4 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
8 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
10 * Removed a lot of unnecessary code and simplified things now that
11 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
13 * Speed up hash, lru, and free list operations. Use gfp() for allocating
14 * hash table, use SLAB cache for buffer heads. SMP threading. -DaveM
16 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
18 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/export.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
45 static int fsync_buffers_list(spinlock_t
*lock
, struct list_head
*list
);
47 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
50 init_buffer(struct buffer_head
*bh
, bh_end_io_t
*handler
, void *private)
52 bh
->b_end_io
= handler
;
53 bh
->b_private
= private;
55 EXPORT_SYMBOL(init_buffer
);
57 static int sleep_on_buffer(void *word
)
63 void __lock_buffer(struct buffer_head
*bh
)
65 wait_on_bit_lock(&bh
->b_state
, BH_Lock
, sleep_on_buffer
,
66 TASK_UNINTERRUPTIBLE
);
68 EXPORT_SYMBOL(__lock_buffer
);
70 void unlock_buffer(struct buffer_head
*bh
)
72 clear_bit_unlock(BH_Lock
, &bh
->b_state
);
73 smp_mb__after_clear_bit();
74 wake_up_bit(&bh
->b_state
, BH_Lock
);
76 EXPORT_SYMBOL(unlock_buffer
);
79 * Block until a buffer comes unlocked. This doesn't stop it
80 * from becoming locked again - you have to lock it yourself
81 * if you want to preserve its state.
83 void __wait_on_buffer(struct buffer_head
* bh
)
85 wait_on_bit(&bh
->b_state
, BH_Lock
, sleep_on_buffer
, TASK_UNINTERRUPTIBLE
);
87 EXPORT_SYMBOL(__wait_on_buffer
);
90 __clear_page_buffers(struct page
*page
)
92 ClearPagePrivate(page
);
93 set_page_private(page
, 0);
94 page_cache_release(page
);
98 static int quiet_error(struct buffer_head
*bh
)
100 if (!test_bit(BH_Quiet
, &bh
->b_state
) && printk_ratelimit())
106 static void buffer_io_error(struct buffer_head
*bh
)
108 char b
[BDEVNAME_SIZE
];
109 printk(KERN_ERR
"Buffer I/O error on device %s, logical block %Lu\n",
110 bdevname(bh
->b_bdev
, b
),
111 (unsigned long long)bh
->b_blocknr
);
115 * End-of-IO handler helper function which does not touch the bh after
117 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
118 * a race there is benign: unlock_buffer() only use the bh's address for
119 * hashing after unlocking the buffer, so it doesn't actually touch the bh
122 static void __end_buffer_read_notouch(struct buffer_head
*bh
, int uptodate
)
125 set_buffer_uptodate(bh
);
127 /* This happens, due to failed READA attempts. */
128 clear_buffer_uptodate(bh
);
134 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
135 * unlock the buffer. This is what ll_rw_block uses too.
137 void end_buffer_read_sync(struct buffer_head
*bh
, int uptodate
)
139 __end_buffer_read_notouch(bh
, uptodate
);
142 EXPORT_SYMBOL(end_buffer_read_sync
);
144 void end_buffer_write_sync(struct buffer_head
*bh
, int uptodate
)
146 char b
[BDEVNAME_SIZE
];
149 set_buffer_uptodate(bh
);
151 if (!quiet_error(bh
)) {
153 printk(KERN_WARNING
"lost page write due to "
155 bdevname(bh
->b_bdev
, b
));
157 set_buffer_write_io_error(bh
);
158 clear_buffer_uptodate(bh
);
163 EXPORT_SYMBOL(end_buffer_write_sync
);
166 * Various filesystems appear to want __find_get_block to be non-blocking.
167 * But it's the page lock which protects the buffers. To get around this,
168 * we get exclusion from try_to_free_buffers with the blockdev mapping's
171 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
172 * may be quite high. This code could TryLock the page, and if that
173 * succeeds, there is no need to take private_lock. (But if
174 * private_lock is contended then so is mapping->tree_lock).
176 static struct buffer_head
*
177 __find_get_block_slow(struct block_device
*bdev
, sector_t block
)
179 struct inode
*bd_inode
= bdev
->bd_inode
;
180 struct address_space
*bd_mapping
= bd_inode
->i_mapping
;
181 struct buffer_head
*ret
= NULL
;
183 struct buffer_head
*bh
;
184 struct buffer_head
*head
;
188 index
= block
>> (PAGE_CACHE_SHIFT
- bd_inode
->i_blkbits
);
189 page
= find_get_page(bd_mapping
, index
);
193 spin_lock(&bd_mapping
->private_lock
);
194 if (!page_has_buffers(page
))
196 head
= page_buffers(page
);
199 if (!buffer_mapped(bh
))
201 else if (bh
->b_blocknr
== block
) {
206 bh
= bh
->b_this_page
;
207 } while (bh
!= head
);
209 /* we might be here because some of the buffers on this page are
210 * not mapped. This is due to various races between
211 * file io on the block device and getblk. It gets dealt with
212 * elsewhere, don't buffer_error if we had some unmapped buffers
215 char b
[BDEVNAME_SIZE
];
217 printk("__find_get_block_slow() failed. "
218 "block=%llu, b_blocknr=%llu\n",
219 (unsigned long long)block
,
220 (unsigned long long)bh
->b_blocknr
);
221 printk("b_state=0x%08lx, b_size=%zu\n",
222 bh
->b_state
, bh
->b_size
);
223 printk("device %s blocksize: %d\n", bdevname(bdev
, b
),
224 1 << bd_inode
->i_blkbits
);
227 spin_unlock(&bd_mapping
->private_lock
);
228 page_cache_release(page
);
234 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
236 static void free_more_memory(void)
241 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM
);
244 for_each_online_node(nid
) {
245 (void)first_zones_zonelist(node_zonelist(nid
, GFP_NOFS
),
246 gfp_zone(GFP_NOFS
), NULL
,
249 try_to_free_pages(node_zonelist(nid
, GFP_NOFS
), 0,
255 * I/O completion handler for block_read_full_page() - pages
256 * which come unlocked at the end of I/O.
258 static void end_buffer_async_read(struct buffer_head
*bh
, int uptodate
)
261 struct buffer_head
*first
;
262 struct buffer_head
*tmp
;
264 int page_uptodate
= 1;
266 BUG_ON(!buffer_async_read(bh
));
270 set_buffer_uptodate(bh
);
272 clear_buffer_uptodate(bh
);
273 if (!quiet_error(bh
))
279 * Be _very_ careful from here on. Bad things can happen if
280 * two buffer heads end IO at almost the same time and both
281 * decide that the page is now completely done.
283 first
= page_buffers(page
);
284 local_irq_save(flags
);
285 bit_spin_lock(BH_Uptodate_Lock
, &first
->b_state
);
286 clear_buffer_async_read(bh
);
290 if (!buffer_uptodate(tmp
))
292 if (buffer_async_read(tmp
)) {
293 BUG_ON(!buffer_locked(tmp
));
296 tmp
= tmp
->b_this_page
;
298 bit_spin_unlock(BH_Uptodate_Lock
, &first
->b_state
);
299 local_irq_restore(flags
);
302 * If none of the buffers had errors and they are all
303 * uptodate then we can set the page uptodate.
305 if (page_uptodate
&& !PageError(page
))
306 SetPageUptodate(page
);
311 bit_spin_unlock(BH_Uptodate_Lock
, &first
->b_state
);
312 local_irq_restore(flags
);
317 * Completion handler for block_write_full_page() - pages which are unlocked
318 * during I/O, and which have PageWriteback cleared upon I/O completion.
320 void end_buffer_async_write(struct buffer_head
*bh
, int uptodate
)
322 char b
[BDEVNAME_SIZE
];
324 struct buffer_head
*first
;
325 struct buffer_head
*tmp
;
328 BUG_ON(!buffer_async_write(bh
));
332 set_buffer_uptodate(bh
);
334 if (!quiet_error(bh
)) {
336 printk(KERN_WARNING
"lost page write due to "
338 bdevname(bh
->b_bdev
, b
));
340 set_bit(AS_EIO
, &page
->mapping
->flags
);
341 set_buffer_write_io_error(bh
);
342 clear_buffer_uptodate(bh
);
346 first
= page_buffers(page
);
347 local_irq_save(flags
);
348 bit_spin_lock(BH_Uptodate_Lock
, &first
->b_state
);
350 clear_buffer_async_write(bh
);
352 tmp
= bh
->b_this_page
;
354 if (buffer_async_write(tmp
)) {
355 BUG_ON(!buffer_locked(tmp
));
358 tmp
= tmp
->b_this_page
;
360 bit_spin_unlock(BH_Uptodate_Lock
, &first
->b_state
);
361 local_irq_restore(flags
);
362 end_page_writeback(page
);
366 bit_spin_unlock(BH_Uptodate_Lock
, &first
->b_state
);
367 local_irq_restore(flags
);
370 EXPORT_SYMBOL(end_buffer_async_write
);
373 * If a page's buffers are under async readin (end_buffer_async_read
374 * completion) then there is a possibility that another thread of
375 * control could lock one of the buffers after it has completed
376 * but while some of the other buffers have not completed. This
377 * locked buffer would confuse end_buffer_async_read() into not unlocking
378 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
379 * that this buffer is not under async I/O.
381 * The page comes unlocked when it has no locked buffer_async buffers
384 * PageLocked prevents anyone starting new async I/O reads any of
387 * PageWriteback is used to prevent simultaneous writeout of the same
390 * PageLocked prevents anyone from starting writeback of a page which is
391 * under read I/O (PageWriteback is only ever set against a locked page).
393 static void mark_buffer_async_read(struct buffer_head
*bh
)
395 bh
->b_end_io
= end_buffer_async_read
;
396 set_buffer_async_read(bh
);
399 static void mark_buffer_async_write_endio(struct buffer_head
*bh
,
400 bh_end_io_t
*handler
)
402 bh
->b_end_io
= handler
;
403 set_buffer_async_write(bh
);
406 void mark_buffer_async_write(struct buffer_head
*bh
)
408 mark_buffer_async_write_endio(bh
, end_buffer_async_write
);
410 EXPORT_SYMBOL(mark_buffer_async_write
);
414 * fs/buffer.c contains helper functions for buffer-backed address space's
415 * fsync functions. A common requirement for buffer-based filesystems is
416 * that certain data from the backing blockdev needs to be written out for
417 * a successful fsync(). For example, ext2 indirect blocks need to be
418 * written back and waited upon before fsync() returns.
420 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
421 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
422 * management of a list of dependent buffers at ->i_mapping->private_list.
424 * Locking is a little subtle: try_to_free_buffers() will remove buffers
425 * from their controlling inode's queue when they are being freed. But
426 * try_to_free_buffers() will be operating against the *blockdev* mapping
427 * at the time, not against the S_ISREG file which depends on those buffers.
428 * So the locking for private_list is via the private_lock in the address_space
429 * which backs the buffers. Which is different from the address_space
430 * against which the buffers are listed. So for a particular address_space,
431 * mapping->private_lock does *not* protect mapping->private_list! In fact,
432 * mapping->private_list will always be protected by the backing blockdev's
435 * Which introduces a requirement: all buffers on an address_space's
436 * ->private_list must be from the same address_space: the blockdev's.
438 * address_spaces which do not place buffers at ->private_list via these
439 * utility functions are free to use private_lock and private_list for
440 * whatever they want. The only requirement is that list_empty(private_list)
441 * be true at clear_inode() time.
443 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
444 * filesystems should do that. invalidate_inode_buffers() should just go
445 * BUG_ON(!list_empty).
