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/module.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 sync_buffer(void *word)
59 struct block_device *bd;
60 struct buffer_head *bh
61 = container_of(word, struct buffer_head, b_state);
66 blk_run_address_space(bd->bd_inode->i_mapping);
71 void __lock_buffer(struct buffer_head *bh)
73 wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
74 TASK_UNINTERRUPTIBLE);
76 EXPORT_SYMBOL(__lock_buffer);
78 void unlock_buffer(struct buffer_head *bh)
80 clear_bit_unlock(BH_Lock, &bh->b_state);
81 smp_mb__after_clear_bit();
82 wake_up_bit(&bh->b_state, BH_Lock);
84 EXPORT_SYMBOL(unlock_buffer);
87 * Block until a buffer comes unlocked. This doesn't stop it
88 * from becoming locked again - you have to lock it yourself
89 * if you want to preserve its state.
91 void __wait_on_buffer(struct buffer_head * bh)
93 wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
95 EXPORT_SYMBOL(__wait_on_buffer);
98 __clear_page_buffers(struct page *page)
100 ClearPagePrivate(page);
101 set_page_private(page, 0);
102 page_cache_release(page);
106 static int quiet_error(struct buffer_head *bh)
108 if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
114 static void buffer_io_error(struct buffer_head *bh)
116 char b[BDEVNAME_SIZE];
117 printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
118 bdevname(bh->b_bdev, b),
119 (unsigned long long)bh->b_blocknr);
123 * End-of-IO handler helper function which does not touch the bh after
125 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
126 * a race there is benign: unlock_buffer() only use the bh's address for
127 * hashing after unlocking the buffer, so it doesn't actually touch the bh
130 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
133 set_buffer_uptodate(bh);
135 /* This happens, due to failed READA attempts. */
136 clear_buffer_uptodate(bh);
142 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
143 * unlock the buffer. This is what ll_rw_block uses too.
145 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
147 __end_buffer_read_notouch(bh, uptodate);
150 EXPORT_SYMBOL(end_buffer_read_sync);
152 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
154 char b[BDEVNAME_SIZE];
157 set_buffer_uptodate(bh);
159 if (!buffer_eopnotsupp(bh) && !quiet_error(bh)) {
161 printk(KERN_WARNING "lost page write due to "
163 bdevname(bh->b_bdev, b));
165 set_buffer_write_io_error(bh);
166 clear_buffer_uptodate(bh);
171 EXPORT_SYMBOL(end_buffer_write_sync);
174 * Various filesystems appear to want __find_get_block to be non-blocking.
175 * But it's the page lock which protects the buffers. To get around this,
176 * we get exclusion from try_to_free_buffers with the blockdev mapping's
179 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
180 * may be quite high. This code could TryLock the page, and if that
181 * succeeds, there is no need to take private_lock. (But if
182 * private_lock is contended then so is mapping->tree_lock).
184 static struct buffer_head *
185 __find_get_block_slow(struct block_device *bdev, sector_t block)
187 struct inode *bd_inode = bdev->bd_inode;
188 struct address_space *bd_mapping = bd_inode->i_mapping;
189 struct buffer_head *ret = NULL;
191 struct buffer_head *bh;
192 struct buffer_head *head;
196 index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
197 page = find_get_page(bd_mapping, index);
201 spin_lock(&bd_mapping->private_lock);
202 if (!page_has_buffers(page))
204 head = page_buffers(page);
207 if (!buffer_mapped(bh))
209 else if (bh->b_blocknr == block) {
214 bh = bh->b_this_page;
215 } while (bh != head);
217 /* we might be here because some of the buffers on this page are
218 * not mapped. This is due to various races between
219 * file io on the block device and getblk. It gets dealt with
220 * elsewhere, don't buffer_error if we had some unmapped buffers
223 printk("__find_get_block_slow() failed. "
224 "block=%llu, b_blocknr=%llu\n",
225 (unsigned long long)block,
226 (unsigned long long)bh->b_blocknr);
227 printk("b_state=0x%08lx, b_size=%zu\n",
228 bh->b_state, bh->b_size);
229 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
232 spin_unlock(&bd_mapping->private_lock);
233 page_cache_release(page);
238 /* If invalidate_buffers() will trash dirty buffers, it means some kind
239 of fs corruption is going on. Trashing dirty data always imply losing
240 information that was supposed to be just stored on the physical layer
243 Thus invalidate_buffers in general usage is not allwowed to trash
244 dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
245 be preserved. These buffers are simply skipped.
247 We also skip buffers which are still in use. For example this can
248 happen if a userspace program is reading the block device.
250 NOTE: In the case where the user removed a removable-media-disk even if
251 there's still dirty data not synced on disk (due a bug in the device driver
252 or due an error of the user), by not destroying the dirty buffers we could
253 generate corruption also on the next media inserted, thus a parameter is
254 necessary to handle this case in the most safe way possible (trying
255 to not corrupt also the new disk inserted with the data belonging to
256 the old now corrupted disk). Also for the ramdisk the natural thing
257 to do in order to release the ramdisk memory is to destroy dirty buffers.
259 These are two special cases. Normal usage imply the device driver
260 to issue a sync on the device (without waiting I/O completion) and
261 then an invalidate_buffers call that doesn't trash dirty buffers.
