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 void __mark_inode_dirty(struct inode
*inode
, int flags
)
50 struct super_block
*sb
= inode
->i_sb
;
53 return; /* swapper_space */
56 * Don't do this for I_DIRTY_PAGES - that doesn't actually
57 * dirty the inode itself
59 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) {
60 if (sb
->s_op
->dirty_inode
)
61 sb
->s_op
->dirty_inode(inode
);
65 * make sure that changes are seen by all cpus before we test i_state
70 /* avoid the locking if we can */
71 if ((inode
->i_state
& flags
) == flags
)
74 spin_lock(&inode_lock
);
75 if ((inode
->i_state
& flags
) != flags
) {
76 const int was_dirty
= inode
->i_state
& I_DIRTY
;
77 struct address_space
*mapping
= inode
->i_mapping
;
79 inode
->i_state
|= flags
;
82 * If the inode is locked, just update its dirty state.
83 * The unlocker will place the inode on the appropriate
84 * superblock list, based upon its state.
86 if (inode
->i_state
& I_LOCK
)
90 * Only add valid (hashed) inodes to the superblock's
91 * dirty list. Add blockdev inodes as well.
93 if ((hlist_unhashed(&inode
->i_hash
) || (inode
->i_state
& (I_FREEING
|I_CLEAR
)))
94 && !S_ISBLK(inode
->i_mode
))
98 * If the inode was already on s_dirty or s_io, don't
99 * reposition it (that would break s_dirty time-ordering).
102 mapping
->dirtied_when
= jiffies
|1; /* 0 is special */
103 list_move(&inode
->i_list
, &sb
->s_dirty
);
107 spin_unlock(&inode_lock
);
110 static void write_inode(struct inode
*inode
, int sync
)
112 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
))
113 inode
->i_sb
->s_op
->write_inode(inode
, sync
);
117 * Write a single inode's dirty pages and inode data out to disk.
118 * If `wait' is set, wait on the writeout.
120 * The whole writeout design is quite complex and fragile. We want to avoid
121 * starvation of particular inodes when others are being redirtied, prevent
124 * So what we do is to move all pages which are to be written from dirty_pages
125 * onto io_pages. And keep on writing io_pages until it's empty. Refusing to
126 * move more pages onto io_pages until io_pages is empty. Once that point has
127 * been reached, we are ready to take another pass across the inode's dirty
130 * Called under inode_lock.
133 __sync_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
136 struct address_space
*mapping
= inode
->i_mapping
;
137 struct super_block
*sb
= inode
->i_sb
;
138 int wait
= wbc
->sync_mode
== WB_SYNC_ALL
;
140 BUG_ON(inode
->i_state
& I_LOCK
);
142 /* Set I_LOCK, reset I_DIRTY */
143 dirty
= inode
->i_state
& I_DIRTY
;
144 inode
->i_state
|= I_LOCK
;
145 inode
->i_state
&= ~I_DIRTY
;
148 * smp_rmb(); note: if you remove write_lock below, you must add this.
149 * mark_inode_dirty doesn't take spinlock, make sure that inode is not
150 * read speculatively by this cpu before &= ~I_DIRTY -- mikulas
153 spin_lock(&mapping
->page_lock
);
154 if (wait
|| !wbc
->for_kupdate
|| list_empty(&mapping
->io_pages
))
155 list_splice_init(&mapping
->dirty_pages
, &mapping
->io_pages
);
156 spin_unlock(&mapping
->page_lock
);
157 spin_unlock(&inode_lock
);
159 do_writepages(mapping
, wbc
);
161 /* Don't write the inode if only I_DIRTY_PAGES was set */
162 if (dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
))
163 write_inode(inode
, wait
);
166 filemap_fdatawait(mapping
);
168 spin_lock(&inode_lock
);
169 inode
->i_state
&= ~I_LOCK
;
170 if (!(inode
->i_state
& I_FREEING
)) {
171 if (!list_empty(&mapping
->io_pages
)) {
172 /* Needs more writeback */
173 inode
->i_state
|= I_DIRTY_PAGES
;
174 } else if (!list_empty(&mapping
->dirty_pages
)) {
176 inode
->i_state
|= I_DIRTY_PAGES
;
177 mapping
->dirtied_when
= jiffies
|1;
178 list_move(&inode
->i_list
, &sb
->s_dirty
);
179 } else if (inode
->i_state
& I_DIRTY
) {
181 mapping
->dirtied_when
= jiffies
|1;
182 list_move(&inode
->i_list
, &sb
->s_dirty
);
183 } else if (atomic_read(&inode
->i_count
)) {
184 mapping
->dirtied_when
= 0;
185 list_move(&inode
->i_list
, &inode_in_use
);
187 mapping
->dirtied_when
= 0;
188 list_move(&inode
->i_list
, &inode_unused
);
191 wake_up_inode(inode
);
195 * Write out an inode's dirty pages. Called under inode_lock.
