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 struct super_block *sb = inode->i_sb;
205 int wait = wbc->sync_mode == WB_SYNC_ALL;
208 BUG_ON(inode->i_state & I_LOCK);
210 /* Set I_LOCK, reset I_DIRTY */
211 dirty = inode->i_state & I_DIRTY;
212 inode->i_state |= I_LOCK;
213 inode->i_state &= ~I_DIRTY;
215 spin_unlock(&inode_lock);
217 ret = do_writepages(mapping, wbc);
219 /* Don't write the inode if only I_DIRTY_PAGES was set */
220 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
221 int err = write_inode(inode, wait);
227 int err = filemap_fdatawait(mapping);
232 spin_lock(&inode_lock);
233 inode->i_state &= ~I_LOCK;
234 if (!(inode->i_state & I_FREEING)) {
235 if (!(inode->i_state & I_DIRTY) &&
236 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
238 * We didn't write back all the pages. nfs_writepages()
239 * sometimes bales out without doing anything. Redirty
240 * the inode. It is still on sb->s_io.
242 if (wbc->for_kupdate) {
244 * For the kupdate function we leave the inode
245 * at the head of sb_dirty so it will get more
246 * writeout as soon as the queue becomes
249 inode->i_state |= I_DIRTY_PAGES;
253 * Otherwise fully redirty the inode so that
254 * other inodes on this superblock will get some
255 * writeout. Otherwise heavy writing to one
256 * file would indefinitely suspend writeout of
257 * all the other files.
259 inode->i_state |= I_DIRTY_PAGES;
260 inode->dirtied_when = jiffies;
261 list_move(&inode->i_list, &sb->s_dirty);
263 } else if (inode->i_state & I_DIRTY) {
265 * Someone redirtied the inode while were writing back
269 } else if (atomic_read(&inode->i_count)) {
271 * The inode is clean, inuse
273 list_move(&inode->i_list, &inode_in_use);
276 * The inode is clean, unused
278 list_move(&inode->i_list, &inode_unused);
281 wake_up_inode(inode);
286 * Write out an inode's dirty pages. Called under inode_lock. Either the
287 * caller has ref on the inode (either via __iget or via syscall against an fd)
288 * or the inode has I_WILL_FREE set (via generic_forget_inode)
291 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
293 wait_queue_head_t *wqh;
295 if (!atomic_read(&inode->i_count))
296 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
298 WARN_ON(inode->i_state & I_WILL_FREE);
300 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) {
301 struct address_space *mapping = inode->i_mapping;
304 list_move(&inode->i_list, &inode->i_sb->s_dirty);
307 * Even if we don't actually write the inode itself here,
308 * we can at least start some of the data writeout..
310 spin_unlock(&inode_lock);
311 ret = do_writepages(mapping, wbc);
312 spin_lock(&inode_lock);
317 * It's a data-integrity sync. We must wait.
319 if (inode->i_state & I_LOCK) {
320 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_LOCK);
322 wqh = bit_waitqueue(&inode->i_state, __I_LOCK);
324 spin_unlock(&inode_lock);
325 __wait_on_bit(wqh, &wq, inode_wait,
326 TASK_UNINTERRUPTIBLE);
327 spin_lock(&inode_lock);
328 } while (inode->i_state & I_LOCK);
330 return __sync_single_inode(inode, wbc);
334 * Write out a superblock's list of dirty inodes. A wait will be performed
335 * upon no inodes, all inodes or the final one, depending upon sync_mode.
337 * If older_than_this is non-NULL, then only write out inodes which
338 * had their first dirtying at a time earlier than *older_than_this.
340 * If we're a pdlfush thread, then implement pdflush collision avoidance
341 * against the entire list.
343 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
344 * that it can be located for waiting on in __writeback_single_inode().
346 * Called under inode_lock.
348 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
349 * This function assumes that the blockdev superblock's inodes are backed by
350 * a variety of queues, so all inodes are searched. For other superblocks,
351 * assume that all inodes are backed by the same queue.
353 * FIXME: this linear search could get expensive with many fileystems. But
354 * how to fix? We need to go from an address_space to all inodes which share
355 * a queue with that address_space. (Easy: have a global "dirty superblocks"
358 * The inodes to be written are parked on sb->s_io. They are moved back onto
359 * sb->s_dirty as they are selected for writing. This way, none can be missed
360 * on the writer throttling path, and we get decent balancing between many
361 * throttled threads: we don't want them all piling up on __wait_on_inode.
