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_after_eq(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);
224 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
226 static void move_expired_inodes(struct list_head *delaying_queue,
227 struct list_head *dispatch_queue,
228 unsigned long *older_than_this)
230 while (!list_empty(delaying_queue)) {
231 struct inode *inode = list_entry(delaying_queue->prev,
232 struct inode, i_list);
233 if (older_than_this &&
234 time_after(inode->dirtied_when, *older_than_this))
236 list_move(&inode->i_list, dispatch_queue);
241 * Queue all expired dirty inodes for io, eldest first.
243 static void queue_io(struct super_block *sb,
244 unsigned long *older_than_this)
246 list_splice_init(&sb->s_more_io, sb->s_io.prev);
247 move_expired_inodes(&sb->s_dirty, &sb->s_io, older_than_this);
250 int sb_has_dirty_inodes(struct super_block *sb)
252 return !list_empty(&sb->s_dirty) ||
253 !list_empty(&sb->s_io) ||
254 !list_empty(&sb->s_more_io);
256 EXPORT_SYMBOL(sb_has_dirty_inodes);
259 * Write a single inode's dirty pages and inode data out to disk.
260 * If `wait' is set, wait on the writeout.
262 * The whole writeout design is quite complex and fragile. We want to avoid
263 * starvation of particular inodes when others are being redirtied, prevent
266 * Called under inode_lock.
269 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
272 struct address_space *mapping = inode->i_mapping;
273 int wait = wbc->sync_mode == WB_SYNC_ALL;
276 BUG_ON(inode->i_state & I_SYNC);
277 WARN_ON(inode->i_state & I_NEW);
279 /* Set I_SYNC, reset I_DIRTY */
280 dirty = inode->i_state & I_DIRTY;
281 inode->i_state |= I_SYNC;
282 inode->i_state &= ~I_DIRTY;
284 spin_unlock(&inode_lock);
286 ret = do_writepages(mapping, wbc);
288 /* Don't write the inode if only I_DIRTY_PAGES was set */
289 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
290 int err = write_inode(inode, wait);
296 int err = filemap_fdatawait(mapping);
301 spin_lock(&inode_lock);
302 WARN_ON(inode->i_state & I_NEW);
303 inode->i_state &= ~I_SYNC;
304 if (!(inode->i_state & I_FREEING)) {
305 if (!(inode->i_state & I_DIRTY) &&
306 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
308 * We didn't write back all the pages. nfs_writepages()
309 * sometimes bales out without doing anything. Redirty
310 * the inode; Move it from s_io onto s_more_io/s_dirty.
313 * akpm: if the caller was the kupdate function we put
314 * this inode at the head of s_dirty so it gets first
315 * consideration. Otherwise, move it to the tail, for
316 * the reasons described there. I'm not really sure
317 * how much sense this makes. Presumably I had a good
318 * reasons for doing it this way, and I'd rather not
319 * muck with it at present.
321 if (wbc->for_kupdate) {
323 * For the kupdate function we move the inode
324 * to s_more_io so it will get more writeout as
325 * soon as the queue becomes uncongested.
327 inode->i_state |= I_DIRTY_PAGES;
328 if (wbc->nr_to_write <= 0) {
330 * slice used up: queue for next turn
335 * somehow blocked: retry later
341 * Otherwise fully redirty the inode so that
342 * other inodes on this superblock will get some
343 * writeout. Otherwise heavy writing to one
344 * file would indefinitely suspend writeout of
345 * all the other files.
