]>
git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - 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 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/spinlock.h>
19 #include <linux/sched.h>
22 #include <linux/writeback.h>
23 #include <linux/blkdev.h>
24 #include <linux/backing-dev.h>
25 #include <linux/buffer_head.h>
30 * writeback_acquire - attempt to get exclusive writeback access to a device
31 * @bdi: the device's backing_dev_info structure
33 * It is a waste of resources to have more than one pdflush thread blocked on
34 * a single request queue. Exclusion at the request_queue level is obtained
35 * via a flag in the request_queue's backing_dev_info.state.
37 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
38 * unless they implement their own. Which is somewhat inefficient, as this
39 * may prevent concurrent writeback against multiple devices.
41 static int writeback_acquire(struct backing_dev_info
*bdi
)
43 return !test_and_set_bit(BDI_pdflush
, &bdi
->state
);
47 * writeback_in_progress - determine whether there is writeback in progress
48 * @bdi: the device's backing_dev_info structure.
50 * Determine whether there is writeback in progress against a backing device.
52 int writeback_in_progress(struct backing_dev_info
*bdi
)
54 return test_bit(BDI_pdflush
, &bdi
->state
);
58 * writeback_release - relinquish exclusive writeback access against a device.
59 * @bdi: the device's backing_dev_info structure
61 static void writeback_release(struct backing_dev_info
*bdi
)
63 BUG_ON(!writeback_in_progress(bdi
));
64 clear_bit(BDI_pdflush
, &bdi
->state
);
68 * __mark_inode_dirty - internal function
69 * @inode: inode to mark
70 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
71 * Mark an inode as dirty. Callers should use mark_inode_dirty or
72 * mark_inode_dirty_sync.
74 * Put the inode on the super block's dirty list.
76 * CAREFUL! We mark it dirty unconditionally, but move it onto the
77 * dirty list only if it is hashed or if it refers to a blockdev.
78 * If it was not hashed, it will never be added to the dirty list
79 * even if it is later hashed, as it will have been marked dirty already.
81 * In short, make sure you hash any inodes _before_ you start marking
84 * This function *must* be atomic for the I_DIRTY_PAGES case -
85 * set_page_dirty() is called under spinlock in several places.
87 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
88 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
89 * the kernel-internal blockdev inode represents the dirtying time of the
90 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
91 * page->mapping->host, so the page-dirtying time is recorded in the internal
94 void __mark_inode_dirty(struct inode
*inode
, int flags
)
96 struct super_block
*sb
= inode
->i_sb
;
99 * Don't do this for I_DIRTY_PAGES - that doesn't actually
100 * dirty the inode itself
102 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
103 if (sb
->s_op
->dirty_inode
)
104 sb
->s_op
->dirty_inode(inode
);
108 * make sure that changes are seen by all cpus before we test i_state
113 /* avoid the locking if we can */
114 if ((inode
->i_state
& flags
) == flags
)
117 if (unlikely(block_dump
)) {
118 struct dentry
*dentry
= NULL
;
119 const char *name
= "?";
121 if (!list_empty(&inode
->i_dentry
)) {
122 dentry
= list_entry(inode
->i_dentry
.next
,
123 struct dentry
, d_alias
);
124 if (dentry
&& dentry
->d_name
.name
)
125 name
= (const char *) dentry
->d_name
.name
;
128 if (inode
->i_ino
|| strcmp(inode
->i_sb
->s_id
, "bdev"))
130 "%s(%d): dirtied inode %lu (%s) on %s\n",
131 current
->comm
, task_pid_nr(current
), inode
->i_ino
,
132 name
, inode
->i_sb
->s_id
);
135 spin_lock(&inode_lock
);
136 if ((inode
->i_state
& flags
) != flags
) {
137 const int was_dirty
= inode
->i_state
& I_DIRTY
;
139 inode
->i_state
|= flags
;
142 * If the inode is being synced, just update its dirty state.
143 * The unlocker will place the inode on the appropriate
144 * superblock list, based upon its state.
146 if (inode
->i_state
& I_SYNC
)
150 * Only add valid (hashed) inodes to the superblock's
151 * dirty list. Add blockdev inodes as well.
153 if (!S_ISBLK(inode
->i_mode
)) {
154 if (hlist_unhashed(&inode
->i_hash
))
157 if (inode
->i_state
& (I_FREEING
|I_CLEAR
))
161 * If the inode was already on s_dirty/s_io/s_more_io, don't
162 * reposition it (that would break s_dirty time-ordering).