447 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
448 * take an address_space, not an inode. And it should be called
449 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
452 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
453 * list if it is already on a list. Because if the buffer is on a list,
454 * it *must* already be on the right one. If not, the filesystem is being
455 * silly. This will save a ton of locking. But first we have to ensure
456 * that buffers are taken *off* the old inode's list when they are freed
457 * (presumably in truncate). That requires careful auditing of all
458 * filesystems (do it inside bforget()). It could also be done by bringing
463 * The buffer's backing address_space's private_lock must be held
465 static void __remove_assoc_queue(struct buffer_head
*bh
)
467 list_del_init(&bh
->b_assoc_buffers
);
468 WARN_ON(!bh
->b_assoc_map
);
469 if (buffer_write_io_error(bh
))
470 set_bit(AS_EIO
, &bh
->b_assoc_map
->flags
);
471 bh
->b_assoc_map
= NULL
;
474 int inode_has_buffers(struct inode
*inode
)
476 return !list_empty(&inode
->i_data
.private_list
);
480 * osync is designed to support O_SYNC io. It waits synchronously for
481 * all already-submitted IO to complete, but does not queue any new
482 * writes to the disk.
484 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
485 * you dirty the buffers, and then use osync_inode_buffers to wait for
486 * completion. Any other dirty buffers which are not yet queued for
487 * write will not be flushed to disk by the osync.
489 static int osync_buffers_list(spinlock_t
*lock
, struct list_head
*list
)
491 struct buffer_head
*bh
;
497 list_for_each_prev(p
, list
) {
499 if (buffer_locked(bh
)) {
503 if (!buffer_uptodate(bh
))
514 static void do_thaw_one(struct super_block
*sb
, void *unused
)
516 char b
[BDEVNAME_SIZE
];
517 while (sb
->s_bdev
&& !thaw_bdev(sb
->s_bdev
, sb
))
518 printk(KERN_WARNING
"Emergency Thaw on %s\n",
519 bdevname(sb
->s_bdev
, b
));
522 static void do_thaw_all(struct work_struct
*work
)
524 iterate_supers(do_thaw_one
, NULL
);
526 printk(KERN_WARNING
"Emergency Thaw complete\n");
530 * emergency_thaw_all -- forcibly thaw every frozen filesystem
532 * Used for emergency unfreeze of all filesystems via SysRq
534 void emergency_thaw_all(void)
536 struct work_struct
*work
;
538 work
= kmalloc(sizeof(*work
), GFP_ATOMIC
);
540 INIT_WORK(work
, do_thaw_all
);
546 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
547 * @mapping: the mapping which wants those buffers written
549 * Starts I/O against the buffers at mapping->private_list, and waits upon
552 * Basically, this is a convenience function for fsync().
553 * @mapping is a file or directory which needs those buffers to be written for
554 * a successful fsync().
556 int sync_mapping_buffers(struct address_space
*mapping
)
558 struct address_space
*buffer_mapping
= mapping
->assoc_mapping
;
560 if (buffer_mapping
== NULL
|| list_empty(&mapping
->private_list
))
563 return fsync_buffers_list(&buffer_mapping
->private_lock
,
564 &mapping
->private_list
);
566 EXPORT_SYMBOL(sync_mapping_buffers
);
569 * Called when we've recently written block `bblock', and it is known that
570 * `bblock' was for a buffer_boundary() buffer. This means that the block at
571 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
572 * dirty, schedule it for IO. So that indirects merge nicely with their data.
574 void write_boundary_block(struct block_device
*bdev
,
575 sector_t bblock
, unsigned blocksize
)
577 struct buffer_head
*bh
= __find_get_block(bdev
, bblock
+ 1, blocksize
);
579 if (buffer_dirty(bh
))
580 ll_rw_block(WRITE
, 1, &bh
);
585 void mark_buffer_dirty_inode(struct buffer_head
*bh
, struct inode
*inode
)
587 struct address_space
*mapping
= inode
->i_mapping
;
588 struct address_space
*buffer_mapping
= bh
->b_page
->mapping
;
590 mark_buffer_dirty(bh
);
591 if (!mapping
->assoc_mapping
) {
592 mapping
->assoc_mapping
= buffer_mapping
;
594 BUG_ON(mapping
->assoc_mapping
!= buffer_mapping
);
596 if (!bh
->b_assoc_map
) {
597 spin_lock(&buffer_mapping
->private_lock
);
598 list_move_tail(&bh
->b_assoc_buffers
,
599 &mapping
->private_list
);
600 bh
->b_assoc_map
= mapping
;
601 spin_unlock(&buffer_mapping
->private_lock
);
604 EXPORT_SYMBOL(mark_buffer_dirty_inode
);
607 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
610 * If warn is true, then emit a warning if the page is not uptodate and has
611 * not been truncated.
613 static void __set_page_dirty(struct page
*page
,
614 struct address_space
*mapping
, int warn
)
616 spin_lock_irq(&mapping
->tree_lock
);
617 if (page
->mapping
) { /* Race with truncate? */
618 WARN_ON_ONCE(warn
&& !PageUptodate(page
));
619 account_page_dirtied(page
, mapping
);
620 radix_tree_tag_set(&mapping
->page_tree
,
621 page_index(page
), PAGECACHE_TAG_DIRTY
);
623 spin_unlock_irq(&mapping
->tree_lock
);
624 __mark_inode_dirty(mapping
->host
, I_DIRTY_PAGES
);
628 * Add a page to the dirty page list.
630 * It is a sad fact of life that this function is called from several places
631 * deeply under spinlocking. It may not sleep.
633 * If the page has buffers, the uptodate buffers are set dirty, to preserve
634 * dirty-state coherency between the page and the buffers. It the page does
635 * not have buffers then when they are later attached they will all be set
638 * The buffers are dirtied before the page is dirtied. There's a small race
639 * window in which a writepage caller may see the page cleanness but not the
640 * buffer dirtiness. That's fine. If this code were to set the page dirty
641 * before the buffers, a concurrent writepage caller could clear the page dirty
642 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
643 * page on the dirty page list.
645 * We use private_lock to lock against try_to_free_buffers while using the
646 * page's buffer list. Also use this to protect against clean buffers being
647 * added to the page after it was set dirty.
649 * FIXME: may need to call ->reservepage here as well. That's rather up to the
650 * address_space though.
652 int __set_page_dirty_buffers(struct page
*page
)
655 struct address_space
*mapping
= page_mapping(page
);
657 if (unlikely(!mapping
))
658 return !TestSetPageDirty(page
);
660 spin_lock(&mapping
->private_lock
);
661 if (page_has_buffers(page
)) {
662 struct buffer_head
*head
= page_buffers(page
);
663 struct buffer_head
*bh
= head
;
666 set_buffer_dirty(bh
);
667 bh
= bh
->b_this_page
;
668 } while (bh
!= head
);
670 newly_dirty
= !TestSetPageDirty(page
);
671 spin_unlock(&mapping
->private_lock
);
674 __set_page_dirty(page
, mapping
, 1);
677 EXPORT_SYMBOL(__set_page_dirty_buffers
);
680 * Write out and wait upon a list of buffers.
682 * We have conflicting pressures: we want to make sure that all
683 * initially dirty buffers get waited on, but that any subsequently
684 * dirtied buffers don't. After all, we don't want fsync to last
685 * forever if somebody is actively writing to the file.
687 * Do this in two main stages: first we copy dirty buffers to a
688 * temporary inode list, queueing the writes as we go. Then we clean
689 * up, waiting for those writes to complete.
691 * During this second stage, any subsequent updates to the file may end
692 * up refiling the buffer on the original inode's dirty list again, so
693 * there is a chance we will end up with a buffer queued for write but
694 * not yet completed on that list. So, as a final cleanup we go through
695 * the osync code to catch these locked, dirty buffers without requeuing
696 * any newly dirty buffers for write.
698 static int fsync_buffers_list(spinlock_t
*lock
, struct list_head
*list
)
700 struct buffer_head
*bh
;
701 struct list_head tmp
;
702 struct address_space
*mapping
;
704 struct blk_plug plug
;
706 INIT_LIST_HEAD(&tmp
);
707 blk_start_plug(&plug
);
710 while (!list_empty(list
)) {
711 bh
= BH_ENTRY(list
->next
);
712 mapping
= bh
->b_assoc_map
;
713 __remove_assoc_queue(bh
);
714 /* Avoid race with mark_buffer_dirty_inode() which does
715 * a lockless check and we rely on seeing the dirty bit */
717 if (buffer_dirty(bh
) || buffer_locked(bh
)) {
718 list_add(&bh
->b_assoc_buffers
, &tmp
);
719 bh
->b_assoc_map
= mapping
;
720 if (buffer_dirty(bh
)) {
724 * Ensure any pending I/O completes so that
725 * write_dirty_buffer() actually writes the
726 * current contents - it is a noop if I/O is
727 * still in flight on potentially older
730 write_dirty_buffer(bh
, WRITE_SYNC
);
733 * Kick off IO for the previous mapping. Note
734 * that we will not run the very last mapping,
735 * wait_on_buffer() will do that for us
736 * through sync_buffer().
745 blk_finish_plug(&plug
);
748 while (!list_empty(&tmp
)) {
749 bh
= BH_ENTRY(tmp
.prev
);
751 mapping
= bh
->b_assoc_map
;
752 __remove_assoc_queue(bh
);
753 /* Avoid race with mark_buffer_dirty_inode() which does
754 * a lockless check and we rely on seeing the dirty bit */
756 if (buffer_dirty(bh
)) {
757 list_add(&bh
->b_assoc_buffers
,
758 &mapping
->private_list
);
759 bh
->b_assoc_map
= mapping
;
763 if (!buffer_uptodate(bh
))
770 err2
= osync_buffers_list(lock
, list
);
778 * Invalidate any and all dirty buffers on a given inode. We are
779 * probably unmounting the fs, but that doesn't mean we have already
780 * done a sync(). Just drop the buffers from the inode list.
782 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
783 * assumes that all the buffers are against the blockdev. Not true
786 void invalidate_inode_buffers(struct inode
*inode
)
788 if (inode_has_buffers(inode
)) {
789 struct address_space
*mapping
= &inode
->i_data
;
790 struct list_head
*list
= &mapping
->private_list
;
791 struct address_space
*buffer_mapping
= mapping
->assoc_mapping
;
793 spin_lock(&buffer_mapping
->private_lock
);
794 while (!list_empty(list
))
795 __remove_assoc_queue(BH_ENTRY(list
->next
));
796 spin_unlock(&buffer_mapping
->private_lock
);
799 EXPORT_SYMBOL(invalidate_inode_buffers
);
802 * Remove any clean buffers from the inode's buffer list. This is called
803 * when we're trying to free the inode itself. Those buffers can pin it.
805 * Returns true if all buffers were removed.
807 int remove_inode_buffers(struct inode
*inode
)
811 if (inode_has_buffers(inode
)) {
812 struct address_space
*mapping
= &inode
->i_data
;
813 struct list_head
*list
= &mapping
->private_list
;
814 struct address_space
*buffer_mapping
= mapping
->assoc_mapping
;
816 spin_lock(&buffer_mapping
->private_lock
);
817 while (!list_empty(list
)) {
818 struct buffer_head
*bh
= BH_ENTRY(list
->next
);
819 if (buffer_dirty(bh
)) {
823 __remove_assoc_queue(bh
);
825 spin_unlock(&buffer_mapping
->private_lock
);
831 * Create the appropriate buffers when given a page for data area and
832 * the size of each buffer.. Use the bh->b_this_page linked list to
833 * follow the buffers created. Return NULL if unable to create more
836 * The retry flag is used to differentiate async IO (paging, swapping)
837 * which may not fail from ordinary buffer allocations.
