Merge branch 'v4l_for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mchehab...
[zen-stable.git] / fs / fs-writeback.c
blobac86f8b3e3cb1cf050753148c3f3d065b18e6cf9
1 /*
2 * fs/fs-writeback.c
4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/fs.h>
22 #include <linux/mm.h>
23 #include <linux/kthread.h>
24 #include <linux/freezer.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/buffer_head.h>
29 #include <linux/tracepoint.h>
30 #include "internal.h"
33 * Passed into wb_writeback(), essentially a subset of writeback_control
35 struct wb_writeback_work {
36 long nr_pages;
37 struct super_block *sb;
38 unsigned long *older_than_this;
39 enum writeback_sync_modes sync_mode;
40 unsigned int tagged_writepages:1;
41 unsigned int for_kupdate:1;
42 unsigned int range_cyclic:1;
43 unsigned int for_background:1;
44 enum wb_reason reason; /* why was writeback initiated? */
46 struct list_head list; /* pending work list */
47 struct completion *done; /* set if the caller waits */
50 const char *wb_reason_name[] = {
51 [WB_REASON_BACKGROUND] = "background",
52 [WB_REASON_TRY_TO_FREE_PAGES] = "try_to_free_pages",
53 [WB_REASON_SYNC] = "sync",
54 [WB_REASON_PERIODIC] = "periodic",
55 [WB_REASON_LAPTOP_TIMER] = "laptop_timer",
56 [WB_REASON_FREE_MORE_MEM] = "free_more_memory",
57 [WB_REASON_FS_FREE_SPACE] = "fs_free_space",
58 [WB_REASON_FORKER_THREAD] = "forker_thread"
62 * Include the creation of the trace points after defining the
63 * wb_writeback_work structure so that the definition remains local to this
64 * file.
66 #define CREATE_TRACE_POINTS
67 #include <trace/events/writeback.h>
70 * We don't actually have pdflush, but this one is exported though /proc...
72 int nr_pdflush_threads;
74 /**
75 * writeback_in_progress - determine whether there is writeback in progress
76 * @bdi: the device's backing_dev_info structure.
78 * Determine whether there is writeback waiting to be handled against a
79 * backing device.
81 int writeback_in_progress(struct backing_dev_info *bdi)
83 return test_bit(BDI_writeback_running, &bdi->state);
86 static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
88 struct super_block *sb = inode->i_sb;
90 if (strcmp(sb->s_type->name, "bdev") == 0)
91 return inode->i_mapping->backing_dev_info;
93 return sb->s_bdi;
96 static inline struct inode *wb_inode(struct list_head *head)
98 return list_entry(head, struct inode, i_wb_list);
101 /* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
102 static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
104 if (bdi->wb.task) {
105 wake_up_process(bdi->wb.task);
106 } else {
108 * The bdi thread isn't there, wake up the forker thread which
109 * will create and run it.
111 wake_up_process(default_backing_dev_info.wb.task);
115 static void bdi_queue_work(struct backing_dev_info *bdi,
116 struct wb_writeback_work *work)
118 trace_writeback_queue(bdi, work);
120 spin_lock_bh(&bdi->wb_lock);
121 list_add_tail(&work->list, &bdi->work_list);
122 if (!bdi->wb.task)
123 trace_writeback_nothread(bdi, work);
124 bdi_wakeup_flusher(bdi);
125 spin_unlock_bh(&bdi->wb_lock);
128 static void
129 __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
130 bool range_cyclic, enum wb_reason reason)
132 struct wb_writeback_work *work;
135 * This is WB_SYNC_NONE writeback, so if allocation fails just
136 * wakeup the thread for old dirty data writeback
138 work = kzalloc(sizeof(*work), GFP_ATOMIC);
139 if (!work) {
140 if (bdi->wb.task) {
141 trace_writeback_nowork(bdi);
142 wake_up_process(bdi->wb.task);
144 return;
147 work->sync_mode = WB_SYNC_NONE;
148 work->nr_pages = nr_pages;
149 work->range_cyclic = range_cyclic;
150 work->reason = reason;
152 bdi_queue_work(bdi, work);
156 * bdi_start_writeback - start writeback
157 * @bdi: the backing device to write from
158 * @nr_pages: the number of pages to write
159 * @reason: reason why some writeback work was initiated
161 * Description:
162 * This does WB_SYNC_NONE opportunistic writeback. The IO is only
163 * started when this function returns, we make no guarantees on
164 * completion. Caller need not hold sb s_umount semaphore.
167 void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
168 enum wb_reason reason)
170 __bdi_start_writeback(bdi, nr_pages, true, reason);
174 * bdi_start_background_writeback - start background writeback
175 * @bdi: the backing device to write from
177 * Description:
178 * This makes sure WB_SYNC_NONE background writeback happens. When
179 * this function returns, it is only guaranteed that for given BDI
180 * some IO is happening if we are over background dirty threshold.
181 * Caller need not hold sb s_umount semaphore.
183 void bdi_start_background_writeback(struct backing_dev_info *bdi)
186 * We just wake up the flusher thread. It will perform background
187 * writeback as soon as there is no other work to do.