198 __writeback_single_inode(struct inode
*inode
,
199 struct writeback_control
*wbc
)
201 if ((wbc
->sync_mode
!= WB_SYNC_ALL
) && (inode
->i_state
& I_LOCK
)) {
202 list_move(&inode
->i_list
, &inode
->i_sb
->s_dirty
);
207 * It's a data-integrity sync. We must wait.
209 while (inode
->i_state
& I_LOCK
) {
211 spin_unlock(&inode_lock
);
212 __wait_on_inode(inode
);
214 spin_lock(&inode_lock
);
216 __sync_single_inode(inode
, wbc
);
220 * Write out a superblock's list of dirty inodes. A wait will be performed
221 * upon no inodes, all inodes or the final one, depending upon sync_mode.
223 * If older_than_this is non-NULL, then only write out mappings which
224 * had their first dirtying at a time earlier than *older_than_this.
226 * If we're a pdlfush thread, then implement pdflush collision avoidance
227 * against the entire list.
229 * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so
230 * that it can be located for waiting on in __writeback_single_inode().
232 * Called under inode_lock.
234 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
235 * This function assumes that the blockdev superblock's inodes are backed by
236 * a variety of queues, so all inodes are searched. For other superblocks,
237 * assume that all inodes are backed by the same queue.
239 * FIXME: this linear search could get expensive with many fileystems. But
240 * how to fix? We need to go from an address_space to all inodes which share
241 * a queue with that address_space. (Easy: have a global "dirty superblocks"
244 * The inodes to be written are parked on sb->s_io. They are moved back onto
245 * sb->s_dirty as they are selected for writing. This way, none can be missed
246 * on the writer throttling path, and we get decent balancing between many
247 * throttled threads: we don't want them all piling up on __wait_on_inode.
250 sync_sb_inodes(struct super_block
*sb
, struct writeback_control
*wbc
)
252 const unsigned long start
= jiffies
; /* livelock avoidance */
254 if (!wbc
->for_kupdate
|| list_empty(&sb
->s_io
))
255 list_splice_init(&sb
->s_dirty
, &sb
->s_io
);
257 while (!list_empty(&sb
->s_io
)) {
258 struct inode
*inode
= list_entry(sb
->s_io
.prev
,
259 struct inode
, i_list
);
260 struct address_space
*mapping
= inode
->i_mapping
;
261 struct backing_dev_info
*bdi
= mapping
->backing_dev_info
;
263 if (bdi
->memory_backed
)
266 if (wbc
->nonblocking
&& bdi_write_congested(bdi
)) {
267 wbc
->encountered_congestion
= 1;
268 if (sb
!= blockdev_superblock
)
269 break; /* Skip a congested fs */
270 list_move(&inode
->i_list
, &sb
->s_dirty
);
271 continue; /* Skip a congested blockdev */
274 if (wbc
->bdi
&& bdi
!= wbc
->bdi
) {
275 if (sb
!= blockdev_superblock
)
276 break; /* fs has the wrong queue */
277 list_move(&inode
->i_list
, &sb
->s_dirty
);
278 continue; /* blockdev has wrong queue */
281 /* Was this inode dirtied after sync_sb_inodes was called? */
282 if (time_after(mapping
->dirtied_when
, start
))
285 /* Was this inode dirtied too recently? */
286 if (wbc
->older_than_this
&& time_after(mapping
->dirtied_when
,
287 *wbc
->older_than_this
))
290 /* Is another pdflush already flushing this queue? */
291 if (current_is_pdflush() && !writeback_acquire(bdi
))
294 BUG_ON(inode
->i_state
& I_FREEING
);
296 __writeback_single_inode(inode
, wbc
);
297 if (wbc
->sync_mode
== WB_SYNC_HOLD
) {
298 mapping
->dirtied_when
= jiffies
|1;
299 list_move(&inode
->i_list
, &sb
->s_dirty
);
301 if (current_is_pdflush())
302 writeback_release(bdi
);
303 spin_unlock(&inode_lock
);
305 spin_lock(&inode_lock
);
306 if (wbc
->nr_to_write
<= 0)
309 return; /* Leave any unwritten inodes on s_io */
313 * Start writeback of dirty pagecache data against all unlocked inodes.