263 For handling cache coherency with the blkdev pagecache the 'update' case
264 is been introduced. It is needed to re-read from disk any pinned
265 buffer. NOTE: re-reading from disk is destructive so we can do it only
266 when we assume nobody is changing the buffercache under our I/O and when
267 we think the disk contains more recent information than the buffercache.
268 The update == 1 pass marks the buffers we need to update, the update == 2
269 pass does the actual I/O. */
270 void invalidate_bdev(struct block_device *bdev)
272 struct address_space *mapping = bdev->bd_inode->i_mapping;
274 if (mapping->nrpages == 0)
277 invalidate_bh_lrus();
278 lru_add_drain_all(); /* make sure all lru add caches are flushed */
279 invalidate_mapping_pages(mapping, 0, -1);
281 EXPORT_SYMBOL(invalidate_bdev);
284 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
286 static void free_more_memory(void)
291 wakeup_flusher_threads(1024);
294 for_each_online_node(nid) {
295 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
296 gfp_zone(GFP_NOFS), NULL,
299 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
305 * I/O completion handler for block_read_full_page() - pages
306 * which come unlocked at the end of I/O.
308 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
311 struct buffer_head *first;
312 struct buffer_head *tmp;
314 int page_uptodate = 1;
316 BUG_ON(!buffer_async_read(bh));
320 set_buffer_uptodate(bh);
322 clear_buffer_uptodate(bh);
323 if (!quiet_error(bh))
329 * Be _very_ careful from here on. Bad things can happen if
330 * two buffer heads end IO at almost the same time and both
331 * decide that the page is now completely done.
333 first = page_buffers(page);
334 local_irq_save(flags);
335 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
336 clear_buffer_async_read(bh);
340 if (!buffer_uptodate(tmp))
342 if (buffer_async_read(tmp)) {
343 BUG_ON(!buffer_locked(tmp));
346 tmp = tmp->b_this_page;
348 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
349 local_irq_restore(flags);
352 * If none of the buffers had errors and they are all
353 * uptodate then we can set the page uptodate.
355 if (page_uptodate && !PageError(page))
356 SetPageUptodate(page);
361 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
362 local_irq_restore(flags);
367 * Completion handler for block_write_full_page() - pages which are unlocked
368 * during I/O, and which have PageWriteback cleared upon I/O completion.
370 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
372 char b[BDEVNAME_SIZE];
374 struct buffer_head *first;
375 struct buffer_head *tmp;
378 BUG_ON(!buffer_async_write(bh));
382 set_buffer_uptodate(bh);
384 if (!quiet_error(bh)) {
386 printk(KERN_WARNING "lost page write due to "
388 bdevname(bh->b_bdev, b));
390 set_bit(AS_EIO, &page->mapping->flags);
391 set_buffer_write_io_error(bh);
392 clear_buffer_uptodate(bh);
396 first = page_buffers(page);
397 local_irq_save(flags);
398 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
400 clear_buffer_async_write(bh);
402 tmp = bh->b_this_page;
404 if (buffer_async_write(tmp)) {
405 BUG_ON(!buffer_locked(tmp));
408 tmp = tmp->b_this_page;
410 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
411 local_irq_restore(flags);
412 end_page_writeback(page);
416 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
417 local_irq_restore(flags);
420 EXPORT_SYMBOL(end_buffer_async_write);
423 * If a page's buffers are under async readin (end_buffer_async_read
424 * completion) then there is a possibility that another thread of
425 * control could lock one of the buffers after it has completed
426 * but while some of the other buffers have not completed. This
427 * locked buffer would confuse end_buffer_async_read() into not unlocking
428 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
429 * that this buffer is not under async I/O.
431 * The page comes unlocked when it has no locked buffer_async buffers
434 * PageLocked prevents anyone starting new async I/O reads any of
437 * PageWriteback is used to prevent simultaneous writeout of the same
440 * PageLocked prevents anyone from starting writeback of a page which is
441 * under read I/O (PageWriteback is only ever set against a locked page).
443 static void mark_buffer_async_read(struct buffer_head *bh)
445 bh->b_end_io = end_buffer_async_read;
446 set_buffer_async_read(bh);
449 static void mark_buffer_async_write_endio(struct buffer_head *bh,
450 bh_end_io_t *handler)
452 bh->b_end_io = handler;
453 set_buffer_async_write(bh);
456 void mark_buffer_async_write(struct buffer_head *bh)
458 mark_buffer_async_write_endio(bh, end_buffer_async_write);
460 EXPORT_SYMBOL(mark_buffer_async_write);
464 * fs/buffer.c contains helper functions for buffer-backed address space's
465 * fsync functions. A common requirement for buffer-based filesystems is
466 * that certain data from the backing blockdev needs to be written out for
467 * a successful fsync(). For example, ext2 indirect blocks need to be
468 * written back and waited upon before fsync() returns.
470 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
471 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
472 * management of a list of dependent buffers at ->i_mapping->private_list.
474 * Locking is a little subtle: try_to_free_buffers() will remove buffers
475 * from their controlling inode's queue when they are being freed. But
476 * try_to_free_buffers() will be operating against the *blockdev* mapping
477 * at the time, not against the S_ISREG file which depends on those buffers.
478 * So the locking for private_list is via the private_lock in the address_space
479 * which backs the buffers. Which is different from the address_space
480 * against which the buffers are listed. So for a particular address_space,
481 * mapping->private_lock does *not* protect mapping->private_list! In fact,
482 * mapping->private_list will always be protected by the backing blockdev's
485 * Which introduces a requirement: all buffers on an address_space's
486 * ->private_list must be from the same address_space: the blockdev's.
