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 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 * Write a single inode's dirty pages and inode data out to disk.
145 * If `wait' is set, wait on the writeout.
147 * The whole writeout design is quite complex and fragile. We want to avoid
148 * starvation of particular inodes when others are being redirtied, prevent
151 * Called under inode_lock.
154 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
157 struct address_space *mapping = inode->i_mapping;
158 struct super_block *sb = inode->i_sb;
159 int wait = wbc->sync_mode == WB_SYNC_ALL;
162 BUG_ON(inode->i_state & I_LOCK);
164 /* Set I_LOCK, reset I_DIRTY */
165 dirty = inode->i_state & I_DIRTY;
166 inode->i_state |= I_LOCK;
167 inode->i_state &= ~I_DIRTY;
169 spin_unlock(&inode_lock);
171 ret = do_writepages(mapping, wbc);
173 /* Don't write the inode if only I_DIRTY_PAGES was set */
174 if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
175 int err = write_inode(inode, wait);
181 int err = filemap_fdatawait(mapping);
186 spin_lock(&inode_lock);
187 inode->i_state &= ~I_LOCK;
188 if (!(inode->i_state & I_FREEING)) {
189 if (!(inode->i_state & I_DIRTY) &&
190 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
192 * We didn't write back all the pages. nfs_writepages()
193 * sometimes bales out without doing anything. Redirty
194 * the inode. It is still on sb->s_io.
196 if (wbc->for_kupdate) {
198 * For the kupdate function we leave the inode
199 * at the head of sb_dirty so it will get more
200 * writeout as soon as the queue becomes
203 inode->i_state |= I_DIRTY_PAGES;
204 list_move_tail(&inode->i_list, &sb->s_dirty);
207 * Otherwise fully redirty the inode so that
208 * other inodes on this superblock will get some
209 * writeout. Otherwise heavy writing to one
210 * file would indefinitely suspend writeout of
211 * all the other files.
213 inode->i_state |= I_DIRTY_PAGES;
214 inode->dirtied_when = jiffies;
215 list_move(&inode->i_list, &sb->s_dirty);
217 } else if (inode->i_state & I_DIRTY) {
219 * Someone redirtied the inode while were writing back
222 list_move(&inode->i_list, &sb->s_dirty);
223 } else if (atomic_read(&inode->i_count)) {
225 * The inode is clean, inuse
227 list_move(&inode->i_list, &inode_in_use);
230 * The inode is clean, unused
232 list_move(&inode->i_list, &inode_unused);
233 inodes_stat.nr_unused++;
236 wake_up_inode(inode);
241 * Write out an inode's dirty pages. Called under inode_lock.
244 __writeback_single_inode(struct inode *inode,
245 struct writeback_control *wbc)
247 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) {
248 list_move(&inode->i_list, &inode->i_sb->s_dirty);
253 * It's a data-integrity sync. We must wait.
255 while (inode->i_state & I_LOCK) {
257 spin_unlock(&inode_lock);
258 __wait_on_inode(inode);
260 spin_lock(&inode_lock);
262 return __sync_single_inode(inode, wbc);
266 * Write out a superblock's list of dirty inodes. A wait will be performed
267 * upon no inodes, all inodes or the final one, depending upon sync_mode.
269 * If older_than_this is non-NULL, then only write out inodes which
270 * had their first dirtying at a time earlier than *older_than_this.
272 * If we're a pdlfush thread, then implement pdflush collision avoidance
273 * against the entire list.
275 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
276 * that it can be located for waiting on in __writeback_single_inode().
278 * Called under inode_lock.
280 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
281 * This function assumes that the blockdev superblock's inodes are backed by
282 * a variety of queues, so all inodes are searched. For other superblocks,
283 * assume that all inodes are backed by the same queue.
285 * FIXME: this linear search could get expensive with many fileystems. But
286 * how to fix? We need to go from an address_space to all inodes which share
287 * a queue with that address_space. (Easy: have a global "dirty superblocks"
290 * The inodes to be written are parked on sb->s_io. They are moved back onto
291 * sb->s_dirty as they are selected for writing. This way, none can be missed
292 * on the writer throttling path, and we get decent balancing between many
293 * throttled threads: we don't want them all piling up on __wait_on_inode.