316 * We don't need to grab a reference to superblock here. If it has non-empty
317 * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
318 * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
319 * empty. Since __sync_single_inode() regains inode_lock before it finally moves
320 * inode from superblock lists we are OK.
322 * If `older_than_this' is non-zero then only flush inodes which have a
323 * flushtime older than *older_than_this.
325 * If `bdi' is non-zero then we will scan the first inode against each
326 * superblock until we find the matching ones. One group will be the dirty
327 * inodes against a filesystem. Then when we hit the dummy blockdev superblock,
328 * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not
329 * super-efficient but we're about to do a ton of I/O...
332 writeback_inodes(struct writeback_control
*wbc
)
334 struct super_block
*sb
;
336 spin_lock(&inode_lock
);
338 sb
= sb_entry(super_blocks
.prev
);
339 for (; sb
!= sb_entry(&super_blocks
); sb
= sb_entry(sb
->s_list
.prev
)) {
340 if (!list_empty(&sb
->s_dirty
) || !list_empty(&sb
->s_io
)) {
341 spin_unlock(&sb_lock
);
342 sync_sb_inodes(sb
, wbc
);
345 if (wbc
->nr_to_write
<= 0)
348 spin_unlock(&sb_lock
);
349 spin_unlock(&inode_lock
);
353 * writeback and wait upon the filesystem's dirty inodes. The caller will
354 * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is
355 * used to park the written inodes on sb->s_dirty for the wait pass.
357 * A finite limit is set on the number of pages which will be written.
358 * To prevent infinite livelock of sys_sync().
360 void sync_inodes_sb(struct super_block
*sb
, int wait
)
362 struct page_state ps
;
363 struct writeback_control wbc
= {
365 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_HOLD
,
366 .older_than_this
= NULL
,
371 wbc
.nr_to_write
= ps
.nr_dirty
+ ps
.nr_unstable
+
372 (ps
.nr_dirty
+ ps
.nr_unstable
) / 4;
373 spin_lock(&inode_lock
);
374 sync_sb_inodes(sb
, &wbc
);
375 spin_unlock(&inode_lock
);
379 * Rather lame livelock avoidance.
381 static void set_sb_syncing(int val
)
383 struct super_block
*sb
;
385 sb
= sb_entry(super_blocks
.prev
);
386 for (; sb
!= sb_entry(&super_blocks
); sb
= sb_entry(sb
->s_list
.prev
)) {
389 spin_unlock(&sb_lock
);
393 * Find a superblock with inodes that need to be synced
395 static struct super_block
*get_super_to_sync(void)
397 struct super_block
*sb
;
400 sb
= sb_entry(super_blocks
.prev
);
401 for (; sb
!= sb_entry(&super_blocks
); sb
= sb_entry(sb
->s_list
.prev
)) {
406 spin_unlock(&sb_lock
);
407 down_read(&sb
->s_umount
);
414 spin_unlock(&sb_lock
);
421 * sync_inodes() goes through each super block's dirty inode list, writes the
422 * inodes out, waits on the writeout and puts the inodes back on the normal
425 * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle
426 * part of the sync functions is that the blockdev "superblock" is processed
427 * last. This is because the write_inode() function of a typical fs will
428 * perform no I/O, but will mark buffers in the blockdev mapping as dirty.
