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 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 void write_inode(struct inode *inode, int sync)
138 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
139 inode->i_sb->s_op->write_inode(inode, sync);
143 * Write a single inode's dirty pages and inode data out to disk.
144 * If `wait' is set, wait on the writeout.
146 * The whole writeout design is quite complex and fragile. We want to avoid
147 * starvation of particular inodes when others are being redirtied, prevent
150 * Called under inode_lock.
153 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
156 struct address_space *mapping = inode->i_mapping;
157 struct super_block *sb = inode->i_sb;
158 int wait = wbc->sync_mode == WB_SYNC_ALL;
161 BUG_ON(inode->i_state & I_LOCK);
163 /* Set I_LOCK, reset I_DIRTY */
164 dirty = inode->i_state & I_DIRTY;
165 inode->i_state |= I_LOCK;
166 inode->i_state &= ~I_DIRTY;
168 spin_unlock(&inode_lock);
170 ret = do_writepages(mapping, wbc);
172 /* Don't write the inode if only I_DIRTY_PAGES was set */
173 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC))
174 write_inode(inode, wait);
177 int err = filemap_fdatawait(mapping);
182 spin_lock(&inode_lock);
183 inode->i_state &= ~I_LOCK;
184 if (!(inode->i_state & I_FREEING)) {
185 if (!(inode->i_state & I_DIRTY) &&
186 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
188 * We didn't write back all the pages. nfs_writepages()
189 * sometimes bales out without doing anything. Redirty
190 * the inode. It is still on sb->s_io.
192 if (wbc->for_kupdate) {
194 * For the kupdate function we leave the inode
195 * at the head of sb_dirty so it will get more
196 * writeout as soon as the queue becomes
199 inode->i_state |= I_DIRTY_PAGES;
200 list_move_tail(&inode->i_list, &sb->s_dirty);
203 * Otherwise fully redirty the inode so that
204 * other inodes on this superblock will get some
205 * writeout. Otherwise heavy writing to one
206 * file would indefinitely suspend writeout of
207 * all the other files.
209 inode->i_state |= I_DIRTY_PAGES;
210 inode->dirtied_when = jiffies;
211 list_move(&inode->i_list, &sb->s_dirty);
213 } else if (inode->i_state & I_DIRTY) {
215 * Someone redirtied the inode while were writing back
216 * the pages: nothing to do.
218 } else if (atomic_read(&inode->i_count)) {
220 * The inode is clean, inuse
222 list_move(&inode->i_list, &inode_in_use);
225 * The inode is clean, unused
227 list_move(&inode->i_list, &inode_unused);
230 wake_up_inode(inode);
235 * Write out an inode's dirty pages. Called under inode_lock.
238 __writeback_single_inode(struct inode *inode,
239 struct writeback_control *wbc)
241 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) {
242 list_move(&inode->i_list, &inode->i_sb->s_dirty);
247 * It's a data-integrity sync. We must wait.
249 while (inode->i_state & I_LOCK) {
251 spin_unlock(&inode_lock);
252 __wait_on_inode(inode);
254 spin_lock(&inode_lock);
256 return __sync_single_inode(inode, wbc);
260 * Write out a superblock's list of dirty inodes. A wait will be performed
261 * upon no inodes, all inodes or the final one, depending upon sync_mode.
263 * If older_than_this is non-NULL, then only write out inodes which
264 * had their first dirtying at a time earlier than *older_than_this.
266 * If we're a pdlfush thread, then implement pdflush collision avoidance
267 * against the entire list.
269 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
270 * that it can be located for waiting on in __writeback_single_inode().
272 * Called under inode_lock.
274 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
275 * This function assumes that the blockdev superblock's inodes are backed by
276 * a variety of queues, so all inodes are searched. For other superblocks,
277 * assume that all inodes are backed by the same queue.
279 * FIXME: this linear search could get expensive with many fileystems. But
280 * how to fix? We need to go from an address_space to all inodes which share
281 * a queue with that address_space. (Easy: have a global "dirty superblocks"
284 * The inodes to be written are parked on sb->s_io. They are moved back onto
285 * sb->s_dirty as they are selected for writing. This way, none can be missed
286 * on the writer throttling path, and we get decent balancing between many
287 * throttled threads: we don't want them all piling up on __wait_on_inode.
