initial commit with v2.6.9
[linux-2.6.9-moxart.git] / fs / fs-writeback.c
blobda522d511f2d149e370d15239a016eeed26fbda6
1 /*
2 * fs/fs-writeback.c
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>
19 #include <linux/fs.h>
20 #include <linux/mm.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;
28 /**
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
43 * them dirty.
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
53 * blockdev inode.
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
70 * -- mikulas
72 smp_mb();
74 /* avoid the locking if we can */
75 if ((inode->i_state & flags) == flags)
76 return;
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"))
90 printk(KERN_DEBUG
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)
108 goto out;
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))
116 goto out;
118 if (inode->i_state & (I_FREEING|I_CLEAR))
119 goto out;
122 * If the inode was already on s_dirty or s_io, don't
123 * reposition it (that would break s_dirty time-ordering).
125 if (!was_dirty) {
126 inode->dirtied_when = jiffies;
127 list_move(&inode->i_list, &sb->s_dirty);
130 out:
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);
140 return 0;
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
149 * livelocks, etc.
151 * Called under inode_lock.
153 static int
154 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
156 unsigned dirty;
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;
160 int ret;
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);
176 if (ret == 0)
177 ret = err;
180 if (wait) {
181 int err = filemap_fdatawait(mapping);
182 if (ret == 0)
183 ret = err;
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
201 * uncongested.
203 inode->i_state |= I_DIRTY_PAGES;
204 list_move_tail(&inode->i_list, &sb->s_dirty);
205 } else {
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
220 * the pages.
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);
228 } else {
230 * The inode is clean, unused
232 list_move(&inode->i_list, &inode_unused);
233 inodes_stat.nr_unused++;
236 wake_up_inode(inode);
237 return ret;
241 * Write out an inode's dirty pages. Called under inode_lock.
243 static int
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);
249 return 0;
253 * It's a data-integrity sync. We must wait.
255 while (inode->i_state & I_LOCK) {
256 __iget(inode);
257 spin_unlock(&inode_lock);
258 __wait_on_inode(inode);
259 iput(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"
288 * list).
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.
295 static void
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;
308 long pages_skipped;
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
317 continue;
320 * Dirty memory-backed inode against a filesystem other
321 * than the kernel-internal bdev filesystem. Skip the
322 * entire superblock.
324 break;
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))
344 break;
346 /* Was this inode dirtied too recently? */
347 if (wbc->older_than_this && time_after(inode->dirtied_when,
348 *wbc->older_than_this))
349 break;
351 /* Is another pdflush already flushing this queue? */
352 if (current_is_pdflush() && !writeback_acquire(bdi))
353 break;
355 BUG_ON(inode->i_state & I_FREEING);
356 __iget(inode);
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);
373 iput(inode);
374 spin_lock(&inode_lock);
375 if (wbc->nr_to_write <= 0)
376 break;
378 return; /* Leave any unwritten inodes on s_io */
382 * Start writeback of dirty pagecache data against all unlocked inodes.
384 * Note:
385 * We don't need to grab a reference to superblock here. If it has non-empty
386 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
387 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
388 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
389 * inode from superblock lists we are OK.
391 * If `older_than_this' is non-zero then only flush inodes which have a
392 * flushtime older than *older_than_this.
394 * If `bdi' is non-zero then we will scan the first inode against each
395 * superblock until we find the matching ones. One group will be the dirty
396 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
397 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
398 * super-efficient but we're about to do a ton of I/O...
400 void
401 writeback_inodes(struct writeback_control *wbc)
403 struct super_block *sb;
405 might_sleep();
406 spin_lock(&sb_lock);
407 restart:
408 sb = sb_entry(super_blocks.prev);
409 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
410 if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) {
411 /* we're making our own get_super here */
412 sb->s_count++;
413 spin_unlock(&sb_lock);
415 * If we can't get the readlock, there's no sense in
416 * waiting around, most of the time the FS is going to
417 * be unmounted by the time it is released.
419 if (down_read_trylock(&sb->s_umount)) {
420 if (sb->s_root) {
421 spin_lock(&inode_lock);
422 sync_sb_inodes(sb, wbc);
423 spin_unlock(&inode_lock);
425 up_read(&sb->s_umount);
427 spin_lock(&sb_lock);
428 if (__put_super_and_need_restart(sb))
429 goto restart;
431 if (wbc->nr_to_write <= 0)
432 break;
434 spin_unlock(&sb_lock);
438 * writeback and wait upon the filesystem's dirty inodes. The caller will
439 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
440 * used to park the written inodes on sb->s_dirty for the wait pass.
442 * A finite limit is set on the number of pages which will be written.
443 * To prevent infinite livelock of sys_sync().
445 * We add in the number of potentially dirty inodes, because each inode write
446 * can dirty pagecache in the underlying blockdev.
448 void sync_inodes_sb(struct super_block *sb, int wait)
450 struct writeback_control wbc = {
451 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
453 unsigned long nr_dirty = read_page_state(nr_dirty);
454 unsigned long nr_unstable = read_page_state(nr_unstable);
456 wbc.nr_to_write = nr_dirty + nr_unstable +
457 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
458 nr_dirty + nr_unstable;
459 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
460 spin_lock(&inode_lock);
461 sync_sb_inodes(sb, &wbc);
462 spin_unlock(&inode_lock);
466 * Rather lame livelock avoidance.
