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);
67 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
69 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
70 struct dentry *dentry;
71 const char *name = "?";
73 dentry = d_find_alias(inode);
75 spin_lock(&dentry->d_lock);
76 name = (const char *) dentry->d_name.name;
79 "%s(%d): dirtied inode %lu (%s) on %s\n",
80 current->comm, task_pid_nr(current), inode->i_ino,
81 name, inode->i_sb->s_id);
83 spin_unlock(&dentry->d_lock);
90 * __mark_inode_dirty - internal function
91 * @inode: inode to mark
92 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
93 * Mark an inode as dirty. Callers should use mark_inode_dirty or
94 * mark_inode_dirty_sync.
96 * Put the inode on the super block's dirty list.
98 * CAREFUL! We mark it dirty unconditionally, but move it onto the
99 * dirty list only if it is hashed or if it refers to a blockdev.
100 * If it was not hashed, it will never be added to the dirty list
101 * even if it is later hashed, as it will have been marked dirty already.
103 * In short, make sure you hash any inodes _before_ you start marking
106 * This function *must* be atomic for the I_DIRTY_PAGES case -
107 * set_page_dirty() is called under spinlock in several places.
109 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
110 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
111 * the kernel-internal blockdev inode represents the dirtying time of the
112 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
113 * page->mapping->host, so the page-dirtying time is recorded in the internal
116 void __mark_inode_dirty(struct inode *inode, int flags)
118 struct super_block *sb = inode->i_sb;
121 * Don't do this for I_DIRTY_PAGES - that doesn't actually
122 * dirty the inode itself
124 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
125 if (sb->s_op->dirty_inode)
126 sb->s_op->dirty_inode(inode);
130 * make sure that changes are seen by all cpus before we test i_state
135 /* avoid the locking if we can */
136 if ((inode->i_state & flags) == flags)
139 if (unlikely(block_dump))
140 block_dump___mark_inode_dirty(inode);
142 spin_lock(&inode_lock);
143 if ((inode->i_state & flags) != flags) {
144 const int was_dirty = inode->i_state & I_DIRTY;
146 inode->i_state |= flags;
149 * If the inode is being synced, just update its dirty state.
150 * The unlocker will place the inode on the appropriate
151 * superblock list, based upon its state.
153 if (inode->i_state & I_SYNC)
157 * Only add valid (hashed) inodes to the superblock's
158 * dirty list. Add blockdev inodes as well.
160 if (!S_ISBLK(inode->i_mode)) {
161 if (hlist_unhashed(&inode->i_hash))
164 if (inode->i_state & (I_FREEING|I_CLEAR))
168 * If the inode was already on s_dirty/s_io/s_more_io, don't
169 * reposition it (that would break s_dirty time-ordering).
172 inode->dirtied_when = jiffies;
173 list_move(&inode->i_list, &sb->s_dirty);
177 spin_unlock(&inode_lock);
180 EXPORT_SYMBOL(__mark_inode_dirty);
182 static int write_inode(struct inode *inode, int sync)
184 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
185 return inode->i_sb->s_op->write_inode(inode, sync);
190 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
191 * furthest end of its superblock's dirty-inode list.
193 * Before stamping the inode's ->dirtied_when, we check to see whether it is
194 * already the most-recently-dirtied inode on the s_dirty list. If that is
195 * the case then the inode must have been redirtied while it was being written
196 * out and we don't reset its dirtied_when.
198 static void redirty_tail(struct inode *inode)
200 struct super_block *sb = inode->i_sb;
202 if (!list_empty(&sb->s_dirty)) {
203 struct inode *tail_inode;
205 tail_inode = list_entry(sb->s_dirty.next, struct inode, i_list);
206 if (time_before(inode->dirtied_when,
207 tail_inode->dirtied_when))
208 inode->dirtied_when = jiffies;
210 list_move(&inode->i_list, &sb->s_dirty);
214 * requeue inode for re-scanning after sb->s_io list is exhausted.
216 static void requeue_io(struct inode *inode)
218 list_move(&inode->i_list, &inode->i_sb->s_more_io);
221 static void inode_sync_complete(struct inode *inode)
224 * Prevent speculative execution through spin_unlock(&inode_lock);
227 wake_up_bit(&inode->i_state, __I_SYNC);
230 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
232 bool ret = time_after(inode->dirtied_when, t);
235 * For inodes being constantly redirtied, dirtied_when can get stuck.
236 * It _appears_ to be in the future, but is actually in distant past.
237 * This test is necessary to prevent such wrapped-around relative times
238 * from permanently stopping the whole pdflush writeback.