364 sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
366 const unsigned long start = jiffies; /* livelock avoidance */
368 if (!wbc->for_kupdate || list_empty(&sb->s_io))
369 list_splice_init(&sb->s_dirty, &sb->s_io);
371 while (!list_empty(&sb->s_io)) {
372 struct inode *inode = list_entry(sb->s_io.prev,
373 struct inode, i_list);
374 struct address_space *mapping = inode->i_mapping;
375 struct backing_dev_info *bdi = mapping->backing_dev_info;
378 if (!bdi_cap_writeback_dirty(bdi)) {
380 if (sb_is_blkdev_sb(sb)) {
382 * Dirty memory-backed blockdev: the ramdisk
383 * driver does this. Skip just this inode
388 * Dirty memory-backed inode against a filesystem other
389 * than the kernel-internal bdev filesystem. Skip the
395 if (wbc->nonblocking && bdi_write_congested(bdi)) {
396 wbc->encountered_congestion = 1;
397 if (!sb_is_blkdev_sb(sb))
398 break; /* Skip a congested fs */
399 list_move(&inode->i_list, &sb->s_dirty);
400 continue; /* Skip a congested blockdev */
403 if (wbc->bdi && bdi != wbc->bdi) {
404 if (!sb_is_blkdev_sb(sb))
405 break; /* fs has the wrong queue */
406 list_move(&inode->i_list, &sb->s_dirty);
407 continue; /* blockdev has wrong queue */
410 /* Was this inode dirtied after sync_sb_inodes was called? */
411 if (time_after(inode->dirtied_when, start))
414 /* Was this inode dirtied too recently? */
415 if (wbc->older_than_this && time_after(inode->dirtied_when,
416 *wbc->older_than_this))
419 /* Is another pdflush already flushing this queue? */
420 if (current_is_pdflush() && !writeback_acquire(bdi))
423 BUG_ON(inode->i_state & I_FREEING);
425 pages_skipped = wbc->pages_skipped;
426 __writeback_single_inode(inode, wbc);
427 if (wbc->sync_mode == WB_SYNC_HOLD) {
428 inode->dirtied_when = jiffies;
429 list_move(&inode->i_list, &sb->s_dirty);
431 if (current_is_pdflush())
432 writeback_release(bdi);
433 if (wbc->pages_skipped != pages_skipped) {
435 * writeback is not making progress due to locked
436 * buffers. Skip this inode for now.
440 spin_unlock(&inode_lock);
443 spin_lock(&inode_lock);
444 if (wbc->nr_to_write <= 0)
447 return; /* Leave any unwritten inodes on s_io */
451 * Start writeback of dirty pagecache data against all unlocked inodes.
454 * We don't need to grab a reference to superblock here. If it has non-empty
455 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
456 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
457 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
458 * inode from superblock lists we are OK.
460 * If `older_than_this' is non-zero then only flush inodes which have a
461 * flushtime older than *older_than_this.
463 * If `bdi' is non-zero then we will scan the first inode against each
464 * superblock until we find the matching ones. One group will be the dirty
465 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
466 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
467 * super-efficient but we're about to do a ton of I/O...
470 writeback_inodes(struct writeback_control *wbc)
472 struct super_block *sb;
477 sb = sb_entry(super_blocks.prev);
478 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
479 if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) {
480 /* we're making our own get_super here */
482 spin_unlock(&sb_lock);
484 * If we can't get the readlock, there's no sense in
485 * waiting around, most of the time the FS is going to
486 * be unmounted by the time it is released.
488 if (down_read_trylock(&sb->s_umount)) {
490 spin_lock(&inode_lock);
491 sync_sb_inodes(sb, wbc);
492 spin_unlock(&inode_lock);
494 up_read(&sb->s_umount);
497 if (__put_super_and_need_restart(sb))
500 if (wbc->nr_to_write <= 0)
503 spin_unlock(&sb_lock);
507 * writeback and wait upon the filesystem's dirty inodes. The caller will
508 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
509 * used to park the written inodes on sb->s_dirty for the wait pass.
511 * A finite limit is set on the number of pages which will be written.
512 * To prevent infinite livelock of sys_sync().
514 * We add in the number of potentially dirty inodes, because each inode write
515 * can dirty pagecache in the underlying blockdev.
517 void sync_inodes_sb(struct super_block *sb, int wait)
519 struct writeback_control wbc = {
520 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
522 .range_end = LLONG_MAX,
524 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
525 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
527 wbc.nr_to_write = nr_dirty + nr_unstable +
528 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
529 nr_dirty + nr_unstable;
530 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
531 spin_lock(&inode_lock);
532 sync_sb_inodes(sb, &wbc);
533 spin_unlock(&inode_lock);
537 * Rather lame livelock avoidance.
