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/sched.h>
22 #include <linux/writeback.h>
23 #include <linux/blkdev.h>
24 #include <linux/backing-dev.h>
25 #include <linux/buffer_head.h>
30 * writeback_acquire - attempt to get exclusive writeback access to a device
31 * @bdi: the device's backing_dev_info structure
33 * It is a waste of resources to have more than one pdflush thread blocked on
34 * a single request queue. Exclusion at the request_queue level is obtained
35 * via a flag in the request_queue's backing_dev_info.state.
37 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
38 * unless they implement their own. Which is somewhat inefficient, as this
39 * may prevent concurrent writeback against multiple devices.
41 static int writeback_acquire(struct backing_dev_info *bdi)
43 return !test_and_set_bit(BDI_pdflush, &bdi->state);
47 * writeback_in_progress - determine whether there is writeback in progress
48 * @bdi: the device's backing_dev_info structure.
50 * Determine whether there is writeback in progress against a backing device.
52 int writeback_in_progress(struct backing_dev_info *bdi)
54 return test_bit(BDI_pdflush, &bdi->state);
58 * writeback_release - relinquish exclusive writeback access against a device.
59 * @bdi: the device's backing_dev_info structure
61 static void writeback_release(struct backing_dev_info *bdi)
63 BUG_ON(!writeback_in_progress(bdi));
64 clear_bit(BDI_pdflush, &bdi->state);
68 * __mark_inode_dirty - internal function
69 * @inode: inode to mark
70 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
71 * Mark an inode as dirty. Callers should use mark_inode_dirty or
72 * mark_inode_dirty_sync.
74 * Put the inode on the super block's dirty list.
76 * CAREFUL! We mark it dirty unconditionally, but move it onto the
77 * dirty list only if it is hashed or if it refers to a blockdev.
78 * If it was not hashed, it will never be added to the dirty list
79 * even if it is later hashed, as it will have been marked dirty already.
81 * In short, make sure you hash any inodes _before_ you start marking
84 * This function *must* be atomic for the I_DIRTY_PAGES case -
85 * set_page_dirty() is called under spinlock in several places.
87 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
88 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
89 * the kernel-internal blockdev inode represents the dirtying time of the
90 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
91 * page->mapping->host, so the page-dirtying time is recorded in the internal
94 void __mark_inode_dirty(struct inode *inode, int flags)
96 struct super_block *sb = inode->i_sb;
99 * Don't do this for I_DIRTY_PAGES - that doesn't actually
100 * dirty the inode itself
102 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
103 if (sb->s_op->dirty_inode)
104 sb->s_op->dirty_inode(inode);
108 * make sure that changes are seen by all cpus before we test i_state
113 /* avoid the locking if we can */
114 if ((inode->i_state & flags) == flags)
117 if (unlikely(block_dump)) {
118 struct dentry *dentry = NULL;
119 const char *name = "?";
121 if (!list_empty(&inode->i_dentry)) {
122 dentry = list_entry(inode->i_dentry.next,
123 struct dentry, d_alias);
124 if (dentry && dentry->d_name.name)
125 name = (const char *) dentry->d_name.name;
128 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev"))
130 "%s(%d): dirtied inode %lu (%s) on %s\n",
131 current->comm, task_pid_nr(current), inode->i_ino,
132 name, inode->i_sb->s_id);
135 spin_lock(&inode_lock);
136 if ((inode->i_state & flags) != flags) {
137 const int was_dirty = inode->i_state & I_DIRTY;
139 inode->i_state |= flags;
142 * If the inode is being synced, just update its dirty state.
143 * The unlocker will place the inode on the appropriate
144 * superblock list, based upon its state.
146 if (inode->i_state & I_SYNC)
150 * Only add valid (hashed) inodes to the superblock's
151 * dirty list. Add blockdev inodes as well.
153 if (!S_ISBLK(inode->i_mode)) {
154 if (hlist_unhashed(&inode->i_hash))
157 if (inode->i_state & (I_FREEING|I_CLEAR))
161 * If the inode was already on s_dirty/s_io/s_more_io, don't
162 * reposition it (that would break s_dirty time-ordering).
