pull master
[hh.org.git] / fs / fs-writeback.c
blobc403b66ec83c6525532c25decbe923c32356649b
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>
25 #include "internal.h"
27 /**
28 * __mark_inode_dirty - internal function
29 * @inode: inode to mark
30 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
31 * Mark an inode as dirty. Callers should use mark_inode_dirty or
32 * mark_inode_dirty_sync.
34 * Put the inode on the super block's dirty list.
36 * CAREFUL! We mark it dirty unconditionally, but move it onto the
37 * dirty list only if it is hashed or if it refers to a blockdev.
38 * If it was not hashed, it will never be added to the dirty list
39 * even if it is later hashed, as it will have been marked dirty already.
41 * In short, make sure you hash any inodes _before_ you start marking
42 * them dirty.
44 * This function *must* be atomic for the I_DIRTY_PAGES case -
45 * set_page_dirty() is called under spinlock in several places.
47 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
48 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
49 * the kernel-internal blockdev inode represents the dirtying time of the
50 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
51 * page->mapping->host, so the page-dirtying time is recorded in the internal
52 * blockdev inode.
54 void __mark_inode_dirty(struct inode *inode, int flags)
56 struct super_block *sb = inode->i_sb;
59 * Don't do this for I_DIRTY_PAGES - that doesn't actually
60 * dirty the inode itself
62 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
63 if (sb->s_op->dirty_inode)
64 sb->s_op->dirty_inode(inode);
68 * make sure that changes are seen by all cpus before we test i_state
69 * -- mikulas
71 smp_mb();
73 /* avoid the locking if we can */
74 if ((inode->i_state & flags) == flags)
75 return;
77 if (unlikely(block_dump)) {
78 struct dentry *dentry = NULL;
79 const char *name = "?";
81 if (!list_empty(&inode->i_dentry)) {
82 dentry = list_entry(inode->i_dentry.next,
83 struct dentry, d_alias);
84 if (dentry && dentry->d_name.name)
85 name = (const char *) dentry->d_name.name;
88 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev"))
89 printk(KERN_DEBUG
90 "%s(%d): dirtied inode %lu (%s) on %s\n",
91 current->comm, current->pid, inode->i_ino,
92 name, inode->i_sb->s_id);
95 spin_lock(&inode_lock);
96 if ((inode->i_state & flags) != flags) {
97 const int was_dirty = inode->i_state & I_DIRTY;
99 inode->i_state |= flags;
102 * If the inode is locked, just update its dirty state.
103 * The unlocker will place the inode on the appropriate
104 * superblock list, based upon its state.
106 if (inode->i_state & I_LOCK)
107 goto out;
110 * Only add valid (hashed) inodes to the superblock's
111 * dirty list. Add blockdev inodes as well.
113 if (!S_ISBLK(inode->i_mode)) {
114 if (hlist_unhashed(&inode->i_hash))
115 goto out;
117 if (inode->i_state & (I_FREEING|I_CLEAR))
118 goto out;
121 * If the inode was already on s_dirty or s_io, don't
122 * reposition it (that would break s_dirty time-ordering).
124 if (!was_dirty) {
125 inode->dirtied_when = jiffies;
126 list_move(&inode->i_list, &sb->s_dirty);
129 out:
130 spin_unlock(&inode_lock);
133 EXPORT_SYMBOL(__mark_inode_dirty);
135 static int write_inode(struct inode *inode, int sync)
137 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
138 return inode->i_sb->s_op->write_inode(inode, sync);
139 return 0;
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
148 * livelocks, etc.
150 * Called under inode_lock.
152 static int
153 __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
155 unsigned dirty;
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;
159 int ret;
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 int err = write_inode(inode, wait);
175 if (ret == 0)
176 ret = err;
179 if (wait) {
180 int err = filemap_fdatawait(mapping);
181 if (ret == 0)
182 ret = err;
185 spin_lock(&inode_lock);
186 inode->i_state &= ~I_LOCK;
187 if (!(inode->i_state & I_FREEING)) {
188 if (!(inode->i_state & I_DIRTY) &&
189 mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
191 * We didn't write back all the pages. nfs_writepages()
192 * sometimes bales out without doing anything. Redirty
193 * the inode. It is still on sb->s_io.
195 if (wbc->for_kupdate) {
197 * For the kupdate function we leave the inode
198 * at the head of sb_dirty so it will get more
199 * writeout as soon as the queue becomes
200 * uncongested.
202 inode->i_state |= I_DIRTY_PAGES;
203 list_move_tail(&inode->i_list, &sb->s_dirty);
204 } else {
206 * Otherwise fully redirty the inode so that
207 * other inodes on this superblock will get some
208 * writeout. Otherwise heavy writing to one
209 * file would indefinitely suspend writeout of
210 * all the other files.