539 static void set_sb_syncing(int val)
541 struct super_block *sb;
543 sb = sb_entry(super_blocks.prev);
544 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
547 spin_unlock(&sb_lock);
551 * sync_inodes - writes all inodes to disk
552 * @wait: wait for completion
554 * sync_inodes() goes through each super block's dirty inode list, writes the
555 * inodes out, waits on the writeout and puts the inodes back on the normal
558 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
559 * part of the sync functions is that the blockdev "superblock" is processed
560 * last. This is because the write_inode() function of a typical fs will
561 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
562 * What we want to do is to perform all that dirtying first, and then write
563 * back all those inode blocks via the blockdev mapping in one sweep. So the
564 * additional (somewhat redundant) sync_blockdev() calls here are to make
565 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
566 * outstanding dirty inodes, the writeback goes block-at-a-time within the
567 * filesystem's write_inode(). This is extremely slow.
569 static void __sync_inodes(int wait)
571 struct super_block *sb;
575 list_for_each_entry(sb, &super_blocks, s_list) {
580 spin_unlock(&sb_lock);
581 down_read(&sb->s_umount);
583 sync_inodes_sb(sb, wait);
584 sync_blockdev(sb->s_bdev);
586 up_read(&sb->s_umount);
588 if (__put_super_and_need_restart(sb))
591 spin_unlock(&sb_lock);
594 void sync_inodes(int wait)
606 * write_inode_now - write an inode to disk
607 * @inode: inode to write to disk
608 * @sync: whether the write should be synchronous or not
610 * This function commits an inode to disk immediately if it is dirty. This is
611 * primarily needed by knfsd.
613 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
615 int write_inode_now(struct inode *inode, int sync)
618 struct writeback_control wbc = {
619 .nr_to_write = LONG_MAX,
620 .sync_mode = WB_SYNC_ALL,
622 .range_end = LLONG_MAX,
625 if (!mapping_cap_writeback_dirty(inode->i_mapping))
629 spin_lock(&inode_lock);
630 ret = __writeback_single_inode(inode, &wbc);
631 spin_unlock(&inode_lock);
633 wait_on_inode(inode);
636 EXPORT_SYMBOL(write_inode_now);
639 * sync_inode - write an inode and its pages to disk.
640 * @inode: the inode to sync
641 * @wbc: controls the writeback mode
643 * sync_inode() will write an inode and its pages to disk. It will also
644 * correctly update the inode on its superblock's dirty inode lists and will
645 * update inode->i_state.
647 * The caller must have a ref on the inode.
649 int sync_inode(struct inode *inode, struct writeback_control *wbc)
653 spin_lock(&inode_lock);
654 ret = __writeback_single_inode(inode, wbc);
655 spin_unlock(&inode_lock);
658 EXPORT_SYMBOL(sync_inode);
661 * generic_osync_inode - flush all dirty data for a given inode to disk
662 * @inode: inode to write
663 * @mapping: the address_space that should be flushed
664 * @what: what to write and wait upon
666 * This can be called by file_write functions for files which have the
667 * O_SYNC flag set, to flush dirty writes to disk.
669 * @what is a bitmask, specifying which part of the inode's data should be
670 * written and waited upon.
672 * OSYNC_DATA: i_mapping's dirty data
673 * OSYNC_METADATA: the buffers at i_mapping->private_list
674 * OSYNC_INODE: the inode itself
677 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
680 int need_write_inode_now = 0;
683 if (what & OSYNC_DATA)
684 err = filemap_fdatawrite(mapping);
685 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
686 err2 = sync_mapping_buffers(mapping);
690 if (what & OSYNC_DATA) {
691 err2 = filemap_fdatawait(mapping);
696 spin_lock(&inode_lock);
697 if ((inode->i_state & I_DIRTY) &&
698 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
699 need_write_inode_now = 1;
700 spin_unlock(&inode_lock);
702 if (need_write_inode_now) {
703 err2 = write_inode_now(inode, 1);
708 wait_on_inode(inode);
713 EXPORT_SYMBOL(generic_osync_inode);
716 * writeback_acquire: attempt to get exclusive writeback access to a device
717 * @bdi: the device's backing_dev_info structure
719 * It is a waste of resources to have more than one pdflush thread blocked on
720 * a single request queue. Exclusion at the request_queue level is obtained
721 * via a flag in the request_queue's backing_dev_info.state.
723 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
724 * unless they implement their own. Which is somewhat inefficient, as this
725 * may prevent concurrent writeback against multiple devices.
727 int writeback_acquire(struct backing_dev_info *bdi)
729 return !test_and_set_bit(BDI_pdflush, &bdi->state);
733 * writeback_in_progress: determine whether there is writeback in progress
734 * @bdi: the device's backing_dev_info structure.
736 * Determine whether there is writeback in progress against a backing device.
738 int writeback_in_progress(struct backing_dev_info *bdi)
740 return test_bit(BDI_pdflush, &bdi->state);
744 * writeback_release: relinquish exclusive writeback access against a device.
745 * @bdi: the device's backing_dev_info structure
747 void writeback_release(struct backing_dev_info *bdi)
749 BUG_ON(!writeback_in_progress(bdi));
750 clear_bit(BDI_pdflush, &bdi->state);