189 trace_writeback_wake_background(bdi);
190 spin_lock_bh(&bdi->wb_lock);
191 bdi_wakeup_flusher(bdi);
192 spin_unlock_bh(&bdi->wb_lock);
196 * Remove the inode from the writeback list it is on.
198 void inode_wb_list_del(struct inode *inode)
200 struct backing_dev_info *bdi = inode_to_bdi(inode);
202 spin_lock(&bdi->wb.list_lock);
203 list_del_init(&inode->i_wb_list);
204 spin_unlock(&bdi->wb.list_lock);
208 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
209 * furthest end of its superblock's dirty-inode list.
211 * Before stamping the inode's ->dirtied_when, we check to see whether it is
212 * already the most-recently-dirtied inode on the b_dirty list. If that is
213 * the case then the inode must have been redirtied while it was being written
214 * out and we don't reset its dirtied_when.
216 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
218 assert_spin_locked(&wb->list_lock);
219 if (!list_empty(&wb->b_dirty)) {
220 struct inode *tail;
222 tail = wb_inode(wb->b_dirty.next);
223 if (time_before(inode->dirtied_when, tail->dirtied_when))
224 inode->dirtied_when = jiffies;
226 list_move(&inode->i_wb_list, &wb->b_dirty);
230 * requeue inode for re-scanning after bdi->b_io list is exhausted.
232 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
234 assert_spin_locked(&wb->list_lock);
235 list_move(&inode->i_wb_list, &wb->b_more_io);
238 static void inode_sync_complete(struct inode *inode)
241 * Prevent speculative execution through
242 * spin_unlock(&wb->list_lock);
245 smp_mb();
246 wake_up_bit(&inode->i_state, __I_SYNC);
249 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
251 bool ret = time_after(inode->dirtied_when, t);
252 #ifndef CONFIG_64BIT
254 * For inodes being constantly redirtied, dirtied_when can get stuck.
255 * It _appears_ to be in the future, but is actually in distant past.
256 * This test is necessary to prevent such wrapped-around relative times
257 * from permanently stopping the whole bdi writeback.
259 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
260 #endif
261 return ret;
265 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
267 static int move_expired_inodes(struct list_head *delaying_queue,
268 struct list_head *dispatch_queue,
269 struct wb_writeback_work *work)
271 LIST_HEAD(tmp);
272 struct list_head *pos, *node;
273 struct super_block *sb = NULL;
274 struct inode *inode;
275 int do_sb_sort = 0;
276 int moved = 0;
278 while (!list_empty(delaying_queue)) {
279 inode = wb_inode(delaying_queue->prev);
280 if (work->older_than_this &&
281 inode_dirtied_after(inode, *work->older_than_this))
282 break;
283 if (sb && sb != inode->i_sb)
284 do_sb_sort = 1;
285 sb = inode->i_sb;
286 list_move(&inode->i_wb_list, &tmp);
287 moved++;
290 /* just one sb in list, splice to dispatch_queue and we're done */
291 if (!do_sb_sort) {
292 list_splice(&tmp, dispatch_queue);
293 goto out;
296 /* Move inodes from one superblock together */
297 while (!list_empty(&tmp)) {
298 sb = wb_inode(tmp.prev)->i_sb;
299 list_for_each_prev_safe(pos, node, &tmp) {
300 inode = wb_inode(pos);
301 if (inode->i_sb == sb)
302 list_move(&inode->i_wb_list, dispatch_queue);
305 out:
306 return moved;
310 * Queue all expired dirty inodes for io, eldest first.
311 * Before
312 * newly dirtied b_dirty b_io b_more_io
313 * =============> gf edc BA
314 * After
315 * newly dirtied b_dirty b_io b_more_io
316 * =============> g fBAedc
318 * +--> dequeue for IO
320 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
322 int moved;
323 assert_spin_locked(&wb->list_lock);
324 list_splice_init(&wb->b_more_io, &wb->b_io);
325 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
326 trace_writeback_queue_io(wb, work, moved);
329 static int write_inode(struct inode *inode, struct writeback_control *wbc)
331 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
332 return inode->i_sb->s_op->write_inode(inode, wbc);
333 return 0;
337 * Wait for writeback on an inode to complete.
339 static void inode_wait_for_writeback(struct inode *inode,
340 struct bdi_writeback *wb)
342 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
343 wait_queue_head_t *wqh;
345 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
346 while (inode->i_state & I_SYNC) {
347 spin_unlock(&inode->i_lock);
348 spin_unlock(&wb->list_lock);
349 __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
350 spin_lock(&wb->list_lock);
351 spin_lock(&inode->i_lock);
356 * Write out an inode's dirty pages. Called under wb->list_lock and
357 * inode->i_lock. Either the caller has an active reference on the inode or
358 * the inode has I_WILL_FREE set.
360 * If `wait' is set, wait on the writeout.
362 * The whole writeout design is quite complex and fragile. We want to avoid
363 * starvation of particular inodes when others are being redirtied, prevent
364 * livelocks, etc.