290 sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
292 const unsigned long start = jiffies; /* livelock avoidance */
294 if (!wbc->for_kupdate || list_empty(&sb->s_io))
295 list_splice_init(&sb->s_dirty, &sb->s_io);
297 while (!list_empty(&sb->s_io)) {
298 struct inode *inode = list_entry(sb->s_io.prev,
299 struct inode, i_list);
300 struct address_space *mapping = inode->i_mapping;
301 struct backing_dev_info *bdi = mapping->backing_dev_info;
304 if (bdi->memory_backed) {
305 if (sb == blockdev_superblock) {
307 * Dirty memory-backed blockdev: the ramdisk
310 list_move(&inode->i_list, &sb->s_dirty);
315 if (wbc->nonblocking && bdi_write_congested(bdi)) {
316 wbc->encountered_congestion = 1;
317 if (sb != blockdev_superblock)
318 break; /* Skip a congested fs */
319 list_move(&inode->i_list, &sb->s_dirty);
320 continue; /* Skip a congested blockdev */
323 if (wbc->bdi && bdi != wbc->bdi) {
324 if (sb != blockdev_superblock)
325 break; /* fs has the wrong queue */
326 list_move(&inode->i_list, &sb->s_dirty);
327 continue; /* blockdev has wrong queue */
330 /* Was this inode dirtied after sync_sb_inodes was called? */
331 if (time_after(inode->dirtied_when, start))
334 /* Was this inode dirtied too recently? */
335 if (wbc->older_than_this && time_after(inode->dirtied_when,
336 *wbc->older_than_this))
339 /* Is another pdflush already flushing this queue? */
340 if (current_is_pdflush() && !writeback_acquire(bdi))
343 BUG_ON(inode->i_state & I_FREEING);
345 pages_skipped = wbc->pages_skipped;
346 __writeback_single_inode(inode, wbc);
347 if (wbc->sync_mode == WB_SYNC_HOLD) {
348 inode->dirtied_when = jiffies;
349 list_move(&inode->i_list, &sb->s_dirty);
351 if (current_is_pdflush())
352 writeback_release(bdi);
353 if (wbc->pages_skipped != pages_skipped) {
355 * writeback is not making progress due to locked
356 * buffers. Skip this inode for now.
358 list_move(&inode->i_list, &sb->s_dirty);
360 spin_unlock(&inode_lock);
362 spin_lock(&inode_lock);
363 if (wbc->nr_to_write <= 0)
366 return; /* Leave any unwritten inodes on s_io */
370 * Start writeback of dirty pagecache data against all unlocked inodes.
373 * We don't need to grab a reference to superblock here. If it has non-empty
374 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
375 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
376 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
377 * inode from superblock lists we are OK.
379 * If `older_than_this' is non-zero then only flush inodes which have a
380 * flushtime older than *older_than_this.
382 * If `bdi' is non-zero then we will scan the first inode against each
383 * superblock until we find the matching ones. One group will be the dirty
384 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
385 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
386 * super-efficient but we're about to do a ton of I/O...
389 writeback_inodes(struct writeback_control *wbc)
391 struct super_block *sb;
393 spin_lock(&inode_lock);
396 sb = sb_entry(super_blocks.prev);
397 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
398 if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) {
400 spin_unlock(&sb_lock);
401 sync_sb_inodes(sb, wbc);
406 if (wbc->nr_to_write <= 0)
409 spin_unlock(&sb_lock);
410 spin_unlock(&inode_lock);
414 * writeback and wait upon the filesystem's dirty inodes. The caller will
415 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
416 * used to park the written inodes on sb->s_dirty for the wait pass.
418 * A finite limit is set on the number of pages which will be written.
419 * To prevent infinite livelock of sys_sync().
421 * We add in the number of potentially dirty inodes, because each inode write
422 * can dirty pagecache in the underlying blockdev.
424 void sync_inodes_sb(struct super_block *sb, int wait)
426 struct page_state ps;
427 struct writeback_control wbc = {
429 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
430 .older_than_this = NULL,
435 wbc.nr_to_write = ps.nr_dirty + ps.nr_unstable +
436 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
437 ps.nr_dirty + ps.nr_unstable;
438 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
439 spin_lock(&inode_lock);
440 sync_sb_inodes(sb, &wbc);
441 spin_unlock(&inode_lock);
445 * Rather lame livelock avoidance.
447 static void set_sb_syncing(int val)
449 struct super_block *sb;
451 sb = sb_entry(super_blocks.prev);
452 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
455 spin_unlock(&sb_lock);
459 * Find a superblock with inodes that need to be synced
461 static struct super_block *get_super_to_sync(void)
463 struct super_block *sb;
466 sb = sb_entry(super_blocks.prev);
467 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
472 spin_unlock(&sb_lock);
473 down_read(&sb->s_umount);
480 spin_unlock(&sb_lock);
487 * sync_inodes() goes through each super block's dirty inode list, writes the
488 * inodes out, waits on the writeout and puts the inodes back on the normal
491 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
492 * part of the sync functions is that the blockdev "superblock" is processed
493 * last. This is because the write_inode() function of a typical fs will
494 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
495 * What we want to do is to perform all that dirtying first, and then write
496 * back all those inode blocks via the blockdev mapping in one sweep. So the
497 * additional (somewhat redundant) sync_blockdev() calls here are to make
498 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
499 * outstanding dirty inodes, the writeback goes block-at-a-time within the
500 * filesystem's write_inode(). This is extremely slow.
