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);
228 inodes_stat.nr_unused++;
231 wake_up_inode(inode);
236 * Write out an inode's dirty pages. Called under inode_lock.
239 __writeback_single_inode(struct inode *inode,
240 struct writeback_control *wbc)
242 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) {
243 list_move(&inode->i_list, &inode->i_sb->s_dirty);
248 * It's a data-integrity sync. We must wait.
250 while (inode->i_state & I_LOCK) {
252 spin_unlock(&inode_lock);
253 __wait_on_inode(inode);
255 spin_lock(&inode_lock);
257 return __sync_single_inode(inode, wbc);
261 * Write out a superblock's list of dirty inodes. A wait will be performed
262 * upon no inodes, all inodes or the final one, depending upon sync_mode.
264 * If older_than_this is non-NULL, then only write out inodes which
265 * had their first dirtying at a time earlier than *older_than_this.
267 * If we're a pdlfush thread, then implement pdflush collision avoidance
268 * against the entire list.
270 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
271 * that it can be located for waiting on in __writeback_single_inode().
273 * Called under inode_lock.
275 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
276 * This function assumes that the blockdev superblock's inodes are backed by
277 * a variety of queues, so all inodes are searched. For other superblocks,
278 * assume that all inodes are backed by the same queue.
280 * FIXME: this linear search could get expensive with many fileystems. But
281 * how to fix? We need to go from an address_space to all inodes which share
282 * a queue with that address_space. (Easy: have a global "dirty superblocks"
285 * The inodes to be written are parked on sb->s_io. They are moved back onto
286 * sb->s_dirty as they are selected for writing. This way, none can be missed
287 * on the writer throttling path, and we get decent balancing between many
288 * throttled threads: we don't want them all piling up on __wait_on_inode.
291 sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
293 const unsigned long start = jiffies; /* livelock avoidance */
295 if (!wbc->for_kupdate || list_empty(&sb->s_io))
296 list_splice_init(&sb->s_dirty, &sb->s_io);
298 while (!list_empty(&sb->s_io)) {
299 struct inode *inode = list_entry(sb->s_io.prev,
300 struct inode, i_list);
301 struct address_space *mapping = inode->i_mapping;
302 struct backing_dev_info *bdi = mapping->backing_dev_info;
305 if (bdi->memory_backed) {
306 if (sb == blockdev_superblock) {
308 * Dirty memory-backed blockdev: the ramdisk
311 list_move(&inode->i_list, &sb->s_dirty);
316 if (wbc->nonblocking && bdi_write_congested(bdi)) {
317 wbc->encountered_congestion = 1;
318 if (sb != blockdev_superblock)
319 break; /* Skip a congested fs */
320 list_move(&inode->i_list, &sb->s_dirty);
321 continue; /* Skip a congested blockdev */
324 if (wbc->bdi && bdi != wbc->bdi) {
325 if (sb != blockdev_superblock)
326 break; /* fs has the wrong queue */
327 list_move(&inode->i_list, &sb->s_dirty);
328 continue; /* blockdev has wrong queue */
331 /* Was this inode dirtied after sync_sb_inodes was called? */
332 if (time_after(inode->dirtied_when, start))
335 /* Was this inode dirtied too recently? */
336 if (wbc->older_than_this && time_after(inode->dirtied_when,
337 *wbc->older_than_this))
340 /* Is another pdflush already flushing this queue? */
341 if (current_is_pdflush() && !writeback_acquire(bdi))
344 BUG_ON(inode->i_state & I_FREEING);
346 pages_skipped = wbc->pages_skipped;
347 __writeback_single_inode(inode, wbc);
348 if (wbc->sync_mode == WB_SYNC_HOLD) {
349 inode->dirtied_when = jiffies;
350 list_move(&inode->i_list, &sb->s_dirty);
352 if (current_is_pdflush())
353 writeback_release(bdi);
354 if (wbc->pages_skipped != pages_skipped) {
356 * writeback is not making progress due to locked
357 * buffers. Skip this inode for now.
