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 or s_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 * Redirty an inode, but mark it as the very next-to-be-written inode on its
169 * superblock's dirty-inode list.
170 * We need to preserve s_dirty's reverse-time-orderedness, so we cheat by
171 * setting this inode's dirtied_when to the same value as that of the inode
172 * which is presently head-of-list, if present head-of-list is newer than this
173 * inode. (head-of-list is the least-recently-dirtied inode: the oldest one).
175 static void redirty_head(struct inode *inode)
177 struct super_block *sb = inode->i_sb;
179 if (!list_empty(&sb->s_dirty)) {
180 struct inode *head_inode;
182 head_inode = list_entry(sb->s_dirty.prev, struct inode, i_list);
183 if (time_after(inode->dirtied_when, head_inode->dirtied_when))
184 inode->dirtied_when = head_inode->dirtied_when;
186 list_move_tail(&inode->i_list, &sb->s_dirty);
190 * Write a single inode's dirty pages and inode data out to disk.
191 * If `wait' is set, wait on the writeout.
193 * The whole writeout design is quite complex and fragile. We want to avoid
194 * starvation of particular inodes when others are being redirtied, prevent
197 * Called under inode_lock.
200 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
203 struct address_space *mapping = inode->i_mapping;
204 int wait = wbc->sync_mode == WB_SYNC_ALL;
207 BUG_ON(inode->i_state & I_LOCK);
209 /* Set I_LOCK, reset I_DIRTY */
210 dirty = inode->i_state & I_DIRTY;
211 inode->i_state |= I_LOCK;
212 inode->i_state &= ~I_DIRTY;
214 spin_unlock(&inode_lock);
216 ret = do_writepages(mapping, wbc);
218 /* Don't write the inode if only I_DIRTY_PAGES was set */
219 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
220 int err = write_inode(inode, wait);
226 int err = filemap_fdatawait(mapping);
231 spin_lock(&inode_lock);
232 inode->i_state &= ~I_LOCK;
233 if (!(inode->i_state & I_FREEING)) {
234 if (!(inode->i_state & I_DIRTY) &&
235 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
237 * We didn't write back all the pages. nfs_writepages()
238 * sometimes bales out without doing anything. Redirty
239 * the inode. It is moved from s_io onto s_dirty.
242 * akpm: if the caller was the kupdate function we put
243 * this inode at the head of s_dirty so it gets first
244 * consideration. Otherwise, move it to the tail, for
245 * the reasons described there. I'm not really sure
246 * how much sense this makes. Presumably I had a good
247 * reasons for doing it this way, and I'd rather not
248 * muck with it at present.
250 if (wbc->for_kupdate) {
252 * For the kupdate function we leave the inode
253 * at the head of sb_dirty so it will get more
254 * writeout as soon as the queue becomes
257 inode->i_state |= I_DIRTY_PAGES;
261 * Otherwise fully redirty the inode so that
262 * other inodes on this superblock will get some
263 * writeout. Otherwise heavy writing to one
264 * file would indefinitely suspend writeout of
265 * all the other files.
267 inode->i_state |= I_DIRTY_PAGES;
270 } else if (inode->i_state & I_DIRTY) {
272 * Someone redirtied the inode while were writing back
276 } else if (atomic_read(&inode->i_count)) {
278 * The inode is clean, inuse
280 list_move(&inode->i_list, &inode_in_use);
283 * The inode is clean, unused
285 list_move(&inode->i_list, &inode_unused);
288 wake_up_inode(inode);
293 * Write out an inode's dirty pages. Called under inode_lock. Either the
294 * caller has ref on the inode (either via __iget or via syscall against an fd)
295 * or the inode has I_WILL_FREE set (via generic_forget_inode)
298 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
300 wait_queue_head_t *wqh;
302 if (!atomic_read(&inode->i_count))
303 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
305 WARN_ON(inode->i_state & I_WILL_FREE);
307 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) {
308 struct address_space *mapping = inode->i_mapping;
312 * We're skipping this inode because it's locked, and we're not
313 * doing writeback-for-data-integrity. Move it to the head of
314 * s_dirty so that writeback can proceed with the other inodes
315 * on s_io. We'll have another go at writing back this inode
316 * when the s_dirty iodes get moved back onto s_io.