165 inode->dirtied_when = jiffies;
166 list_move(&inode->i_list, &sb->s_dirty);
170 spin_unlock(&inode_lock);
173 EXPORT_SYMBOL(__mark_inode_dirty);
175 static int write_inode(struct inode *inode, int sync)
177 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
178 return inode->i_sb->s_op->write_inode(inode, sync);
183 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
184 * furthest end of its superblock's dirty-inode list.
186 * Before stamping the inode's ->dirtied_when, we check to see whether it is
187 * already the most-recently-dirtied inode on the s_dirty list. If that is
188 * the case then the inode must have been redirtied while it was being written
189 * out and we don't reset its dirtied_when.
191 static void redirty_tail(struct inode *inode)
193 struct super_block *sb = inode->i_sb;
195 if (!list_empty(&sb->s_dirty)) {
196 struct inode *tail_inode;
198 tail_inode = list_entry(sb->s_dirty.next, struct inode, i_list);
199 if (time_before(inode->dirtied_when,
200 tail_inode->dirtied_when))
201 inode->dirtied_when = jiffies;
203 list_move(&inode->i_list, &sb->s_dirty);
207 * requeue inode for re-scanning after sb->s_io list is exhausted.
209 static void requeue_io(struct inode *inode)
211 list_move(&inode->i_list, &inode->i_sb->s_more_io);
214 static void inode_sync_complete(struct inode *inode)
217 * Prevent speculative execution through spin_unlock(&inode_lock);
220 wake_up_bit(&inode->i_state, __I_SYNC);
223 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
225 bool ret = time_after(inode->dirtied_when, t);
228 * For inodes being constantly redirtied, dirtied_when can get stuck.
229 * It _appears_ to be in the future, but is actually in distant past.
230 * This test is necessary to prevent such wrapped-around relative times
231 * from permanently stopping the whole pdflush writeback.
233 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
239 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
241 static void move_expired_inodes(struct list_head *delaying_queue,
242 struct list_head *dispatch_queue,
243 unsigned long *older_than_this)
245 while (!list_empty(delaying_queue)) {
246 struct inode *inode = list_entry(delaying_queue->prev,
247 struct inode, i_list);
248 if (older_than_this &&
249 inode_dirtied_after(inode, *older_than_this))
251 list_move(&inode->i_list, dispatch_queue);
256 * Queue all expired dirty inodes for io, eldest first.
258 static void queue_io(struct super_block *sb,
259 unsigned long *older_than_this)
261 list_splice_init(&sb->s_more_io, sb->s_io.prev);
262 move_expired_inodes(&sb->s_dirty, &sb->s_io, older_than_this);
265 int sb_has_dirty_inodes(struct super_block *sb)
267 return !list_empty(&sb->s_dirty) ||
268 !list_empty(&sb->s_io) ||
269 !list_empty(&sb->s_more_io);
271 EXPORT_SYMBOL(sb_has_dirty_inodes);
274 * Write a single inode's dirty pages and inode data out to disk.
275 * If `wait' is set, wait on the writeout.
277 * The whole writeout design is quite complex and fragile. We want to avoid
278 * starvation of particular inodes when others are being redirtied, prevent
281 * Called under inode_lock.
284 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
287 struct address_space *mapping = inode->i_mapping;
288 int wait = wbc->sync_mode == WB_SYNC_ALL;
291 BUG_ON(inode->i_state & I_SYNC);
293 /* Set I_SYNC, reset I_DIRTY */
294 dirty = inode->i_state & I_DIRTY;
295 inode->i_state |= I_SYNC;
296 inode->i_state &= ~I_DIRTY;
298 spin_unlock(&inode_lock);
300 ret = do_writepages(mapping, wbc);
302 /* Don't write the inode if only I_DIRTY_PAGES was set */
303 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
304 int err = write_inode(inode, wait);
310 int err = filemap_fdatawait(mapping);
315 spin_lock(&inode_lock);
316 inode->i_state &= ~I_SYNC;
317 if (!(inode->i_state & I_FREEING)) {
318 if (!(inode->i_state & I_DIRTY) &&
319 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
321 * We didn't write back all the pages. nfs_writepages()
322 * sometimes bales out without doing anything. Redirty
323 * the inode; Move it from s_io onto s_more_io/s_dirty.