366 static int
367 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
368 struct writeback_control *wbc)
370 struct address_space *mapping = inode->i_mapping;
371 long nr_to_write = wbc->nr_to_write;
372 unsigned dirty;
373 int ret;
375 assert_spin_locked(&wb->list_lock);
376 assert_spin_locked(&inode->i_lock);
378 if (!atomic_read(&inode->i_count))
379 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
380 else
381 WARN_ON(inode->i_state & I_WILL_FREE);
383 if (inode->i_state & I_SYNC) {
385 * If this inode is locked for writeback and we are not doing
386 * writeback-for-data-integrity, move it to b_more_io so that
387 * writeback can proceed with the other inodes on s_io.
389 * We'll have another go at writing back this inode when we
390 * completed a full scan of b_io.
392 if (wbc->sync_mode != WB_SYNC_ALL) {
393 requeue_io(inode, wb);
394 trace_writeback_single_inode_requeue(inode, wbc,
395 nr_to_write);
396 return 0;
400 * It's a data-integrity sync. We must wait.
402 inode_wait_for_writeback(inode, wb);
405 BUG_ON(inode->i_state & I_SYNC);
407 /* Set I_SYNC, reset I_DIRTY_PAGES */
408 inode->i_state |= I_SYNC;
409 inode->i_state &= ~I_DIRTY_PAGES;
410 spin_unlock(&inode->i_lock);
411 spin_unlock(&wb->list_lock);
413 ret = do_writepages(mapping, wbc);
416 * Make sure to wait on the data before writing out the metadata.
417 * This is important for filesystems that modify metadata on data
418 * I/O completion.
420 if (wbc->sync_mode == WB_SYNC_ALL) {
421 int err = filemap_fdatawait(mapping);
422 if (ret == 0)
423 ret = err;
427 * Some filesystems may redirty the inode during the writeback
428 * due to delalloc, clear dirty metadata flags right before
429 * write_inode()
431 spin_lock(&inode->i_lock);
432 dirty = inode->i_state & I_DIRTY;
433 inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
434 spin_unlock(&inode->i_lock);
435 /* Don't write the inode if only I_DIRTY_PAGES was set */
436 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
437 int err = write_inode(inode, wbc);
438 if (ret == 0)
439 ret = err;
442 spin_lock(&wb->list_lock);
443 spin_lock(&inode->i_lock);
444 inode->i_state &= ~I_SYNC;
445 if (!(inode->i_state & I_FREEING)) {
447 * Sync livelock prevention. Each inode is tagged and synced in
448 * one shot. If still dirty, it will be redirty_tail()'ed below.
449 * Update the dirty time to prevent enqueue and sync it again.
451 if ((inode->i_state & I_DIRTY) &&
452 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
453 inode->dirtied_when = jiffies;
455 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
457 * We didn't write back all the pages. nfs_writepages()
458 * sometimes bales out without doing anything.
460 inode->i_state |= I_DIRTY_PAGES;
461 if (wbc->nr_to_write <= 0) {
463 * slice used up: queue for next turn
465 requeue_io(inode, wb);
466 } else {
468 * Writeback blocked by something other than
469 * congestion. Delay the inode for some time to
470 * avoid spinning on the CPU (100% iowait)
471 * retrying writeback of the dirty page/inode
472 * that cannot be performed immediately.
474 redirty_tail(inode, wb);
476 } else if (inode->i_state & I_DIRTY) {
478 * Filesystems can dirty the inode during writeback
479 * operations, such as delayed allocation during
480 * submission or metadata updates after data IO
481 * completion.
483 redirty_tail(inode, wb);
484 } else {
486 * The inode is clean. At this point we either have
487 * a reference to the inode or it's on it's way out.
488 * No need to add it back to the LRU.
490 list_del_init(&inode->i_wb_list);
493 inode_sync_complete(inode);
494 trace_writeback_single_inode(inode, wbc, nr_to_write);
495 return ret;
498 static long writeback_chunk_size(struct backing_dev_info *bdi,
499 struct wb_writeback_work *work)
501 long pages;
504 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
505 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
506 * here avoids calling into writeback_inodes_wb() more than once.
508 * The intended call sequence for WB_SYNC_ALL writeback is:
510 * wb_writeback()
511 * writeback_sb_inodes() <== called only once
512 * write_cache_pages() <== called once for each inode
513 * (quickly) tag currently dirty pages
514 * (maybe slowly) sync all tagged pages
516 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
517 pages = LONG_MAX;
518 else {
519 pages = min(bdi->avg_write_bandwidth / 2,
520 global_dirty_limit / DIRTY_SCOPE);
521 pages = min(pages, work->nr_pages);
522 pages = round_down(pages + MIN_WRITEBACK_PAGES,
523 MIN_WRITEBACK_PAGES);
526 return pages;
530 * Write a portion of b_io inodes which belong to @sb.
532 * If @only_this_sb is true, then find and write all such
533 * inodes. Otherwise write only ones which go sequentially
534 * in reverse order.