347 inode->i_state |= I_DIRTY_PAGES;
350 } else if (inode->i_state & I_DIRTY) {
352 * Someone redirtied the inode while were writing back
356 } else if (atomic_read(&inode->i_count)) {
358 * The inode is clean, inuse
360 list_move(&inode->i_list, &inode_in_use);
363 * The inode is clean, unused
365 list_move(&inode->i_list, &inode_unused);
368 inode_sync_complete(inode);
373 * Write out an inode's dirty pages. Called under inode_lock. Either the
374 * caller has ref on the inode (either via __iget or via syscall against an fd)
375 * or the inode has I_WILL_FREE set (via generic_forget_inode)
378 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
380 wait_queue_head_t *wqh;
382 if (!atomic_read(&inode->i_count))
383 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
385 WARN_ON(inode->i_state & I_WILL_FREE);
387 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_SYNC)) {
389 * We're skipping this inode because it's locked, and we're not
390 * doing writeback-for-data-integrity. Move it to s_more_io so
391 * that writeback can proceed with the other inodes on s_io.
392 * We'll have another go at writing back this inode when we
393 * completed a full scan of s_io.
400 * It's a data-integrity sync. We must wait.
402 if (inode->i_state & I_SYNC) {
403 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
405 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
407 spin_unlock(&inode_lock);
408 __wait_on_bit(wqh, &wq, inode_wait,
409 TASK_UNINTERRUPTIBLE);
410 spin_lock(&inode_lock);
411 } while (inode->i_state & I_SYNC);
413 return __sync_single_inode(inode, wbc);
417 * Write out a superblock's list of dirty inodes. A wait will be performed
418 * upon no inodes, all inodes or the final one, depending upon sync_mode.
420 * If older_than_this is non-NULL, then only write out inodes which
421 * had their first dirtying at a time earlier than *older_than_this.
423 * If we're a pdlfush thread, then implement pdflush collision avoidance
424 * against the entire list.
426 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
427 * This function assumes that the blockdev superblock's inodes are backed by
428 * a variety of queues, so all inodes are searched. For other superblocks,
429 * assume that all inodes are backed by the same queue.
431 * FIXME: this linear search could get expensive with many fileystems. But
432 * how to fix? We need to go from an address_space to all inodes which share
433 * a queue with that address_space. (Easy: have a global "dirty superblocks"
436 * The inodes to be written are parked on sb->s_io. They are moved back onto
437 * sb->s_dirty as they are selected for writing. This way, none can be missed
438 * on the writer throttling path, and we get decent balancing between many
439 * throttled threads: we don't want them all piling up on inode_sync_wait.
441 void generic_sync_sb_inodes(struct super_block *sb,
442 struct writeback_control *wbc)
444 const unsigned long start = jiffies; /* livelock avoidance */
445 int sync = wbc->sync_mode == WB_SYNC_ALL;
447 spin_lock(&inode_lock);
448 if (!wbc->for_kupdate || list_empty(&sb->s_io))
449 queue_io(sb, wbc->older_than_this);
451 while (!list_empty(&sb->s_io)) {
452 struct inode *inode = list_entry(sb->s_io.prev,
453 struct inode, i_list);
454 struct address_space *mapping = inode->i_mapping;
455 struct backing_dev_info *bdi = mapping->backing_dev_info;
458 if (!bdi_cap_writeback_dirty(bdi)) {
460 if (sb_is_blkdev_sb(sb)) {
462 * Dirty memory-backed blockdev: the ramdisk
463 * driver does this. Skip just this inode
468 * Dirty memory-backed inode against a filesystem other
469 * than the kernel-internal bdev filesystem. Skip the
475 if (inode->i_state & I_NEW) {
480 if (wbc->nonblocking && bdi_write_congested(bdi)) {
481 wbc->encountered_congestion = 1;
482 if (!sb_is_blkdev_sb(sb))
483 break; /* Skip a congested fs */
485 continue; /* Skip a congested blockdev */
488 if (wbc->bdi && bdi != wbc->bdi) {
489 if (!sb_is_blkdev_sb(sb))
490 break; /* fs has the wrong queue */
492 continue; /* blockdev has wrong queue */
495 /* Was this inode dirtied after sync_sb_inodes was called? */
496 if (time_after(inode->dirtied_when, start))
499 /* Is another pdflush already flushing this queue? */
500 if (current_is_pdflush() && !writeback_acquire(bdi))
503 BUG_ON(inode->i_state & I_FREEING);
505 pages_skipped = wbc->pages_skipped;
506 __writeback_single_inode(inode, wbc);
507 if (current_is_pdflush())
508 writeback_release(bdi);
509 if (wbc->pages_skipped != pages_skipped) {
511 * writeback is not making progress due to locked
512 * buffers. Skip this inode for now.