212 inode->i_state |= I_DIRTY_PAGES;
213 inode->dirtied_when = jiffies;
214 list_move(&inode->i_list, &sb->s_dirty);
216 } else if (inode->i_state & I_DIRTY) {
218 * Someone redirtied the inode while were writing back
219 * the pages.
221 list_move(&inode->i_list, &sb->s_dirty);
222 } else if (atomic_read(&inode->i_count)) {
224 * The inode is clean, inuse
226 list_move(&inode->i_list, &inode_in_use);
227 } else {
229 * The inode is clean, unused
231 list_move(&inode->i_list, &inode_unused);
234 wake_up_inode(inode);
235 return ret;
239 * Write out an inode's dirty pages. Called under inode_lock. Either the
240 * caller has ref on the inode (either via __iget or via syscall against an fd)
241 * or the inode has I_WILL_FREE set (via generic_forget_inode)
243 static int
244 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
246 wait_queue_head_t *wqh;
248 if (!atomic_read(&inode->i_count))
249 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
250 else
251 WARN_ON(inode->i_state & I_WILL_FREE);
253 if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) {
254 list_move(&inode->i_list, &inode->i_sb->s_dirty);
255 return 0;
259 * It's a data-integrity sync. We must wait.
261 if (inode->i_state & I_LOCK) {
262 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_LOCK);
264 wqh = bit_waitqueue(&inode->i_state, __I_LOCK);
265 do {
266 spin_unlock(&inode_lock);
267 __wait_on_bit(wqh, &wq, inode_wait,
268 TASK_UNINTERRUPTIBLE);
269 spin_lock(&inode_lock);
270 } while (inode->i_state & I_LOCK);
272 return __sync_single_inode(inode, wbc);
276 * Write out a superblock's list of dirty inodes. A wait will be performed
277 * upon no inodes, all inodes or the final one, depending upon sync_mode.
279 * If older_than_this is non-NULL, then only write out inodes which
280 * had their first dirtying at a time earlier than *older_than_this.
282 * If we're a pdlfush thread, then implement pdflush collision avoidance
283 * against the entire list.
285 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
286 * that it can be located for waiting on in __writeback_single_inode().
288 * Called under inode_lock.
290 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
291 * This function assumes that the blockdev superblock's inodes are backed by
292 * a variety of queues, so all inodes are searched. For other superblocks,
293 * assume that all inodes are backed by the same queue.
295 * FIXME: this linear search could get expensive with many fileystems. But
296 * how to fix? We need to go from an address_space to all inodes which share
297 * a queue with that address_space. (Easy: have a global "dirty superblocks"
298 * list).
300 * The inodes to be written are parked on sb->s_io. They are moved back onto
301 * sb->s_dirty as they are selected for writing. This way, none can be missed
302 * on the writer throttling path, and we get decent balancing between many
303 * throttled threads: we don't want them all piling up on __wait_on_inode.
305 static void
306 sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
308 const unsigned long start = jiffies; /* livelock avoidance */
310 if (!wbc->for_kupdate || list_empty(&sb->s_io))
311 list_splice_init(&sb->s_dirty, &sb->s_io);
313 while (!list_empty(&sb->s_io)) {
314 struct inode *inode = list_entry(sb->s_io.prev,
315 struct inode, i_list);
316 struct address_space *mapping = inode->i_mapping;
317 struct backing_dev_info *bdi = mapping->backing_dev_info;
318 long pages_skipped;
320 if (!bdi_cap_writeback_dirty(bdi)) {
321 list_move(&inode->i_list, &sb->s_dirty);
322 if (sb_is_blkdev_sb(sb)) {
324 * Dirty memory-backed blockdev: the ramdisk
325 * driver does this. Skip just this inode
327 continue;
330 * Dirty memory-backed inode against a filesystem other
331 * than the kernel-internal bdev filesystem. Skip the
332 * entire superblock.
334 break;
337 if (wbc->nonblocking && bdi_write_congested(bdi)) {
338 wbc->encountered_congestion = 1;
339 if (!sb_is_blkdev_sb(sb))
340 break; /* Skip a congested fs */
341 list_move(&inode->i_list, &sb->s_dirty);
342 continue; /* Skip a congested blockdev */
345 if (wbc->bdi && bdi != wbc->bdi) {
346 if (!sb_is_blkdev_sb(sb))
347 break; /* fs has the wrong queue */
348 list_move(&inode->i_list, &sb->s_dirty);
349 continue; /* blockdev has wrong queue */
352 /* Was this inode dirtied after sync_sb_inodes was called? */
353 if (time_after(inode->dirtied_when, start))
354 break;
356 /* Was this inode dirtied too recently? */
357 if (wbc->older_than_this && time_after(inode->dirtied_when,
358 *wbc->older_than_this))
359 break;
361 /* Is another pdflush already flushing this queue? */
362 if (current_is_pdflush() && !writeback_acquire(bdi))
363 break;
365 BUG_ON(inode->i_state & I_FREEING);
366 __iget(inode);
367 pages_skipped = wbc->pages_skipped;
368 __writeback_single_inode(inode, wbc);
369 if (wbc->sync_mode == WB_SYNC_HOLD) {
370 inode->dirtied_when = jiffies;
371 list_move(&inode->i_list, &sb->s_dirty);
373 if (current_is_pdflush())
374 writeback_release(bdi);
375 if (wbc->pages_skipped != pages_skipped) {
377 * writeback is not making progress due to locked
378 * buffers. Skip this inode for now.