240 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
246 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
248 static void move_expired_inodes(struct list_head *delaying_queue,
249 struct list_head *dispatch_queue,
250 unsigned long *older_than_this)
252 while (!list_empty(delaying_queue)) {
253 struct inode *inode = list_entry(delaying_queue->prev,
254 struct inode, i_list);
255 if (older_than_this &&
256 inode_dirtied_after(inode, *older_than_this))
258 list_move(&inode->i_list, dispatch_queue);
263 * Queue all expired dirty inodes for io, eldest first.
265 static void queue_io(struct super_block *sb,
266 unsigned long *older_than_this)
268 list_splice_init(&sb->s_more_io, sb->s_io.prev);
269 move_expired_inodes(&sb->s_dirty, &sb->s_io, older_than_this);
272 int sb_has_dirty_inodes(struct super_block *sb)
274 return !list_empty(&sb->s_dirty) ||
275 !list_empty(&sb->s_io) ||
276 !list_empty(&sb->s_more_io);
278 EXPORT_SYMBOL(sb_has_dirty_inodes);
281 * Write a single inode's dirty pages and inode data out to disk.
282 * If `wait' is set, wait on the writeout.
284 * The whole writeout design is quite complex and fragile. We want to avoid
285 * starvation of particular inodes when others are being redirtied, prevent
288 * Called under inode_lock.
291 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
294 struct address_space *mapping = inode->i_mapping;
295 int wait = wbc->sync_mode == WB_SYNC_ALL;
298 BUG_ON(inode->i_state & I_SYNC);
300 /* Set I_SYNC, reset I_DIRTY */
301 dirty = inode->i_state & I_DIRTY;
302 inode->i_state |= I_SYNC;
303 inode->i_state &= ~I_DIRTY;
305 spin_unlock(&inode_lock);
307 ret = do_writepages(mapping, wbc);
309 /* Don't write the inode if only I_DIRTY_PAGES was set */
310 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
311 int err = write_inode(inode, wait);
317 int err = filemap_fdatawait(mapping);
322 spin_lock(&inode_lock);
323 inode->i_state &= ~I_SYNC;
324 if (!(inode->i_state & I_FREEING)) {
325 if (!(inode->i_state & I_DIRTY) &&
326 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
328 * We didn't write back all the pages. nfs_writepages()
329 * sometimes bales out without doing anything. Redirty
330 * the inode; Move it from s_io onto s_more_io/s_dirty.
333 * akpm: if the caller was the kupdate function we put
334 * this inode at the head of s_dirty so it gets first
335 * consideration. Otherwise, move it to the tail, for
336 * the reasons described there. I'm not really sure
337 * how much sense this makes. Presumably I had a good
338 * reasons for doing it this way, and I'd rather not
339 * muck with it at present.
341 if (wbc->for_kupdate) {
343 * For the kupdate function we move the inode
344 * to s_more_io so it will get more writeout as
345 * soon as the queue becomes uncongested.
347 inode->i_state |= I_DIRTY_PAGES;
348 if (wbc->nr_to_write <= 0) {
350 * slice used up: queue for next turn
355 * somehow blocked: retry later
361 * Otherwise fully redirty the inode so that
362 * other inodes on this superblock will get some
363 * writeout. Otherwise heavy writing to one
364 * file would indefinitely suspend writeout of
365 * all the other files.
367 inode->i_state |= I_DIRTY_PAGES;
370 } else if (inode->i_state & I_DIRTY) {
372 * Someone redirtied the inode while were writing back
376 } else if (atomic_read(&inode->i_count)) {
378 * The inode is clean, inuse
380 list_move(&inode->i_list, &inode_in_use);
383 * The inode is clean, unused
385 list_move(&inode->i_list, &inode_unused);
388 inode_sync_complete(inode);
393 * Write out an inode's dirty pages. Called under inode_lock. Either the
394 * caller has ref on the inode (either via __iget or via syscall against an fd)
395 * or the inode has I_WILL_FREE set (via generic_forget_inode)
398 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
400 wait_queue_head_t *wqh;
402 if (!atomic_read(&inode->i_count))
403 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
405 WARN_ON(inode->i_state & I_WILL_FREE);
407 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_SYNC)) {
409 * We're skipping this inode because it's locked, and we're not
410 * doing writeback-for-data-integrity. Move it to s_more_io so
411 * that writeback can proceed with the other inodes on s_io.
412 * We'll have another go at writing back this inode when we
413 * completed a full scan of s_io.
420 * It's a data-integrity sync. We must wait.
422 if (inode->i_state & I_SYNC) {
423 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
425 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
427 spin_unlock(&inode_lock);
428 __wait_on_bit(wqh, &wq, inode_wait,
429 TASK_UNINTERRUPTIBLE);
430 spin_lock(&inode_lock);
431 } while (inode->i_state & I_SYNC);
433 return __sync_single_inode(inode, wbc);
437 * Write out a superblock's list of dirty inodes. A wait will be performed
438 * upon no inodes, all inodes or the final one, depending upon sync_mode.