536 * Return the number of pages and/or inodes written.
538 static long writeback_sb_inodes(struct super_block *sb,
539 struct bdi_writeback *wb,
540 struct wb_writeback_work *work)
542 struct writeback_control wbc = {
543 .sync_mode = work->sync_mode,
544 .tagged_writepages = work->tagged_writepages,
545 .for_kupdate = work->for_kupdate,
546 .for_background = work->for_background,
547 .range_cyclic = work->range_cyclic,
548 .range_start = 0,
549 .range_end = LLONG_MAX,
551 unsigned long start_time = jiffies;
552 long write_chunk;
553 long wrote = 0; /* count both pages and inodes */
555 while (!list_empty(&wb->b_io)) {
556 struct inode *inode = wb_inode(wb->b_io.prev);
558 if (inode->i_sb != sb) {
559 if (work->sb) {
561 * We only want to write back data for this
562 * superblock, move all inodes not belonging
563 * to it back onto the dirty list.
565 redirty_tail(inode, wb);
566 continue;
570 * The inode belongs to a different superblock.
571 * Bounce back to the caller to unpin this and
572 * pin the next superblock.
574 break;
578 * Don't bother with new inodes or inodes beeing freed, first
579 * kind does not need peridic writeout yet, and for the latter
580 * kind writeout is handled by the freer.
582 spin_lock(&inode->i_lock);
583 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
584 spin_unlock(&inode->i_lock);
585 redirty_tail(inode, wb);
586 continue;
588 __iget(inode);
589 write_chunk = writeback_chunk_size(wb->bdi, work);
590 wbc.nr_to_write = write_chunk;
591 wbc.pages_skipped = 0;
593 writeback_single_inode(inode, wb, &wbc);
595 work->nr_pages -= write_chunk - wbc.nr_to_write;
596 wrote += write_chunk - wbc.nr_to_write;
597 if (!(inode->i_state & I_DIRTY))
598 wrote++;
599 if (wbc.pages_skipped) {
601 * writeback is not making progress due to locked
602 * buffers. Skip this inode for now.
604 redirty_tail(inode, wb);
606 spin_unlock(&inode->i_lock);
607 spin_unlock(&wb->list_lock);
608 iput(inode);
609 cond_resched();
610 spin_lock(&wb->list_lock);
612 * bail out to wb_writeback() often enough to check
613 * background threshold and other termination conditions.
615 if (wrote) {
616 if (time_is_before_jiffies(start_time + HZ / 10UL))
617 break;
618 if (work->nr_pages <= 0)
619 break;
622 return wrote;
625 static long __writeback_inodes_wb(struct bdi_writeback *wb,
626 struct wb_writeback_work *work)
628 unsigned long start_time = jiffies;
629 long wrote = 0;
631 while (!list_empty(&wb->b_io)) {
632 struct inode *inode = wb_inode(wb->b_io.prev);
633 struct super_block *sb = inode->i_sb;
635 if (!grab_super_passive(sb)) {
637 * grab_super_passive() may fail consistently due to
638 * s_umount being grabbed by someone else. Don't use
639 * requeue_io() to avoid busy retrying the inode/sb.
641 redirty_tail(inode, wb);
642 continue;
644 wrote += writeback_sb_inodes(sb, wb, work);
645 drop_super(sb);
647 /* refer to the same tests at the end of writeback_sb_inodes */
648 if (wrote) {
649 if (time_is_before_jiffies(start_time + HZ / 10UL))
650 break;
651 if (work->nr_pages <= 0)
652 break;
655 /* Leave any unwritten inodes on b_io */
656 return wrote;
659 long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
660 enum wb_reason reason)
662 struct wb_writeback_work work = {
663 .nr_pages = nr_pages,
664 .sync_mode = WB_SYNC_NONE,
665 .range_cyclic = 1,
666 .reason = reason,
669 spin_lock(&wb->list_lock);
670 if (list_empty(&wb->b_io))
671 queue_io(wb, &work);
672 __writeback_inodes_wb(wb, &work);
673 spin_unlock(&wb->list_lock);
675 return nr_pages - work.nr_pages;
678 static bool over_bground_thresh(struct backing_dev_info *bdi)
680 unsigned long background_thresh, dirty_thresh;
682 global_dirty_limits(&background_thresh, &dirty_thresh);
684 if (global_page_state(NR_FILE_DIRTY) +
685 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
686 return true;
688 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
689 bdi_dirty_limit(bdi, background_thresh))
690 return true;
692 return false;
696 * Called under wb->list_lock. If there are multiple wb per bdi,
697 * only the flusher working on the first wb should do it.
699 static void wb_update_bandwidth(struct bdi_writeback *wb,
700 unsigned long start_time)
702 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
706 * Explicit flushing or periodic writeback of "old" data.
708 * Define "old": the first time one of an inode's pages is dirtied, we mark the
709 * dirtying-time in the inode's address_space. So this periodic writeback code
710 * just walks the superblock inode list, writing back any inodes which are
711 * older than a specific point in time.