359 list_move(&inode->i_list, &sb->s_dirty);
361 spin_unlock(&inode_lock);
364 spin_lock(&inode_lock);
365 if (wbc->nr_to_write <= 0)
368 return; /* Leave any unwritten inodes on s_io */
372 * Start writeback of dirty pagecache data against all unlocked inodes.
375 * We don't need to grab a reference to superblock here. If it has non-empty
376 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
377 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
378 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
379 * inode from superblock lists we are OK.
381 * If `older_than_this' is non-zero then only flush inodes which have a
382 * flushtime older than *older_than_this.
384 * If `bdi' is non-zero then we will scan the first inode against each
385 * superblock until we find the matching ones. One group will be the dirty
386 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
387 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
388 * super-efficient but we're about to do a ton of I/O...
391 writeback_inodes(struct writeback_control *wbc)
393 struct super_block *sb;
396 spin_lock(&inode_lock);
399 sb = sb_entry(super_blocks.prev);
400 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
401 if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) {
402 /* we're making our own get_super here */
404 spin_unlock(&sb_lock);
406 * If we can't get the readlock, there's no sense in
407 * waiting around, most of the time the FS is going to
408 * be unmounted by the time it is released.
410 if (down_read_trylock(&sb->s_umount)) {
412 sync_sb_inodes(sb, wbc);
413 up_read(&sb->s_umount);
419 if (wbc->nr_to_write <= 0)
422 spin_unlock(&sb_lock);
423 spin_unlock(&inode_lock);
427 * writeback and wait upon the filesystem's dirty inodes. The caller will
428 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
429 * used to park the written inodes on sb->s_dirty for the wait pass.
431 * A finite limit is set on the number of pages which will be written.
432 * To prevent infinite livelock of sys_sync().
434 * We add in the number of potentially dirty inodes, because each inode write
435 * can dirty pagecache in the underlying blockdev.
437 void sync_inodes_sb(struct super_block *sb, int wait)
439 struct writeback_control wbc = {
440 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
442 unsigned long nr_dirty = read_page_state(nr_dirty);
443 unsigned long nr_unstable = read_page_state(nr_unstable);
445 wbc.nr_to_write = nr_dirty + nr_unstable +
446 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
447 nr_dirty + nr_unstable;
448 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
449 spin_lock(&inode_lock);
450 sync_sb_inodes(sb, &wbc);
451 spin_unlock(&inode_lock);
455 * Rather lame livelock avoidance.
457 static void set_sb_syncing(int val)
459 struct super_block *sb;
461 sb = sb_entry(super_blocks.prev);
462 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
465 spin_unlock(&sb_lock);
469 * Find a superblock with inodes that need to be synced
471 static struct super_block *get_super_to_sync(void)
473 struct super_block *sb;
476 sb = sb_entry(super_blocks.prev);
477 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
482 spin_unlock(&sb_lock);
483 down_read(&sb->s_umount);
490 spin_unlock(&sb_lock);
497 * sync_inodes() goes through each super block's dirty inode list, writes the
498 * inodes out, waits on the writeout and puts the inodes back on the normal
501 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
502 * part of the sync functions is that the blockdev "superblock" is processed
503 * last. This is because the write_inode() function of a typical fs will
504 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
505 * What we want to do is to perform all that dirtying first, and then write
506 * back all those inode blocks via the blockdev mapping in one sweep. So the
507 * additional (somewhat redundant) sync_blockdev() calls here are to make
508 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
509 * outstanding dirty inodes, the writeback goes block-at-a-time within the
510 * filesystem's write_inode(). This is extremely slow.