440 * If older_than_this is non-NULL, then only write out inodes which
441 * had their first dirtying at a time earlier than *older_than_this.
443 * If we're a pdflush thread, then implement pdflush collision avoidance
444 * against the entire list.
446 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
447 * This function assumes that the blockdev superblock's inodes are backed by
448 * a variety of queues, so all inodes are searched. For other superblocks,
449 * assume that all inodes are backed by the same queue.
451 * FIXME: this linear search could get expensive with many fileystems. But
452 * how to fix? We need to go from an address_space to all inodes which share
453 * a queue with that address_space. (Easy: have a global "dirty superblocks"
456 * The inodes to be written are parked on sb->s_io. They are moved back onto
457 * sb->s_dirty as they are selected for writing. This way, none can be missed
458 * on the writer throttling path, and we get decent balancing between many
459 * throttled threads: we don't want them all piling up on inode_sync_wait.
461 void generic_sync_sb_inodes(struct super_block *sb,
462 struct writeback_control *wbc)
464 const unsigned long start = jiffies; /* livelock avoidance */
465 int sync = wbc->sync_mode == WB_SYNC_ALL;
467 spin_lock(&inode_lock);
468 if (!wbc->for_kupdate || list_empty(&sb->s_io))
469 queue_io(sb, wbc->older_than_this);
471 while (!list_empty(&sb->s_io)) {
472 struct inode *inode = list_entry(sb->s_io.prev,
473 struct inode, i_list);
474 struct address_space *mapping = inode->i_mapping;
475 struct backing_dev_info *bdi = mapping->backing_dev_info;
478 if (!bdi_cap_writeback_dirty(bdi)) {
480 if (sb_is_blkdev_sb(sb)) {
482 * Dirty memory-backed blockdev: the ramdisk
483 * driver does this. Skip just this inode
488 * Dirty memory-backed inode against a filesystem other
489 * than the kernel-internal bdev filesystem. Skip the
495 if (inode->i_state & I_NEW) {
500 if (wbc->nonblocking && bdi_write_congested(bdi)) {
501 wbc->encountered_congestion = 1;
502 if (!sb_is_blkdev_sb(sb))
503 break; /* Skip a congested fs */
505 continue; /* Skip a congested blockdev */
508 if (wbc->bdi && bdi != wbc->bdi) {
509 if (!sb_is_blkdev_sb(sb))
510 break; /* fs has the wrong queue */
512 continue; /* blockdev has wrong queue */
516 * Was this inode dirtied after sync_sb_inodes was called?
517 * This keeps sync from extra jobs and livelock.
519 if (inode_dirtied_after(inode, start))
522 /* Is another pdflush already flushing this queue? */
523 if (current_is_pdflush() && !writeback_acquire(bdi))
526 BUG_ON(inode->i_state & I_FREEING);
528 pages_skipped = wbc->pages_skipped;
529 __writeback_single_inode(inode, wbc);
530 if (current_is_pdflush())
531 writeback_release(bdi);
532 if (wbc->pages_skipped != pages_skipped) {
534 * writeback is not making progress due to locked
535 * buffers. Skip this inode for now.
539 spin_unlock(&inode_lock);
542 spin_lock(&inode_lock);
543 if (wbc->nr_to_write <= 0) {
547 if (!list_empty(&sb->s_more_io))
552 struct inode *inode, *old_inode = NULL;
555 * Data integrity sync. Must wait for all pages under writeback,
556 * because there may have been pages dirtied before our sync
557 * call, but which had writeout started before we write it out.
558 * In which case, the inode may not be on the dirty list, but
559 * we still have to wait for that writeout.
561 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
562 struct address_space *mapping;
565 (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW))
567 mapping = inode->i_mapping;
568 if (mapping->nrpages == 0)
571 spin_unlock(&inode_lock);
573 * We hold a reference to 'inode' so it couldn't have
574 * been removed from s_inodes list while we dropped the
575 * inode_lock. We cannot iput the inode now as we can
576 * be holding the last reference and we cannot iput it
577 * under inode_lock. So we keep the reference and iput
583 filemap_fdatawait(mapping);
587 spin_lock(&inode_lock);
589 spin_unlock(&inode_lock);
592 spin_unlock(&inode_lock);
594 return; /* Leave any unwritten inodes on s_io */
596 EXPORT_SYMBOL_GPL(generic_sync_sb_inodes);
598 static void sync_sb_inodes(struct super_block *sb,
599 struct writeback_control *wbc)
601 generic_sync_sb_inodes(sb, wbc);
605 * Start writeback of dirty pagecache data against all unlocked inodes.