713 * Try to run once per dirty_writeback_interval. But if a writeback event
714 * takes longer than a dirty_writeback_interval interval, then leave a
715 * one-second gap.
717 * older_than_this takes precedence over nr_to_write. So we'll only write back
718 * all dirty pages if they are all attached to "old" mappings.
720 static long wb_writeback(struct bdi_writeback *wb,
721 struct wb_writeback_work *work)
723 unsigned long wb_start = jiffies;
724 long nr_pages = work->nr_pages;
725 unsigned long oldest_jif;
726 struct inode *inode;
727 long progress;
729 oldest_jif = jiffies;
730 work->older_than_this = &oldest_jif;
732 spin_lock(&wb->list_lock);
733 for (;;) {
735 * Stop writeback when nr_pages has been consumed
737 if (work->nr_pages <= 0)
738 break;
741 * Background writeout and kupdate-style writeback may
742 * run forever. Stop them if there is other work to do
743 * so that e.g. sync can proceed. They'll be restarted
744 * after the other works are all done.
746 if ((work->for_background || work->for_kupdate) &&
747 !list_empty(&wb->bdi->work_list))
748 break;
751 * For background writeout, stop when we are below the
752 * background dirty threshold
754 if (work->for_background && !over_bground_thresh(wb->bdi))
755 break;
757 if (work->for_kupdate) {
758 oldest_jif = jiffies -
759 msecs_to_jiffies(dirty_expire_interval * 10);
760 work->older_than_this = &oldest_jif;
763 trace_writeback_start(wb->bdi, work);
764 if (list_empty(&wb->b_io))
765 queue_io(wb, work);
766 if (work->sb)
767 progress = writeback_sb_inodes(work->sb, wb, work);
768 else
769 progress = __writeback_inodes_wb(wb, work);
770 trace_writeback_written(wb->bdi, work);
772 wb_update_bandwidth(wb, wb_start);
775 * Did we write something? Try for more
777 * Dirty inodes are moved to b_io for writeback in batches.
778 * The completion of the current batch does not necessarily
779 * mean the overall work is done. So we keep looping as long
780 * as made some progress on cleaning pages or inodes.
782 if (progress)
783 continue;
785 * No more inodes for IO, bail
787 if (list_empty(&wb->b_more_io))
788 break;
790 * Nothing written. Wait for some inode to
791 * become available for writeback. Otherwise
792 * we'll just busyloop.
794 if (!list_empty(&wb->b_more_io)) {
795 trace_writeback_wait(wb->bdi, work);
796 inode = wb_inode(wb->b_more_io.prev);
797 spin_lock(&inode->i_lock);
798 inode_wait_for_writeback(inode, wb);
799 spin_unlock(&inode->i_lock);
802 spin_unlock(&wb->list_lock);
804 return nr_pages - work->nr_pages;
808 * Return the next wb_writeback_work struct that hasn't been processed yet.
810 static struct wb_writeback_work *
811 get_next_work_item(struct backing_dev_info *bdi)
813 struct wb_writeback_work *work = NULL;
815 spin_lock_bh(&bdi->wb_lock);
816 if (!list_empty(&bdi->work_list)) {
817 work = list_entry(bdi->work_list.next,
818 struct wb_writeback_work, list);
819 list_del_init(&work->list);
821 spin_unlock_bh(&bdi->wb_lock);
822 return work;
826 * Add in the number of potentially dirty inodes, because each inode
827 * write can dirty pagecache in the underlying blockdev.
829 static unsigned long get_nr_dirty_pages(void)
831 return global_page_state(NR_FILE_DIRTY) +
832 global_page_state(NR_UNSTABLE_NFS) +
833 get_nr_dirty_inodes();
836 static long wb_check_background_flush(struct bdi_writeback *wb)
838 if (over_bground_thresh(wb->bdi)) {
840 struct wb_writeback_work work = {
841 .nr_pages = LONG_MAX,
842 .sync_mode = WB_SYNC_NONE,
843 .for_background = 1,
844 .range_cyclic = 1,
845 .reason = WB_REASON_BACKGROUND,
848 return wb_writeback(wb, &work);
851 return 0;
854 static long wb_check_old_data_flush(struct bdi_writeback *wb)
856 unsigned long expired;
857 long nr_pages;
860 * When set to zero, disable periodic writeback
862 if (!dirty_writeback_interval)
863 return 0;
865 expired = wb->last_old_flush +
866 msecs_to_jiffies(dirty_writeback_interval * 10);
867 if (time_before(jiffies, expired))
868 return 0;
870 wb->last_old_flush = jiffies;
871 nr_pages = get_nr_dirty_pages();
873 if (nr_pages) {
874 struct wb_writeback_work work = {
875 .nr_pages = nr_pages,
876 .sync_mode = WB_SYNC_NONE,
877 .for_kupdate = 1,
878 .range_cyclic = 1,
879 .reason = WB_REASON_PERIODIC,
882 return wb_writeback(wb, &work);
885 return 0;
889 * Retrieve work items and do the writeback they describe
891 long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
893 struct backing_dev_info *bdi = wb->bdi;
894 struct wb_writeback_work *work;
895 long wrote = 0;
897 set_bit(BDI_writeback_running, &wb->bdi->state);
898 while ((work = get_next_work_item(bdi)) != NULL) {
900 * Override sync mode, in case we must wait for completion
901 * because this thread is exiting now.