488 * address_spaces which do not place buffers at ->private_list via these
489 * utility functions are free to use private_lock and private_list for
490 * whatever they want. The only requirement is that list_empty(private_list)
491 * be true at clear_inode() time.
493 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
494 * filesystems should do that. invalidate_inode_buffers() should just go
495 * BUG_ON(!list_empty).
497 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
498 * take an address_space, not an inode. And it should be called
499 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
502 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
503 * list if it is already on a list. Because if the buffer is on a list,
504 * it *must* already be on the right one. If not, the filesystem is being
505 * silly. This will save a ton of locking. But first we have to ensure
506 * that buffers are taken *off* the old inode's list when they are freed
507 * (presumably in truncate). That requires careful auditing of all
508 * filesystems (do it inside bforget()). It could also be done by bringing
513 * The buffer's backing address_space's private_lock must be held
515 static void __remove_assoc_queue(struct buffer_head *bh)
517 list_del_init(&bh->b_assoc_buffers);
518 WARN_ON(!bh->b_assoc_map);
519 if (buffer_write_io_error(bh))
520 set_bit(AS_EIO, &bh->b_assoc_map->flags);
521 bh->b_assoc_map = NULL;
524 int inode_has_buffers(struct inode *inode)
526 return !list_empty(&inode->i_data.private_list);
530 * osync is designed to support O_SYNC io. It waits synchronously for
531 * all already-submitted IO to complete, but does not queue any new
532 * writes to the disk.
534 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
535 * you dirty the buffers, and then use osync_inode_buffers to wait for
536 * completion. Any other dirty buffers which are not yet queued for
537 * write will not be flushed to disk by the osync.
539 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
541 struct buffer_head *bh;
547 list_for_each_prev(p, list) {
549 if (buffer_locked(bh)) {
553 if (!buffer_uptodate(bh))
564 static void do_thaw_one(struct super_block *sb, void *unused)
566 char b[BDEVNAME_SIZE];
567 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
568 printk(KERN_WARNING "Emergency Thaw on %s\n",
569 bdevname(sb->s_bdev, b));
572 static void do_thaw_all(struct work_struct *work)
574 iterate_supers(do_thaw_one, NULL);
576 printk(KERN_WARNING "Emergency Thaw complete\n");
580 * emergency_thaw_all -- forcibly thaw every frozen filesystem
582 * Used for emergency unfreeze of all filesystems via SysRq
584 void emergency_thaw_all(void)
586 struct work_struct *work;
588 work = kmalloc(sizeof(*work), GFP_ATOMIC);
590 INIT_WORK(work, do_thaw_all);
596 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
597 * @mapping: the mapping which wants those buffers written
599 * Starts I/O against the buffers at mapping->private_list, and waits upon
602 * Basically, this is a convenience function for fsync().
603 * @mapping is a file or directory which needs those buffers to be written for
604 * a successful fsync().
606 int sync_mapping_buffers(struct address_space *mapping)
608 struct address_space *buffer_mapping = mapping->assoc_mapping;
610 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
613 return fsync_buffers_list(&buffer_mapping->private_lock,
614 &mapping->private_list);
616 EXPORT_SYMBOL(sync_mapping_buffers);
619 * Called when we've recently written block `bblock', and it is known that
620 * `bblock' was for a buffer_boundary() buffer. This means that the block at
621 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
622 * dirty, schedule it for IO. So that indirects merge nicely with their data.
624 void write_boundary_block(struct block_device *bdev,
625 sector_t bblock, unsigned blocksize)
627 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
629 if (buffer_dirty(bh))
630 ll_rw_block(WRITE, 1, &bh);
635 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
637 struct address_space *mapping = inode->i_mapping;
638 struct address_space *buffer_mapping = bh->b_page->mapping;
640 mark_buffer_dirty(bh);
641 if (!mapping->assoc_mapping) {
642 mapping->assoc_mapping = buffer_mapping;
644 BUG_ON(mapping->assoc_mapping != buffer_mapping);
646 if (!bh->b_assoc_map) {
647 spin_lock(&buffer_mapping->private_lock);
648 list_move_tail(&bh->b_assoc_buffers,
649 &mapping->private_list);
650 bh->b_assoc_map = mapping;
651 spin_unlock(&buffer_mapping->private_lock);
654 EXPORT_SYMBOL(mark_buffer_dirty_inode);
657 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
660 * If warn is true, then emit a warning if the page is not uptodate and has
661 * not been truncated.
663 static void __set_page_dirty(struct page *page,
664 struct address_space *mapping, int warn)
666 spin_lock_irq(&mapping->tree_lock);
667 if (page->mapping) { /* Race with truncate? */
668 WARN_ON_ONCE(warn && !PageUptodate(page));
669 account_page_dirtied(page, mapping);
670 radix_tree_tag_set(&mapping->page_tree,
671 page_index(page), PAGECACHE_TAG_DIRTY);
673 spin_unlock_irq(&mapping->tree_lock);
674 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
678 * Add a page to the dirty page list.
680 * It is a sad fact of life that this function is called from several places
681 * deeply under spinlocking. It may not sleep.