326 * akpm: if the caller was the kupdate function we put
327 * this inode at the head of s_dirty so it gets first
328 * consideration. Otherwise, move it to the tail, for
329 * the reasons described there. I'm not really sure
330 * how much sense this makes. Presumably I had a good
331 * reasons for doing it this way, and I'd rather not
332 * muck with it at present.
334 if (wbc->for_kupdate) {
336 * For the kupdate function we move the inode
337 * to s_more_io so it will get more writeout as
338 * soon as the queue becomes uncongested.
340 inode->i_state |= I_DIRTY_PAGES;
341 if (wbc->nr_to_write <= 0) {
343 * slice used up: queue for next turn
348 * somehow blocked: retry later
354 * Otherwise fully redirty the inode so that
355 * other inodes on this superblock will get some
356 * writeout. Otherwise heavy writing to one
357 * file would indefinitely suspend writeout of
358 * all the other files.
360 inode->i_state |= I_DIRTY_PAGES;
363 } else if (inode->i_state & I_DIRTY) {
365 * Someone redirtied the inode while were writing back
369 } else if (atomic_read(&inode->i_count)) {
371 * The inode is clean, inuse
373 list_move(&inode->i_list, &inode_in_use);
376 * The inode is clean, unused
378 list_move(&inode->i_list, &inode_unused);
381 inode_sync_complete(inode);
386 * Write out an inode's dirty pages. Called under inode_lock. Either the
387 * caller has ref on the inode (either via __iget or via syscall against an fd)
388 * or the inode has I_WILL_FREE set (via generic_forget_inode)
391 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
393 wait_queue_head_t *wqh;
395 if (!atomic_read(&inode->i_count))
396 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
398 WARN_ON(inode->i_state & I_WILL_FREE);
400 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_SYNC)) {
402 * We're skipping this inode because it's locked, and we're not
403 * doing writeback-for-data-integrity. Move it to s_more_io so
404 * that writeback can proceed with the other inodes on s_io.
405 * We'll have another go at writing back this inode when we
406 * completed a full scan of s_io.
413 * It's a data-integrity sync. We must wait.
415 if (inode->i_state & I_SYNC) {
416 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
418 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
420 spin_unlock(&inode_lock);
421 __wait_on_bit(wqh, &wq, inode_wait,
422 TASK_UNINTERRUPTIBLE);
423 spin_lock(&inode_lock);
424 } while (inode->i_state & I_SYNC);
426 return __sync_single_inode(inode, wbc);
430 * Write out a superblock's list of dirty inodes. A wait will be performed
431 * upon no inodes, all inodes or the final one, depending upon sync_mode.
433 * If older_than_this is non-NULL, then only write out inodes which
434 * had their first dirtying at a time earlier than *older_than_this.
436 * If we're a pdflush thread, then implement pdflush collision avoidance
437 * against the entire list.
439 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
440 * This function assumes that the blockdev superblock's inodes are backed by
441 * a variety of queues, so all inodes are searched. For other superblocks,
442 * assume that all inodes are backed by the same queue.
444 * FIXME: this linear search could get expensive with many fileystems. But
445 * how to fix? We need to go from an address_space to all inodes which share
446 * a queue with that address_space. (Easy: have a global "dirty superblocks"
449 * The inodes to be written are parked on sb->s_io. They are moved back onto
450 * sb->s_dirty as they are selected for writing. This way, none can be missed
451 * on the writer throttling path, and we get decent balancing between many
452 * throttled threads: we don't want them all piling up on inode_sync_wait.