165 inode
->dirtied_when
= jiffies
;
166 list_move(&inode
->i_list
, &sb
->s_dirty
);
170 spin_unlock(&inode_lock
);
173 EXPORT_SYMBOL(__mark_inode_dirty
);
175 static int write_inode(struct inode
*inode
, int sync
)
177 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
))
178 return inode
->i_sb
->s_op
->write_inode(inode
, sync
);
183 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
184 * furthest end of its superblock's dirty-inode list.
186 * Before stamping the inode's ->dirtied_when, we check to see whether it is
187 * already the most-recently-dirtied inode on the s_dirty list. If that is
188 * the case then the inode must have been redirtied while it was being written
189 * out and we don't reset its dirtied_when.
191 static void redirty_tail(struct inode
*inode
)
193 struct super_block
*sb
= inode
->i_sb
;
195 if (!list_empty(&sb
->s_dirty
)) {
196 struct inode
*tail_inode
;
198 tail_inode
= list_entry(sb
->s_dirty
.next
, struct inode
, i_list
);
199 if (!time_after_eq(inode
->dirtied_when
,
200 tail_inode
->dirtied_when
))
201 inode
->dirtied_when
= jiffies
;
203 list_move(&inode
->i_list
, &sb
->s_dirty
);
207 * requeue inode for re-scanning after sb->s_io list is exhausted.
209 static void requeue_io(struct inode
*inode
)
211 list_move(&inode
->i_list
, &inode
->i_sb
->s_more_io
);
214 static void inode_sync_complete(struct inode
*inode
)
217 * Prevent speculative execution through spin_unlock(&inode_lock);
220 wake_up_bit(&inode
->i_state
, __I_SYNC
);
224 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
226 static void move_expired_inodes(struct list_head
*delaying_queue
,
227 struct list_head
*dispatch_queue
,
228 unsigned long *older_than_this
)
230 while (!list_empty(delaying_queue
)) {
231 struct inode
*inode
= list_entry(delaying_queue
->prev
,
232 struct inode
, i_list
);
233 if (older_than_this
&&
234 time_after(inode
->dirtied_when
, *older_than_this
))
236 list_move(&inode
->i_list
, dispatch_queue
);
241 * Queue all expired dirty inodes for io, eldest first.
243 static void queue_io(struct super_block
*sb
,
244 unsigned long *older_than_this
)
246 list_splice_init(&sb
->s_more_io
, sb
->s_io
.prev
);
247 move_expired_inodes(&sb
->s_dirty
, &sb
->s_io
, older_than_this
);
250 int sb_has_dirty_inodes(struct super_block
*sb
)
252 return !list_empty(&sb
->s_dirty
) ||
253 !list_empty(&sb
->s_io
) ||
254 !list_empty(&sb
->s_more_io
);
256 EXPORT_SYMBOL(sb_has_dirty_inodes
);
259 * Write a single inode's dirty pages and inode data out to disk.
260 * If `wait' is set, wait on the writeout.
262 * The whole writeout design is quite complex and fragile. We want to avoid
263 * starvation of particular inodes when others are being redirtied, prevent
266 * Called under inode_lock.
269 __sync_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
272 struct address_space
*mapping
= inode
->i_mapping
;
273 int wait
= wbc
->sync_mode
== WB_SYNC_ALL
;
276 BUG_ON(inode
->i_state
& I_SYNC
);
278 /* Set I_SYNC, reset I_DIRTY */
279 dirty
= inode
->i_state
& I_DIRTY
;
280 inode
->i_state
|= I_SYNC
;
281 inode
->i_state
&= ~I_DIRTY
;
283 spin_unlock(&inode_lock
);
285 ret
= do_writepages(mapping
, wbc
);
287 /* Don't write the inode if only I_DIRTY_PAGES was set */
288 if (dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
289 int err
= write_inode(inode
, wait
);
295 int err
= filemap_fdatawait(mapping
);
300 spin_lock(&inode_lock
);
301 inode
->i_state
&= ~I_SYNC
;
302 if (!(inode
->i_state
& I_FREEING
)) {
303 if (!(inode
->i_state
& I_DIRTY
) &&
304 mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
)) {
306 * We didn't write back all the pages. nfs_writepages()
307 * sometimes bales out without doing anything. Redirty
308 * the inode; Move it from s_io onto s_more_io/s_dirty.
311 * akpm: if the caller was the kupdate function we put
312 * this inode at the head of s_dirty so it gets first
313 * consideration. Otherwise, move it to the tail, for
314 * the reasons described there. I'm not really sure
315 * how much sense this makes. Presumably I had a good
316 * reasons for doing it this way, and I'd rather not
317 * muck with it at present.