683 * If the page has buffers, the uptodate buffers are set dirty, to preserve
684 * dirty-state coherency between the page and the buffers. It the page does
685 * not have buffers then when they are later attached they will all be set
688 * The buffers are dirtied before the page is dirtied. There's a small race
689 * window in which a writepage caller may see the page cleanness but not the
690 * buffer dirtiness. That's fine. If this code were to set the page dirty
691 * before the buffers, a concurrent writepage caller could clear the page dirty
692 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
693 * page on the dirty page list.
695 * We use private_lock to lock against try_to_free_buffers while using the
696 * page's buffer list. Also use this to protect against clean buffers being
697 * added to the page after it was set dirty.
699 * FIXME: may need to call ->reservepage here as well. That's rather up to the
700 * address_space though.
702 int __set_page_dirty_buffers(struct page *page)
705 struct address_space *mapping = page_mapping(page);
707 if (unlikely(!mapping))
708 return !TestSetPageDirty(page);
710 spin_lock(&mapping->private_lock);
711 if (page_has_buffers(page)) {
712 struct buffer_head *head = page_buffers(page);
713 struct buffer_head *bh = head;
716 set_buffer_dirty(bh);
717 bh = bh->b_this_page;
718 } while (bh != head);
720 newly_dirty = !TestSetPageDirty(page);
721 spin_unlock(&mapping->private_lock);
724 __set_page_dirty(page, mapping, 1);
727 EXPORT_SYMBOL(__set_page_dirty_buffers);
730 * Write out and wait upon a list of buffers.
732 * We have conflicting pressures: we want to make sure that all
733 * initially dirty buffers get waited on, but that any subsequently
734 * dirtied buffers don't. After all, we don't want fsync to last
735 * forever if somebody is actively writing to the file.
737 * Do this in two main stages: first we copy dirty buffers to a
738 * temporary inode list, queueing the writes as we go. Then we clean
739 * up, waiting for those writes to complete.
741 * During this second stage, any subsequent updates to the file may end
742 * up refiling the buffer on the original inode's dirty list again, so
743 * there is a chance we will end up with a buffer queued for write but
744 * not yet completed on that list. So, as a final cleanup we go through
745 * the osync code to catch these locked, dirty buffers without requeuing
746 * any newly dirty buffers for write.
748 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
750 struct buffer_head *bh;
751 struct list_head tmp;
752 struct address_space *mapping, *prev_mapping = NULL;
755 INIT_LIST_HEAD(&tmp);
758 while (!list_empty(list)) {
759 bh = BH_ENTRY(list->next);
760 mapping = bh->b_assoc_map;
761 __remove_assoc_queue(bh);
762 /* Avoid race with mark_buffer_dirty_inode() which does
763 * a lockless check and we rely on seeing the dirty bit */
765 if (buffer_dirty(bh) || buffer_locked(bh)) {
766 list_add(&bh->b_assoc_buffers, &tmp);
767 bh->b_assoc_map = mapping;
768 if (buffer_dirty(bh)) {
772 * Ensure any pending I/O completes so that
773 * ll_rw_block() actually writes the current
774 * contents - it is a noop if I/O is still in
775 * flight on potentially older contents.
777 ll_rw_block(SWRITE_SYNC_PLUG, 1, &bh);
780 * Kick off IO for the previous mapping. Note
781 * that we will not run the very last mapping,
782 * wait_on_buffer() will do that for us
783 * through sync_buffer().
785 if (prev_mapping && prev_mapping != mapping)
786 blk_run_address_space(prev_mapping);
787 prev_mapping = mapping;
795 while (!list_empty(&tmp)) {
796 bh = BH_ENTRY(tmp.prev);
798 mapping = bh->b_assoc_map;
799 __remove_assoc_queue(bh);
800 /* Avoid race with mark_buffer_dirty_inode() which does
801 * a lockless check and we rely on seeing the dirty bit */
803 if (buffer_dirty(bh)) {
804 list_add(&bh->b_assoc_buffers,
805 &mapping->private_list);
806 bh->b_assoc_map = mapping;
810 if (!buffer_uptodate(bh))
817 err2 = osync_buffers_list(lock, list);
825 * Invalidate any and all dirty buffers on a given inode. We are
826 * probably unmounting the fs, but that doesn't mean we have already
827 * done a sync(). Just drop the buffers from the inode list.
829 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
830 * assumes that all the buffers are against the blockdev. Not true
833 void invalidate_inode_buffers(struct inode *inode)
835 if (inode_has_buffers(inode)) {
836 struct address_space *mapping = &inode->i_data;
837 struct list_head *list = &mapping->private_list;
838 struct address_space *buffer_mapping = mapping->assoc_mapping;
840 spin_lock(&buffer_mapping->private_lock);
841 while (!list_empty(list))
842 __remove_assoc_queue(BH_ENTRY(list->next));
843 spin_unlock(&buffer_mapping->private_lock);
846 EXPORT_SYMBOL(invalidate_inode_buffers);
849 * Remove any clean buffers from the inode's buffer list. This is called
850 * when we're trying to free the inode itself. Those buffers can pin it.
852 * Returns true if all buffers were removed.
