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/spinlock.h>
18 #include <linux/sched.h>
21 #include <linux/writeback.h>
22 #include <linux/blkdev.h>
23 #include <linux/backing-dev.h>
24 #include <linux/buffer_head.h>
26 extern struct super_block *blockdev_superblock;
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 printk("%s(%d): dirtied file\n", current->comm, current->pid);
81 spin_lock(&inode_lock);
82 if ((inode->i_state & flags) != flags) {
83 const int was_dirty = inode->i_state & I_DIRTY;
85 inode->i_state |= flags;
88 * If the inode is locked, just update its dirty state.
89 * The unlocker will place the inode on the appropriate
90 * superblock list, based upon its state.
92 if (inode->i_state & I_LOCK)
96 * Only add valid (hashed) inodes to the superblock's
97 * dirty list. Add blockdev inodes as well.
99 if (!S_ISBLK(inode->i_mode)) {
100 if (hlist_unhashed(&inode->i_hash))
103 if (inode->i_state & (I_FREEING|I_CLEAR))
107 * If the inode was already on s_dirty or s_io, don't
108 * reposition it (that would break s_dirty time-ordering).
111 inode->dirtied_when = jiffies;
112 list_move(&inode->i_list, &sb->s_dirty);
116 spin_unlock(&inode_lock);
119 EXPORT_SYMBOL(__mark_inode_dirty);
121 static void write_inode(struct inode *inode, int sync)
123 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
124 inode->i_sb->s_op->write_inode(inode, sync);
128 * Write a single inode's dirty pages and inode data out to disk.
129 * If `wait' is set, wait on the writeout.
131 * The whole writeout design is quite complex and fragile. We want to avoid
132 * starvation of particular inodes when others are being redirtied, prevent
135 * Called under inode_lock.
138 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
141 struct address_space *mapping = inode->i_mapping;
142 struct super_block *sb = inode->i_sb;
143 int wait = wbc->sync_mode == WB_SYNC_ALL;
146 BUG_ON(inode->i_state & I_LOCK);
148 /* Set I_LOCK, reset I_DIRTY */
149 dirty = inode->i_state & I_DIRTY;
150 inode->i_state |= I_LOCK;
151 inode->i_state &= ~I_DIRTY;
153 spin_unlock(&inode_lock);
155 ret = do_writepages(mapping, wbc);
157 /* Don't write the inode if only I_DIRTY_PAGES was set */
158 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC))
159 write_inode(inode, wait);
162 int err = filemap_fdatawait(mapping);
167 spin_lock(&inode_lock);
168 inode->i_state &= ~I_LOCK;
169 if (!(inode->i_state & I_FREEING)) {
170 if (!(inode->i_state & I_DIRTY) &&
171 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
173 * We didn't write back all the pages. nfs_writepages()
174 * sometimes bales out without doing anything. Redirty
175 * the inode. It is still on sb->s_io.
177 if (wbc->for_kupdate) {
179 * For the kupdate function we leave the inode
180 * at the head of sb_dirty so it will get more
181 * writeout as soon as the queue becomes
184 inode->i_state |= I_DIRTY_PAGES;
185 list_move_tail(&inode->i_list, &sb->s_dirty);
188 * Otherwise fully redirty the inode so that
189 * other inodes on this superblock will get some
190 * writeout. Otherwise heavy writing to one
191 * file would indefinitely suspend writeout of
192 * all the other files.
194 inode->i_state |= I_DIRTY_PAGES;
195 inode->dirtied_when = jiffies;
196 list_move(&inode->i_list, &sb->s_dirty);
198 } else if (inode->i_state & I_DIRTY) {
200 * Someone redirtied the inode while were writing back
201 * the pages: nothing to do.
203 } else if (atomic_read(&inode->i_count)) {
205 * The inode is clean, inuse
207 list_move(&inode->i_list, &inode_in_use);
210 * The inode is clean, unused
212 list_move(&inode->i_list, &inode_unused);
215 wake_up_inode(inode);
220 * Write out an inode's dirty pages. Called under inode_lock.
223 __writeback_single_inode(struct inode *inode,
224 struct writeback_control *wbc)
226 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) {
227 list_move(&inode->i_list, &inode->i_sb->s_dirty);
232 * It's a data-integrity sync. We must wait.
234 while (inode->i_state & I_LOCK) {
236 spin_unlock(&inode_lock);
237 __wait_on_inode(inode);
239 spin_lock(&inode_lock);
241 return __sync_single_inode(inode, wbc);
245 * Write out a superblock's list of dirty inodes. A wait will be performed
246 * upon no inodes, all inodes or the final one, depending upon sync_mode.
248 * If older_than_this is non-NULL, then only write out inodes which
249 * had their first dirtying at a time earlier than *older_than_this.
251 * If we're a pdlfush thread, then implement pdflush collision avoidance
252 * against the entire list.
254 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
255 * that it can be located for waiting on in __writeback_single_inode().
257 * Called under inode_lock.
259 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
260 * This function assumes that the blockdev superblock's inodes are backed by
261 * a variety of queues, so all inodes are searched. For other superblocks,
262 * assume that all inodes are backed by the same queue.
