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 akpm@zip.com.au
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>
29 * __mark_inode_dirty - internal function
30 * @inode: inode to mark
31 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
32 * Mark an inode as dirty. Callers should use mark_inode_dirty or
33 * mark_inode_dirty_sync.
35 * Put the inode on the super block's dirty list.
37 * CAREFUL! We mark it dirty unconditionally, but move it onto the
38 * dirty list only if it is hashed or if it refers to a blockdev.
39 * If it was not hashed, it will never be added to the dirty list
40 * even if it is later hashed, as it will have been marked dirty already.
42 * In short, make sure you hash any inodes _before_ you start marking
45 * This function *must* be atomic for the I_DIRTY_PAGES case -
46 * set_page_dirty() is called under spinlock in several places.
48 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
49 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
50 * the kernel-internal blockdev inode represents the dirtying time of the
51 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
52 * page->mapping->host, so the page-dirtying time is recorded in the internal
55 void __mark_inode_dirty(struct inode *inode, int flags)
57 struct super_block *sb = inode->i_sb;
60 * Don't do this for I_DIRTY_PAGES - that doesn't actually
61 * dirty the inode itself
63 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
64 if (sb->s_op->dirty_inode)
65 sb->s_op->dirty_inode(inode);
69 * make sure that changes are seen by all cpus before we test i_state
74 /* avoid the locking if we can */
75 if ((inode->i_state & flags) == flags)
78 if (unlikely(block_dump)) {
79 struct dentry *dentry = NULL;
80 const char *name = "?";
82 if (!list_empty(&inode->i_dentry)) {
83 dentry = list_entry(inode->i_dentry.next,
84 struct dentry, d_alias);
85 if (dentry && dentry->d_name.name)
86 name = (const char *) dentry->d_name.name;
89 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev"))
91 "%s(%d): dirtied inode %lu (%s) on %s\n",
92 current->comm, current->pid, inode->i_ino,
93 name, inode->i_sb->s_id);
96 spin_lock(&inode_lock);
97 if ((inode->i_state & flags) != flags) {
98 const int was_dirty = inode->i_state & I_DIRTY;
100 inode->i_state |= flags;
103 * If the inode is locked, just update its dirty state.
104 * The unlocker will place the inode on the appropriate
105 * superblock list, based upon its state.
107 if (inode->i_state & I_LOCK)
111 * Only add valid (hashed) inodes to the superblock's
112 * dirty list. Add blockdev inodes as well.
114 if (!S_ISBLK(inode->i_mode)) {
115 if (hlist_unhashed(&inode->i_hash))
118 if (inode->i_state & (I_FREEING|I_CLEAR))
122 * If the inode was already on s_dirty/s_io/s_more_io, don't
123 * reposition it (that would break s_dirty time-ordering).
126 inode->dirtied_when = jiffies;
127 list_move(&inode->i_list, &sb->s_dirty);
131 spin_unlock(&inode_lock);
134 EXPORT_SYMBOL(__mark_inode_dirty);
136 static int write_inode(struct inode *inode, int sync)
138 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
139 return inode->i_sb->s_op->write_inode(inode, sync);
144 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
145 * furthest end of its superblock's dirty-inode list.
147 * Before stamping the inode's ->dirtied_when, we check to see whether it is
148 * already the most-recently-dirtied inode on the s_dirty list. If that is
149 * the case then the inode must have been redirtied while it was being written
150 * out and we don't reset its dirtied_when.
152 static void redirty_tail(struct inode *inode)
154 struct super_block *sb = inode->i_sb;
156 if (!list_empty(&sb->s_dirty)) {
157 struct inode *tail_inode;
159 tail_inode = list_entry(sb->s_dirty.next, struct inode, i_list);
160 if (!time_after_eq(inode->dirtied_when,
161 tail_inode->dirtied_when))
162 inode->dirtied_when = jiffies;
164 list_move(&inode->i_list, &sb->s_dirty);
168 * requeue inode for re-scanning after sb->s_io list is exhausted.