839 struct buffer_head
*alloc_page_buffers(struct page
*page
, unsigned long size
,
842 struct buffer_head
*bh
, *head
;
848 while ((offset
-= size
) >= 0) {
849 bh
= alloc_buffer_head(GFP_NOFS
);
854 bh
->b_this_page
= head
;
859 atomic_set(&bh
->b_count
, 0);
862 /* Link the buffer to its page */
863 set_bh_page(bh
, page
, offset
);
865 init_buffer(bh
, NULL
, NULL
);
869 * In case anything failed, we just free everything we got.
875 head
= head
->b_this_page
;
876 free_buffer_head(bh
);
881 * Return failure for non-async IO requests. Async IO requests
882 * are not allowed to fail, so we have to wait until buffer heads
883 * become available. But we don't want tasks sleeping with
884 * partially complete buffers, so all were released above.
889 /* We're _really_ low on memory. Now we just
890 * wait for old buffer heads to become free due to
891 * finishing IO. Since this is an async request and
892 * the reserve list is empty, we're sure there are
893 * async buffer heads in use.
898 EXPORT_SYMBOL_GPL(alloc_page_buffers
);
901 link_dev_buffers(struct page
*page
, struct buffer_head
*head
)
903 struct buffer_head
*bh
, *tail
;
908 bh
= bh
->b_this_page
;
910 tail
->b_this_page
= head
;
911 attach_page_buffers(page
, head
);
914 static sector_t
blkdev_max_block(struct block_device
*bdev
, unsigned int size
)
916 sector_t retval
= ~((sector_t
)0);
917 loff_t sz
= i_size_read(bdev
->bd_inode
);
920 unsigned int sizebits
= blksize_bits(size
);
921 retval
= (sz
>> sizebits
);
927 * Initialise the state of a blockdev page's buffers.
930 init_page_buffers(struct page
*page
, struct block_device
*bdev
,
931 sector_t block
, int size
)
933 struct buffer_head
*head
= page_buffers(page
);
934 struct buffer_head
*bh
= head
;
935 int uptodate
= PageUptodate(page
);
936 sector_t end_block
= blkdev_max_block(I_BDEV(bdev
->bd_inode
), size
);
939 if (!buffer_mapped(bh
)) {
940 init_buffer(bh
, NULL
, NULL
);
942 bh
->b_blocknr
= block
;
944 set_buffer_uptodate(bh
);
945 if (block
< end_block
)
946 set_buffer_mapped(bh
);
949 bh
= bh
->b_this_page
;
950 } while (bh
!= head
);
953 * Caller needs to validate requested block against end of device.
959 * Create the page-cache page that contains the requested block.
961 * This is used purely for blockdev mappings.
964 grow_dev_page(struct block_device
*bdev
, sector_t block
,
965 pgoff_t index
, int size
, int sizebits
)
967 struct inode
*inode
= bdev
->bd_inode
;
969 struct buffer_head
*bh
;
971 int ret
= 0; /* Will call free_more_memory() */
973 page
= find_or_create_page(inode
->i_mapping
, index
,
974 (mapping_gfp_mask(inode
->i_mapping
) & ~__GFP_FS
)|__GFP_MOVABLE
);
978 BUG_ON(!PageLocked(page
));
980 if (page_has_buffers(page
)) {
981 bh
= page_buffers(page
);
982 if (bh
->b_size
== size
) {
983 end_block
= init_page_buffers(page
, bdev
,
984 index
<< sizebits
, size
);
987 if (!try_to_free_buffers(page
))
992 * Allocate some buffers for this page
994 bh
= alloc_page_buffers(page
, size
, 0);
999 * Link the page to the buffers and initialise them. Take the
1000 * lock to be atomic wrt __find_get_block(), which does not
1001 * run under the page lock.
1003 spin_lock(&inode
->i_mapping
->private_lock
);
1004 link_dev_buffers(page
, bh
);
1005 end_block
= init_page_buffers(page
, bdev
, index
<< sizebits
, size
);
1006 spin_unlock(&inode
->i_mapping
->private_lock
);
1008 ret
= (block
< end_block
) ? 1 : -ENXIO
;
1011 page_cache_release(page
);
1016 * Create buffers for the specified block device block's page. If
1017 * that page was dirty, the buffers are set dirty also.
1020 grow_buffers(struct block_device
*bdev
, sector_t block
, int size
)
1028 } while ((size
<< sizebits
) < PAGE_SIZE
);
1030 index
= block
>> sizebits
;
1033 * Check for a block which wants to lie outside our maximum possible
1034 * pagecache index. (this comparison is done using sector_t types).
1036 if (unlikely(index
!= block
>> sizebits
)) {
1037 char b
[BDEVNAME_SIZE
];
1039 printk(KERN_ERR
"%s: requested out-of-range block %llu for "
1041 __func__
, (unsigned long long)block
,
1046 /* Create a page with the proper size buffers.. */
1047 return grow_dev_page(bdev
, block
, index
, size
, sizebits
);
1050 static struct buffer_head
*
1051 __getblk_slow(struct block_device
*bdev
, sector_t block
, int size
)
1053 /* Size must be multiple of hard sectorsize */
1054 if (unlikely(size
& (bdev_logical_block_size(bdev
)-1) ||
1055 (size
< 512 || size
> PAGE_SIZE
))) {
1056 printk(KERN_ERR
"getblk(): invalid block size %d requested\n",
1058 printk(KERN_ERR
"logical block size: %d\n",
1059 bdev_logical_block_size(bdev
));
1066 struct buffer_head
*bh
;
1069 bh
= __find_get_block(bdev
, block
, size
);
1073 ret
= grow_buffers(bdev
, block
, size
);
1082 * The relationship between dirty buffers and dirty pages:
1084 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1085 * the page is tagged dirty in its radix tree.
1087 * At all times, the dirtiness of the buffers represents the dirtiness of
1088 * subsections of the page. If the page has buffers, the page dirty bit is
1089 * merely a hint about the true dirty state.
1091 * When a page is set dirty in its entirety, all its buffers are marked dirty
1092 * (if the page has buffers).
1094 * When a buffer is marked dirty, its page is dirtied, but the page's other
1097 * Also. When blockdev buffers are explicitly read with bread(), they
1098 * individually become uptodate. But their backing page remains not
1099 * uptodate - even if all of its buffers are uptodate. A subsequent
1100 * block_read_full_page() against that page will discover all the uptodate
1101 * buffers, will set the page uptodate and will perform no I/O.
1105 * mark_buffer_dirty - mark a buffer_head as needing writeout
1106 * @bh: the buffer_head to mark dirty
1108 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1109 * backing page dirty, then tag the page as dirty in its address_space's radix
1110 * tree and then attach the address_space's inode to its superblock's dirty
1113 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1114 * mapping->tree_lock and mapping->host->i_lock.
1116 void mark_buffer_dirty(struct buffer_head
*bh
)
1118 WARN_ON_ONCE(!buffer_uptodate(bh
));
1121 * Very *carefully* optimize the it-is-already-dirty case.
1123 * Don't let the final "is it dirty" escape to before we
1124 * perhaps modified the buffer.
1126 if (buffer_dirty(bh
)) {
1128 if (buffer_dirty(bh
))
1132 if (!test_set_buffer_dirty(bh
)) {
1133 struct page
*page
= bh
->b_page
;
1134 if (!TestSetPageDirty(page
)) {
1135 struct address_space
*mapping
= page_mapping(page
);
1137 __set_page_dirty(page
, mapping
, 0);
1141 EXPORT_SYMBOL(mark_buffer_dirty
);
1144 * Decrement a buffer_head's reference count. If all buffers against a page
1145 * have zero reference count, are clean and unlocked, and if the page is clean
1146 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1147 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1148 * a page but it ends up not being freed, and buffers may later be reattached).
1150 void __brelse(struct buffer_head
* buf
)
1152 if (atomic_read(&buf
->b_count
)) {
1156 WARN(1, KERN_ERR
"VFS: brelse: Trying to free free buffer\n");
1158 EXPORT_SYMBOL(__brelse
);
1161 * bforget() is like brelse(), except it discards any
1162 * potentially dirty data.
1164 void __bforget(struct buffer_head
*bh
)
1166 clear_buffer_dirty(bh
);
1167 if (bh
->b_assoc_map
) {
1168 struct address_space
*buffer_mapping
= bh
->b_page
->mapping
;
1170 spin_lock(&buffer_mapping
->private_lock
);
1171 list_del_init(&bh
->b_assoc_buffers
);
1172 bh
->b_assoc_map
= NULL
;
1173 spin_unlock(&buffer_mapping
->private_lock
);
1177 EXPORT_SYMBOL(__bforget
);
1179 static struct buffer_head
*__bread_slow(struct buffer_head
*bh
)
1182 if (buffer_uptodate(bh
)) {
1187 bh
->b_end_io
= end_buffer_read_sync
;
1188 submit_bh(READ
, bh
);
1190 if (buffer_uptodate(bh
))
1198 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1199 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1200 * refcount elevated by one when they're in an LRU. A buffer can only appear
1201 * once in a particular CPU's LRU. A single buffer can be present in multiple
1202 * CPU's LRUs at the same time.
1204 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1205 * sb_find_get_block().
1207 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1208 * a local interrupt disable for that.
1211 #define BH_LRU_SIZE 8
1214 struct buffer_head
*bhs
[BH_LRU_SIZE
];
1217 static DEFINE_PER_CPU(struct bh_lru
, bh_lrus
) = {{ NULL
}};
1220 #define bh_lru_lock() local_irq_disable()
1221 #define bh_lru_unlock() local_irq_enable()
1223 #define bh_lru_lock() preempt_disable()
1224 #define bh_lru_unlock() preempt_enable()
1227 static inline void check_irqs_on(void)
1229 #ifdef irqs_disabled
1230 BUG_ON(irqs_disabled());
1235 * The LRU management algorithm is dopey-but-simple. Sorry.
1237 static void bh_lru_install(struct buffer_head
*bh
)
1239 struct buffer_head
*evictee
= NULL
;
1243 if (__this_cpu_read(bh_lrus
.bhs
[0]) != bh
) {
1244 struct buffer_head
*bhs
[BH_LRU_SIZE
];
1250 for (in
= 0; in
< BH_LRU_SIZE
; in
++) {
1251 struct buffer_head
*bh2
=
1252 __this_cpu_read(bh_lrus
.bhs
[in
]);
1257 if (out
>= BH_LRU_SIZE
) {
1258 BUG_ON(evictee
!= NULL
);
1265 while (out
< BH_LRU_SIZE
)
1267 memcpy(__this_cpu_ptr(&bh_lrus
.bhs
), bhs
, sizeof(bhs
));
1276 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1278 static struct buffer_head
*
1279 lookup_bh_lru(struct block_device
*bdev
, sector_t block
, unsigned size
)
1281 struct buffer_head
*ret
= NULL
;
1286 for (i
= 0; i
< BH_LRU_SIZE
; i
++) {
1287 struct buffer_head
*bh
= __this_cpu_read(bh_lrus
.bhs
[i
]);
1289 if (bh
&& bh
->b_bdev
== bdev
&&
1290 bh
->b_blocknr
== block
&& bh
->b_size
== size
) {
1293 __this_cpu_write(bh_lrus
.bhs
[i
],
1294 __this_cpu_read(bh_lrus
.bhs
[i
- 1]));
1297 __this_cpu_write(bh_lrus
.bhs
[0], bh
);
1309 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1310 * it in the LRU and mark it as accessed. If it is not present then return
1313 struct buffer_head
*
1314 __find_get_block(struct block_device
*bdev
, sector_t block
, unsigned size
)
1316 struct buffer_head
*bh
= lookup_bh_lru(bdev
, block
, size
);
1319 bh
= __find_get_block_slow(bdev
, block
);
1327 EXPORT_SYMBOL(__find_get_block
);
1330 * __getblk will locate (and, if necessary, create) the buffer_head
1331 * which corresponds to the passed block_device, block and size. The
1332 * returned buffer has its reference count incremented.