468 static void set_sb_syncing(int val)
470 struct super_block *sb;
471 spin_lock(&sb_lock);
472 sb = sb_entry(super_blocks.prev);
473 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
474 sb->s_syncing = val;
476 spin_unlock(&sb_lock);
480 * Find a superblock with inodes that need to be synced
482 static struct super_block *get_super_to_sync(void)
484 struct super_block *sb;
485 restart:
486 spin_lock(&sb_lock);
487 sb = sb_entry(super_blocks.prev);
488 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
489 if (sb->s_syncing)
490 continue;
491 sb->s_syncing = 1;
492 sb->s_count++;
493 spin_unlock(&sb_lock);
494 down_read(&sb->s_umount);
495 if (!sb->s_root) {
496 drop_super(sb);
497 goto restart;
499 return sb;
501 spin_unlock(&sb_lock);
502 return NULL;
506 * sync_inodes
508 * sync_inodes() goes through each super block's dirty inode list, writes the
509 * inodes out, waits on the writeout and puts the inodes back on the normal
510 * list.
512 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
513 * part of the sync functions is that the blockdev "superblock" is processed
514 * last. This is because the write_inode() function of a typical fs will
515 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
516 * What we want to do is to perform all that dirtying first, and then write
517 * back all those inode blocks via the blockdev mapping in one sweep. So the
518 * additional (somewhat redundant) sync_blockdev() calls here are to make
519 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
520 * outstanding dirty inodes, the writeback goes block-at-a-time within the
521 * filesystem's write_inode(). This is extremely slow.
523 void sync_inodes(int wait)
525 struct super_block *sb;
527 set_sb_syncing(0);
528 while ((sb = get_super_to_sync()) != NULL) {
529 sync_inodes_sb(sb, 0);
530 sync_blockdev(sb->s_bdev);
531 drop_super(sb);
533 if (wait) {
534 set_sb_syncing(0);
535 while ((sb = get_super_to_sync()) != NULL) {
536 sync_inodes_sb(sb, 1);
537 sync_blockdev(sb->s_bdev);
538 drop_super(sb);
544 * write_inode_now - write an inode to disk
545 * @inode: inode to write to disk
546 * @sync: whether the write should be synchronous or not
548 * This function commits an inode to disk immediately if it is
549 * dirty. This is primarily needed by knfsd.
552 void write_inode_now(struct inode *inode, int sync)
554 struct writeback_control wbc = {
555 .nr_to_write = LONG_MAX,
556 .sync_mode = WB_SYNC_ALL,
559 if (inode->i_mapping->backing_dev_info->memory_backed)
560 return;
562 might_sleep();
563 spin_lock(&inode_lock);
564 __writeback_single_inode(inode, &wbc);
565 spin_unlock(&inode_lock);
566 if (sync)
567 wait_on_inode(inode);
569 EXPORT_SYMBOL(write_inode_now);
572 * sync_inode - write an inode and its pages to disk.
573 * @inode: the inode to sync
574 * @wbc: controls the writeback mode
576 * sync_inode() will write an inode and its pages to disk. It will also
577 * correctly update the inode on its superblock's dirty inode lists and will
578 * update inode->i_state.
580 * The caller must have a ref on the inode.
582 int sync_inode(struct inode *inode, struct writeback_control *wbc)
584 int ret;
586 spin_lock(&inode_lock);
587 ret = __writeback_single_inode(inode, wbc);
588 spin_unlock(&inode_lock);
589 return ret;
591 EXPORT_SYMBOL(sync_inode);
594 * generic_osync_inode - flush all dirty data for a given inode to disk
595 * @inode: inode to write
596 * @what: what to write and wait upon
598 * This can be called by file_write functions for files which have the
599 * O_SYNC flag set, to flush dirty writes to disk.
601 * @what is a bitmask, specifying which part of the inode's data should be
602 * written and waited upon:
604 * OSYNC_DATA: i_mapping's dirty data
605 * OSYNC_METADATA: the buffers at i_mapping->private_list
606 * OSYNC_INODE: the inode itself
609 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
611 int err = 0;
612 int need_write_inode_now = 0;
613 int err2;
615 current->flags |= PF_SYNCWRITE;
616 if (what & OSYNC_DATA)
617 err = filemap_fdatawrite(mapping);
618 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
619 err2 = sync_mapping_buffers(mapping);
620 if (!err)
621 err = err2;
623 if (what & OSYNC_DATA) {
624 err2 = filemap_fdatawait(mapping);
625 if (!err)
626 err = err2;
628 current->flags &= ~PF_SYNCWRITE;
630 spin_lock(&inode_lock);
631 if ((inode->i_state & I_DIRTY) &&
632 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
633 need_write_inode_now = 1;
634 spin_unlock(&inode_lock);
636 if (need_write_inode_now)
637 write_inode_now(inode, 1);
638 else
639 wait_on_inode(inode);
641 return err;
644 EXPORT_SYMBOL(generic_osync_inode);
647 * writeback_acquire: attempt to get exclusive writeback access to a device
648 * @bdi: the device's backing_dev_info structure
650 * It is a waste of resources to have more than one pdflush thread blocked on
651 * a single request queue. Exclusion at the request_queue level is obtained
652 * via a flag in the request_queue's backing_dev_info.state.
654 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
655 * unless they implement their own. Which is somewhat inefficient, as this
656 * may prevent concurrent writeback against multiple devices.
658 int writeback_acquire(struct backing_dev_info *bdi)
660 return !test_and_set_bit(BDI_pdflush, &bdi->state);
664 * writeback_in_progress: determine whether there is writeback in progress
665 * against a backing device.
666 * @bdi: the device's backing_dev_info structure.
668 int writeback_in_progress(struct backing_dev_info *bdi)
670 return test_bit(BDI_pdflush, &bdi->state);
674 * writeback_release: relinquish exclusive writeback access against a device.
675 * @bdi: the device's backing_dev_info structure
677 void writeback_release(struct backing_dev_info *bdi)
679 BUG_ON(!writeback_in_progress(bdi));
680 clear_bit(BDI_pdflush, &bdi->state);