319 if (wbc
->for_kupdate
) {
321 * For the kupdate function we move the inode
322 * to s_more_io so it will get more writeout as
323 * soon as the queue becomes uncongested.
325 inode
->i_state
|= I_DIRTY_PAGES
;
326 if (wbc
->nr_to_write
<= 0) {
328 * slice used up: queue for next turn
333 * somehow blocked: retry later
339 * Otherwise fully redirty the inode so that
340 * other inodes on this superblock will get some
341 * writeout. Otherwise heavy writing to one
342 * file would indefinitely suspend writeout of
343 * all the other files.
345 inode
->i_state
|= I_DIRTY_PAGES
;
348 } else if (inode
->i_state
& I_DIRTY
) {
350 * Someone redirtied the inode while were writing back
354 } else if (atomic_read(&inode
->i_count
)) {
356 * The inode is clean, inuse
358 list_move(&inode
->i_list
, &inode_in_use
);
361 * The inode is clean, unused
363 list_move(&inode
->i_list
, &inode_unused
);
366 inode_sync_complete(inode
);
371 * Write out an inode's dirty pages. Called under inode_lock. Either the
372 * caller has ref on the inode (either via __iget or via syscall against an fd)
373 * or the inode has I_WILL_FREE set (via generic_forget_inode)
376 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
378 wait_queue_head_t
*wqh
;
380 if (!atomic_read(&inode
->i_count
))
381 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
383 WARN_ON(inode
->i_state
& I_WILL_FREE
);
385 if ((wbc
->sync_mode
!= WB_SYNC_ALL
) && (inode
->i_state
& I_SYNC
)) {
387 * We're skipping this inode because it's locked, and we're not
388 * doing writeback-for-data-integrity. Move it to s_more_io so
389 * that writeback can proceed with the other inodes on s_io.
390 * We'll have another go at writing back this inode when we
391 * completed a full scan of s_io.
398 * It's a data-integrity sync. We must wait.
400 if (inode
->i_state
& I_SYNC
) {
401 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
403 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
405 spin_unlock(&inode_lock
);
406 __wait_on_bit(wqh
, &wq
, inode_wait
,
407 TASK_UNINTERRUPTIBLE
);
408 spin_lock(&inode_lock
);
409 } while (inode
->i_state
& I_SYNC
);
411 return __sync_single_inode(inode
, wbc
);
415 * Write out a superblock's list of dirty inodes. A wait will be performed
416 * upon no inodes, all inodes or the final one, depending upon sync_mode.
418 * If older_than_this is non-NULL, then only write out inodes which
419 * had their first dirtying at a time earlier than *older_than_this.
421 * If we're a pdlfush thread, then implement pdflush collision avoidance
422 * against the entire list.
424 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
425 * that it can be located for waiting on in __writeback_single_inode().
427 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
428 * This function assumes that the blockdev superblock's inodes are backed by
429 * a variety of queues, so all inodes are searched. For other superblocks,
430 * assume that all inodes are backed by the same queue.
432 * FIXME: this linear search could get expensive with many fileystems. But
433 * how to fix? We need to go from an address_space to all inodes which share
434 * a queue with that address_space. (Easy: have a global "dirty superblocks"
437 * The inodes to be written are parked on sb->s_io. They are moved back onto
438 * sb->s_dirty as they are selected for writing. This way, none can be missed
439 * on the writer throttling path, and we get decent balancing between many
440 * throttled threads: we don't want them all piling up on inode_sync_wait.