903 if (force_wait)
904 work->sync_mode = WB_SYNC_ALL;
906 trace_writeback_exec(bdi, work);
908 wrote += wb_writeback(wb, work);
911 * Notify the caller of completion if this is a synchronous
912 * work item, otherwise just free it.
914 if (work->done)
915 complete(work->done);
916 else
917 kfree(work);
921 * Check for periodic writeback, kupdated() style
923 wrote += wb_check_old_data_flush(wb);
924 wrote += wb_check_background_flush(wb);
925 clear_bit(BDI_writeback_running, &wb->bdi->state);
927 return wrote;
931 * Handle writeback of dirty data for the device backed by this bdi. Also
932 * wakes up periodically and does kupdated style flushing.
934 int bdi_writeback_thread(void *data)
936 struct bdi_writeback *wb = data;
937 struct backing_dev_info *bdi = wb->bdi;
938 long pages_written;
940 current->flags |= PF_SWAPWRITE;
941 set_freezable();
942 wb->last_active = jiffies;
945 * Our parent may run at a different priority, just set us to normal
947 set_user_nice(current, 0);
949 trace_writeback_thread_start(bdi);
951 while (!kthread_should_stop()) {
953 * Remove own delayed wake-up timer, since we are already awake
954 * and we'll take care of the preriodic write-back.
956 del_timer(&wb->wakeup_timer);
958 pages_written = wb_do_writeback(wb, 0);
960 trace_writeback_pages_written(pages_written);
962 if (pages_written)
963 wb->last_active = jiffies;
965 set_current_state(TASK_INTERRUPTIBLE);
966 if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
967 __set_current_state(TASK_RUNNING);
968 continue;
971 if (wb_has_dirty_io(wb) && dirty_writeback_interval)
972 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
973 else {
975 * We have nothing to do, so can go sleep without any
976 * timeout and save power. When a work is queued or
977 * something is made dirty - we will be woken up.
979 schedule();
982 try_to_freeze();
985 /* Flush any work that raced with us exiting */
986 if (!list_empty(&bdi->work_list))
987 wb_do_writeback(wb, 1);
989 trace_writeback_thread_stop(bdi);
990 return 0;
995 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
996 * the whole world.
998 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1000 struct backing_dev_info *bdi;
1002 if (!nr_pages) {
1003 nr_pages = global_page_state(NR_FILE_DIRTY) +
1004 global_page_state(NR_UNSTABLE_NFS);
1007 rcu_read_lock();
1008 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1009 if (!bdi_has_dirty_io(bdi))
1010 continue;
1011 __bdi_start_writeback(bdi, nr_pages, false, reason);
1013 rcu_read_unlock();
1016 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1018 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1019 struct dentry *dentry;
1020 const char *name = "?";
1022 dentry = d_find_alias(inode);
1023 if (dentry) {
1024 spin_lock(&dentry->d_lock);
1025 name = (const char *) dentry->d_name.name;
1027 printk(KERN_DEBUG
1028 "%s(%d): dirtied inode %lu (%s) on %s\n",
1029 current->comm, task_pid_nr(current), inode->i_ino,
1030 name, inode->i_sb->s_id);
1031 if (dentry) {
1032 spin_unlock(&dentry->d_lock);
1033 dput(dentry);
1039 * __mark_inode_dirty - internal function
1040 * @inode: inode to mark
1041 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1042 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1043 * mark_inode_dirty_sync.
1045 * Put the inode on the super block's dirty list.
1047 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1048 * dirty list only if it is hashed or if it refers to a blockdev.
1049 * If it was not hashed, it will never be added to the dirty list
1050 * even if it is later hashed, as it will have been marked dirty already.
1052 * In short, make sure you hash any inodes _before_ you start marking
1053 * them dirty.
1055 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1056 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1057 * the kernel-internal blockdev inode represents the dirtying time of the
1058 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1059 * page->mapping->host, so the page-dirtying time is recorded in the internal
1060 * blockdev inode.
1062 void __mark_inode_dirty(struct inode *inode, int flags)
1064 struct super_block *sb = inode->i_sb;
1065 struct backing_dev_info *bdi = NULL;
1068 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1069 * dirty the inode itself
1071 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1072 if (sb->s_op->dirty_inode)
1073 sb->s_op->dirty_inode(inode, flags);
1077 * make sure that changes are seen by all cpus before we test i_state
1078 * -- mikulas
1080 smp_mb();
1082 /* avoid the locking if we can */
1083 if ((inode->i_state & flags) == flags)
1084 return;
1086 if (unlikely(block_dump))
1087 block_dump___mark_inode_dirty(inode);
1089 spin_lock(&inode->i_lock);
1090 if ((inode->i_state & flags) != flags) {
1091 const int was_dirty = inode->i_state & I_DIRTY;
1093 inode->i_state |= flags;
1096 * If the inode is being synced, just update its dirty state.