516 spin_unlock(&inode_lock);
519 spin_lock(&inode_lock);
520 if (wbc->nr_to_write <= 0) {
524 if (!list_empty(&sb->s_more_io))
529 struct inode *inode, *old_inode = NULL;
532 * Data integrity sync. Must wait for all pages under writeback,
533 * because there may have been pages dirtied before our sync
534 * call, but which had writeout started before we write it out.
535 * In which case, the inode may not be on the dirty list, but
536 * we still have to wait for that writeout.
538 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
539 struct address_space *mapping;
541 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW))
543 mapping = inode->i_mapping;
544 if (mapping->nrpages == 0)
547 spin_unlock(&inode_lock);
549 * We hold a reference to 'inode' so it couldn't have
550 * been removed from s_inodes list while we dropped the
551 * inode_lock. We cannot iput the inode now as we can
552 * be holding the last reference and we cannot iput it
553 * under inode_lock. So we keep the reference and iput
559 filemap_fdatawait(mapping);
563 spin_lock(&inode_lock);
565 spin_unlock(&inode_lock);
568 spin_unlock(&inode_lock);
570 return; /* Leave any unwritten inodes on s_io */
572 EXPORT_SYMBOL_GPL(generic_sync_sb_inodes);
574 static void sync_sb_inodes(struct super_block *sb,
575 struct writeback_control *wbc)
577 generic_sync_sb_inodes(sb, wbc);
581 * Start writeback of dirty pagecache data against all unlocked inodes.
584 * We don't need to grab a reference to superblock here. If it has non-empty
585 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
586 * past sync_inodes_sb() until the ->s_dirty/s_io/s_more_io lists are all
587 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
588 * inode from superblock lists we are OK.
590 * If `older_than_this' is non-zero then only flush inodes which have a
591 * flushtime older than *older_than_this.
593 * If `bdi' is non-zero then we will scan the first inode against each
594 * superblock until we find the matching ones. One group will be the dirty
595 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
596 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
597 * super-efficient but we're about to do a ton of I/O...
600 writeback_inodes(struct writeback_control *wbc)
602 struct super_block *sb;
607 list_for_each_entry_reverse(sb, &super_blocks, s_list) {
608 if (sb_has_dirty_inodes(sb)) {
609 /* we're making our own get_super here */
611 spin_unlock(&sb_lock);
613 * If we can't get the readlock, there's no sense in
614 * waiting around, most of the time the FS is going to
615 * be unmounted by the time it is released.
617 if (down_read_trylock(&sb->s_umount)) {
619 sync_sb_inodes(sb, wbc);
620 up_read(&sb->s_umount);
623 if (__put_super_and_need_restart(sb))
626 if (wbc->nr_to_write <= 0)
629 spin_unlock(&sb_lock);
633 * writeback and wait upon the filesystem's dirty inodes. The caller will
634 * do this in two passes - one to write, and one to wait.
636 * A finite limit is set on the number of pages which will be written.
637 * To prevent infinite livelock of sys_sync().
639 * We add in the number of potentially dirty inodes, because each inode write
640 * can dirty pagecache in the underlying blockdev.