442 void generic_sync_sb_inodes(struct super_block
*sb
,
443 struct writeback_control
*wbc
)
445 const unsigned long start
= jiffies
; /* livelock avoidance */
447 spin_lock(&inode_lock
);
448 if (!wbc
->for_kupdate
|| list_empty(&sb
->s_io
))
449 queue_io(sb
, wbc
->older_than_this
);
451 while (!list_empty(&sb
->s_io
)) {
452 struct inode
*inode
= list_entry(sb
->s_io
.prev
,
453 struct inode
, i_list
);
454 struct address_space
*mapping
= inode
->i_mapping
;
455 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
458 if (!bdi_cap_writeback_dirty(bdi
)) {
460 if (sb_is_blkdev_sb(sb
)) {
462 * Dirty memory-backed blockdev: the ramdisk
463 * driver does this. Skip just this inode
468 * Dirty memory-backed inode against a filesystem other
469 * than the kernel-internal bdev filesystem. Skip the
475 if (wbc
->nonblocking
&& bdi_write_congested(bdi
)) {
476 wbc
->encountered_congestion
= 1;
477 if (!sb_is_blkdev_sb(sb
))
478 break; /* Skip a congested fs */
480 continue; /* Skip a congested blockdev */
483 if (wbc
->bdi
&& bdi
!= wbc
->bdi
) {
484 if (!sb_is_blkdev_sb(sb
))
485 break; /* fs has the wrong queue */
487 continue; /* blockdev has wrong queue */
490 /* Was this inode dirtied after sync_sb_inodes was called? */
491 if (time_after(inode
->dirtied_when
, start
))
494 /* Is another pdflush already flushing this queue? */
495 if (current_is_pdflush() && !writeback_acquire(bdi
))
498 BUG_ON(inode
->i_state
& I_FREEING
);
500 pages_skipped
= wbc
->pages_skipped
;
501 __writeback_single_inode(inode
, wbc
);
502 if (wbc
->sync_mode
== WB_SYNC_HOLD
) {
503 inode
->dirtied_when
= jiffies
;
504 list_move(&inode
->i_list
, &sb
->s_dirty
);
506 if (current_is_pdflush())
507 writeback_release(bdi
);
508 if (wbc
->pages_skipped
!= pages_skipped
) {
510 * writeback is not making progress due to locked
511 * buffers. Skip this inode for now.
515 spin_unlock(&inode_lock
);
518 spin_lock(&inode_lock
);
519 if (wbc
->nr_to_write
<= 0) {
523 if (!list_empty(&sb
->s_more_io
))
526 spin_unlock(&inode_lock
);
527 return; /* Leave any unwritten inodes on s_io */
529 EXPORT_SYMBOL_GPL(generic_sync_sb_inodes
);
531 static void sync_sb_inodes(struct super_block
*sb
,
532 struct writeback_control
*wbc
)
534 generic_sync_sb_inodes(sb
, wbc
);
538 * Start writeback of dirty pagecache data against all unlocked inodes.
541 * We don't need to grab a reference to superblock here. If it has non-empty
542 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
543 * past sync_inodes_sb() until the ->s_dirty/s_io/s_more_io lists are all
544 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
545 * inode from superblock lists we are OK.
547 * If `older_than_this' is non-zero then only flush inodes which have a
548 * flushtime older than *older_than_this.
550 * If `bdi' is non-zero then we will scan the first inode against each
551 * superblock until we find the matching ones. One group will be the dirty
552 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
553 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
554 * super-efficient but we're about to do a ton of I/O...
557 writeback_inodes(struct writeback_control
*wbc
)
559 struct super_block
*sb
;
564 list_for_each_entry_reverse(sb
, &super_blocks
, s_list
) {
565 if (sb_has_dirty_inodes(sb
)) {
566 /* we're making our own get_super here */
568 spin_unlock(&sb_lock
);
570 * If we can't get the readlock, there's no sense in
571 * waiting around, most of the time the FS is going to
572 * be unmounted by the time it is released.
574 if (down_read_trylock(&sb
->s_umount
)) {
576 sync_sb_inodes(sb
, wbc
);
577 up_read(&sb
->s_umount
);
580 if (__put_super_and_need_restart(sb
))
583 if (wbc
->nr_to_write
<= 0)
586 spin_unlock(&sb_lock
);
590 * writeback and wait upon the filesystem's dirty inodes. The caller will
591 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
592 * used to park the written inodes on sb->s_dirty for the wait pass.
594 * A finite limit is set on the number of pages which will be written.
595 * To prevent infinite livelock of sys_sync().
597 * We add in the number of potentially dirty inodes, because each inode write
598 * can dirty pagecache in the underlying blockdev.
600 void sync_inodes_sb(struct super_block
*sb
, int wait
)
602 struct writeback_control wbc
= {
603 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_HOLD
,
605 .range_end
= LLONG_MAX
,
607 unsigned long nr_dirty
= global_page_state(NR_FILE_DIRTY
);
608 unsigned long nr_unstable
= global_page_state(NR_UNSTABLE_NFS
);
610 wbc
.nr_to_write
= nr_dirty
+ nr_unstable
+
611 (inodes_stat
.nr_inodes
- inodes_stat
.nr_unused
) +
612 nr_dirty
+ nr_unstable
;
613 wbc
.nr_to_write
+= wbc
.nr_to_write
/ 2; /* Bit more for luck */
614 sync_sb_inodes(sb
, &wbc
);
618 * Rather lame livelock avoidance.