1334 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1335 * attempt is failing. FIXME, perhaps?
1337 struct buffer_head
*
1338 __getblk(struct block_device
*bdev
, sector_t block
, unsigned size
)
1340 struct buffer_head
*bh
= __find_get_block(bdev
, block
, size
);
1344 bh
= __getblk_slow(bdev
, block
, size
);
1347 EXPORT_SYMBOL(__getblk
);
1350 * Do async read-ahead on a buffer..
1352 void __breadahead(struct block_device
*bdev
, sector_t block
, unsigned size
)
1354 struct buffer_head
*bh
= __getblk(bdev
, block
, size
);
1356 ll_rw_block(READA
, 1, &bh
);
1360 EXPORT_SYMBOL(__breadahead
);
1363 * __bread() - reads a specified block and returns the bh
1364 * @bdev: the block_device to read from
1365 * @block: number of block
1366 * @size: size (in bytes) to read
1368 * Reads a specified block, and returns buffer head that contains it.
1369 * It returns NULL if the block was unreadable.
1371 struct buffer_head
*
1372 __bread(struct block_device
*bdev
, sector_t block
, unsigned size
)
1374 struct buffer_head
*bh
= __getblk(bdev
, block
, size
);
1376 if (likely(bh
) && !buffer_uptodate(bh
))
1377 bh
= __bread_slow(bh
);
1380 EXPORT_SYMBOL(__bread
);
1383 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1384 * This doesn't race because it runs in each cpu either in irq
1385 * or with preempt disabled.
1387 static void invalidate_bh_lru(void *arg
)
1389 struct bh_lru
*b
= &get_cpu_var(bh_lrus
);
1392 for (i
= 0; i
< BH_LRU_SIZE
; i
++) {
1396 put_cpu_var(bh_lrus
);
1399 static bool has_bh_in_lru(int cpu
, void *dummy
)
1401 struct bh_lru
*b
= per_cpu_ptr(&bh_lrus
, cpu
);
1404 for (i
= 0; i
< BH_LRU_SIZE
; i
++) {
1412 void invalidate_bh_lrus(void)
1414 on_each_cpu_cond(has_bh_in_lru
, invalidate_bh_lru
, NULL
, 1, GFP_KERNEL
);
1416 EXPORT_SYMBOL_GPL(invalidate_bh_lrus
);
1418 void set_bh_page(struct buffer_head
*bh
,
1419 struct page
*page
, unsigned long offset
)
1422 BUG_ON(offset
>= PAGE_SIZE
);
1423 if (PageHighMem(page
))
1425 * This catches illegal uses and preserves the offset:
1427 bh
->b_data
= (char *)(0 + offset
);
1429 bh
->b_data
= page_address(page
) + offset
;
1431 EXPORT_SYMBOL(set_bh_page
);
1434 * Called when truncating a buffer on a page completely.
1436 static void discard_buffer(struct buffer_head
* bh
)
1439 clear_buffer_dirty(bh
);
1441 clear_buffer_mapped(bh
);
1442 clear_buffer_req(bh
);
1443 clear_buffer_new(bh
);
1444 clear_buffer_delay(bh
);
1445 clear_buffer_unwritten(bh
);
1450 * block_invalidatepage - invalidate part or all of a buffer-backed page
1452 * @page: the page which is affected
1453 * @offset: the index of the truncation point
1455 * block_invalidatepage() is called when all or part of the page has become
1456 * invalidated by a truncate operation.
1458 * block_invalidatepage() does not have to release all buffers, but it must
1459 * ensure that no dirty buffer is left outside @offset and that no I/O
1460 * is underway against any of the blocks which are outside the truncation
1461 * point. Because the caller is about to free (and possibly reuse) those
1464 void block_invalidatepage(struct page
*page
, unsigned long offset
)
1466 struct buffer_head
*head
, *bh
, *next
;
1467 unsigned int curr_off
= 0;
1469 BUG_ON(!PageLocked(page
));
1470 if (!page_has_buffers(page
))
1473 head
= page_buffers(page
);
1476 unsigned int next_off
= curr_off
+ bh
->b_size
;
1477 next
= bh
->b_this_page
;
1480 * is this block fully invalidated?
1482 if (offset
<= curr_off
)
1484 curr_off
= next_off
;
1486 } while (bh
!= head
);
1489 * We release buffers only if the entire page is being invalidated.
1490 * The get_block cached value has been unconditionally invalidated,
1491 * so real IO is not possible anymore.
1494 try_to_release_page(page
, 0);
1498 EXPORT_SYMBOL(block_invalidatepage
);
1501 * We attach and possibly dirty the buffers atomically wrt
1502 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1503 * is already excluded via the page lock.
1505 void create_empty_buffers(struct page
*page
,
1506 unsigned long blocksize
, unsigned long b_state
)
1508 struct buffer_head
*bh
, *head
, *tail
;
1510 head
= alloc_page_buffers(page
, blocksize
, 1);
1513 bh
->b_state
|= b_state
;
1515 bh
= bh
->b_this_page
;
1517 tail
->b_this_page
= head
;
1519 spin_lock(&page
->mapping
->private_lock
);
1520 if (PageUptodate(page
) || PageDirty(page
)) {
1523 if (PageDirty(page
))
1524 set_buffer_dirty(bh
);
1525 if (PageUptodate(page
))
1526 set_buffer_uptodate(bh
);
1527 bh
= bh
->b_this_page
;
1528 } while (bh
!= head
);
1530 attach_page_buffers(page
, head
);
1531 spin_unlock(&page
->mapping
->private_lock
);
1533 EXPORT_SYMBOL(create_empty_buffers
);
1536 * We are taking a block for data and we don't want any output from any
1537 * buffer-cache aliases starting from return from that function and
1538 * until the moment when something will explicitly mark the buffer
1539 * dirty (hopefully that will not happen until we will free that block ;-)
1540 * We don't even need to mark it not-uptodate - nobody can expect
1541 * anything from a newly allocated buffer anyway. We used to used
1542 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1543 * don't want to mark the alias unmapped, for example - it would confuse
1544 * anyone who might pick it with bread() afterwards...
1546 * Also.. Note that bforget() doesn't lock the buffer. So there can
1547 * be writeout I/O going on against recently-freed buffers. We don't
1548 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1549 * only if we really need to. That happens here.
1551 void unmap_underlying_metadata(struct block_device
*bdev
, sector_t block
)
1553 struct buffer_head
*old_bh
;
1557 old_bh
= __find_get_block_slow(bdev
, block
);
1559 clear_buffer_dirty(old_bh
);
1560 wait_on_buffer(old_bh
);
1561 clear_buffer_req(old_bh
);
1565 EXPORT_SYMBOL(unmap_underlying_metadata
);
1568 * Size is a power-of-two in the range 512..PAGE_SIZE,
1569 * and the case we care about most is PAGE_SIZE.
1571 * So this *could* possibly be written with those
1572 * constraints in mind (relevant mostly if some
1573 * architecture has a slow bit-scan instruction)
1575 static inline int block_size_bits(unsigned int blocksize
)
1577 return ilog2(blocksize
);
1580 static struct buffer_head
*create_page_buffers(struct page
*page
, struct inode
*inode
, unsigned int b_state
)
1582 BUG_ON(!PageLocked(page
));
1584 if (!page_has_buffers(page
))
1585 create_empty_buffers(page
, 1 << ACCESS_ONCE(inode
->i_blkbits
), b_state
);
1586 return page_buffers(page
);
1590 * NOTE! All mapped/uptodate combinations are valid:
1592 * Mapped Uptodate Meaning
1594 * No No "unknown" - must do get_block()
1595 * No Yes "hole" - zero-filled
1596 * Yes No "allocated" - allocated on disk, not read in
1597 * Yes Yes "valid" - allocated and up-to-date in memory.
1599 * "Dirty" is valid only with the last case (mapped+uptodate).
1603 * While block_write_full_page is writing back the dirty buffers under
1604 * the page lock, whoever dirtied the buffers may decide to clean them
1605 * again at any time. We handle that by only looking at the buffer
1606 * state inside lock_buffer().
1608 * If block_write_full_page() is called for regular writeback
1609 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1610 * locked buffer. This only can happen if someone has written the buffer
1611 * directly, with submit_bh(). At the address_space level PageWriteback
1612 * prevents this contention from occurring.
1614 * If block_write_full_page() is called with wbc->sync_mode ==
1615 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1616 * causes the writes to be flagged as synchronous writes.
1618 static int __block_write_full_page(struct inode
*inode
, struct page
*page
,
1619 get_block_t
*get_block
, struct writeback_control
*wbc
,
1620 bh_end_io_t
*handler
)
1624 sector_t last_block
;
1625 struct buffer_head
*bh
, *head
;
1626 unsigned int blocksize
, bbits
;
1627 int nr_underway
= 0;
1628 int write_op
= (wbc
->sync_mode
== WB_SYNC_ALL
?
1629 WRITE_SYNC
: WRITE
);
1631 head
= create_page_buffers(page
, inode
,
1632 (1 << BH_Dirty
)|(1 << BH_Uptodate
));
1635 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1636 * here, and the (potentially unmapped) buffers may become dirty at
1637 * any time. If a buffer becomes dirty here after we've inspected it
1638 * then we just miss that fact, and the page stays dirty.
1640 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1641 * handle that here by just cleaning them.
1645 blocksize
= bh
->b_size
;
1646 bbits
= block_size_bits(blocksize
);
1648 block
= (sector_t
)page
->index
<< (PAGE_CACHE_SHIFT
- bbits
);
1649 last_block
= (i_size_read(inode
) - 1) >> bbits
;
1652 * Get all the dirty buffers mapped to disk addresses and
1653 * handle any aliases from the underlying blockdev's mapping.
1656 if (block
> last_block
) {
1658 * mapped buffers outside i_size will occur, because
1659 * this page can be outside i_size when there is a
1660 * truncate in progress.
1663 * The buffer was zeroed by block_write_full_page()
1665 clear_buffer_dirty(bh
);
1666 set_buffer_uptodate(bh
);
1667 } else if ((!buffer_mapped(bh
) || buffer_delay(bh
)) &&
1669 WARN_ON(bh
->b_size
!= blocksize
);
1670 err
= get_block(inode
, block
, bh
, 1);
1673 clear_buffer_delay(bh
);
1674 if (buffer_new(bh
)) {
1675 /* blockdev mappings never come here */
1676 clear_buffer_new(bh
);
1677 unmap_underlying_metadata(bh
->b_bdev
,
1681 bh
= bh
->b_this_page
;
1683 } while (bh
!= head
);
1686 if (!buffer_mapped(bh
))
1689 * If it's a fully non-blocking write attempt and we cannot
1690 * lock the buffer then redirty the page. Note that this can
1691 * potentially cause a busy-wait loop from writeback threads
1692 * and kswapd activity, but those code paths have their own
1693 * higher-level throttling.