170 static void requeue_io(struct inode *inode)
172 list_move(&inode->i_list, &inode->i_sb->s_more_io);
176 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
178 static void move_expired_inodes(struct list_head *delaying_queue,
179 struct list_head *dispatch_queue,
180 unsigned long *older_than_this)
182 while (!list_empty(delaying_queue)) {
183 struct inode *inode = list_entry(delaying_queue->prev,
184 struct inode, i_list);
185 if (older_than_this &&
186 time_after(inode->dirtied_when, *older_than_this))
188 list_move(&inode->i_list, dispatch_queue);
193 * Queue all expired dirty inodes for io, eldest first.
195 static void queue_io(struct super_block *sb,
196 unsigned long *older_than_this)
198 list_splice_init(&sb->s_more_io, sb->s_io.prev);
199 move_expired_inodes(&sb->s_dirty, &sb->s_io, older_than_this);
202 int sb_has_dirty_inodes(struct super_block *sb)
204 return !list_empty(&sb->s_dirty) ||
205 !list_empty(&sb->s_io) ||
206 !list_empty(&sb->s_more_io);
208 EXPORT_SYMBOL(sb_has_dirty_inodes);
211 * Write a single inode's dirty pages and inode data out to disk.
212 * If `wait' is set, wait on the writeout.
214 * The whole writeout design is quite complex and fragile. We want to avoid
215 * starvation of particular inodes when others are being redirtied, prevent
218 * Called under inode_lock.
221 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
224 struct address_space *mapping = inode->i_mapping;
225 int wait = wbc->sync_mode == WB_SYNC_ALL;
228 BUG_ON(inode->i_state & I_LOCK);
230 /* Set I_LOCK, reset I_DIRTY */
231 dirty = inode->i_state & I_DIRTY;
232 inode->i_state |= I_LOCK;
233 inode->i_state &= ~I_DIRTY;
235 spin_unlock(&inode_lock);
237 ret = do_writepages(mapping, wbc);
239 /* Don't write the inode if only I_DIRTY_PAGES was set */
240 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
241 int err = write_inode(inode, wait);
247 int err = filemap_fdatawait(mapping);
252 spin_lock(&inode_lock);
253 inode->i_state &= ~I_LOCK;
254 if (!(inode->i_state & I_FREEING)) {
255 if (!(inode->i_state & I_DIRTY) &&
256 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
258 * We didn't write back all the pages. nfs_writepages()
259 * sometimes bales out without doing anything. Redirty
260 * the inode; Move it from s_io onto s_more_io/s_dirty.
263 * akpm: if the caller was the kupdate function we put
264 * this inode at the head of s_dirty so it gets first
265 * consideration. Otherwise, move it to the tail, for
266 * the reasons described there. I'm not really sure
267 * how much sense this makes. Presumably I had a good
268 * reasons for doing it this way, and I'd rather not
269 * muck with it at present.
271 if (wbc->for_kupdate) {
273 * For the kupdate function we move the inode
274 * to s_more_io so it will get more writeout as
275 * soon as the queue becomes uncongested.
277 inode->i_state |= I_DIRTY_PAGES;
281 * Otherwise fully redirty the inode so that
282 * other inodes on this superblock will get some
283 * writeout. Otherwise heavy writing to one
284 * file would indefinitely suspend writeout of
285 * all the other files.
287 inode->i_state |= I_DIRTY_PAGES;
290 } else if (inode->i_state & I_DIRTY) {
292 * Someone redirtied the inode while were writing back
296 } else if (atomic_read(&inode->i_count)) {
298 * The inode is clean, inuse
300 list_move(&inode->i_list, &inode_in_use);
303 * The inode is clean, unused
305 list_move(&inode->i_list, &inode_unused);
308 wake_up_inode(inode);
313 * Write out an inode's dirty pages. Called under inode_lock. Either the
314 * caller has ref on the inode (either via __iget or via syscall against an fd)
315 * or the inode has I_WILL_FREE set (via generic_forget_inode)
318 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
320 wait_queue_head_t *wqh;
322 if (!atomic_read(&inode->i_count))
323 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
325 WARN_ON(inode->i_state & I_WILL_FREE);
327 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) {
328 struct address_space *mapping = inode->i_mapping;
332 * We're skipping this inode because it's locked, and we're not
333 * doing writeback-for-data-integrity. Move it to s_more_io so
334 * that writeback can proceed with the other inodes on s_io.