620 static void set_sb_syncing(int val
)
622 struct super_block
*sb
;
624 list_for_each_entry_reverse(sb
, &super_blocks
, s_list
)
626 spin_unlock(&sb_lock
);
630 * sync_inodes - writes all inodes to disk
631 * @wait: wait for completion
633 * sync_inodes() goes through each super block's dirty inode list, writes the
634 * inodes out, waits on the writeout and puts the inodes back on the normal
637 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
638 * part of the sync functions is that the blockdev "superblock" is processed
639 * last. This is because the write_inode() function of a typical fs will
640 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
641 * What we want to do is to perform all that dirtying first, and then write
642 * back all those inode blocks via the blockdev mapping in one sweep. So the
643 * additional (somewhat redundant) sync_blockdev() calls here are to make
644 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
645 * outstanding dirty inodes, the writeback goes block-at-a-time within the
646 * filesystem's write_inode(). This is extremely slow.
648 static void __sync_inodes(int wait
)
650 struct super_block
*sb
;
654 list_for_each_entry(sb
, &super_blocks
, s_list
) {
659 spin_unlock(&sb_lock
);
660 down_read(&sb
->s_umount
);
662 sync_inodes_sb(sb
, wait
);
663 sync_blockdev(sb
->s_bdev
);
665 up_read(&sb
->s_umount
);
667 if (__put_super_and_need_restart(sb
))
670 spin_unlock(&sb_lock
);
673 void sync_inodes(int wait
)
685 * write_inode_now - write an inode to disk
686 * @inode: inode to write to disk
687 * @sync: whether the write should be synchronous or not
689 * This function commits an inode to disk immediately if it is dirty. This is
690 * primarily needed by knfsd.
692 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
694 int write_inode_now(struct inode
*inode
, int sync
)
697 struct writeback_control wbc
= {
698 .nr_to_write
= LONG_MAX
,
699 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
701 .range_end
= LLONG_MAX
,
704 if (!mapping_cap_writeback_dirty(inode
->i_mapping
))
708 spin_lock(&inode_lock
);
709 ret
= __writeback_single_inode(inode
, &wbc
);
710 spin_unlock(&inode_lock
);
712 inode_sync_wait(inode
);
715 EXPORT_SYMBOL(write_inode_now
);
718 * sync_inode - write an inode and its pages to disk.
719 * @inode: the inode to sync
720 * @wbc: controls the writeback mode
722 * sync_inode() will write an inode and its pages to disk. It will also
723 * correctly update the inode on its superblock's dirty inode lists and will
724 * update inode->i_state.
726 * The caller must have a ref on the inode.
728 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
732 spin_lock(&inode_lock
);
733 ret
= __writeback_single_inode(inode
, wbc
);
734 spin_unlock(&inode_lock
);
737 EXPORT_SYMBOL(sync_inode
);
740 * generic_osync_inode - flush all dirty data for a given inode to disk
741 * @inode: inode to write
742 * @mapping: the address_space that should be flushed
743 * @what: what to write and wait upon
745 * This can be called by file_write functions for files which have the
746 * O_SYNC flag set, to flush dirty writes to disk.
748 * @what is a bitmask, specifying which part of the inode's data should be
749 * written and waited upon.
751 * OSYNC_DATA: i_mapping's dirty data
752 * OSYNC_METADATA: the buffers at i_mapping->private_list
753 * OSYNC_INODE: the inode itself
756 int generic_osync_inode(struct inode
*inode
, struct address_space
*mapping
, int what
)
759 int need_write_inode_now
= 0;
762 if (what
& OSYNC_DATA
)
763 err
= filemap_fdatawrite(mapping
);
764 if (what
& (OSYNC_METADATA
|OSYNC_DATA
)) {
765 err2
= sync_mapping_buffers(mapping
);
769 if (what
& OSYNC_DATA
) {
770 err2
= filemap_fdatawait(mapping
);
775 spin_lock(&inode_lock
);
776 if ((inode
->i_state
& I_DIRTY
) &&
777 ((what
& OSYNC_INODE
) || (inode
->i_state
& I_DIRTY_DATASYNC
)))
778 need_write_inode_now
= 1;
779 spin_unlock(&inode_lock
);
781 if (need_write_inode_now
) {
782 err2
= write_inode_now(inode
, 1);
787 inode_sync_wait(inode
);
791 EXPORT_SYMBOL(generic_osync_inode
);