264 * FIXME: this linear search could get expensive with many fileystems. But
265 * how to fix? We need to go from an address_space to all inodes which share
266 * a queue with that address_space. (Easy: have a global "dirty superblocks"
269 * The inodes to be written are parked on sb->s_io. They are moved back onto
270 * sb->s_dirty as they are selected for writing. This way, none can be missed
271 * on the writer throttling path, and we get decent balancing between many
272 * throttled threads: we don't want them all piling up on __wait_on_inode.
275 sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
277 const unsigned long start = jiffies; /* livelock avoidance */
279 if (!wbc->for_kupdate || list_empty(&sb->s_io))
280 list_splice_init(&sb->s_dirty, &sb->s_io);
282 while (!list_empty(&sb->s_io)) {
283 struct inode *inode = list_entry(sb->s_io.prev,
284 struct inode, i_list);
285 struct address_space *mapping = inode->i_mapping;
286 struct backing_dev_info *bdi = mapping->backing_dev_info;
289 if (bdi->memory_backed) {
290 if (sb == blockdev_superblock) {
292 * Dirty memory-backed blockdev: the ramdisk
295 list_move(&inode->i_list, &sb->s_dirty);
299 * Assume that all inodes on this superblock are memory
300 * backed. Skip the superblock.
305 if (wbc->nonblocking && bdi_write_congested(bdi)) {
306 wbc->encountered_congestion = 1;
307 if (sb != blockdev_superblock)
308 break; /* Skip a congested fs */
309 list_move(&inode->i_list, &sb->s_dirty);
310 continue; /* Skip a congested blockdev */
313 if (wbc->bdi && bdi != wbc->bdi) {
314 if (sb != blockdev_superblock)
315 break; /* fs has the wrong queue */
316 list_move(&inode->i_list, &sb->s_dirty);
317 continue; /* blockdev has wrong queue */
320 /* Was this inode dirtied after sync_sb_inodes was called? */
321 if (time_after(inode->dirtied_when, start))
324 /* Was this inode dirtied too recently? */
325 if (wbc->older_than_this && time_after(inode->dirtied_when,
326 *wbc->older_than_this))
329 /* Is another pdflush already flushing this queue? */
330 if (current_is_pdflush() && !writeback_acquire(bdi))
333 BUG_ON(inode->i_state & I_FREEING);
335 pages_skipped = wbc->pages_skipped;
336 __writeback_single_inode(inode, wbc);
337 if (wbc->sync_mode == WB_SYNC_HOLD) {
338 inode->dirtied_when = jiffies;
339 list_move(&inode->i_list, &sb->s_dirty);
341 if (current_is_pdflush())
342 writeback_release(bdi);
343 if (wbc->pages_skipped != pages_skipped) {
345 * writeback is not making progress due to locked
346 * buffers. Skip this inode for now.
348 list_move(&inode->i_list, &sb->s_dirty);
350 spin_unlock(&inode_lock);
352 spin_lock(&inode_lock);
353 if (wbc->nr_to_write <= 0)
356 return; /* Leave any unwritten inodes on s_io */
360 * Start writeback of dirty pagecache data against all unlocked inodes.
363 * We don't need to grab a reference to superblock here. If it has non-empty
364 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
365 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
366 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
367 * inode from superblock lists we are OK.
369 * If `older_than_this' is non-zero then only flush inodes which have a
370 * flushtime older than *older_than_this.
372 * If `bdi' is non-zero then we will scan the first inode against each
373 * superblock until we find the matching ones. One group will be the dirty
374 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
375 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
376 * super-efficient but we're about to do a ton of I/O...
379 writeback_inodes(struct writeback_control *wbc)
381 struct super_block *sb;
383 spin_lock(&inode_lock);
385 sb = sb_entry(super_blocks.prev);
386 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
387 if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) {
388 spin_unlock(&sb_lock);
389 sync_sb_inodes(sb, wbc);
392 if (wbc->nr_to_write <= 0)
395 spin_unlock(&sb_lock);
396 spin_unlock(&inode_lock);
400 * writeback and wait upon the filesystem's dirty inodes. The caller will
401 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
402 * used to park the written inodes on sb->s_dirty for the wait pass.
404 * A finite limit is set on the number of pages which will be written.
405 * To prevent infinite livelock of sys_sync().
407 * We add in the number of potentially dirty inodes, because each inode write
408 * can dirty pagecache in the underlying blockdev.
410 void sync_inodes_sb(struct super_block *sb, int wait)
412 struct page_state ps;
413 struct writeback_control wbc = {
415 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
416 .older_than_this = NULL,
421 wbc.nr_to_write = ps.nr_dirty + ps.nr_unstable +
422 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
423 ps.nr_dirty + ps.nr_unstable;
424 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
425 spin_lock(&inode_lock);
426 sync_sb_inodes(sb, &wbc);
427 spin_unlock(&inode_lock);
431 * Rather lame livelock avoidance.