512 void sync_inodes(int wait)
514 struct super_block *sb;
517 while ((sb = get_super_to_sync()) != NULL) {
518 sync_inodes_sb(sb, 0);
519 sync_blockdev(sb->s_bdev);
524 while ((sb = get_super_to_sync()) != NULL) {
525 sync_inodes_sb(sb, 1);
526 sync_blockdev(sb->s_bdev);
533 * write_inode_now - write an inode to disk
534 * @inode: inode to write to disk
535 * @sync: whether the write should be synchronous or not
537 * This function commits an inode to disk immediately if it is
538 * dirty. This is primarily needed by knfsd.
541 void write_inode_now(struct inode *inode, int sync)
543 struct writeback_control wbc = {
544 .nr_to_write = LONG_MAX,
545 .sync_mode = WB_SYNC_ALL,
548 if (inode->i_mapping->backing_dev_info->memory_backed)
552 spin_lock(&inode_lock);
553 __writeback_single_inode(inode, &wbc);
554 spin_unlock(&inode_lock);
556 wait_on_inode(inode);
558 EXPORT_SYMBOL(write_inode_now);
561 * sync_inode - write an inode and its pages to disk.
562 * @inode: the inode to sync
563 * @wbc: controls the writeback mode
565 * sync_inode() will write an inode and its pages to disk. It will also
566 * correctly update the inode on its superblock's dirty inode lists and will
567 * update inode->i_state.
569 * The caller must have a ref on the inode.
571 int sync_inode(struct inode *inode, struct writeback_control *wbc)
575 spin_lock(&inode_lock);
576 ret = __writeback_single_inode(inode, wbc);
577 spin_unlock(&inode_lock);
580 EXPORT_SYMBOL(sync_inode);
583 * generic_osync_inode - flush all dirty data for a given inode to disk
584 * @inode: inode to write
585 * @what: what to write and wait upon
587 * This can be called by file_write functions for files which have the
588 * O_SYNC flag set, to flush dirty writes to disk.
590 * @what is a bitmask, specifying which part of the inode's data should be
591 * written and waited upon:
593 * OSYNC_DATA: i_mapping's dirty data
594 * OSYNC_METADATA: the buffers at i_mapping->private_list
595 * OSYNC_INODE: the inode itself
598 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
601 int need_write_inode_now = 0;
604 current->flags |= PF_SYNCWRITE;
605 if (what & OSYNC_DATA)
606 err = filemap_fdatawrite(mapping);
607 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
608 err2 = sync_mapping_buffers(mapping);
612 if (what & OSYNC_DATA) {
613 err2 = filemap_fdatawait(mapping);
617 current->flags &= ~PF_SYNCWRITE;
619 spin_lock(&inode_lock);
620 if ((inode->i_state & I_DIRTY) &&
621 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
622 need_write_inode_now = 1;
623 spin_unlock(&inode_lock);
625 if (need_write_inode_now)
626 write_inode_now(inode, 1);
628 wait_on_inode(inode);
633 EXPORT_SYMBOL(generic_osync_inode);
636 * writeback_acquire: attempt to get exclusive writeback access to a device
637 * @bdi: the device's backing_dev_info structure
639 * It is a waste of resources to have more than one pdflush thread blocked on
640 * a single request queue. Exclusion at the request_queue level is obtained
641 * via a flag in the request_queue's backing_dev_info.state.
643 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
644 * unless they implement their own. Which is somewhat inefficient, as this
645 * may prevent concurrent writeback against multiple devices.
647 int writeback_acquire(struct backing_dev_info *bdi)
649 return !test_and_set_bit(BDI_pdflush, &bdi->state);
653 * writeback_in_progress: determine whether there is writeback in progress
654 * against a backing device.
655 * @bdi: the device's backing_dev_info structure.
657 int writeback_in_progress(struct backing_dev_info *bdi)
659 return test_bit(BDI_pdflush, &bdi->state);
663 * writeback_release: relinquish exclusive writeback access against a device.
664 * @bdi: the device's backing_dev_info structure
666 void writeback_release(struct backing_dev_info *bdi)
668 BUG_ON(!writeback_in_progress(bdi));
669 clear_bit(BDI_pdflush, &bdi->state);