854 int remove_inode_buffers(struct inode *inode)
858 if (inode_has_buffers(inode)) {
859 struct address_space *mapping = &inode->i_data;
860 struct list_head *list = &mapping->private_list;
861 struct address_space *buffer_mapping = mapping->assoc_mapping;
863 spin_lock(&buffer_mapping->private_lock);
864 while (!list_empty(list)) {
865 struct buffer_head *bh = BH_ENTRY(list->next);
866 if (buffer_dirty(bh)) {
870 __remove_assoc_queue(bh);
872 spin_unlock(&buffer_mapping->private_lock);
878 * Create the appropriate buffers when given a page for data area and
879 * the size of each buffer.. Use the bh->b_this_page linked list to
880 * follow the buffers created. Return NULL if unable to create more
883 * The retry flag is used to differentiate async IO (paging, swapping)
884 * which may not fail from ordinary buffer allocations.
886 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
889 struct buffer_head *bh, *head;
895 while ((offset -= size) >= 0) {
896 bh = alloc_buffer_head(GFP_NOFS);
901 bh->b_this_page = head;
906 atomic_set(&bh->b_count, 0);
907 bh->b_private = NULL;
910 /* Link the buffer to its page */
911 set_bh_page(bh, page, offset);
913 init_buffer(bh, NULL, NULL);
917 * In case anything failed, we just free everything we got.
923 head = head->b_this_page;
924 free_buffer_head(bh);
929 * Return failure for non-async IO requests. Async IO requests
930 * are not allowed to fail, so we have to wait until buffer heads
931 * become available. But we don't want tasks sleeping with
932 * partially complete buffers, so all were released above.
937 /* We're _really_ low on memory. Now we just
938 * wait for old buffer heads to become free due to
939 * finishing IO. Since this is an async request and
940 * the reserve list is empty, we're sure there are
941 * async buffer heads in use.
946 EXPORT_SYMBOL_GPL(alloc_page_buffers);
949 link_dev_buffers(struct page *page, struct buffer_head *head)
951 struct buffer_head *bh, *tail;
956 bh = bh->b_this_page;
958 tail->b_this_page = head;
959 attach_page_buffers(page, head);
963 * Initialise the state of a blockdev page's buffers.
966 init_page_buffers(struct page *page, struct block_device *bdev,
967 sector_t block, int size)
969 struct buffer_head *head = page_buffers(page);
970 struct buffer_head *bh = head;
971 int uptodate = PageUptodate(page);
974 if (!buffer_mapped(bh)) {
975 init_buffer(bh, NULL, NULL);
977 bh->b_blocknr = block;
979 set_buffer_uptodate(bh);
980 set_buffer_mapped(bh);
983 bh = bh->b_this_page;
984 } while (bh != head);
988 * Create the page-cache page that contains the requested block.
990 * This is user purely for blockdev mappings.
993 grow_dev_page(struct block_device *bdev, sector_t block,
994 pgoff_t index, int size)
996 struct inode *inode = bdev->bd_inode;
998 struct buffer_head *bh;
1000 page = find_or_create_page(inode->i_mapping, index,
1001 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1005 BUG_ON(!PageLocked(page));
1007 if (page_has_buffers(page)) {
1008 bh = page_buffers(page);
1009 if (bh->b_size == size) {
1010 init_page_buffers(page, bdev, block, size);
1013 if (!try_to_free_buffers(page))
1018 * Allocate some buffers for this page
1020 bh = alloc_page_buffers(page, size, 0);
1025 * Link the page to the buffers and initialise them. Take the
1026 * lock to be atomic wrt __find_get_block(), which does not
1027 * run under the page lock.
1029 spin_lock(&inode->i_mapping->private_lock);
1030 link_dev_buffers(page, bh);
1031 init_page_buffers(page, bdev, block, size);
1032 spin_unlock(&inode->i_mapping->private_lock);
1038 page_cache_release(page);
1043 * Create buffers for the specified block device block's page. If
1044 * that page was dirty, the buffers are set dirty also.
1047 grow_buffers(struct block_device *bdev, sector_t block, int size)
1056 } while ((size << sizebits) < PAGE_SIZE);
1058 index = block >> sizebits;
1061 * Check for a block which wants to lie outside our maximum possible
1062 * pagecache index. (this comparison is done using sector_t types).
1064 if (unlikely(index != block >> sizebits)) {
1065 char b[BDEVNAME_SIZE];
1067 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1069 __func__, (unsigned long long)block,
1073 block = index << sizebits;
1074 /* Create a page with the proper size buffers.. */
1075 page = grow_dev_page(bdev, block, index, size);
1079 page_cache_release(page);
1083 static struct buffer_head *
1084 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1086 /* Size must be multiple of hard sectorsize */
1087 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1088 (size < 512 || size > PAGE_SIZE))) {
1089 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1091 printk(KERN_ERR "logical block size: %d\n",
1092 bdev_logical_block_size(bdev));
1099 struct buffer_head * bh;
1102 bh = __find_get_block(bdev, block, size);
1106 ret = grow_buffers(bdev, block, size);
1115 * The relationship between dirty buffers and dirty pages:
1117 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1118 * the page is tagged dirty in its radix tree.
1120 * At all times, the dirtiness of the buffers represents the dirtiness of
1121 * subsections of the page. If the page has buffers, the page dirty bit is
1122 * merely a hint about the true dirty state.
1124 * When a page is set dirty in its entirety, all its buffers are marked dirty
1125 * (if the page has buffers).
1127 * When a buffer is marked dirty, its page is dirtied, but the page's other
1130 * Also. When blockdev buffers are explicitly read with bread(), they
1131 * individually become uptodate. But their backing page remains not
1132 * uptodate - even if all of its buffers are uptodate. A subsequent
1133 * block_read_full_page() against that page will discover all the uptodate
1134 * buffers, will set the page uptodate and will perform no I/O.