429 * What we want to do is to perform all that dirtying first, and then write
430 * back all those inode blocks via the blockdev mapping in one sweep. So the
431 * additional (somewhat redundant) sync_blockdev() calls here are to make
432 * sure that really happens. Because if we call sync_inodes_sb(wait=1) with
433 * outstanding dirty inodes, the writeback goes block-at-a-time within the
434 * filesystem's write_inode(). This is extremely slow.
436 void sync_inodes(int wait
)
438 struct super_block
*sb
;
441 while ((sb
= get_super_to_sync()) != NULL
) {
442 sync_inodes_sb(sb
, 0);
443 sync_blockdev(sb
->s_bdev
);
448 while ((sb
= get_super_to_sync()) != NULL
) {
449 sync_inodes_sb(sb
, 1);
450 sync_blockdev(sb
->s_bdev
);
457 * write_inode_now - write an inode to disk
458 * @inode: inode to write to disk
459 * @sync: whether the write should be synchronous or not
461 * This function commits an inode to disk immediately if it is
462 * dirty. This is primarily needed by knfsd.
465 void write_inode_now(struct inode
*inode
, int sync
)
467 struct writeback_control wbc
= {
468 .nr_to_write
= LONG_MAX
,
469 .sync_mode
= WB_SYNC_ALL
,
472 spin_lock(&inode_lock
);
473 __writeback_single_inode(inode
, &wbc
);
474 spin_unlock(&inode_lock
);
476 wait_on_inode(inode
);
480 * generic_osync_inode - flush all dirty data for a given inode to disk
481 * @inode: inode to write
482 * @what: what to write and wait upon
484 * This can be called by file_write functions for files which have the
485 * O_SYNC flag set, to flush dirty writes to disk.
487 * @what is a bitmask, specifying which part of the inode's data should be
488 * written and waited upon:
490 * OSYNC_DATA: i_mapping's dirty data
491 * OSYNC_METADATA: the buffers at i_mapping->private_list
492 * OSYNC_INODE: the inode itself
495 int generic_osync_inode(struct inode
*inode
, int what
)
498 int need_write_inode_now
= 0;
501 if (what
& OSYNC_DATA
)
502 err
= filemap_fdatawrite(inode
->i_mapping
);
503 if (what
& (OSYNC_METADATA
|OSYNC_DATA
)) {
504 err2
= sync_mapping_buffers(inode
->i_mapping
);
508 if (what
& OSYNC_DATA
) {
509 err2
= filemap_fdatawait(inode
->i_mapping
);
514 spin_lock(&inode_lock
);
515 if ((inode
->i_state
& I_DIRTY
) &&
516 ((what
& OSYNC_INODE
) || (inode
->i_state
& I_DIRTY_DATASYNC
)))
517 need_write_inode_now
= 1;
518 spin_unlock(&inode_lock
);
520 if (need_write_inode_now
)
521 write_inode_now(inode
, 1);
523 wait_on_inode(inode
);
529 * writeback_acquire: attempt to get exclusive writeback access to a device
530 * @bdi: the device's backing_dev_info structure
532 * It is a waste of resources to have more than one pdflush thread blocked on
533 * a single request queue. Exclusion at the request_queue level is obtained
534 * via a flag in the request_queue's backing_dev_info.state.
536 * Non-request_queue-backed address_spaces will share default_backing_dev_info,
537 * unless they implement their own. Which is somewhat inefficient, as this
538 * may prevent concurrent writeback against multiple devices.
540 int writeback_acquire(struct backing_dev_info
*bdi
)
542 return !test_and_set_bit(BDI_pdflush
, &bdi
->state
);
546 * writeback_in_progress: determine whether there is writeback in progress
547 * against a backing device.
548 * @bdi: the device's backing_dev_info structure.
550 int writeback_in_progress(struct backing_dev_info
*bdi
)
552 return test_bit(BDI_pdflush
, &bdi
->state
);
556 * writeback_release: relinquish exclusive writeback access against a device.
557 * @bdi: the device's backing_dev_info structure
559 void writeback_release(struct backing_dev_info
*bdi
)
561 BUG_ON(!writeback_in_progress(bdi
));
562 clear_bit(BDI_pdflush
, &bdi
->state
);