1695 if (wbc
->sync_mode
!= WB_SYNC_NONE
) {
1697 } else if (!trylock_buffer(bh
)) {
1698 redirty_page_for_writepage(wbc
, page
);
1701 if (test_clear_buffer_dirty(bh
)) {
1702 mark_buffer_async_write_endio(bh
, handler
);
1706 } while ((bh
= bh
->b_this_page
) != head
);
1709 * The page and its buffers are protected by PageWriteback(), so we can
1710 * drop the bh refcounts early.
1712 BUG_ON(PageWriteback(page
));
1713 set_page_writeback(page
);
1716 struct buffer_head
*next
= bh
->b_this_page
;
1717 if (buffer_async_write(bh
)) {
1718 submit_bh(write_op
, bh
);
1722 } while (bh
!= head
);
1727 if (nr_underway
== 0) {
1729 * The page was marked dirty, but the buffers were
1730 * clean. Someone wrote them back by hand with
1731 * ll_rw_block/submit_bh. A rare case.
1733 end_page_writeback(page
);
1736 * The page and buffer_heads can be released at any time from
1744 * ENOSPC, or some other error. We may already have added some
1745 * blocks to the file, so we need to write these out to avoid
1746 * exposing stale data.
1747 * The page is currently locked and not marked for writeback
1750 /* Recovery: lock and submit the mapped buffers */
1752 if (buffer_mapped(bh
) && buffer_dirty(bh
) &&
1753 !buffer_delay(bh
)) {
1755 mark_buffer_async_write_endio(bh
, handler
);
1758 * The buffer may have been set dirty during
1759 * attachment to a dirty page.
1761 clear_buffer_dirty(bh
);
1763 } while ((bh
= bh
->b_this_page
) != head
);
1765 BUG_ON(PageWriteback(page
));
1766 mapping_set_error(page
->mapping
, err
);
1767 set_page_writeback(page
);
1769 struct buffer_head
*next
= bh
->b_this_page
;
1770 if (buffer_async_write(bh
)) {
1771 clear_buffer_dirty(bh
);
1772 submit_bh(write_op
, bh
);
1776 } while (bh
!= head
);
1782 * If a page has any new buffers, zero them out here, and mark them uptodate
1783 * and dirty so they'll be written out (in order to prevent uninitialised
1784 * block data from leaking). And clear the new bit.
1786 void page_zero_new_buffers(struct page
*page
, unsigned from
, unsigned to
)
1788 unsigned int block_start
, block_end
;
1789 struct buffer_head
*head
, *bh
;
1791 BUG_ON(!PageLocked(page
));
1792 if (!page_has_buffers(page
))
1795 bh
= head
= page_buffers(page
);
1798 block_end
= block_start
+ bh
->b_size
;
1800 if (buffer_new(bh
)) {
1801 if (block_end
> from
&& block_start
< to
) {
1802 if (!PageUptodate(page
)) {
1803 unsigned start
, size
;
1805 start
= max(from
, block_start
);
1806 size
= min(to
, block_end
) - start
;
1808 zero_user(page
, start
, size
);
1809 set_buffer_uptodate(bh
);
1812 clear_buffer_new(bh
);
1813 mark_buffer_dirty(bh
);
1817 block_start
= block_end
;
1818 bh
= bh
->b_this_page
;
1819 } while (bh
!= head
);
1821 EXPORT_SYMBOL(page_zero_new_buffers
);
1823 int __block_write_begin(struct page
*page
, loff_t pos
, unsigned len
,
1824 get_block_t
*get_block
)
1826 unsigned from
= pos
& (PAGE_CACHE_SIZE
- 1);
1827 unsigned to
= from
+ len
;
1828 struct inode
*inode
= page
->mapping
->host
;
1829 unsigned block_start
, block_end
;
1832 unsigned blocksize
, bbits
;
1833 struct buffer_head
*bh
, *head
, *wait
[2], **wait_bh
=wait
;
1835 BUG_ON(!PageLocked(page
));
1836 BUG_ON(from
> PAGE_CACHE_SIZE
);
1837 BUG_ON(to
> PAGE_CACHE_SIZE
);
1840 head
= create_page_buffers(page
, inode
, 0);
1841 blocksize
= head
->b_size
;
1842 bbits
= block_size_bits(blocksize
);
1844 block
= (sector_t
)page
->index
<< (PAGE_CACHE_SHIFT
- bbits
);
1846 for(bh
= head
, block_start
= 0; bh
!= head
|| !block_start
;
1847 block
++, block_start
=block_end
, bh
= bh
->b_this_page
) {
1848 block_end
= block_start
+ blocksize
;
1849 if (block_end
<= from
|| block_start
>= to
) {
1850 if (PageUptodate(page
)) {
1851 if (!buffer_uptodate(bh
))
1852 set_buffer_uptodate(bh
);
1857 clear_buffer_new(bh
);
1858 if (!buffer_mapped(bh
)) {
1859 WARN_ON(bh
->b_size
!= blocksize
);
1860 err
= get_block(inode
, block
, bh
, 1);
1863 if (buffer_new(bh
)) {
1864 unmap_underlying_metadata(bh
->b_bdev
,
1866 if (PageUptodate(page
)) {
1867 clear_buffer_new(bh
);
1868 set_buffer_uptodate(bh
);
1869 mark_buffer_dirty(bh
);
1872 if (block_end
> to
|| block_start
< from
)
1873 zero_user_segments(page
,
1879 if (PageUptodate(page
)) {
1880 if (!buffer_uptodate(bh
))
1881 set_buffer_uptodate(bh
);
1884 if (!buffer_uptodate(bh
) && !buffer_delay(bh
) &&
1885 !buffer_unwritten(bh
) &&
1886 (block_start
< from
|| block_end
> to
)) {
1887 ll_rw_block(READ
, 1, &bh
);
1892 * If we issued read requests - let them complete.
1894 while(wait_bh
> wait
) {
1895 wait_on_buffer(*--wait_bh
);
1896 if (!buffer_uptodate(*wait_bh
))
1900 page_zero_new_buffers(page
, from
, to
);
1903 EXPORT_SYMBOL(__block_write_begin
);
1905 static int __block_commit_write(struct inode
*inode
, struct page
*page
,
1906 unsigned from
, unsigned to
)
1908 unsigned block_start
, block_end
;
1911 struct buffer_head
*bh
, *head
;
1913 bh
= head
= page_buffers(page
);
1914 blocksize
= bh
->b_size
;
1918 block_end
= block_start
+ blocksize
;
1919 if (block_end
<= from
|| block_start
>= to
) {
1920 if (!buffer_uptodate(bh
))
1923 set_buffer_uptodate(bh
);
1924 mark_buffer_dirty(bh
);
1926 clear_buffer_new(bh
);
1928 block_start
= block_end
;
1929 bh
= bh
->b_this_page
;
1930 } while (bh
!= head
);
1933 * If this is a partial write which happened to make all buffers
1934 * uptodate then we can optimize away a bogus readpage() for
1935 * the next read(). Here we 'discover' whether the page went
1936 * uptodate as a result of this (potentially partial) write.
1939 SetPageUptodate(page
);
1944 * block_write_begin takes care of the basic task of block allocation and
1945 * bringing partial write blocks uptodate first.
1947 * The filesystem needs to handle block truncation upon failure.
1949 int block_write_begin(struct address_space
*mapping
, loff_t pos
, unsigned len
,
1950 unsigned flags
, struct page
**pagep
, get_block_t
*get_block
)
1952 pgoff_t index
= pos
>> PAGE_CACHE_SHIFT
;
1956 page
= grab_cache_page_write_begin(mapping
, index
, flags
);
1960 status
= __block_write_begin(page
, pos
, len
, get_block
);
1961 if (unlikely(status
)) {
1963 page_cache_release(page
);
1970 EXPORT_SYMBOL(block_write_begin
);
1972 int block_write_end(struct file
*file
, struct address_space
*mapping
,
1973 loff_t pos
, unsigned len
, unsigned copied
,
1974 struct page
*page
, void *fsdata
)
1976 struct inode
*inode
= mapping
->host
;
1979 start
= pos
& (PAGE_CACHE_SIZE
- 1);
1981 if (unlikely(copied
< len
)) {
1983 * The buffers that were written will now be uptodate, so we
1984 * don't have to worry about a readpage reading them and
1985 * overwriting a partial write. However if we have encountered
1986 * a short write and only partially written into a buffer, it
1987 * will not be marked uptodate, so a readpage might come in and
1988 * destroy our partial write.
1990 * Do the simplest thing, and just treat any short write to a
1991 * non uptodate page as a zero-length write, and force the
1992 * caller to redo the whole thing.
1994 if (!PageUptodate(page
))
1997 page_zero_new_buffers(page
, start
+copied
, start
+len
);
1999 flush_dcache_page(page
);
2001 /* This could be a short (even 0-length) commit */
2002 __block_commit_write(inode
, page
, start
, start
+copied
);
2006 EXPORT_SYMBOL(block_write_end
);
2008 int generic_write_end(struct file
*file
, struct address_space
*mapping
,
2009 loff_t pos
, unsigned len
, unsigned copied
,
2010 struct page
*page
, void *fsdata
)
2012 struct inode
*inode
= mapping
->host
;
2013 int i_size_changed
= 0;
2015 copied
= block_write_end(file
, mapping
, pos
, len
, copied
, page
, fsdata
);
2018 * No need to use i_size_read() here, the i_size
2019 * cannot change under us because we hold i_mutex.
2021 * But it's important to update i_size while still holding page lock:
2022 * page writeout could otherwise come in and zero beyond i_size.
2024 if (pos
+copied
> inode
->i_size
) {
2025 i_size_write(inode
, pos
+copied
);
2030 page_cache_release(page
);
2033 * Don't mark the inode dirty under page lock. First, it unnecessarily
2034 * makes the holding time of page lock longer. Second, it forces lock
2035 * ordering of page lock and transaction start for journaling
2039 mark_inode_dirty(inode
);
2043 EXPORT_SYMBOL(generic_write_end
);
2046 * block_is_partially_uptodate checks whether buffers within a page are
2049 * Returns true if all buffers which correspond to a file portion
2050 * we want to read are uptodate.
2052 int block_is_partially_uptodate(struct page
*page
, read_descriptor_t
*desc
,
2055 unsigned block_start
, block_end
, blocksize
;
2057 struct buffer_head
*bh
, *head
;
2060 if (!page_has_buffers(page
))
2063 head
= page_buffers(page
);
2064 blocksize
= head
->b_size
;
2065 to
= min_t(unsigned, PAGE_CACHE_SIZE
- from
, desc
->count
);
2067 if (from
< blocksize
&& to
> PAGE_CACHE_SIZE
- blocksize
)
2073 block_end
= block_start
+ blocksize
;
2074 if (block_end
> from
&& block_start
< to
) {
2075 if (!buffer_uptodate(bh
)) {
2079 if (block_end
>= to
)
2082 block_start
= block_end
;
2083 bh
= bh
->b_this_page
;
2084 } while (bh
!= head
);
2088 EXPORT_SYMBOL(block_is_partially_uptodate
);
2091 * Generic "read page" function for block devices that have the normal
2092 * get_block functionality. This is most of the block device filesystems.
2093 * Reads the page asynchronously --- the unlock_buffer() and
2094 * set/clear_buffer_uptodate() functions propagate buffer state into the
2095 * page struct once IO has completed.