642 void sync_inodes_sb(struct super_block *sb, int wait)
644 struct writeback_control wbc = {
645 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
647 .range_end = LLONG_MAX,
651 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
652 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
654 wbc.nr_to_write = nr_dirty + nr_unstable +
655 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
657 wbc.nr_to_write = LONG_MAX; /* doesn't actually matter */
659 sync_sb_inodes(sb, &wbc);
663 * sync_inodes - writes all inodes to disk
664 * @wait: wait for completion
666 * sync_inodes() goes through each super block's dirty inode list, writes the
667 * inodes out, waits on the writeout and puts the inodes back on the normal
670 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
671 * part of the sync functions is that the blockdev "superblock" is processed
672 * last. This is because the write_inode() function of a typical fs will
673 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
674 * What we want to do is to perform all that dirtying first, and then write
675 * back all those inode blocks via the blockdev mapping in one sweep. So the
676 * additional (somewhat redundant) sync_blockdev() calls here are to make
677 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
678 * outstanding dirty inodes, the writeback goes block-at-a-time within the
679 * filesystem's write_inode(). This is extremely slow.
681 static void __sync_inodes(int wait)
683 struct super_block *sb;
687 list_for_each_entry(sb, &super_blocks, s_list) {
689 spin_unlock(&sb_lock);
690 down_read(&sb->s_umount);
692 sync_inodes_sb(sb, wait);
693 sync_blockdev(sb->s_bdev);
695 up_read(&sb->s_umount);
697 if (__put_super_and_need_restart(sb))
700 spin_unlock(&sb_lock);
703 void sync_inodes(int wait)
712 * write_inode_now - write an inode to disk
713 * @inode: inode to write to disk
714 * @sync: whether the write should be synchronous or not
716 * This function commits an inode to disk immediately if it is dirty. This is
717 * primarily needed by knfsd.
719 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
721 int write_inode_now(struct inode *inode, int sync)
724 struct writeback_control wbc = {
725 .nr_to_write = LONG_MAX,
726 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
728 .range_end = LLONG_MAX,
731 if (!mapping_cap_writeback_dirty(inode->i_mapping))
735 spin_lock(&inode_lock);
736 ret = __writeback_single_inode(inode, &wbc);
737 spin_unlock(&inode_lock);
739 inode_sync_wait(inode);
742 EXPORT_SYMBOL(write_inode_now);
745 * sync_inode - write an inode and its pages to disk.
746 * @inode: the inode to sync
747 * @wbc: controls the writeback mode
749 * sync_inode() will write an inode and its pages to disk. It will also
750 * correctly update the inode on its superblock's dirty inode lists and will
751 * update inode->i_state.
753 * The caller must have a ref on the inode.
755 int sync_inode(struct inode *inode, struct writeback_control *wbc)
759 spin_lock(&inode_lock);
760 ret = __writeback_single_inode(inode, wbc);
761 spin_unlock(&inode_lock);
764 EXPORT_SYMBOL(sync_inode);
767 * generic_osync_inode - flush all dirty data for a given inode to disk
768 * @inode: inode to write
769 * @mapping: the address_space that should be flushed
770 * @what: what to write and wait upon
772 * This can be called by file_write functions for files which have the
773 * O_SYNC flag set, to flush dirty writes to disk.
775 * @what is a bitmask, specifying which part of the inode's data should be
776 * written and waited upon.
778 * OSYNC_DATA: i_mapping's dirty data
779 * OSYNC_METADATA: the buffers at i_mapping->private_list
780 * OSYNC_INODE: the inode itself
783 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
786 int need_write_inode_now = 0;
789 if (what & OSYNC_DATA)
790 err = filemap_fdatawrite(mapping);
791 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
792 err2 = sync_mapping_buffers(mapping);
796 if (what & OSYNC_DATA) {
797 err2 = filemap_fdatawait(mapping);
802 spin_lock(&inode_lock);
803 if ((inode->i_state & I_DIRTY) &&
804 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
805 need_write_inode_now = 1;
806 spin_unlock(&inode_lock);
808 if (need_write_inode_now) {
809 err2 = write_inode_now(inode, 1);
814 inode_sync_wait(inode);
818 EXPORT_SYMBOL(generic_osync_inode);