380 list_move(&inode->i_list, &sb->s_dirty);
382 spin_unlock(&inode_lock);
383 iput(inode);
384 cond_resched();
385 spin_lock(&inode_lock);
386 if (wbc->nr_to_write <= 0)
387 break;
389 return; /* Leave any unwritten inodes on s_io */
393 * Start writeback of dirty pagecache data against all unlocked inodes.
395 * Note:
396 * We don't need to grab a reference to superblock here. If it has non-empty
397 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
398 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
399 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
400 * inode from superblock lists we are OK.
402 * If `older_than_this' is non-zero then only flush inodes which have a
403 * flushtime older than *older_than_this.
405 * If `bdi' is non-zero then we will scan the first inode against each
406 * superblock until we find the matching ones. One group will be the dirty
407 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
408 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
409 * super-efficient but we're about to do a ton of I/O...
411 void
412 writeback_inodes(struct writeback_control *wbc)
414 struct super_block *sb;
416 might_sleep();
417 spin_lock(&sb_lock);
418 restart:
419 sb = sb_entry(super_blocks.prev);
420 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
421 if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) {
422 /* we're making our own get_super here */
423 sb->s_count++;
424 spin_unlock(&sb_lock);
426 * If we can't get the readlock, there's no sense in
427 * waiting around, most of the time the FS is going to
428 * be unmounted by the time it is released.
430 if (down_read_trylock(&sb->s_umount)) {
431 if (sb->s_root) {
432 spin_lock(&inode_lock);
433 sync_sb_inodes(sb, wbc);
434 spin_unlock(&inode_lock);
436 up_read(&sb->s_umount);
438 spin_lock(&sb_lock);
439 if (__put_super_and_need_restart(sb))
440 goto restart;
442 if (wbc->nr_to_write <= 0)
443 break;
445 spin_unlock(&sb_lock);
449 * writeback and wait upon the filesystem's dirty inodes. The caller will
450 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
451 * used to park the written inodes on sb->s_dirty for the wait pass.
453 * A finite limit is set on the number of pages which will be written.
454 * To prevent infinite livelock of sys_sync().
456 * We add in the number of potentially dirty inodes, because each inode write
457 * can dirty pagecache in the underlying blockdev.
459 void sync_inodes_sb(struct super_block *sb, int wait)
461 struct writeback_control wbc = {
462 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD,
463 .range_start = 0,
464 .range_end = LLONG_MAX,
466 unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
467 unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
469 wbc.nr_to_write = nr_dirty + nr_unstable +
470 (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
471 nr_dirty + nr_unstable;
472 wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */
473 spin_lock(&inode_lock);
474 sync_sb_inodes(sb, &wbc);
475 spin_unlock(&inode_lock);
479 * Rather lame livelock avoidance.
481 static void set_sb_syncing(int val)
483 struct super_block *sb;
484 spin_lock(&sb_lock);
485 sb = sb_entry(super_blocks.prev);
486 for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
487 sb->s_syncing = val;
489 spin_unlock(&sb_lock);
493 * sync_inodes - writes all inodes to disk
494 * @wait: wait for completion
496 * sync_inodes() goes through each super block's dirty inode list, writes the
497 * inodes out, waits on the writeout and puts the inodes back on the normal
498 * list.
500 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
501 * part of the sync functions is that the blockdev "superblock" is processed
502 * last. This is because the write_inode() function of a typical fs will
503 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
504 * What we want to do is to perform all that dirtying first, and then write
505 * back all those inode blocks via the blockdev mapping in one sweep. So the
506 * additional (somewhat redundant) sync_blockdev() calls here are to make
507 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
508 * outstanding dirty inodes, the writeback goes block-at-a-time within the
509 * filesystem's write_inode(). This is extremely slow.