2097 int block_read_full_page(struct page
*page
, get_block_t
*get_block
)
2099 struct inode
*inode
= page
->mapping
->host
;
2100 sector_t iblock
, lblock
;
2101 struct buffer_head
*bh
, *head
, *arr
[MAX_BUF_PER_PAGE
];
2102 unsigned int blocksize
, bbits
;
2104 int fully_mapped
= 1;
2106 head
= create_page_buffers(page
, inode
, 0);
2107 blocksize
= head
->b_size
;
2108 bbits
= block_size_bits(blocksize
);
2110 iblock
= (sector_t
)page
->index
<< (PAGE_CACHE_SHIFT
- bbits
);
2111 lblock
= (i_size_read(inode
)+blocksize
-1) >> bbits
;
2117 if (buffer_uptodate(bh
))
2120 if (!buffer_mapped(bh
)) {
2124 if (iblock
< lblock
) {
2125 WARN_ON(bh
->b_size
!= blocksize
);
2126 err
= get_block(inode
, iblock
, bh
, 0);
2130 if (!buffer_mapped(bh
)) {
2131 zero_user(page
, i
* blocksize
, blocksize
);
2133 set_buffer_uptodate(bh
);
2137 * get_block() might have updated the buffer
2140 if (buffer_uptodate(bh
))
2144 } while (i
++, iblock
++, (bh
= bh
->b_this_page
) != head
);
2147 SetPageMappedToDisk(page
);
2151 * All buffers are uptodate - we can set the page uptodate
2152 * as well. But not if get_block() returned an error.
2154 if (!PageError(page
))
2155 SetPageUptodate(page
);
2160 /* Stage two: lock the buffers */
2161 for (i
= 0; i
< nr
; i
++) {
2164 mark_buffer_async_read(bh
);
2168 * Stage 3: start the IO. Check for uptodateness
2169 * inside the buffer lock in case another process reading
2170 * the underlying blockdev brought it uptodate (the sct fix).
2172 for (i
= 0; i
< nr
; i
++) {
2174 if (buffer_uptodate(bh
))
2175 end_buffer_async_read(bh
, 1);
2177 submit_bh(READ
, bh
);
2181 EXPORT_SYMBOL(block_read_full_page
);
2183 /* utility function for filesystems that need to do work on expanding
2184 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2185 * deal with the hole.
2187 int generic_cont_expand_simple(struct inode
*inode
, loff_t size
)
2189 struct address_space
*mapping
= inode
->i_mapping
;
2194 err
= inode_newsize_ok(inode
, size
);
2198 err
= pagecache_write_begin(NULL
, mapping
, size
, 0,
2199 AOP_FLAG_UNINTERRUPTIBLE
|AOP_FLAG_CONT_EXPAND
,
2204 err
= pagecache_write_end(NULL
, mapping
, size
, 0, 0, page
, fsdata
);
2210 EXPORT_SYMBOL(generic_cont_expand_simple
);
2212 static int cont_expand_zero(struct file
*file
, struct address_space
*mapping
,
2213 loff_t pos
, loff_t
*bytes
)
2215 struct inode
*inode
= mapping
->host
;
2216 unsigned blocksize
= 1 << inode
->i_blkbits
;
2219 pgoff_t index
, curidx
;
2221 unsigned zerofrom
, offset
, len
;
2224 index
= pos
>> PAGE_CACHE_SHIFT
;
2225 offset
= pos
& ~PAGE_CACHE_MASK
;
2227 while (index
> (curidx
= (curpos
= *bytes
)>>PAGE_CACHE_SHIFT
)) {
2228 zerofrom
= curpos
& ~PAGE_CACHE_MASK
;
2229 if (zerofrom
& (blocksize
-1)) {
2230 *bytes
|= (blocksize
-1);
2233 len
= PAGE_CACHE_SIZE
- zerofrom
;
2235 err
= pagecache_write_begin(file
, mapping
, curpos
, len
,
2236 AOP_FLAG_UNINTERRUPTIBLE
,
2240 zero_user(page
, zerofrom
, len
);
2241 err
= pagecache_write_end(file
, mapping
, curpos
, len
, len
,
2248 balance_dirty_pages_ratelimited(mapping
);
2251 /* page covers the boundary, find the boundary offset */
2252 if (index
== curidx
) {
2253 zerofrom
= curpos
& ~PAGE_CACHE_MASK
;
2254 /* if we will expand the thing last block will be filled */
2255 if (offset
<= zerofrom
) {
2258 if (zerofrom
& (blocksize
-1)) {
2259 *bytes
|= (blocksize
-1);
2262 len
= offset
- zerofrom
;
2264 err
= pagecache_write_begin(file
, mapping
, curpos
, len
,
2265 AOP_FLAG_UNINTERRUPTIBLE
,
2269 zero_user(page
, zerofrom
, len
);
2270 err
= pagecache_write_end(file
, mapping
, curpos
, len
, len
,
2282 * For moronic filesystems that do not allow holes in file.
2283 * We may have to extend the file.
2285 int cont_write_begin(struct file
*file
, struct address_space
*mapping
,
2286 loff_t pos
, unsigned len
, unsigned flags
,
2287 struct page
**pagep
, void **fsdata
,
2288 get_block_t
*get_block
, loff_t
*bytes
)
2290 struct inode
*inode
= mapping
->host
;
2291 unsigned blocksize
= 1 << inode
->i_blkbits
;
2295 err
= cont_expand_zero(file
, mapping
, pos
, bytes
);
2299 zerofrom
= *bytes
& ~PAGE_CACHE_MASK
;
2300 if (pos
+len
> *bytes
&& zerofrom
& (blocksize
-1)) {
2301 *bytes
|= (blocksize
-1);
2305 return block_write_begin(mapping
, pos
, len
, flags
, pagep
, get_block
);
2307 EXPORT_SYMBOL(cont_write_begin
);
2309 int block_commit_write(struct page
*page
, unsigned from
, unsigned to
)
2311 struct inode
*inode
= page
->mapping
->host
;
2312 __block_commit_write(inode
,page
,from
,to
);
2315 EXPORT_SYMBOL(block_commit_write
);
2318 * block_page_mkwrite() is not allowed to change the file size as it gets
2319 * called from a page fault handler when a page is first dirtied. Hence we must
2320 * be careful to check for EOF conditions here. We set the page up correctly
2321 * for a written page which means we get ENOSPC checking when writing into
2322 * holes and correct delalloc and unwritten extent mapping on filesystems that
2323 * support these features.
2325 * We are not allowed to take the i_mutex here so we have to play games to
2326 * protect against truncate races as the page could now be beyond EOF. Because
2327 * truncate writes the inode size before removing pages, once we have the
2328 * page lock we can determine safely if the page is beyond EOF. If it is not
2329 * beyond EOF, then the page is guaranteed safe against truncation until we
2332 * Direct callers of this function should protect against filesystem freezing
2333 * using sb_start_write() - sb_end_write() functions.
2335 int __block_page_mkwrite(struct vm_area_struct
*vma
, struct vm_fault
*vmf
,
2336 get_block_t get_block
)
2338 struct page
*page
= vmf
->page
;
2339 struct inode
*inode
= vma
->vm_file
->f_path
.dentry
->d_inode
;
2345 size
= i_size_read(inode
);
2346 if ((page
->mapping
!= inode
->i_mapping
) ||
2347 (page_offset(page
) > size
)) {
2348 /* We overload EFAULT to mean page got truncated */
2353 /* page is wholly or partially inside EOF */
2354 if (((page
->index
+ 1) << PAGE_CACHE_SHIFT
) > size
)
2355 end
= size
& ~PAGE_CACHE_MASK
;
2357 end
= PAGE_CACHE_SIZE
;
2359 ret
= __block_write_begin(page
, 0, end
, get_block
);
2361 ret
= block_commit_write(page
, 0, end
);
2363 if (unlikely(ret
< 0))
2365 set_page_dirty(page
);
2366 wait_on_page_writeback(page
);
2372 EXPORT_SYMBOL(__block_page_mkwrite
);
2374 int block_page_mkwrite(struct vm_area_struct
*vma
, struct vm_fault
*vmf
,
2375 get_block_t get_block
)
2378 struct super_block
*sb
= vma
->vm_file
->f_path
.dentry
->d_inode
->i_sb
;
2380 sb_start_pagefault(sb
);
2383 * Update file times before taking page lock. We may end up failing the
2384 * fault so this update may be superfluous but who really cares...
2386 file_update_time(vma
->vm_file
);
2388 ret
= __block_page_mkwrite(vma
, vmf
, get_block
);
2389 sb_end_pagefault(sb
);
2390 return block_page_mkwrite_return(ret
);
2392 EXPORT_SYMBOL(block_page_mkwrite
);
2395 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2396 * immediately, while under the page lock. So it needs a special end_io
2397 * handler which does not touch the bh after unlocking it.
2399 static void end_buffer_read_nobh(struct buffer_head
*bh
, int uptodate
)
2401 __end_buffer_read_notouch(bh
, uptodate
);
2405 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2406 * the page (converting it to circular linked list and taking care of page
2409 static void attach_nobh_buffers(struct page
*page
, struct buffer_head
*head
)
2411 struct buffer_head
*bh
;
2413 BUG_ON(!PageLocked(page
));
2415 spin_lock(&page
->mapping
->private_lock
);
2418 if (PageDirty(page
))
2419 set_buffer_dirty(bh
);
2420 if (!bh
->b_this_page
)
2421 bh
->b_this_page
= head
;
2422 bh
= bh
->b_this_page
;
2423 } while (bh
!= head
);
2424 attach_page_buffers(page
, head
);
2425 spin_unlock(&page
->mapping
->private_lock
);
2429 * On entry, the page is fully not uptodate.
2430 * On exit the page is fully uptodate in the areas outside (from,to)
2431 * The filesystem needs to handle block truncation upon failure.
2433 int nobh_write_begin(struct address_space
*mapping
,
2434 loff_t pos
, unsigned len
, unsigned flags
,
2435 struct page
**pagep
, void **fsdata
,
2436 get_block_t
*get_block
)
2438 struct inode
*inode
= mapping
->host
;
2439 const unsigned blkbits
= inode
->i_blkbits
;
2440 const unsigned blocksize
= 1 << blkbits
;
2441 struct buffer_head
*head
, *bh
;
2445 unsigned block_in_page
;
2446 unsigned block_start
, block_end
;
2447 sector_t block_in_file
;
2450 int is_mapped_to_disk
= 1;
2452 index
= pos
>> PAGE_CACHE_SHIFT
;
2453 from
= pos
& (PAGE_CACHE_SIZE
- 1);
2456 page
= grab_cache_page_write_begin(mapping
, index
, flags
);
2462 if (page_has_buffers(page
)) {
2463 ret
= __block_write_begin(page
, pos
, len
, get_block
);
2469 if (PageMappedToDisk(page
))
2473 * Allocate buffers so that we can keep track of state, and potentially
2474 * attach them to the page if an error occurs. In the common case of
2475 * no error, they will just be freed again without ever being attached
2476 * to the page (which is all OK, because we're under the page lock).
2478 * Be careful: the buffer linked list is a NULL terminated one, rather
2479 * than the circular one we're used to.
2481 head
= alloc_page_buffers(page
, blocksize
, 0);
2487 block_in_file
= (sector_t
)page
->index
<< (PAGE_CACHE_SHIFT
- blkbits
);
2490 * We loop across all blocks in the page, whether or not they are
2491 * part of the affected region. This is so we can discover if the
2492 * page is fully mapped-to-disk.