296 sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
298 const unsigned long start = jiffies; /* livelock avoidance */
300 if (!wbc->for_kupdate || list_empty(&sb->s_io))
301 list_splice_init(&sb->s_dirty, &sb->s_io);
303 while (!list_empty(&sb->s_io)) {
304 struct inode *inode = list_entry(sb->s_io.prev,
305 struct inode, i_list);
306 struct address_space *mapping = inode->i_mapping;
307 struct backing_dev_info *bdi = mapping->backing_dev_info;
310 if (bdi->memory_backed) {
311 list_move(&inode->i_list, &sb->s_dirty);
312 if (sb == blockdev_superblock) {
314 * Dirty memory-backed blockdev: the ramdisk
315 * driver does this. Skip just this inode
320 * Dirty memory-backed inode against a filesystem other
321 * than the kernel-internal bdev filesystem. Skip the
327 if (wbc->nonblocking && bdi_write_congested(bdi)) {
328 wbc->encountered_congestion = 1;
329 if (sb != blockdev_superblock)
330 break; /* Skip a congested fs */
331 list_move(&inode->i_list, &sb->s_dirty);
332 continue; /* Skip a congested blockdev */
335 if (wbc->bdi && bdi != wbc->bdi) {
336 if (sb != blockdev_superblock)
337 break; /* fs has the wrong queue */
338 list_move(&inode->i_list, &sb->s_dirty);
339 continue; /* blockdev has wrong queue */
342 /* Was this inode dirtied after sync_sb_inodes was called? */
343 if (time_after(inode->dirtied_when, start))
346 /* Was this inode dirtied too recently? */
347 if (wbc->older_than_this && time_after(inode->dirtied_when,
348 *wbc->older_than_this))
351 /* Is another pdflush already flushing this queue? */
352 if (current_is_pdflush() && !writeback_acquire(bdi))
355 BUG_ON(inode->i_state & I_FREEING);
357 pages_skipped = wbc->pages_skipped;
358 __writeback_single_inode(inode, wbc);
359 if (wbc->sync_mode == WB_SYNC_HOLD) {
360 inode->dirtied_when = jiffies;
361 list_move(&inode->i_list, &sb->s_dirty);
363 if (current_is_pdflush())
364 writeback_release(bdi);
365 if (wbc->pages_skipped != pages_skipped) {
367 * writeback is not making progress due to locked
368 * buffers. Skip this inode for now.
370 list_move(&inode->i_list, &sb->s_dirty);
372 spin_unlock(&inode_lock);
375 spin_lock(&inode_lock);
376 if (wbc->nr_to_write <= 0)
379 return; /* Leave any unwritten inodes on s_io */
383 * Start writeback of dirty pagecache data against all unlocked inodes.
386 * We don't need to grab a reference to superblock here. If it has non-empty
387 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
388 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
389 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
390 * inode from superblock lists we are OK.
392 * If `older_than_this' is non-zero then only flush inodes which have a
393 * flushtime older than *older_than_this.
395 * If `bdi' is non-zero then we will scan the first inode against each
396 * superblock until we find the matching ones. One group will be the dirty
397 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
398 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
399 * super-efficient but we're about to do a ton of I/O...
402 writeback_inodes(struct writeback_control *wbc)
404 struct super_block *sb;
409 sb = sb_entry(super_blocks.prev);
410 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
411 if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) {
412 /* we're making our own get_super here */
414 spin_unlock(&sb_lock);
416 * If we can't get the readlock, there's no sense in
417 * waiting around, most of the time the FS is going to
418 * be unmounted by the time it is released.
420 if (down_read_trylock(&sb->s_umount)) {
422 spin_lock(&inode_lock);
423 sync_sb_inodes(sb, wbc);
424 spin_unlock(&inode_lock);
426 up_read(&sb->s_umount);
429 if (__put_super_and_need_restart(sb))
432 if (wbc->nr_to_write <= 0)
435 spin_unlock(&sb_lock);
439 * writeback and wait upon the filesystem's dirty inodes. The caller will
440 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
441 * used to park the written inodes on sb->s_dirty for the wait pass.
443 * A finite limit is set on the number of pages which will be written.
444 * To prevent infinite livelock of sys_sync().
446 * We add in the number of potentially dirty inodes, because each inode write
447 * can dirty pagecache in the underlying blockdev.
449 void sync_inodes_sb(struct super_block *sb, int wait)
451 struct writeback_control wbc = {
452 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
454 unsigned long nr_dirty = read_page_state(nr_dirty);
455 unsigned long nr_unstable = read_page_state(nr_unstable);
457 wbc.nr_to_write = nr_dirty + nr_unstable +
458 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
459 nr_dirty + nr_unstable;
460 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
461 spin_lock(&inode_lock);
462 sync_sb_inodes(sb, &wbc);
463 spin_unlock(&inode_lock);
467 * Rather lame livelock avoidance.