335 * We'll have another go at writing back this inode when we
336 * completed a full scan of s_io.
341 * Even if we don't actually write the inode itself here,
342 * we can at least start some of the data writeout..
344 spin_unlock(&inode_lock);
345 ret = do_writepages(mapping, wbc);
346 spin_lock(&inode_lock);
351 * It's a data-integrity sync. We must wait.
353 if (inode->i_state & I_LOCK) {
354 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_LOCK);
356 wqh = bit_waitqueue(&inode->i_state, __I_LOCK);
358 spin_unlock(&inode_lock);
359 __wait_on_bit(wqh, &wq, inode_wait,
360 TASK_UNINTERRUPTIBLE);
361 spin_lock(&inode_lock);
362 } while (inode->i_state & I_LOCK);
364 return __sync_single_inode(inode, wbc);
368 * Write out a superblock's list of dirty inodes. A wait will be performed
369 * upon no inodes, all inodes or the final one, depending upon sync_mode.
371 * If older_than_this is non-NULL, then only write out inodes which
372 * had their first dirtying at a time earlier than *older_than_this.
374 * If we're a pdlfush thread, then implement pdflush collision avoidance
375 * against the entire list.
377 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
378 * that it can be located for waiting on in __writeback_single_inode().
380 * Called under inode_lock.
382 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
383 * This function assumes that the blockdev superblock's inodes are backed by
384 * a variety of queues, so all inodes are searched. For other superblocks,
385 * assume that all inodes are backed by the same queue.
387 * FIXME: this linear search could get expensive with many fileystems. But
388 * how to fix? We need to go from an address_space to all inodes which share
389 * a queue with that address_space. (Easy: have a global "dirty superblocks"
392 * The inodes to be written are parked on sb->s_io. They are moved back onto
393 * sb->s_dirty as they are selected for writing. This way, none can be missed
394 * on the writer throttling path, and we get decent balancing between many
395 * throttled threads: we don't want them all piling up on __wait_on_inode.
398 sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
400 const unsigned long start = jiffies; /* livelock avoidance */
402 if (!wbc->for_kupdate || list_empty(&sb->s_io))
403 queue_io(sb, wbc->older_than_this);
405 while (!list_empty(&sb->s_io)) {
406 struct inode *inode = list_entry(sb->s_io.prev,
407 struct inode, i_list);
408 struct address_space *mapping = inode->i_mapping;
409 struct backing_dev_info *bdi = mapping->backing_dev_info;
412 if (!bdi_cap_writeback_dirty(bdi)) {
414 if (sb_is_blkdev_sb(sb)) {
416 * Dirty memory-backed blockdev: the ramdisk
417 * driver does this. Skip just this inode
422 * Dirty memory-backed inode against a filesystem other
423 * than the kernel-internal bdev filesystem. Skip the
429 if (wbc->nonblocking && bdi_write_congested(bdi)) {
430 wbc->encountered_congestion = 1;
431 if (!sb_is_blkdev_sb(sb))
432 break; /* Skip a congested fs */
434 continue; /* Skip a congested blockdev */
437 if (wbc->bdi && bdi != wbc->bdi) {
438 if (!sb_is_blkdev_sb(sb))
439 break; /* fs has the wrong queue */
441 continue; /* blockdev has wrong queue */
444 /* Was this inode dirtied after sync_sb_inodes was called? */
445 if (time_after(inode->dirtied_when, start))
448 /* Is another pdflush already flushing this queue? */
449 if (current_is_pdflush() && !writeback_acquire(bdi))
452 BUG_ON(inode->i_state & I_FREEING);
454 pages_skipped = wbc->pages_skipped;
455 __writeback_single_inode(inode, wbc);
456 if (wbc->sync_mode == WB_SYNC_HOLD) {
457 inode->dirtied_when = jiffies;
458 list_move(&inode->i_list, &sb->s_dirty);
460 if (current_is_pdflush())
461 writeback_release(bdi);
462 if (wbc->pages_skipped != pages_skipped) {
464 * writeback is not making progress due to locked
465 * buffers. Skip this inode for now.