2494 for (block_start
= 0, block_in_page
= 0, bh
= head
;
2495 block_start
< PAGE_CACHE_SIZE
;
2496 block_in_page
++, block_start
+= blocksize
, bh
= bh
->b_this_page
) {
2499 block_end
= block_start
+ blocksize
;
2502 if (block_start
>= to
)
2504 ret
= get_block(inode
, block_in_file
+ block_in_page
,
2508 if (!buffer_mapped(bh
))
2509 is_mapped_to_disk
= 0;
2511 unmap_underlying_metadata(bh
->b_bdev
, bh
->b_blocknr
);
2512 if (PageUptodate(page
)) {
2513 set_buffer_uptodate(bh
);
2516 if (buffer_new(bh
) || !buffer_mapped(bh
)) {
2517 zero_user_segments(page
, block_start
, from
,
2521 if (buffer_uptodate(bh
))
2522 continue; /* reiserfs does this */
2523 if (block_start
< from
|| block_end
> to
) {
2525 bh
->b_end_io
= end_buffer_read_nobh
;
2526 submit_bh(READ
, bh
);
2533 * The page is locked, so these buffers are protected from
2534 * any VM or truncate activity. Hence we don't need to care
2535 * for the buffer_head refcounts.
2537 for (bh
= head
; bh
; bh
= bh
->b_this_page
) {
2539 if (!buffer_uptodate(bh
))
2546 if (is_mapped_to_disk
)
2547 SetPageMappedToDisk(page
);
2549 *fsdata
= head
; /* to be released by nobh_write_end */
2556 * Error recovery is a bit difficult. We need to zero out blocks that
2557 * were newly allocated, and dirty them to ensure they get written out.
2558 * Buffers need to be attached to the page at this point, otherwise
2559 * the handling of potential IO errors during writeout would be hard
2560 * (could try doing synchronous writeout, but what if that fails too?)
2562 attach_nobh_buffers(page
, head
);
2563 page_zero_new_buffers(page
, from
, to
);
2567 page_cache_release(page
);
2572 EXPORT_SYMBOL(nobh_write_begin
);
2574 int nobh_write_end(struct file
*file
, struct address_space
*mapping
,
2575 loff_t pos
, unsigned len
, unsigned copied
,
2576 struct page
*page
, void *fsdata
)
2578 struct inode
*inode
= page
->mapping
->host
;
2579 struct buffer_head
*head
= fsdata
;
2580 struct buffer_head
*bh
;
2581 BUG_ON(fsdata
!= NULL
&& page_has_buffers(page
));
2583 if (unlikely(copied
< len
) && head
)
2584 attach_nobh_buffers(page
, head
);
2585 if (page_has_buffers(page
))
2586 return generic_write_end(file
, mapping
, pos
, len
,
2587 copied
, page
, fsdata
);
2589 SetPageUptodate(page
);
2590 set_page_dirty(page
);
2591 if (pos
+copied
> inode
->i_size
) {
2592 i_size_write(inode
, pos
+copied
);
2593 mark_inode_dirty(inode
);
2597 page_cache_release(page
);
2601 head
= head
->b_this_page
;
2602 free_buffer_head(bh
);
2607 EXPORT_SYMBOL(nobh_write_end
);
2610 * nobh_writepage() - based on block_full_write_page() except
2611 * that it tries to operate without attaching bufferheads to
2614 int nobh_writepage(struct page
*page
, get_block_t
*get_block
,
2615 struct writeback_control
*wbc
)
2617 struct inode
* const inode
= page
->mapping
->host
;
2618 loff_t i_size
= i_size_read(inode
);
2619 const pgoff_t end_index
= i_size
>> PAGE_CACHE_SHIFT
;
2623 /* Is the page fully inside i_size? */
2624 if (page
->index
< end_index
)
2627 /* Is the page fully outside i_size? (truncate in progress) */
2628 offset
= i_size
& (PAGE_CACHE_SIZE
-1);
2629 if (page
->index
>= end_index
+1 || !offset
) {
2631 * The page may have dirty, unmapped buffers. For example,
2632 * they may have been added in ext3_writepage(). Make them
2633 * freeable here, so the page does not leak.
2636 /* Not really sure about this - do we need this ? */
2637 if (page
->mapping
->a_ops
->invalidatepage
)
2638 page
->mapping
->a_ops
->invalidatepage(page
, offset
);
2641 return 0; /* don't care */
2645 * The page straddles i_size. It must be zeroed out on each and every
2646 * writepage invocation because it may be mmapped. "A file is mapped
2647 * in multiples of the page size. For a file that is not a multiple of
2648 * the page size, the remaining memory is zeroed when mapped, and
2649 * writes to that region are not written out to the file."
2651 zero_user_segment(page
, offset
, PAGE_CACHE_SIZE
);
2653 ret
= mpage_writepage(page
, get_block
, wbc
);
2655 ret
= __block_write_full_page(inode
, page
, get_block
, wbc
,
2656 end_buffer_async_write
);
2659 EXPORT_SYMBOL(nobh_writepage
);
2661 int nobh_truncate_page(struct address_space
*mapping
,
2662 loff_t from
, get_block_t
*get_block
)
2664 pgoff_t index
= from
>> PAGE_CACHE_SHIFT
;
2665 unsigned offset
= from
& (PAGE_CACHE_SIZE
-1);
2668 unsigned length
, pos
;
2669 struct inode
*inode
= mapping
->host
;
2671 struct buffer_head map_bh
;
2674 blocksize
= 1 << inode
->i_blkbits
;
2675 length
= offset
& (blocksize
- 1);
2677 /* Block boundary? Nothing to do */
2681 length
= blocksize
- length
;
2682 iblock
= (sector_t
)index
<< (PAGE_CACHE_SHIFT
- inode
->i_blkbits
);
2684 page
= grab_cache_page(mapping
, index
);
2689 if (page_has_buffers(page
)) {
2692 page_cache_release(page
);
2693 return block_truncate_page(mapping
, from
, get_block
);
2696 /* Find the buffer that contains "offset" */
2698 while (offset
>= pos
) {
2703 map_bh
.b_size
= blocksize
;
2705 err
= get_block(inode
, iblock
, &map_bh
, 0);
2708 /* unmapped? It's a hole - nothing to do */
2709 if (!buffer_mapped(&map_bh
))
2712 /* Ok, it's mapped. Make sure it's up-to-date */
2713 if (!PageUptodate(page
)) {
2714 err
= mapping
->a_ops
->readpage(NULL
, page
);
2716 page_cache_release(page
);
2720 if (!PageUptodate(page
)) {
2724 if (page_has_buffers(page
))
2727 zero_user(page
, offset
, length
);
2728 set_page_dirty(page
);
2733 page_cache_release(page
);
2737 EXPORT_SYMBOL(nobh_truncate_page
);
2739 int block_truncate_page(struct address_space
*mapping
,
2740 loff_t from
, get_block_t
*get_block
)
2742 pgoff_t index
= from
>> PAGE_CACHE_SHIFT
;
2743 unsigned offset
= from
& (PAGE_CACHE_SIZE
-1);
2746 unsigned length
, pos
;
2747 struct inode
*inode
= mapping
->host
;
2749 struct buffer_head
*bh
;
2752 blocksize
= 1 << inode
->i_blkbits
;
2753 length
= offset
& (blocksize
- 1);
2755 /* Block boundary? Nothing to do */
2759 length
= blocksize
- length
;
2760 iblock
= (sector_t
)index
<< (PAGE_CACHE_SHIFT
- inode
->i_blkbits
);
2762 page
= grab_cache_page(mapping
, index
);
2767 if (!page_has_buffers(page
))
2768 create_empty_buffers(page
, blocksize
, 0);
2770 /* Find the buffer that contains "offset" */
2771 bh
= page_buffers(page
);
2773 while (offset
>= pos
) {
2774 bh
= bh
->b_this_page
;
2780 if (!buffer_mapped(bh
)) {
2781 WARN_ON(bh
->b_size
!= blocksize
);
2782 err
= get_block(inode
, iblock
, bh
, 0);
2785 /* unmapped? It's a hole - nothing to do */
2786 if (!buffer_mapped(bh
))
2790 /* Ok, it's mapped. Make sure it's up-to-date */
2791 if (PageUptodate(page
))
2792 set_buffer_uptodate(bh
);
2794 if (!buffer_uptodate(bh
) && !buffer_delay(bh
) && !buffer_unwritten(bh
)) {
2796 ll_rw_block(READ
, 1, &bh
);
2798 /* Uhhuh. Read error. Complain and punt. */
2799 if (!buffer_uptodate(bh
))
2803 zero_user(page
, offset
, length
);
2804 mark_buffer_dirty(bh
);
2809 page_cache_release(page
);
2813 EXPORT_SYMBOL(block_truncate_page
);
2816 * The generic ->writepage function for buffer-backed address_spaces
2817 * this form passes in the end_io handler used to finish the IO.
2819 int block_write_full_page_endio(struct page
*page
, get_block_t
*get_block
,
2820 struct writeback_control
*wbc
, bh_end_io_t
*handler
)
2822 struct inode
* const inode
= page
->mapping
->host
;
2823 loff_t i_size
= i_size_read(inode
);
2824 const pgoff_t end_index
= i_size
>> PAGE_CACHE_SHIFT
;
2827 /* Is the page fully inside i_size? */
2828 if (page
->index
< end_index
)
2829 return __block_write_full_page(inode
, page
, get_block
, wbc
,
2832 /* Is the page fully outside i_size? (truncate in progress) */
2833 offset
= i_size
& (PAGE_CACHE_SIZE
-1);
2834 if (page
->index
>= end_index
+1 || !offset
) {
2836 * The page may have dirty, unmapped buffers. For example,
2837 * they may have been added in ext3_writepage(). Make them
2838 * freeable here, so the page does not leak.
2840 do_invalidatepage(page
, 0);
2842 return 0; /* don't care */
2846 * The page straddles i_size. It must be zeroed out on each and every
2847 * writepage invocation because it may be mmapped. "A file is mapped
2848 * in multiples of the page size. For a file that is not a multiple of
2849 * the page size, the remaining memory is zeroed when mapped, and
2850 * writes to that region are not written out to the file."
2852 zero_user_segment(page
, offset
, PAGE_CACHE_SIZE
);
2853 return __block_write_full_page(inode
, page
, get_block
, wbc
, handler
);
2855 EXPORT_SYMBOL(block_write_full_page_endio
);
2858 * The generic ->writepage function for buffer-backed address_spaces
2860 int block_write_full_page(struct page
*page
, get_block_t
*get_block
,
2861 struct writeback_control
*wbc
)
2863 return block_write_full_page_endio(page
, get_block
, wbc
,
2864 end_buffer_async_write
);
2866 EXPORT_SYMBOL(block_write_full_page
);
2868 sector_t
generic_block_bmap(struct address_space
*mapping
, sector_t block
,
2869 get_block_t
*get_block
)
2871 struct buffer_head tmp
;
2872 struct inode
*inode
= mapping
->host
;
2875 tmp
.b_size
= 1 << inode
->i_blkbits
;
2876 get_block(inode
, block
, &tmp
, 0);
2877 return tmp
.b_blocknr
;
2879 EXPORT_SYMBOL(generic_block_bmap
);
2881 static void end_bio_bh_io_sync(struct bio
*bio
, int err
)
2883 struct buffer_head
*bh
= bio
->bi_private
;
2885 if (err
== -EOPNOTSUPP
) {
2886 set_bit(BIO_EOPNOTSUPP
, &bio
->bi_flags
);
2889 if (unlikely (test_bit(BIO_QUIET
,&bio
->bi_flags
)))
2890 set_bit(BH_Quiet
, &bh
->b_state
);
2892 bh
->b_end_io(bh
, test_bit(BIO_UPTODATE
, &bio
->bi_flags
));
2896 int submit_bh(int rw
, struct buffer_head
* bh
)
2901 BUG_ON(!buffer_locked(bh
));
2902 BUG_ON(!buffer_mapped(bh
));
2903 BUG_ON(!bh
->b_end_io
);
2904 BUG_ON(buffer_delay(bh
));
2905 BUG_ON(buffer_unwritten(bh
));
2908 * Only clear out a write error when rewriting
2910 if (test_set_buffer_req(bh
) && (rw
& WRITE
))
2911 clear_buffer_write_io_error(bh
);
2914 * from here on down, it's all bio -- do the initial mapping,
2915 * submit_bio -> generic_make_request may further map this bio around
2917 bio
= bio_alloc(GFP_NOIO
, 1);
2919 bio
->bi_sector
= bh
->b_blocknr
* (bh
->b_size
>> 9);
2920 bio
->bi_bdev
= bh
->b_bdev
;
2921 bio
->bi_io_vec
[0].bv_page
= bh
->b_page
;
2922 bio
->bi_io_vec
[0].bv_len
= bh
->b_size
;
2923 bio
->bi_io_vec
[0].bv_offset
= bh_offset(bh
);
2927 bio
->bi_size
= bh
->b_size
;
2929 bio
->bi_end_io
= end_bio_bh_io_sync
;
2930 bio
->bi_private
= bh
;
2933 submit_bio(rw
, bio
);
2935 if (bio_flagged(bio
, BIO_EOPNOTSUPP
))
2941 EXPORT_SYMBOL(submit_bh
);
2944 * ll_rw_block: low-level access to block devices (DEPRECATED)
2945 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2946 * @nr: number of &struct buffer_heads in the array
2947 * @bhs: array of pointers to &struct buffer_head
2949 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2950 * requests an I/O operation on them, either a %READ or a %WRITE. The third
2951 * %READA option is described in the documentation for generic_make_request()
2952 * which ll_rw_block() calls.