608 * We don't need to grab a reference to superblock here. If it has non-empty
609 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
610 * past sync_inodes_sb() until the ->s_dirty/s_io/s_more_io lists are all
611 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
612 * inode from superblock lists we are OK.
614 * If `older_than_this' is non-zero then only flush inodes which have a
615 * flushtime older than *older_than_this.
617 * If `bdi' is non-zero then we will scan the first inode against each
618 * superblock until we find the matching ones. One group will be the dirty
619 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
620 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
621 * super-efficient but we're about to do a ton of I/O...
624 writeback_inodes(struct writeback_control *wbc)
626 struct super_block *sb;
631 list_for_each_entry_reverse(sb, &super_blocks, s_list) {
632 if (sb_has_dirty_inodes(sb)) {
633 /* we're making our own get_super here */
635 spin_unlock(&sb_lock);
637 * If we can't get the readlock, there's no sense in
638 * waiting around, most of the time the FS is going to
639 * be unmounted by the time it is released.
641 if (down_read_trylock(&sb->s_umount)) {
643 sync_sb_inodes(sb, wbc);
644 up_read(&sb->s_umount);
647 if (__put_super_and_need_restart(sb))
650 if (wbc->nr_to_write <= 0)
653 spin_unlock(&sb_lock);
657 * writeback and wait upon the filesystem's dirty inodes. The caller will
658 * do this in two passes - one to write, and one to wait.
660 * A finite limit is set on the number of pages which will be written.
661 * To prevent infinite livelock of sys_sync().
663 * We add in the number of potentially dirty inodes, because each inode write
664 * can dirty pagecache in the underlying blockdev.
666 void sync_inodes_sb(struct super_block *sb, int wait)
668 struct writeback_control wbc = {
669 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
671 .range_end = LLONG_MAX,
675 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
676 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
678 wbc.nr_to_write = nr_dirty + nr_unstable +
679 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
681 wbc.nr_to_write = LONG_MAX; /* doesn't actually matter */
683 sync_sb_inodes(sb, &wbc);
687 * write_inode_now - write an inode to disk
688 * @inode: inode to write to disk
689 * @sync: whether the write should be synchronous or not
691 * This function commits an inode to disk immediately if it is dirty. This is
692 * primarily needed by knfsd.
694 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
696 int write_inode_now(struct inode *inode, int sync)
699 struct writeback_control wbc = {
700 .nr_to_write = LONG_MAX,
701 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
703 .range_end = LLONG_MAX,
706 if (!mapping_cap_writeback_dirty(inode->i_mapping))
710 spin_lock(&inode_lock);
711 ret = __writeback_single_inode(inode, &wbc);
712 spin_unlock(&inode_lock);
714 inode_sync_wait(inode);
717 EXPORT_SYMBOL(write_inode_now);
720 * sync_inode - write an inode and its pages to disk.
721 * @inode: the inode to sync
722 * @wbc: controls the writeback mode
724 * sync_inode() will write an inode and its pages to disk. It will also
725 * correctly update the inode on its superblock's dirty inode lists and will
726 * update inode->i_state.
728 * The caller must have a ref on the inode.
730 int sync_inode(struct inode *inode, struct writeback_control *wbc)
734 spin_lock(&inode_lock);
735 ret = __writeback_single_inode(inode, wbc);
736 spin_unlock(&inode_lock);
739 EXPORT_SYMBOL(sync_inode);
742 * generic_osync_inode - flush all dirty data for a given inode to disk
743 * @inode: inode to write
744 * @mapping: the address_space that should be flushed
745 * @what: what to write and wait upon
747 * This can be called by file_write functions for files which have the
748 * O_SYNC flag set, to flush dirty writes to disk.
750 * @what is a bitmask, specifying which part of the inode's data should be
751 * written and waited upon.
753 * OSYNC_DATA: i_mapping's dirty data
754 * OSYNC_METADATA: the buffers at i_mapping->private_list
755 * OSYNC_INODE: the inode itself
758 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
761 int need_write_inode_now = 0;
764 if (what & OSYNC_DATA)
765 err = filemap_fdatawrite(mapping);
766 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
767 err2 = sync_mapping_buffers(mapping);
771 if (what & OSYNC_DATA) {
772 err2 = filemap_fdatawait(mapping);
777 spin_lock(&inode_lock);
778 if ((inode->i_state & I_DIRTY) &&
779 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
780 need_write_inode_now = 1;
781 spin_unlock(&inode_lock);
783 if (need_write_inode_now) {
784 err2 = write_inode_now(inode, 1);
789 inode_sync_wait(inode);
793 EXPORT_SYMBOL(generic_osync_inode);