469 spin_unlock(&inode_lock);
472 spin_lock(&inode_lock);
473 if (wbc->nr_to_write <= 0)
476 if (!list_empty(&sb->s_more_io))
478 return; /* Leave any unwritten inodes on s_io */
482 * Start writeback of dirty pagecache data against all unlocked inodes.
485 * We don't need to grab a reference to superblock here. If it has non-empty
486 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
487 * past sync_inodes_sb() until the ->s_dirty/s_io/s_more_io lists are all
488 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
489 * inode from superblock lists we are OK.
491 * If `older_than_this' is non-zero then only flush inodes which have a
492 * flushtime older than *older_than_this.
494 * If `bdi' is non-zero then we will scan the first inode against each
495 * superblock until we find the matching ones. One group will be the dirty
496 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
497 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
498 * super-efficient but we're about to do a ton of I/O...
501 writeback_inodes(struct writeback_control *wbc)
503 struct super_block *sb;
508 sb = sb_entry(super_blocks.prev);
509 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
510 if (sb_has_dirty_inodes(sb)) {
511 /* we're making our own get_super here */
513 spin_unlock(&sb_lock);
515 * If we can't get the readlock, there's no sense in
516 * waiting around, most of the time the FS is going to
517 * be unmounted by the time it is released.
519 if (down_read_trylock(&sb->s_umount)) {
521 spin_lock(&inode_lock);
522 sync_sb_inodes(sb, wbc);
523 spin_unlock(&inode_lock);
525 up_read(&sb->s_umount);
528 if (__put_super_and_need_restart(sb))
531 if (wbc->nr_to_write <= 0)
534 spin_unlock(&sb_lock);
538 * writeback and wait upon the filesystem's dirty inodes. The caller will
539 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
540 * used to park the written inodes on sb->s_dirty for the wait pass.
542 * A finite limit is set on the number of pages which will be written.
543 * To prevent infinite livelock of sys_sync().
545 * We add in the number of potentially dirty inodes, because each inode write
546 * can dirty pagecache in the underlying blockdev.
548 void sync_inodes_sb(struct super_block *sb, int wait)
550 struct writeback_control wbc = {
551 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
553 .range_end = LLONG_MAX,
555 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
556 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
558 wbc.nr_to_write = nr_dirty + nr_unstable +
559 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
560 nr_dirty + nr_unstable;
561 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
562 spin_lock(&inode_lock);
563 sync_sb_inodes(sb, &wbc);
564 spin_unlock(&inode_lock);
568 * Rather lame livelock avoidance.
570 static void set_sb_syncing(int val)
572 struct super_block *sb;
574 sb = sb_entry(super_blocks.prev);
575 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
578 spin_unlock(&sb_lock);
582 * sync_inodes - writes all inodes to disk
583 * @wait: wait for completion
585 * sync_inodes() goes through each super block's dirty inode list, writes the
586 * inodes out, waits on the writeout and puts the inodes back on the normal
589 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
590 * part of the sync functions is that the blockdev "superblock" is processed
591 * last. This is because the write_inode() function of a typical fs will
592 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
593 * What we want to do is to perform all that dirtying first, and then write
594 * back all those inode blocks via the blockdev mapping in one sweep. So the
595 * additional (somewhat redundant) sync_blockdev() calls here are to make
596 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
597 * outstanding dirty inodes, the writeback goes block-at-a-time within the
598 * filesystem's write_inode(). This is extremely slow.