454 void generic_sync_sb_inodes(struct super_block *sb,
455 struct writeback_control *wbc)
457 const unsigned long start = jiffies; /* livelock avoidance */
458 int sync = wbc->sync_mode == WB_SYNC_ALL;
460 spin_lock(&inode_lock);
461 if (!wbc->for_kupdate || list_empty(&sb->s_io))
462 queue_io(sb, wbc->older_than_this);
464 while (!list_empty(&sb->s_io)) {
465 struct inode *inode = list_entry(sb->s_io.prev,
466 struct inode, i_list);
467 struct address_space *mapping = inode->i_mapping;
468 struct backing_dev_info *bdi = mapping->backing_dev_info;
471 if (!bdi_cap_writeback_dirty(bdi)) {
473 if (sb_is_blkdev_sb(sb)) {
475 * Dirty memory-backed blockdev: the ramdisk
476 * driver does this. Skip just this inode
481 * Dirty memory-backed inode against a filesystem other
482 * than the kernel-internal bdev filesystem. Skip the
488 if (inode->i_state & I_NEW) {
493 if (wbc->nonblocking && bdi_write_congested(bdi)) {
494 wbc->encountered_congestion = 1;
495 if (!sb_is_blkdev_sb(sb))
496 break; /* Skip a congested fs */
498 continue; /* Skip a congested blockdev */
501 if (wbc->bdi && bdi != wbc->bdi) {
502 if (!sb_is_blkdev_sb(sb))
503 break; /* fs has the wrong queue */
505 continue; /* blockdev has wrong queue */
509 * Was this inode dirtied after sync_sb_inodes was called?
510 * This keeps sync from extra jobs and livelock.
512 if (inode_dirtied_after(inode, start))
515 /* Is another pdflush already flushing this queue? */
516 if (current_is_pdflush() && !writeback_acquire(bdi))
519 BUG_ON(inode->i_state & I_FREEING);
521 pages_skipped = wbc->pages_skipped;
522 __writeback_single_inode(inode, wbc);
523 if (current_is_pdflush())
524 writeback_release(bdi);
525 if (wbc->pages_skipped != pages_skipped) {
527 * writeback is not making progress due to locked
528 * buffers. Skip this inode for now.
532 spin_unlock(&inode_lock);
535 spin_lock(&inode_lock);
536 if (wbc->nr_to_write <= 0) {
540 if (!list_empty(&sb->s_more_io))
545 struct inode *inode, *old_inode = NULL;
548 * Data integrity sync. Must wait for all pages under writeback,
549 * because there may have been pages dirtied before our sync
550 * call, but which had writeout started before we write it out.
551 * In which case, the inode may not be on the dirty list, but
552 * we still have to wait for that writeout.
554 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
555 struct address_space *mapping;
558 (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
560 mapping = inode->i_mapping;
561 if (mapping->nrpages == 0)
564 spin_unlock(&inode_lock);
566 * We hold a reference to 'inode' so it couldn't have
567 * been removed from s_inodes list while we dropped the
568 * inode_lock. We cannot iput the inode now as we can
569 * be holding the last reference and we cannot iput it
570 * under inode_lock. So we keep the reference and iput
576 filemap_fdatawait(mapping);
580 spin_lock(&inode_lock);
582 spin_unlock(&inode_lock);
585 spin_unlock(&inode_lock);
587 return; /* Leave any unwritten inodes on s_io */
589 EXPORT_SYMBOL_GPL(generic_sync_sb_inodes);
591 static void sync_sb_inodes(struct super_block *sb,
592 struct writeback_control *wbc)
594 generic_sync_sb_inodes(sb, wbc);
598 * Start writeback of dirty pagecache data against all unlocked inodes.
601 * We don't need to grab a reference to superblock here. If it has non-empty
602 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
603 * past sync_inodes_sb() until the ->s_dirty/s_io/s_more_io lists are all
604 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
605 * inode from superblock lists we are OK.
607 * If `older_than_this' is non-zero then only flush inodes which have a
608 * flushtime older than *older_than_this.
610 * If `bdi' is non-zero then we will scan the first inode against each
611 * superblock until we find the matching ones. One group will be the dirty
612 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
613 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
614 * super-efficient but we're about to do a ton of I/O...