1138 * mark_buffer_dirty - mark a buffer_head as needing writeout
1139 * @bh: the buffer_head to mark dirty
1141 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1142 * backing page dirty, then tag the page as dirty in its address_space's radix
1143 * tree and then attach the address_space's inode to its superblock's dirty
1146 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1147 * mapping->tree_lock and the global inode_lock.
1149 void mark_buffer_dirty(struct buffer_head *bh)
1151 WARN_ON_ONCE(!buffer_uptodate(bh));
1154 * Very *carefully* optimize the it-is-already-dirty case.
1156 * Don't let the final "is it dirty" escape to before we
1157 * perhaps modified the buffer.
1159 if (buffer_dirty(bh)) {
1161 if (buffer_dirty(bh))
1165 if (!test_set_buffer_dirty(bh)) {
1166 struct page *page = bh->b_page;
1167 if (!TestSetPageDirty(page)) {
1168 struct address_space *mapping = page_mapping(page);
1170 __set_page_dirty(page, mapping, 0);
1174 EXPORT_SYMBOL(mark_buffer_dirty);
1177 * Decrement a buffer_head's reference count. If all buffers against a page
1178 * have zero reference count, are clean and unlocked, and if the page is clean
1179 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1180 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1181 * a page but it ends up not being freed, and buffers may later be reattached).
1183 void __brelse(struct buffer_head * buf)
1185 if (atomic_read(&buf->b_count)) {
1189 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1191 EXPORT_SYMBOL(__brelse);
1194 * bforget() is like brelse(), except it discards any
1195 * potentially dirty data.
1197 void __bforget(struct buffer_head *bh)
1199 clear_buffer_dirty(bh);
1200 if (bh->b_assoc_map) {
1201 struct address_space *buffer_mapping = bh->b_page->mapping;
1203 spin_lock(&buffer_mapping->private_lock);
1204 list_del_init(&bh->b_assoc_buffers);
1205 bh->b_assoc_map = NULL;
1206 spin_unlock(&buffer_mapping->private_lock);
1210 EXPORT_SYMBOL(__bforget);
1212 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1215 if (buffer_uptodate(bh)) {
1220 bh->b_end_io = end_buffer_read_sync;
1221 submit_bh(READ, bh);
1223 if (buffer_uptodate(bh))
1231 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1232 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1233 * refcount elevated by one when they're in an LRU. A buffer can only appear
1234 * once in a particular CPU's LRU. A single buffer can be present in multiple
1235 * CPU's LRUs at the same time.
1237 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1238 * sb_find_get_block().
1240 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1241 * a local interrupt disable for that.
1244 #define BH_LRU_SIZE 8
1247 struct buffer_head *bhs[BH_LRU_SIZE];
1250 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1253 #define bh_lru_lock() local_irq_disable()
1254 #define bh_lru_unlock() local_irq_enable()
1256 #define bh_lru_lock() preempt_disable()
1257 #define bh_lru_unlock() preempt_enable()
1260 static inline void check_irqs_on(void)
1262 #ifdef irqs_disabled
1263 BUG_ON(irqs_disabled());
1268 * The LRU management algorithm is dopey-but-simple. Sorry.
1270 static void bh_lru_install(struct buffer_head *bh)
1272 struct buffer_head *evictee = NULL;
1277 lru = &__get_cpu_var(bh_lrus);
1278 if (lru->bhs[0] != bh) {
1279 struct buffer_head *bhs[BH_LRU_SIZE];
1285 for (in = 0; in < BH_LRU_SIZE; in++) {
1286 struct buffer_head *bh2 = lru->bhs[in];
1291 if (out >= BH_LRU_SIZE) {
1292 BUG_ON(evictee != NULL);
1299 while (out < BH_LRU_SIZE)
1301 memcpy(lru->bhs, bhs, sizeof(bhs));
1310 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1312 static struct buffer_head *
1313 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1315 struct buffer_head *ret = NULL;
1321 lru = &__get_cpu_var(bh_lrus);
1322 for (i = 0; i < BH_LRU_SIZE; i++) {
1323 struct buffer_head *bh = lru->bhs[i];
1325 if (bh && bh->b_bdev == bdev &&
1326 bh->b_blocknr == block && bh->b_size == size) {
1329 lru->bhs[i] = lru->bhs[i - 1];
1344 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1345 * it in the LRU and mark it as accessed. If it is not present then return
1348 struct buffer_head *
1349 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1351 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1354 bh = __find_get_block_slow(bdev, block);
1362 EXPORT_SYMBOL(__find_get_block);
1365 * __getblk will locate (and, if necessary, create) the buffer_head
1366 * which corresponds to the passed block_device, block and size. The
1367 * returned buffer has its reference count incremented.
1369 * __getblk() cannot fail - it just keeps trying. If you pass it an
1370 * illegal block number, __getblk() will happily return a buffer_head
1371 * which represents the non-existent block. Very weird.
1373 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1374 * attempt is failing. FIXME, perhaps?
1376 struct buffer_head *
1377 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1379 struct buffer_head *bh = __find_get_block(bdev, block, size);
1383 bh = __getblk_slow(bdev, block, size);
1386 EXPORT_SYMBOL(__getblk);
1389 * Do async read-ahead on a buffer..