1097 * The unlocker will place the inode on the appropriate
1098 * superblock list, based upon its state.
1100 if (inode->i_state & I_SYNC)
1101 goto out_unlock_inode;
1104 * Only add valid (hashed) inodes to the superblock's
1105 * dirty list. Add blockdev inodes as well.
1107 if (!S_ISBLK(inode->i_mode)) {
1108 if (inode_unhashed(inode))
1109 goto out_unlock_inode;
1111 if (inode->i_state & I_FREEING)
1112 goto out_unlock_inode;
1115 * If the inode was already on b_dirty/b_io/b_more_io, don't
1116 * reposition it (that would break b_dirty time-ordering).
1118 if (!was_dirty) {
1119 bool wakeup_bdi = false;
1120 bdi = inode_to_bdi(inode);
1122 if (bdi_cap_writeback_dirty(bdi)) {
1123 WARN(!test_bit(BDI_registered, &bdi->state),
1124 "bdi-%s not registered\n", bdi->name);
1127 * If this is the first dirty inode for this
1128 * bdi, we have to wake-up the corresponding
1129 * bdi thread to make sure background
1130 * write-back happens later.
1132 if (!wb_has_dirty_io(&bdi->wb))
1133 wakeup_bdi = true;
1136 spin_unlock(&inode->i_lock);
1137 spin_lock(&bdi->wb.list_lock);
1138 inode->dirtied_when = jiffies;
1139 list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1140 spin_unlock(&bdi->wb.list_lock);
1142 if (wakeup_bdi)
1143 bdi_wakeup_thread_delayed(bdi);
1144 return;
1147 out_unlock_inode:
1148 spin_unlock(&inode->i_lock);
1151 EXPORT_SYMBOL(__mark_inode_dirty);
1154 * Write out a superblock's list of dirty inodes. A wait will be performed
1155 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1157 * If older_than_this is non-NULL, then only write out inodes which
1158 * had their first dirtying at a time earlier than *older_than_this.
1160 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1161 * This function assumes that the blockdev superblock's inodes are backed by
1162 * a variety of queues, so all inodes are searched. For other superblocks,
1163 * assume that all inodes are backed by the same queue.
1165 * The inodes to be written are parked on bdi->b_io. They are moved back onto
1166 * bdi->b_dirty as they are selected for writing. This way, none can be missed
1167 * on the writer throttling path, and we get decent balancing between many
1168 * throttled threads: we don't want them all piling up on inode_sync_wait.
1170 static void wait_sb_inodes(struct super_block *sb)
1172 struct inode *inode, *old_inode = NULL;
1175 * We need to be protected against the filesystem going from
1176 * r/o to r/w or vice versa.
1178 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1180 spin_lock(&inode_sb_list_lock);
1183 * Data integrity sync. Must wait for all pages under writeback,
1184 * because there may have been pages dirtied before our sync
1185 * call, but which had writeout started before we write it out.
1186 * In which case, the inode may not be on the dirty list, but
1187 * we still have to wait for that writeout.
1189 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1190 struct address_space *mapping = inode->i_mapping;
1192 spin_lock(&inode->i_lock);
1193 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1194 (mapping->nrpages == 0)) {
1195 spin_unlock(&inode->i_lock);
1196 continue;
1198 __iget(inode);
1199 spin_unlock(&inode->i_lock);
1200 spin_unlock(&inode_sb_list_lock);
1203 * We hold a reference to 'inode' so it couldn't have been
1204 * removed from s_inodes list while we dropped the
1205 * inode_sb_list_lock. We cannot iput the inode now as we can
1206 * be holding the last reference and we cannot iput it under
1207 * inode_sb_list_lock. So we keep the reference and iput it
1208 * later.
1210 iput(old_inode);
1211 old_inode = inode;
1213 filemap_fdatawait(mapping);
1215 cond_resched();
1217 spin_lock(&inode_sb_list_lock);
1219 spin_unlock(&inode_sb_list_lock);
1220 iput(old_inode);
1224 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1225 * @sb: the superblock
1226 * @nr: the number of pages to write
1227 * @reason: reason why some writeback work initiated
1229 * Start writeback on some inodes on this super_block. No guarantees are made
1230 * on how many (if any) will be written, and this function does not wait
1231 * for IO completion of submitted IO.
1233 void writeback_inodes_sb_nr(struct super_block *sb,
1234 unsigned long nr,
1235 enum wb_reason reason)
1237 DECLARE_COMPLETION_ONSTACK(done);
1238 struct wb_writeback_work work = {
1239 .sb = sb,
1240 .sync_mode = WB_SYNC_NONE,
1241 .tagged_writepages = 1,
1242 .done = &done,
1243 .nr_pages = nr,
1244 .reason = reason,
1247 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1248 bdi_queue_work(sb->s_bdi, &work);
1249 wait_for_completion(&done);
1251 EXPORT_SYMBOL(writeback_inodes_sb_nr);
1254 * writeback_inodes_sb - writeback dirty inodes from given super_block
1255 * @sb: the superblock
1256 * @reason: reason why some writeback work was initiated
1258 * Start writeback on some inodes on this super_block. No guarantees are made
1259 * on how many (if any) will be written, and this function does not wait
1260 * for IO completion of submitted IO.