321 * Even if we don't actually write the inode itself here,
322 * we can at least start some of the data writeout..
324 spin_unlock(&inode_lock);
325 ret = do_writepages(mapping, wbc);
326 spin_lock(&inode_lock);
331 * It's a data-integrity sync. We must wait.
333 if (inode->i_state & I_LOCK) {
334 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_LOCK);
336 wqh = bit_waitqueue(&inode->i_state, __I_LOCK);
338 spin_unlock(&inode_lock);
339 __wait_on_bit(wqh, &wq, inode_wait,
340 TASK_UNINTERRUPTIBLE);
341 spin_lock(&inode_lock);
342 } while (inode->i_state & I_LOCK);
344 return __sync_single_inode(inode, wbc);
348 * Write out a superblock's list of dirty inodes. A wait will be performed
349 * upon no inodes, all inodes or the final one, depending upon sync_mode.
351 * If older_than_this is non-NULL, then only write out inodes which
352 * had their first dirtying at a time earlier than *older_than_this.
354 * If we're a pdlfush thread, then implement pdflush collision avoidance
355 * against the entire list.
357 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
358 * that it can be located for waiting on in __writeback_single_inode().
360 * Called under inode_lock.
362 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
363 * This function assumes that the blockdev superblock's inodes are backed by
364 * a variety of queues, so all inodes are searched. For other superblocks,
365 * assume that all inodes are backed by the same queue.
367 * FIXME: this linear search could get expensive with many fileystems. But
368 * how to fix? We need to go from an address_space to all inodes which share
369 * a queue with that address_space. (Easy: have a global "dirty superblocks"
372 * The inodes to be written are parked on sb->s_io. They are moved back onto
373 * sb->s_dirty as they are selected for writing. This way, none can be missed
374 * on the writer throttling path, and we get decent balancing between many
375 * throttled threads: we don't want them all piling up on __wait_on_inode.
378 sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
380 const unsigned long start = jiffies; /* livelock avoidance */
382 if (!wbc->for_kupdate || list_empty(&sb->s_io))
383 list_splice_init(&sb->s_dirty, &sb->s_io);
385 while (!list_empty(&sb->s_io)) {
386 struct inode *inode = list_entry(sb->s_io.prev,
387 struct inode, i_list);
388 struct address_space *mapping = inode->i_mapping;
389 struct backing_dev_info *bdi = mapping->backing_dev_info;
392 if (!bdi_cap_writeback_dirty(bdi)) {
394 if (sb_is_blkdev_sb(sb)) {
396 * Dirty memory-backed blockdev: the ramdisk
397 * driver does this. Skip just this inode
402 * Dirty memory-backed inode against a filesystem other
403 * than the kernel-internal bdev filesystem. Skip the
409 if (wbc->nonblocking && bdi_write_congested(bdi)) {
410 wbc->encountered_congestion = 1;
411 if (!sb_is_blkdev_sb(sb))
412 break; /* Skip a congested fs */
414 continue; /* Skip a congested blockdev */
417 if (wbc->bdi && bdi != wbc->bdi) {
418 if (!sb_is_blkdev_sb(sb))
419 break; /* fs has the wrong queue */
421 continue; /* blockdev has wrong queue */
424 /* Was this inode dirtied after sync_sb_inodes was called? */
425 if (time_after(inode->dirtied_when, start))
428 /* Was this inode dirtied too recently? */
429 if (wbc->older_than_this && time_after(inode->dirtied_when,
430 *wbc->older_than_this))
433 /* Is another pdflush already flushing this queue? */
434 if (current_is_pdflush() && !writeback_acquire(bdi))
437 BUG_ON(inode->i_state & I_FREEING);
439 pages_skipped = wbc->pages_skipped;
440 __writeback_single_inode(inode, wbc);
441 if (wbc->sync_mode == WB_SYNC_HOLD) {
442 inode->dirtied_when = jiffies;
443 list_move(&inode->i_list, &sb->s_dirty);
445 if (current_is_pdflush())
446 writeback_release(bdi);
447 if (wbc->pages_skipped != pages_skipped) {
449 * writeback is not making progress due to locked
450 * buffers. Skip this inode for now.