433 static void set_sb_syncing(int val)
435 struct super_block *sb;
437 sb = sb_entry(super_blocks.prev);
438 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
441 spin_unlock(&sb_lock);
445 * Find a superblock with inodes that need to be synced
447 static struct super_block *get_super_to_sync(void)
449 struct super_block *sb;
452 sb = sb_entry(super_blocks.prev);
453 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
458 spin_unlock(&sb_lock);
459 down_read(&sb->s_umount);
466 spin_unlock(&sb_lock);
473 * sync_inodes() goes through each super block's dirty inode list, writes the
474 * inodes out, waits on the writeout and puts the inodes back on the normal
477 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
478 * part of the sync functions is that the blockdev "superblock" is processed
479 * last. This is because the write_inode() function of a typical fs will
480 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
481 * What we want to do is to perform all that dirtying first, and then write
482 * back all those inode blocks via the blockdev mapping in one sweep. So the
483 * additional (somewhat redundant) sync_blockdev() calls here are to make
484 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
485 * outstanding dirty inodes, the writeback goes block-at-a-time within the
486 * filesystem's write_inode(). This is extremely slow.
488 void sync_inodes(int wait)
490 struct super_block *sb;
493 while ((sb = get_super_to_sync()) != NULL) {
494 sync_inodes_sb(sb, 0);
495 sync_blockdev(sb->s_bdev);
500 while ((sb = get_super_to_sync()) != NULL) {
501 sync_inodes_sb(sb, 1);
502 sync_blockdev(sb->s_bdev);
509 * write_inode_now - write an inode to disk
510 * @inode: inode to write to disk
511 * @sync: whether the write should be synchronous or not
513 * This function commits an inode to disk immediately if it is
514 * dirty. This is primarily needed by knfsd.
517 void write_inode_now(struct inode *inode, int sync)
519 struct writeback_control wbc = {
520 .nr_to_write = LONG_MAX,
521 .sync_mode = WB_SYNC_ALL,
524 spin_lock(&inode_lock);
525 __writeback_single_inode(inode, &wbc);
526 spin_unlock(&inode_lock);
528 wait_on_inode(inode);
530 EXPORT_SYMBOL(write_inode_now);
533 * sync_inode - write an inode and its pages to disk.
534 * @inode: the inode to sync
535 * @wbc: controls the writeback mode
537 * sync_inode() will write an inode and its pages to disk. It will also
538 * correctly update the inode on its superblock's dirty inode lists and will
539 * update inode->i_state.
541 * The caller must have a ref on the inode.
543 int sync_inode(struct inode *inode, struct writeback_control *wbc)
547 spin_lock(&inode_lock);
548 ret = __writeback_single_inode(inode, wbc);
549 spin_unlock(&inode_lock);
552 EXPORT_SYMBOL(sync_inode);
555 * generic_osync_inode - flush all dirty data for a given inode to disk
556 * @inode: inode to write
557 * @what: what to write and wait upon
559 * This can be called by file_write functions for files which have the
560 * O_SYNC flag set, to flush dirty writes to disk.
562 * @what is a bitmask, specifying which part of the inode's data should be
563 * written and waited upon:
565 * OSYNC_DATA: i_mapping's dirty data
566 * OSYNC_METADATA: the buffers at i_mapping->private_list
567 * OSYNC_INODE: the inode itself
570 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
573 int need_write_inode_now = 0;
576 current->flags |= PF_SYNCWRITE;
577 if (what & OSYNC_DATA)
578 err = filemap_fdatawrite(mapping);
579 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
580 err2 = sync_mapping_buffers(mapping);
584 if (what & OSYNC_DATA) {
585 err2 = filemap_fdatawait(mapping);
589 current->flags &= ~PF_SYNCWRITE;
591 spin_lock(&inode_lock);
592 if ((inode->i_state & I_DIRTY) &&
593 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
594 need_write_inode_now = 1;
595 spin_unlock(&inode_lock);
597 if (need_write_inode_now)
598 write_inode_now(inode, 1);
600 wait_on_inode(inode);
605 EXPORT_SYMBOL(generic_osync_inode);
608 * writeback_acquire: attempt to get exclusive writeback access to a device
609 * @bdi: the device's backing_dev_info structure
611 * It is a waste of resources to have more than one pdflush thread blocked on
612 * a single request queue. Exclusion at the request_queue level is obtained
613 * via a flag in the request_queue's backing_dev_info.state.
615 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
616 * unless they implement their own. Which is somewhat inefficient, as this
617 * may prevent concurrent writeback against multiple devices.
619 int writeback_acquire(struct backing_dev_info *bdi)
621 return !test_and_set_bit(BDI_pdflush, &bdi->state);
625 * writeback_in_progress: determine whether there is writeback in progress
626 * against a backing device.
627 * @bdi: the device's backing_dev_info structure.
629 int writeback_in_progress(struct backing_dev_info *bdi)
631 return test_bit(BDI_pdflush, &bdi->state);
635 * writeback_release: relinquish exclusive writeback access against a device.
636 * @bdi: the device's backing_dev_info structure
638 void writeback_release(struct backing_dev_info *bdi)
640 BUG_ON(!writeback_in_progress(bdi));
641 clear_bit(BDI_pdflush, &bdi->state);