511 static void __sync_inodes(int wait)
513 struct super_block *sb;
515 spin_lock(&sb_lock);
516 restart:
517 list_for_each_entry(sb, &super_blocks, s_list) {
518 if (sb->s_syncing)
519 continue;
520 sb->s_syncing = 1;
521 sb->s_count++;
522 spin_unlock(&sb_lock);
523 down_read(&sb->s_umount);
524 if (sb->s_root) {
525 sync_inodes_sb(sb, wait);
526 sync_blockdev(sb->s_bdev);
528 up_read(&sb->s_umount);
529 spin_lock(&sb_lock);
530 if (__put_super_and_need_restart(sb))
531 goto restart;
533 spin_unlock(&sb_lock);
536 void sync_inodes(int wait)
538 set_sb_syncing(0);
539 __sync_inodes(0);
541 if (wait) {
542 set_sb_syncing(0);
543 __sync_inodes(1);
548 * write_inode_now - write an inode to disk
549 * @inode: inode to write to disk
550 * @sync: whether the write should be synchronous or not
552 * This function commits an inode to disk immediately if it is dirty. This is
553 * primarily needed by knfsd.
555 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
557 int write_inode_now(struct inode *inode, int sync)
559 int ret;
560 struct writeback_control wbc = {
561 .nr_to_write = LONG_MAX,
562 .sync_mode = WB_SYNC_ALL,
563 .range_start = 0,
564 .range_end = LLONG_MAX,
567 if (!mapping_cap_writeback_dirty(inode->i_mapping))
568 wbc.nr_to_write = 0;
570 might_sleep();
571 spin_lock(&inode_lock);
572 ret = __writeback_single_inode(inode, &wbc);
573 spin_unlock(&inode_lock);
574 if (sync)
575 wait_on_inode(inode);
576 return ret;
578 EXPORT_SYMBOL(write_inode_now);
581 * sync_inode - write an inode and its pages to disk.
582 * @inode: the inode to sync
583 * @wbc: controls the writeback mode
585 * sync_inode() will write an inode and its pages to disk. It will also
586 * correctly update the inode on its superblock's dirty inode lists and will
587 * update inode->i_state.
589 * The caller must have a ref on the inode.
591 int sync_inode(struct inode *inode, struct writeback_control *wbc)
593 int ret;
595 spin_lock(&inode_lock);
596 ret = __writeback_single_inode(inode, wbc);
597 spin_unlock(&inode_lock);
598 return ret;
600 EXPORT_SYMBOL(sync_inode);
603 * generic_osync_inode - flush all dirty data for a given inode to disk
604 * @inode: inode to write
605 * @mapping: the address_space that should be flushed
606 * @what: what to write and wait upon
608 * This can be called by file_write functions for files which have the
609 * O_SYNC flag set, to flush dirty writes to disk.
611 * @what is a bitmask, specifying which part of the inode's data should be
612 * written and waited upon.
614 * OSYNC_DATA: i_mapping's dirty data
615 * OSYNC_METADATA: the buffers at i_mapping->private_list
616 * OSYNC_INODE: the inode itself
619 int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what)
621 int err = 0;
622 int need_write_inode_now = 0;
623 int err2;
625 if (what & OSYNC_DATA)
626 err = filemap_fdatawrite(mapping);
627 if (what & (OSYNC_METADATA|OSYNC_DATA)) {
628 err2 = sync_mapping_buffers(mapping);
629 if (!err)
630 err = err2;
632 if (what & OSYNC_DATA) {
633 err2 = filemap_fdatawait(mapping);
634 if (!err)
635 err = err2;
638 spin_lock(&inode_lock);
639 if ((inode->i_state & I_DIRTY) &&
640 ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC)))
641 need_write_inode_now = 1;
642 spin_unlock(&inode_lock);
644 if (need_write_inode_now) {
645 err2 = write_inode_now(inode, 1);
646 if (!err)
647 err = err2;
649 else
650 wait_on_inode(inode);
652 return err;
655 EXPORT_SYMBOL(generic_osync_inode);
658 * writeback_acquire: attempt to get exclusive writeback access to a device
659 * @bdi: the device's backing_dev_info structure
661 * It is a waste of resources to have more than one pdflush thread blocked on
662 * a single request queue. Exclusion at the request_queue level is obtained
663 * via a flag in the request_queue's backing_dev_info.state.
665 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
666 * unless they implement their own. Which is somewhat inefficient, as this
667 * may prevent concurrent writeback against multiple devices.
669 int writeback_acquire(struct backing_dev_info *bdi)
671 return !test_and_set_bit(BDI_pdflush, &bdi->state);
675 * writeback_in_progress: determine whether there is writeback in progress
676 * @bdi: the device's backing_dev_info structure.
678 * Determine whether there is writeback in progress against a backing device.
680 int writeback_in_progress(struct backing_dev_info *bdi)
682 return test_bit(BDI_pdflush, &bdi->state);
686 * writeback_release: relinquish exclusive writeback access against a device.
687 * @bdi: the device's backing_dev_info structure
689 void writeback_release(struct backing_dev_info *bdi)
691 BUG_ON(!writeback_in_progress(bdi));
692 clear_bit(BDI_pdflush, &bdi->state);