469 static void set_sb_syncing(int val)
471 struct super_block *sb;
473 sb = sb_entry(super_blocks.prev);
474 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
477 spin_unlock(&sb_lock);
481 * Find a superblock with inodes that need to be synced
483 static struct super_block *get_super_to_sync(void)
485 struct super_block *sb;
488 sb = sb_entry(super_blocks.prev);
489 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
494 spin_unlock(&sb_lock);
495 down_read(&sb->s_umount);
502 spin_unlock(&sb_lock);
509 * sync_inodes() goes through each super block's dirty inode list, writes the
510 * inodes out, waits on the writeout and puts the inodes back on the normal
513 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
514 * part of the sync functions is that the blockdev "superblock" is processed
515 * last. This is because the write_inode() function of a typical fs will
516 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
517 * What we want to do is to perform all that dirtying first, and then write
518 * back all those inode blocks via the blockdev mapping in one sweep. So the
519 * additional (somewhat redundant) sync_blockdev() calls here are to make
520 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
521 * outstanding dirty inodes, the writeback goes block-at-a-time within the
522 * filesystem's write_inode(). This is extremely slow.
524 void sync_inodes(int wait)
526 struct super_block *sb;
529 while ((sb = get_super_to_sync()) != NULL) {
530 sync_inodes_sb(sb, 0);
531 sync_blockdev(sb->s_bdev);
536 while ((sb = get_super_to_sync()) != NULL) {
537 sync_inodes_sb(sb, 1);
538 sync_blockdev(sb->s_bdev);
545 * write_inode_now - write an inode to disk
546 * @inode: inode to write to disk
547 * @sync: whether the write should be synchronous or not
549 * This function commits an inode to disk immediately if it is
550 * dirty. This is primarily needed by knfsd.
553 void write_inode_now(struct inode *inode, int sync)
555 struct writeback_control wbc = {
556 .nr_to_write = LONG_MAX,
557 .sync_mode = WB_SYNC_ALL,
560 if (inode->i_mapping->backing_dev_info->memory_backed)
564 spin_lock(&inode_lock);
565 __writeback_single_inode(inode, &wbc);
566 spin_unlock(&inode_lock);
568 wait_on_inode(inode);
570 EXPORT_SYMBOL(write_inode_now);
573 * sync_inode - write an inode and its pages to disk.
574 * @inode: the inode to sync
575 * @wbc: controls the writeback mode
577 * sync_inode() will write an inode and its pages to disk. It will also
578 * correctly update the inode on its superblock's dirty inode lists and will
579 * update inode->i_state.
581 * The caller must have a ref on the inode.
583 int sync_inode(struct inode *inode, struct writeback_control *wbc)
588 spin_lock(&inode_lock);
589 ret = __writeback_single_inode(inode, wbc);
590 spin_unlock(&inode_lock);
593 EXPORT_SYMBOL(sync_inode);
596 * generic_osync_inode - flush all dirty data for a given inode to disk
597 * @inode: inode to write
598 * @what: what to write and wait upon
600 * This can be called by file_write functions for files which have the
601 * O_SYNC flag set, to flush dirty writes to disk.
603 * @what is a bitmask, specifying which part of the inode's data should be
604 * written and waited upon:
606 * OSYNC_DATA: i_mapping's dirty data
607 * OSYNC_METADATA: the buffers at i_mapping->private_list
608 * OSYNC_INODE: the inode itself
611 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
614 int need_write_inode_now = 0;
617 current->flags |= PF_SYNCWRITE;
618 if (what & OSYNC_DATA)
619 err = filemap_fdatawrite(mapping);
620 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
621 err2 = sync_mapping_buffers(mapping);
625 if (what & OSYNC_DATA) {
626 err2 = filemap_fdatawait(mapping);
630 current->flags &= ~PF_SYNCWRITE;
632 spin_lock(&inode_lock);
633 if ((inode->i_state & I_DIRTY) &&
634 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
635 need_write_inode_now = 1;
636 spin_unlock(&inode_lock);
638 if (need_write_inode_now)
639 write_inode_now(inode, 1);
641 wait_on_inode(inode);
646 EXPORT_SYMBOL(generic_osync_inode);
649 * writeback_acquire: attempt to get exclusive writeback access to a device
650 * @bdi: the device's backing_dev_info structure
652 * It is a waste of resources to have more than one pdflush thread blocked on
653 * a single request queue. Exclusion at the request_queue level is obtained
654 * via a flag in the request_queue's backing_dev_info.state.
656 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
657 * unless they implement their own. Which is somewhat inefficient, as this
658 * may prevent concurrent writeback against multiple devices.
660 int writeback_acquire(struct backing_dev_info *bdi)
662 return !test_and_set_bit(BDI_pdflush, &bdi->state);
666 * writeback_in_progress: determine whether there is writeback in progress
667 * against a backing device.
668 * @bdi: the device's backing_dev_info structure.
670 int writeback_in_progress(struct backing_dev_info *bdi)
672 return test_bit(BDI_pdflush, &bdi->state);
676 * writeback_release: relinquish exclusive writeback access against a device.
677 * @bdi: the device's backing_dev_info structure
679 void writeback_release(struct backing_dev_info *bdi)
681 BUG_ON(!writeback_in_progress(bdi));
682 clear_bit(BDI_pdflush, &bdi->state);