2954 * This function drops any buffer that it cannot get a lock on (with the
2955 * BH_Lock state bit), any buffer that appears to be clean when doing a write
2956 * request, and any buffer that appears to be up-to-date when doing read
2957 * request. Further it marks as clean buffers that are processed for
2958 * writing (the buffer cache won't assume that they are actually clean
2959 * until the buffer gets unlocked).
2961 * ll_rw_block sets b_end_io to simple completion handler that marks
2962 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2965 * All of the buffers must be for the same device, and must also be a
2966 * multiple of the current approved size for the device.
2968 void ll_rw_block(int rw
, int nr
, struct buffer_head
*bhs
[])
2972 for (i
= 0; i
< nr
; i
++) {
2973 struct buffer_head
*bh
= bhs
[i
];
2975 if (!trylock_buffer(bh
))
2978 if (test_clear_buffer_dirty(bh
)) {
2979 bh
->b_end_io
= end_buffer_write_sync
;
2981 submit_bh(WRITE
, bh
);
2985 if (!buffer_uptodate(bh
)) {
2986 bh
->b_end_io
= end_buffer_read_sync
;
2995 EXPORT_SYMBOL(ll_rw_block
);
2997 void write_dirty_buffer(struct buffer_head
*bh
, int rw
)
3000 if (!test_clear_buffer_dirty(bh
)) {
3004 bh
->b_end_io
= end_buffer_write_sync
;
3008 EXPORT_SYMBOL(write_dirty_buffer
);
3011 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3012 * and then start new I/O and then wait upon it. The caller must have a ref on
3015 int __sync_dirty_buffer(struct buffer_head
*bh
, int rw
)
3019 WARN_ON(atomic_read(&bh
->b_count
) < 1);
3021 if (test_clear_buffer_dirty(bh
)) {
3023 bh
->b_end_io
= end_buffer_write_sync
;
3024 ret
= submit_bh(rw
, bh
);
3026 if (!ret
&& !buffer_uptodate(bh
))
3033 EXPORT_SYMBOL(__sync_dirty_buffer
);
3035 int sync_dirty_buffer(struct buffer_head
*bh
)
3037 return __sync_dirty_buffer(bh
, WRITE_SYNC
);
3039 EXPORT_SYMBOL(sync_dirty_buffer
);
3042 * try_to_free_buffers() checks if all the buffers on this particular page
3043 * are unused, and releases them if so.
3045 * Exclusion against try_to_free_buffers may be obtained by either
3046 * locking the page or by holding its mapping's private_lock.
3048 * If the page is dirty but all the buffers are clean then we need to
3049 * be sure to mark the page clean as well. This is because the page
3050 * may be against a block device, and a later reattachment of buffers
3051 * to a dirty page will set *all* buffers dirty. Which would corrupt
3052 * filesystem data on the same device.
3054 * The same applies to regular filesystem pages: if all the buffers are
3055 * clean then we set the page clean and proceed. To do that, we require
3056 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3059 * try_to_free_buffers() is non-blocking.
3061 static inline int buffer_busy(struct buffer_head
*bh
)
3063 return atomic_read(&bh
->b_count
) |
3064 (bh
->b_state
& ((1 << BH_Dirty
) | (1 << BH_Lock
)));
3068 drop_buffers(struct page
*page
, struct buffer_head
**buffers_to_free
)
3070 struct buffer_head
*head
= page_buffers(page
);
3071 struct buffer_head
*bh
;
3075 if (buffer_write_io_error(bh
) && page
->mapping
)
3076 set_bit(AS_EIO
, &page
->mapping
->flags
);
3077 if (buffer_busy(bh
))
3079 bh
= bh
->b_this_page
;
3080 } while (bh
!= head
);
3083 struct buffer_head
*next
= bh
->b_this_page
;
3085 if (bh
->b_assoc_map
)
3086 __remove_assoc_queue(bh
);
3088 } while (bh
!= head
);
3089 *buffers_to_free
= head
;
3090 __clear_page_buffers(page
);
3096 int try_to_free_buffers(struct page
*page
)
3098 struct address_space
* const mapping
= page
->mapping
;
3099 struct buffer_head
*buffers_to_free
= NULL
;
3102 BUG_ON(!PageLocked(page
));
3103 if (PageWriteback(page
))
3106 if (mapping
== NULL
) { /* can this still happen? */
3107 ret
= drop_buffers(page
, &buffers_to_free
);
3111 spin_lock(&mapping
->private_lock
);
3112 ret
= drop_buffers(page
, &buffers_to_free
);
3115 * If the filesystem writes its buffers by hand (eg ext3)
3116 * then we can have clean buffers against a dirty page. We
3117 * clean the page here; otherwise the VM will never notice
3118 * that the filesystem did any IO at all.
3120 * Also, during truncate, discard_buffer will have marked all
3121 * the page's buffers clean. We discover that here and clean
3124 * private_lock must be held over this entire operation in order
3125 * to synchronise against __set_page_dirty_buffers and prevent the
3126 * dirty bit from being lost.
3129 cancel_dirty_page(page
, PAGE_CACHE_SIZE
);
3130 spin_unlock(&mapping
->private_lock
);
3132 if (buffers_to_free
) {
3133 struct buffer_head
*bh
= buffers_to_free
;
3136 struct buffer_head
*next
= bh
->b_this_page
;
3137 free_buffer_head(bh
);
3139 } while (bh
!= buffers_to_free
);
3143 EXPORT_SYMBOL(try_to_free_buffers
);
3146 * There are no bdflush tunables left. But distributions are
3147 * still running obsolete flush daemons, so we terminate them here.
3149 * Use of bdflush() is deprecated and will be removed in a future kernel.
3150 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3152 SYSCALL_DEFINE2(bdflush
, int, func
, long, data
)
3154 static int msg_count
;
3156 if (!capable(CAP_SYS_ADMIN
))
3159 if (msg_count
< 5) {
3162 "warning: process `%s' used the obsolete bdflush"
3163 " system call\n", current
->comm
);
3164 printk(KERN_INFO
"Fix your initscripts?\n");
3173 * Buffer-head allocation
3175 static struct kmem_cache
*bh_cachep __read_mostly
;
3178 * Once the number of bh's in the machine exceeds this level, we start
3179 * stripping them in writeback.
3181 static int max_buffer_heads
;
3183 int buffer_heads_over_limit
;
3185 struct bh_accounting
{
3186 int nr
; /* Number of live bh's */
3187 int ratelimit
; /* Limit cacheline bouncing */
3190 static DEFINE_PER_CPU(struct bh_accounting
, bh_accounting
) = {0, 0};
3192 static void recalc_bh_state(void)
3197 if (__this_cpu_inc_return(bh_accounting
.ratelimit
) - 1 < 4096)
3199 __this_cpu_write(bh_accounting
.ratelimit
, 0);
3200 for_each_online_cpu(i
)
3201 tot
+= per_cpu(bh_accounting
, i
).nr
;
3202 buffer_heads_over_limit
= (tot
> max_buffer_heads
);
3205 struct buffer_head
*alloc_buffer_head(gfp_t gfp_flags
)
3207 struct buffer_head
*ret
= kmem_cache_zalloc(bh_cachep
, gfp_flags
);
3209 INIT_LIST_HEAD(&ret
->b_assoc_buffers
);
3211 __this_cpu_inc(bh_accounting
.nr
);
3217 EXPORT_SYMBOL(alloc_buffer_head
);
3219 void free_buffer_head(struct buffer_head
*bh
)
3221 BUG_ON(!list_empty(&bh
->b_assoc_buffers
));
3222 kmem_cache_free(bh_cachep
, bh
);
3224 __this_cpu_dec(bh_accounting
.nr
);
3228 EXPORT_SYMBOL(free_buffer_head
);
3230 static void buffer_exit_cpu(int cpu
)
3233 struct bh_lru
*b
= &per_cpu(bh_lrus
, cpu
);
3235 for (i
= 0; i
< BH_LRU_SIZE
; i
++) {
3239 this_cpu_add(bh_accounting
.nr
, per_cpu(bh_accounting
, cpu
).nr
);
3240 per_cpu(bh_accounting
, cpu
).nr
= 0;
3243 static int buffer_cpu_notify(struct notifier_block
*self
,
3244 unsigned long action
, void *hcpu
)
3246 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
)
3247 buffer_exit_cpu((unsigned long)hcpu
);
3252 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3253 * @bh: struct buffer_head
3255 * Return true if the buffer is up-to-date and false,
3256 * with the buffer locked, if not.
3258 int bh_uptodate_or_lock(struct buffer_head
*bh
)
3260 if (!buffer_uptodate(bh
)) {
3262 if (!buffer_uptodate(bh
))
3268 EXPORT_SYMBOL(bh_uptodate_or_lock
);
3271 * bh_submit_read - Submit a locked buffer for reading
3272 * @bh: struct buffer_head
3274 * Returns zero on success and -EIO on error.
3276 int bh_submit_read(struct buffer_head
*bh
)
3278 BUG_ON(!buffer_locked(bh
));
3280 if (buffer_uptodate(bh
)) {
3286 bh
->b_end_io
= end_buffer_read_sync
;
3287 submit_bh(READ
, bh
);
3289 if (buffer_uptodate(bh
))
3293 EXPORT_SYMBOL(bh_submit_read
);
3295 void __init
buffer_init(void)
3299 bh_cachep
= kmem_cache_create("buffer_head",
3300 sizeof(struct buffer_head
), 0,
3301 (SLAB_RECLAIM_ACCOUNT
|SLAB_PANIC
|
3306 * Limit the bh occupancy to 10% of ZONE_NORMAL
3308 nrpages
= (nr_free_buffer_pages() * 10) / 100;
3309 max_buffer_heads
= nrpages
* (PAGE_SIZE
/ sizeof(struct buffer_head
));
3310 hotcpu_notifier(buffer_cpu_notify
, 0);