454 spin_unlock(&inode_lock);
457 spin_lock(&inode_lock);
458 if (wbc->nr_to_write <= 0)
461 return; /* Leave any unwritten inodes on s_io */
465 * Start writeback of dirty pagecache data against all unlocked inodes.
468 * We don't need to grab a reference to superblock here. If it has non-empty
469 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
470 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
471 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
472 * inode from superblock lists we are OK.
474 * If `older_than_this' is non-zero then only flush inodes which have a
475 * flushtime older than *older_than_this.
477 * If `bdi' is non-zero then we will scan the first inode against each
478 * superblock until we find the matching ones. One group will be the dirty
479 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
480 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
481 * super-efficient but we're about to do a ton of I/O...
484 writeback_inodes(struct writeback_control *wbc)
486 struct super_block *sb;
491 sb = sb_entry(super_blocks.prev);
492 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
493 if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) {
494 /* we're making our own get_super here */
496 spin_unlock(&sb_lock);
498 * If we can't get the readlock, there's no sense in
499 * waiting around, most of the time the FS is going to
500 * be unmounted by the time it is released.
502 if (down_read_trylock(&sb->s_umount)) {
504 spin_lock(&inode_lock);
505 sync_sb_inodes(sb, wbc);
506 spin_unlock(&inode_lock);
508 up_read(&sb->s_umount);
511 if (__put_super_and_need_restart(sb))
514 if (wbc->nr_to_write <= 0)
517 spin_unlock(&sb_lock);
521 * writeback and wait upon the filesystem's dirty inodes. The caller will
522 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
523 * used to park the written inodes on sb->s_dirty for the wait pass.
525 * A finite limit is set on the number of pages which will be written.
526 * To prevent infinite livelock of sys_sync().
528 * We add in the number of potentially dirty inodes, because each inode write
529 * can dirty pagecache in the underlying blockdev.
531 void sync_inodes_sb(struct super_block *sb, int wait)
533 struct writeback_control wbc = {
534 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
536 .range_end = LLONG_MAX,
538 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
539 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
541 wbc.nr_to_write = nr_dirty + nr_unstable +
542 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
543 nr_dirty + nr_unstable;
544 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
545 spin_lock(&inode_lock);
546 sync_sb_inodes(sb, &wbc);
547 spin_unlock(&inode_lock);
551 * Rather lame livelock avoidance.
553 static void set_sb_syncing(int val)
555 struct super_block *sb;
557 sb = sb_entry(super_blocks.prev);
558 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
561 spin_unlock(&sb_lock);
565 * sync_inodes - writes all inodes to disk
566 * @wait: wait for completion
568 * sync_inodes() goes through each super block's dirty inode list, writes the
569 * inodes out, waits on the writeout and puts the inodes back on the normal
572 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
573 * part of the sync functions is that the blockdev "superblock" is processed
574 * last. This is because the write_inode() function of a typical fs will
575 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
576 * What we want to do is to perform all that dirtying first, and then write
577 * back all those inode blocks via the blockdev mapping in one sweep. So the
578 * additional (somewhat redundant) sync_blockdev() calls here are to make
579 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
580 * outstanding dirty inodes, the writeback goes block-at-a-time within the
581 * filesystem's write_inode(). This is extremely slow.