1391 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1393 struct buffer_head *bh = __getblk(bdev, block, size);
1395 ll_rw_block(READA, 1, &bh);
1399 EXPORT_SYMBOL(__breadahead);
1402 * __bread() - reads a specified block and returns the bh
1403 * @bdev: the block_device to read from
1404 * @block: number of block
1405 * @size: size (in bytes) to read
1407 * Reads a specified block, and returns buffer head that contains it.
1408 * It returns NULL if the block was unreadable.
1410 struct buffer_head *
1411 __bread(struct block_device *bdev, sector_t block, unsigned size)
1413 struct buffer_head *bh = __getblk(bdev, block, size);
1415 if (likely(bh) && !buffer_uptodate(bh))
1416 bh = __bread_slow(bh);
1419 EXPORT_SYMBOL(__bread);
1422 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1423 * This doesn't race because it runs in each cpu either in irq
1424 * or with preempt disabled.
1426 static void invalidate_bh_lru(void *arg)
1428 struct bh_lru *b = &get_cpu_var(bh_lrus);
1431 for (i = 0; i < BH_LRU_SIZE; i++) {
1435 put_cpu_var(bh_lrus);
1438 void invalidate_bh_lrus(void)
1440 on_each_cpu(invalidate_bh_lru, NULL, 1);
1442 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1444 void set_bh_page(struct buffer_head *bh,
1445 struct page *page, unsigned long offset)
1448 BUG_ON(offset >= PAGE_SIZE);
1449 if (PageHighMem(page))
1451 * This catches illegal uses and preserves the offset:
1453 bh->b_data = (char *)(0 + offset);
1455 bh->b_data = page_address(page) + offset;
1457 EXPORT_SYMBOL(set_bh_page);
1460 * Called when truncating a buffer on a page completely.
1462 static void discard_buffer(struct buffer_head * bh)
1465 clear_buffer_dirty(bh);
1467 clear_buffer_mapped(bh);
1468 clear_buffer_req(bh);
1469 clear_buffer_new(bh);
1470 clear_buffer_delay(bh);
1471 clear_buffer_unwritten(bh);
1476 * block_invalidatepage - invalidate part of all of a buffer-backed page
1478 * @page: the page which is affected
1479 * @offset: the index of the truncation point
1481 * block_invalidatepage() is called when all or part of the page has become
1482 * invalidatedby a truncate operation.
1484 * block_invalidatepage() does not have to release all buffers, but it must
1485 * ensure that no dirty buffer is left outside @offset and that no I/O
1486 * is underway against any of the blocks which are outside the truncation
1487 * point. Because the caller is about to free (and possibly reuse) those
1490 void block_invalidatepage(struct page *page, unsigned long offset)
1492 struct buffer_head *head, *bh, *next;
1493 unsigned int curr_off = 0;
1495 BUG_ON(!PageLocked(page));
1496 if (!page_has_buffers(page))
1499 head = page_buffers(page);
1502 unsigned int next_off = curr_off + bh->b_size;
1503 next = bh->b_this_page;
1506 * is this block fully invalidated?
1508 if (offset <= curr_off)
1510 curr_off = next_off;
1512 } while (bh != head);
1515 * We release buffers only if the entire page is being invalidated.
1516 * The get_block cached value has been unconditionally invalidated,
1517 * so real IO is not possible anymore.
1520 try_to_release_page(page, 0);
1524 EXPORT_SYMBOL(block_invalidatepage);
1527 * We attach and possibly dirty the buffers atomically wrt
1528 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1529 * is already excluded via the page lock.
1531 void create_empty_buffers(struct page *page,
1532 unsigned long blocksize, unsigned long b_state)
1534 struct buffer_head *bh, *head, *tail;
1536 head = alloc_page_buffers(page, blocksize, 1);
1539 bh->b_state |= b_state;
1541 bh = bh->b_this_page;
1543 tail->b_this_page = head;
1545 spin_lock(&page->mapping->private_lock);
1546 if (PageUptodate(page) || PageDirty(page)) {
1549 if (PageDirty(page))
1550 set_buffer_dirty(bh);
1551 if (PageUptodate(page))
1552 set_buffer_uptodate(bh);
1553 bh = bh->b_this_page;
1554 } while (bh != head);
1556 attach_page_buffers(page, head);
1557 spin_unlock(&page->mapping->private_lock);
1559 EXPORT_SYMBOL(create_empty_buffers);
1562 * We are taking a block for data and we don't want any output from any
1563 * buffer-cache aliases starting from return from that function and
1564 * until the moment when something will explicitly mark the buffer
1565 * dirty (hopefully that will not happen until we will free that block ;-)
1566 * We don't even need to mark it not-uptodate - nobody can expect
1567 * anything from a newly allocated buffer anyway. We used to used
1568 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1569 * don't want to mark the alias unmapped, for example - it would confuse
1570 * anyone who might pick it with bread() afterwards...
1572 * Also.. Note that bforget() doesn't lock the buffer. So there can
1573 * be writeout I/O going on against recently-freed buffers. We don't
1574 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1575 * only if we really need to. That happens here.