502 void sync_inodes(int wait)
504 struct super_block *sb;
507 while ((sb = get_super_to_sync()) != NULL) {
508 sync_inodes_sb(sb, 0);
509 sync_blockdev(sb->s_bdev);
514 while ((sb = get_super_to_sync()) != NULL) {
515 sync_inodes_sb(sb, 1);
516 sync_blockdev(sb->s_bdev);
523 * write_inode_now - write an inode to disk
524 * @inode: inode to write to disk
525 * @sync: whether the write should be synchronous or not
527 * This function commits an inode to disk immediately if it is
528 * dirty. This is primarily needed by knfsd.
531 void write_inode_now(struct inode *inode, int sync)
533 struct writeback_control wbc = {
534 .nr_to_write = LONG_MAX,
535 .sync_mode = WB_SYNC_ALL,
538 spin_lock(&inode_lock);
539 __writeback_single_inode(inode, &wbc);
540 spin_unlock(&inode_lock);
542 wait_on_inode(inode);
544 EXPORT_SYMBOL(write_inode_now);
547 * sync_inode - write an inode and its pages to disk.
548 * @inode: the inode to sync
549 * @wbc: controls the writeback mode
551 * sync_inode() will write an inode and its pages to disk. It will also
552 * correctly update the inode on its superblock's dirty inode lists and will
553 * update inode->i_state.
555 * The caller must have a ref on the inode.
557 int sync_inode(struct inode *inode, struct writeback_control *wbc)
561 spin_lock(&inode_lock);
562 ret = __writeback_single_inode(inode, wbc);
563 spin_unlock(&inode_lock);
566 EXPORT_SYMBOL(sync_inode);
569 * generic_osync_inode - flush all dirty data for a given inode to disk
570 * @inode: inode to write
571 * @what: what to write and wait upon
573 * This can be called by file_write functions for files which have the
574 * O_SYNC flag set, to flush dirty writes to disk.
576 * @what is a bitmask, specifying which part of the inode's data should be
577 * written and waited upon:
579 * OSYNC_DATA: i_mapping's dirty data
580 * OSYNC_METADATA: the buffers at i_mapping->private_list
581 * OSYNC_INODE: the inode itself
584 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
587 int need_write_inode_now = 0;
590 current->flags |= PF_SYNCWRITE;
591 if (what & OSYNC_DATA)
592 err = filemap_fdatawrite(mapping);
593 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
594 err2 = sync_mapping_buffers(mapping);
598 if (what & OSYNC_DATA) {
599 err2 = filemap_fdatawait(mapping);
603 current->flags &= ~PF_SYNCWRITE;
605 spin_lock(&inode_lock);
606 if ((inode->i_state & I_DIRTY) &&
607 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
608 need_write_inode_now = 1;
609 spin_unlock(&inode_lock);
611 if (need_write_inode_now)
612 write_inode_now(inode, 1);
614 wait_on_inode(inode);
619 EXPORT_SYMBOL(generic_osync_inode);
622 * writeback_acquire: attempt to get exclusive writeback access to a device
623 * @bdi: the device's backing_dev_info structure
625 * It is a waste of resources to have more than one pdflush thread blocked on
626 * a single request queue. Exclusion at the request_queue level is obtained
627 * via a flag in the request_queue's backing_dev_info.state.
629 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
630 * unless they implement their own. Which is somewhat inefficient, as this
631 * may prevent concurrent writeback against multiple devices.
633 int writeback_acquire(struct backing_dev_info *bdi)
635 return !test_and_set_bit(BDI_pdflush, &bdi->state);
639 * writeback_in_progress: determine whether there is writeback in progress
640 * against a backing device.
641 * @bdi: the device's backing_dev_info structure.
643 int writeback_in_progress(struct backing_dev_info *bdi)
645 return test_bit(BDI_pdflush, &bdi->state);
649 * writeback_release: relinquish exclusive writeback access against a device.
650 * @bdi: the device's backing_dev_info structure
652 void writeback_release(struct backing_dev_info *bdi)
654 BUG_ON(!writeback_in_progress(bdi));
655 clear_bit(BDI_pdflush, &bdi->state);