1262 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1264 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1266 EXPORT_SYMBOL(writeback_inodes_sb);
1269 * writeback_inodes_sb_if_idle - start writeback if none underway
1270 * @sb: the superblock
1271 * @reason: reason why some writeback work was initiated
1273 * Invoke writeback_inodes_sb if no writeback is currently underway.
1274 * Returns 1 if writeback was started, 0 if not.
1276 int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1278 if (!writeback_in_progress(sb->s_bdi)) {
1279 down_read(&sb->s_umount);
1280 writeback_inodes_sb(sb, reason);
1281 up_read(&sb->s_umount);
1282 return 1;
1283 } else
1284 return 0;
1286 EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1289 * writeback_inodes_sb_if_idle - start writeback if none underway
1290 * @sb: the superblock
1291 * @nr: the number of pages to write
1292 * @reason: reason why some writeback work was initiated
1294 * Invoke writeback_inodes_sb if no writeback is currently underway.
1295 * Returns 1 if writeback was started, 0 if not.
1297 int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1298 unsigned long nr,
1299 enum wb_reason reason)
1301 if (!writeback_in_progress(sb->s_bdi)) {
1302 down_read(&sb->s_umount);
1303 writeback_inodes_sb_nr(sb, nr, reason);
1304 up_read(&sb->s_umount);
1305 return 1;
1306 } else
1307 return 0;
1309 EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1312 * sync_inodes_sb - sync sb inode pages
1313 * @sb: the superblock
1315 * This function writes and waits on any dirty inode belonging to this
1316 * super_block.
1318 void sync_inodes_sb(struct super_block *sb)
1320 DECLARE_COMPLETION_ONSTACK(done);
1321 struct wb_writeback_work work = {
1322 .sb = sb,
1323 .sync_mode = WB_SYNC_ALL,
1324 .nr_pages = LONG_MAX,
1325 .range_cyclic = 0,
1326 .done = &done,
1327 .reason = WB_REASON_SYNC,
1330 WARN_ON(!rwsem_is_locked(&sb->s_umount));
1332 bdi_queue_work(sb->s_bdi, &work);
1333 wait_for_completion(&done);
1335 wait_sb_inodes(sb);
1337 EXPORT_SYMBOL(sync_inodes_sb);
1340 * write_inode_now - write an inode to disk
1341 * @inode: inode to write to disk
1342 * @sync: whether the write should be synchronous or not
1344 * This function commits an inode to disk immediately if it is dirty. This is
1345 * primarily needed by knfsd.
1347 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1349 int write_inode_now(struct inode *inode, int sync)
1351 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1352 int ret;
1353 struct writeback_control wbc = {
1354 .nr_to_write = LONG_MAX,
1355 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1356 .range_start = 0,
1357 .range_end = LLONG_MAX,
1360 if (!mapping_cap_writeback_dirty(inode->i_mapping))
1361 wbc.nr_to_write = 0;
1363 might_sleep();
1364 spin_lock(&wb->list_lock);
1365 spin_lock(&inode->i_lock);
1366 ret = writeback_single_inode(inode, wb, &wbc);
1367 spin_unlock(&inode->i_lock);
1368 spin_unlock(&wb->list_lock);
1369 if (sync)
1370 inode_sync_wait(inode);
1371 return ret;
1373 EXPORT_SYMBOL(write_inode_now);
1376 * sync_inode - write an inode and its pages to disk.
1377 * @inode: the inode to sync
1378 * @wbc: controls the writeback mode
1380 * sync_inode() will write an inode and its pages to disk. It will also
1381 * correctly update the inode on its superblock's dirty inode lists and will
1382 * update inode->i_state.
1384 * The caller must have a ref on the inode.
1386 int sync_inode(struct inode *inode, struct writeback_control *wbc)
1388 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
1389 int ret;
1391 spin_lock(&wb->list_lock);
1392 spin_lock(&inode->i_lock);
1393 ret = writeback_single_inode(inode, wb, wbc);
1394 spin_unlock(&inode->i_lock);
1395 spin_unlock(&wb->list_lock);
1396 return ret;
1398 EXPORT_SYMBOL(sync_inode);
1401 * sync_inode_metadata - write an inode to disk
1402 * @inode: the inode to sync
1403 * @wait: wait for I/O to complete.
1405 * Write an inode to disk and adjust its dirty state after completion.
1407 * Note: only writes the actual inode, no associated data or other metadata.
1409 int sync_inode_metadata(struct inode *inode, int wait)
1411 struct writeback_control wbc = {
1412 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1413 .nr_to_write = 0, /* metadata-only */
1416 return sync_inode(inode, &wbc);
1418 EXPORT_SYMBOL(sync_inode_metadata);