583 static void __sync_inodes(int wait)
585 struct super_block *sb;
589 list_for_each_entry(sb, &super_blocks, s_list) {
594 spin_unlock(&sb_lock);
595 down_read(&sb->s_umount);
597 sync_inodes_sb(sb, wait);
598 sync_blockdev(sb->s_bdev);
600 up_read(&sb->s_umount);
602 if (__put_super_and_need_restart(sb))
605 spin_unlock(&sb_lock);
608 void sync_inodes(int wait)
620 * write_inode_now - write an inode to disk
621 * @inode: inode to write to disk
622 * @sync: whether the write should be synchronous or not
624 * This function commits an inode to disk immediately if it is dirty. This is
625 * primarily needed by knfsd.
627 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
629 int write_inode_now(struct inode *inode, int sync)
632 struct writeback_control wbc = {
633 .nr_to_write = LONG_MAX,
634 .sync_mode = WB_SYNC_ALL,
636 .range_end = LLONG_MAX,
639 if (!mapping_cap_writeback_dirty(inode->i_mapping))
643 spin_lock(&inode_lock);
644 ret = __writeback_single_inode(inode, &wbc);
645 spin_unlock(&inode_lock);
647 wait_on_inode(inode);
650 EXPORT_SYMBOL(write_inode_now);
653 * sync_inode - write an inode and its pages to disk.
654 * @inode: the inode to sync
655 * @wbc: controls the writeback mode
657 * sync_inode() will write an inode and its pages to disk. It will also
658 * correctly update the inode on its superblock's dirty inode lists and will
659 * update inode->i_state.
661 * The caller must have a ref on the inode.
663 int sync_inode(struct inode *inode, struct writeback_control *wbc)
667 spin_lock(&inode_lock);
668 ret = __writeback_single_inode(inode, wbc);
669 spin_unlock(&inode_lock);
672 EXPORT_SYMBOL(sync_inode);
675 * generic_osync_inode - flush all dirty data for a given inode to disk
676 * @inode: inode to write
677 * @mapping: the address_space that should be flushed
678 * @what: what to write and wait upon
680 * This can be called by file_write functions for files which have the
681 * O_SYNC flag set, to flush dirty writes to disk.
683 * @what is a bitmask, specifying which part of the inode's data should be
684 * written and waited upon.
686 * OSYNC_DATA: i_mapping's dirty data
687 * OSYNC_METADATA: the buffers at i_mapping->private_list
688 * OSYNC_INODE: the inode itself
691 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
694 int need_write_inode_now = 0;
697 if (what & OSYNC_DATA)
698 err = filemap_fdatawrite(mapping);
699 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
700 err2 = sync_mapping_buffers(mapping);
704 if (what & OSYNC_DATA) {
705 err2 = filemap_fdatawait(mapping);
710 spin_lock(&inode_lock);
711 if ((inode->i_state & I_DIRTY) &&
712 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
713 need_write_inode_now = 1;
714 spin_unlock(&inode_lock);
716 if (need_write_inode_now) {
717 err2 = write_inode_now(inode, 1);
722 wait_on_inode(inode);
727 EXPORT_SYMBOL(generic_osync_inode);
730 * writeback_acquire: attempt to get exclusive writeback access to a device
731 * @bdi: the device's backing_dev_info structure
733 * It is a waste of resources to have more than one pdflush thread blocked on
734 * a single request queue. Exclusion at the request_queue level is obtained
735 * via a flag in the request_queue's backing_dev_info.state.
737 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
738 * unless they implement their own. Which is somewhat inefficient, as this
739 * may prevent concurrent writeback against multiple devices.
741 int writeback_acquire(struct backing_dev_info *bdi)
743 return !test_and_set_bit(BDI_pdflush, &bdi->state);
747 * writeback_in_progress: determine whether there is writeback in progress
748 * @bdi: the device's backing_dev_info structure.
750 * Determine whether there is writeback in progress against a backing device.
752 int writeback_in_progress(struct backing_dev_info *bdi)
754 return test_bit(BDI_pdflush, &bdi->state);
758 * writeback_release: relinquish exclusive writeback access against a device.
759 * @bdi: the device's backing_dev_info structure
761 void writeback_release(struct backing_dev_info *bdi)
763 BUG_ON(!writeback_in_progress(bdi));
764 clear_bit(BDI_pdflush, &bdi->state);