617 writeback_inodes(struct writeback_control *wbc)
619 struct super_block *sb;
624 list_for_each_entry_reverse(sb, &super_blocks, s_list) {
625 if (sb_has_dirty_inodes(sb)) {
626 /* we're making our own get_super here */
628 spin_unlock(&sb_lock);
630 * If we can't get the readlock, there's no sense in
631 * waiting around, most of the time the FS is going to
632 * be unmounted by the time it is released.
634 if (down_read_trylock(&sb->s_umount)) {
636 sync_sb_inodes(sb, wbc);
637 up_read(&sb->s_umount);
640 if (__put_super_and_need_restart(sb))
643 if (wbc->nr_to_write <= 0)
646 spin_unlock(&sb_lock);
650 * writeback and wait upon the filesystem's dirty inodes. The caller will
651 * do this in two passes - one to write, and one to wait.
653 * A finite limit is set on the number of pages which will be written.
654 * To prevent infinite livelock of sys_sync().
656 * We add in the number of potentially dirty inodes, because each inode write
657 * can dirty pagecache in the underlying blockdev.
659 void sync_inodes_sb(struct super_block *sb, int wait)
661 struct writeback_control wbc = {
662 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
664 .range_end = LLONG_MAX,
668 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
669 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
671 wbc.nr_to_write = nr_dirty + nr_unstable +
672 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
674 wbc.nr_to_write = LONG_MAX; /* doesn't actually matter */
676 sync_sb_inodes(sb, &wbc);
680 * write_inode_now - write an inode to disk
681 * @inode: inode to write to disk
682 * @sync: whether the write should be synchronous or not
684 * This function commits an inode to disk immediately if it is dirty. This is
685 * primarily needed by knfsd.
687 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
689 int write_inode_now(struct inode *inode, int sync)
692 struct writeback_control wbc = {
693 .nr_to_write = LONG_MAX,
694 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
696 .range_end = LLONG_MAX,
699 if (!mapping_cap_writeback_dirty(inode->i_mapping))
703 spin_lock(&inode_lock);
704 ret = __writeback_single_inode(inode, &wbc);
705 spin_unlock(&inode_lock);
707 inode_sync_wait(inode);
710 EXPORT_SYMBOL(write_inode_now);
713 * sync_inode - write an inode and its pages to disk.
714 * @inode: the inode to sync
715 * @wbc: controls the writeback mode
717 * sync_inode() will write an inode and its pages to disk. It will also
718 * correctly update the inode on its superblock's dirty inode lists and will
719 * update inode->i_state.
721 * The caller must have a ref on the inode.
723 int sync_inode(struct inode *inode, struct writeback_control *wbc)
727 spin_lock(&inode_lock);
728 ret = __writeback_single_inode(inode, wbc);
729 spin_unlock(&inode_lock);
732 EXPORT_SYMBOL(sync_inode);
735 * generic_osync_inode - flush all dirty data for a given inode to disk
736 * @inode: inode to write
737 * @mapping: the address_space that should be flushed
738 * @what: what to write and wait upon
740 * This can be called by file_write functions for files which have the
741 * O_SYNC flag set, to flush dirty writes to disk.
743 * @what is a bitmask, specifying which part of the inode's data should be
744 * written and waited upon.
746 * OSYNC_DATA: i_mapping's dirty data
747 * OSYNC_METADATA: the buffers at i_mapping->private_list
748 * OSYNC_INODE: the inode itself
751 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
754 int need_write_inode_now = 0;
757 if (what & OSYNC_DATA)
758 err = filemap_fdatawrite(mapping);
759 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
760 err2 = sync_mapping_buffers(mapping);
764 if (what & OSYNC_DATA) {
765 err2 = filemap_fdatawait(mapping);
770 spin_lock(&inode_lock);
771 if ((inode->i_state & I_DIRTY) &&
772 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
773 need_write_inode_now = 1;
774 spin_unlock(&inode_lock);
776 if (need_write_inode_now) {
777 err2 = write_inode_now(inode, 1);
782 inode_sync_wait(inode);
786 EXPORT_SYMBOL(generic_osync_inode);