 return writeback_single_inode(inode, wb, &wbc);
1568 EXPORT_SYMBOL(write_inode_now);
1571 * sync_inode - write an inode and its pages to disk.
1572 * @inode: the inode to sync
1573 * @wbc: controls the writeback mode
1575 * sync_inode() will write an inode and its pages to disk. It will also
1576 * correctly update the inode on its superblock's dirty inode lists and will
1577 * update inode->i_state.
1579 * The caller must have a ref on the inode.
1581 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1583 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
1585 EXPORT_SYMBOL(sync_inode);
1588 * sync_inode_metadata - write an inode to disk
1589 * @inode: the inode to sync
1590 * @wait: wait for I/O to complete.
1592 * Write an inode to disk and adjust its dirty state after completion.
1594 * Note: only writes the actual inode, no associated data or other metadata.
1596 int sync_inode_metadata(struct inode *inode, int wait)
1598 struct writeback_control wbc = {
1599 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1600 .nr_to_write = 0, /* metadata-only */
1603 return sync_inode(inode, &wbc);
1605 EXPORT_SYMBOL(sync_inode_metadata);