600 static void __sync_inodes(int wait)
602 struct super_block *sb;
606 list_for_each_entry(sb, &super_blocks, s_list) {
611 spin_unlock(&sb_lock);
612 down_read(&sb->s_umount);
614 sync_inodes_sb(sb, wait);
615 sync_blockdev(sb->s_bdev);
617 up_read(&sb->s_umount);
619 if (__put_super_and_need_restart(sb))
622 spin_unlock(&sb_lock);
625 void sync_inodes(int wait)
637 * write_inode_now - write an inode to disk
638 * @inode: inode to write to disk
639 * @sync: whether the write should be synchronous or not
641 * This function commits an inode to disk immediately if it is dirty. This is
642 * primarily needed by knfsd.
644 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
646 int write_inode_now(struct inode *inode, int sync)
649 struct writeback_control wbc = {
650 .nr_to_write = LONG_MAX,
651 .sync_mode = WB_SYNC_ALL,
653 .range_end = LLONG_MAX,
656 if (!mapping_cap_writeback_dirty(inode->i_mapping))
660 spin_lock(&inode_lock);
661 ret = __writeback_single_inode(inode, &wbc);
662 spin_unlock(&inode_lock);
664 wait_on_inode(inode);
667 EXPORT_SYMBOL(write_inode_now);
670 * sync_inode - write an inode and its pages to disk.
671 * @inode: the inode to sync
672 * @wbc: controls the writeback mode
674 * sync_inode() will write an inode and its pages to disk. It will also
675 * correctly update the inode on its superblock's dirty inode lists and will
676 * update inode->i_state.
678 * The caller must have a ref on the inode.
680 int sync_inode(struct inode *inode, struct writeback_control *wbc)
684 spin_lock(&inode_lock);
685 ret = __writeback_single_inode(inode, wbc);
686 spin_unlock(&inode_lock);
689 EXPORT_SYMBOL(sync_inode);
692 * generic_osync_inode - flush all dirty data for a given inode to disk
693 * @inode: inode to write
694 * @mapping: the address_space that should be flushed
695 * @what: what to write and wait upon
697 * This can be called by file_write functions for files which have the
698 * O_SYNC flag set, to flush dirty writes to disk.
700 * @what is a bitmask, specifying which part of the inode's data should be
701 * written and waited upon.
703 * OSYNC_DATA: i_mapping's dirty data
704 * OSYNC_METADATA: the buffers at i_mapping->private_list
705 * OSYNC_INODE: the inode itself
708 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
711 int need_write_inode_now = 0;
714 if (what & OSYNC_DATA)
715 err = filemap_fdatawrite(mapping);
716 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
717 err2 = sync_mapping_buffers(mapping);
721 if (what & OSYNC_DATA) {
722 err2 = filemap_fdatawait(mapping);
727 spin_lock(&inode_lock);
728 if ((inode->i_state & I_DIRTY) &&
729 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
730 need_write_inode_now = 1;
731 spin_unlock(&inode_lock);
733 if (need_write_inode_now) {
734 err2 = write_inode_now(inode, 1);
739 wait_on_inode(inode);
744 EXPORT_SYMBOL(generic_osync_inode);
747 * writeback_acquire: attempt to get exclusive writeback access to a device
748 * @bdi: the device's backing_dev_info structure
750 * It is a waste of resources to have more than one pdflush thread blocked on
751 * a single request queue. Exclusion at the request_queue level is obtained
752 * via a flag in the request_queue's backing_dev_info.state.
754 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
755 * unless they implement their own. Which is somewhat inefficient, as this
756 * may prevent concurrent writeback against multiple devices.
758 int writeback_acquire(struct backing_dev_info *bdi)
760 return !test_and_set_bit(BDI_pdflush, &bdi->state);
764 * writeback_in_progress: determine whether there is writeback in progress
765 * @bdi: the device's backing_dev_info structure.
767 * Determine whether there is writeback in progress against a backing device.
769 int writeback_in_progress(struct backing_dev_info *bdi)
771 return test_bit(BDI_pdflush, &bdi->state);
775 * writeback_release: relinquish exclusive writeback access against a device.
776 * @bdi: the device's backing_dev_info structure
778 void writeback_release(struct backing_dev_info *bdi)
780 BUG_ON(!writeback_in_progress(bdi));
781 clear_bit(BDI_pdflush, &bdi->state);