1577 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1579 struct buffer_head *old_bh;
1583 old_bh = __find_get_block_slow(bdev, block);
1585 clear_buffer_dirty(old_bh);
1586 wait_on_buffer(old_bh);
1587 clear_buffer_req(old_bh);
1591 EXPORT_SYMBOL(unmap_underlying_metadata);
1594 * NOTE! All mapped/uptodate combinations are valid:
1596 * Mapped Uptodate Meaning
1598 * No No "unknown" - must do get_block()
1599 * No Yes "hole" - zero-filled
1600 * Yes No "allocated" - allocated on disk, not read in
1601 * Yes Yes "valid" - allocated and up-to-date in memory.
1603 * "Dirty" is valid only with the last case (mapped+uptodate).
1607 * While block_write_full_page is writing back the dirty buffers under
1608 * the page lock, whoever dirtied the buffers may decide to clean them
1609 * again at any time. We handle that by only looking at the buffer
1610 * state inside lock_buffer().
1612 * If block_write_full_page() is called for regular writeback
1613 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1614 * locked buffer. This only can happen if someone has written the buffer
1615 * directly, with submit_bh(). At the address_space level PageWriteback
1616 * prevents this contention from occurring.
1618 * If block_write_full_page() is called with wbc->sync_mode ==
1619 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC_PLUG; this
1620 * causes the writes to be flagged as synchronous writes, but the
1621 * block device queue will NOT be unplugged, since usually many pages
1622 * will be pushed to the out before the higher-level caller actually
1623 * waits for the writes to be completed. The various wait functions,
1624 * such as wait_on_writeback_range() will ultimately call sync_page()
1625 * which will ultimately call blk_run_backing_dev(), which will end up
1626 * unplugging the device queue.
1628 static int __block_write_full_page(struct inode *inode, struct page *page,
1629 get_block_t *get_block, struct writeback_control *wbc,
1630 bh_end_io_t *handler)
1634 sector_t last_block;
1635 struct buffer_head *bh, *head;
1636 const unsigned blocksize = 1 << inode->i_blkbits;
1637 int nr_underway = 0;
1638 int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1639 WRITE_SYNC_PLUG : WRITE);
1641 BUG_ON(!PageLocked(page));
1643 last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1645 if (!page_has_buffers(page)) {
1646 create_empty_buffers(page, blocksize,
1647 (1 << BH_Dirty)|(1 << BH_Uptodate));
1651 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1652 * here, and the (potentially unmapped) buffers may become dirty at
1653 * any time. If a buffer becomes dirty here after we've inspected it
1654 * then we just miss that fact, and the page stays dirty.
1656 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1657 * handle that here by just cleaning them.
1660 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1661 head = page_buffers(page);
1665 * Get all the dirty buffers mapped to disk addresses and
1666 * handle any aliases from the underlying blockdev's mapping.
1669 if (block > last_block) {
1671 * mapped buffers outside i_size will occur, because
1672 * this page can be outside i_size when there is a
1673 * truncate in progress.
1676 * The buffer was zeroed by block_write_full_page()
1678 clear_buffer_dirty(bh);
1679 set_buffer_uptodate(bh);
1680 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1682 WARN_ON(bh->b_size != blocksize);
1683 err = get_block(inode, block, bh, 1);
1686 clear_buffer_delay(bh);
1687 if (buffer_new(bh)) {
1688 /* blockdev mappings never come here */
1689 clear_buffer_new(bh);
1690 unmap_underlying_metadata(bh->b_bdev,
1694 bh = bh->b_this_page;
1696 } while (bh != head);
1699 if (!buffer_mapped(bh))
1702 * If it's a fully non-blocking write attempt and we cannot
1703 * lock the buffer then redirty the page. Note that this can
1704 * potentially cause a busy-wait loop from writeback threads
1705 * and kswapd activity, but those code paths have their own
1706 * higher-level throttling.
1708 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1710 } else if (!trylock_buffer(bh)) {
1711 redirty_page_for_writepage(wbc, page);
1714 if (test_clear_buffer_dirty(bh)) {
1715 mark_buffer_async_write_endio(bh, handler);
1719 } while ((bh = bh->b_this_page) != head);
1722 * The page and its buffers are protected by PageWriteback(), so we can
1723 * drop the bh refcounts early.
1725 BUG_ON(PageWriteback(page));
1726 set_page_writeback(page);
1729 struct buffer_head *next = bh->b_this_page;
1730 if (buffer_async_write(bh)) {
1731 submit_bh(write_op, bh);
1735 } while (bh != head);
1740 if (nr_underway == 0) {
1742 * The page was marked dirty, but the buffers were
1743 * clean. Someone wrote them back by hand with
1744 * ll_rw_block/submit_bh. A rare case.
1746 end_page_writeback(page);
1749 * The page and buffer_heads can be released at any time from
1757 * ENOSPC, or some other error. We may already have added some
1758 * blocks to the file, so we need to write these out to avoid
1759 * exposing stale data.
1760 * The page is currently locked and not marked for writeback
1763 /* Recovery: lock and submit the mapped buffers */
1765 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1766 !buffer_delay(bh)) {
1768 mark_buffer_async_write_endio(bh, handler);
1771 * The buffer may have been set dirty during
1772 * attachment to a dirty page.
1774 clear_buffer_dirty(bh);
1776 } while ((bh = bh->b_this_page) != head);
1778 BUG_ON(PageWriteback(page));
1779 mapping_set_error(page->mapping, err);
1780 set_page_writeback(page);
1782 struct buffer_head *next = bh->b_this_page;
1783 if (buffer_async_write(bh)) {
1784 clear_buffer_dirty(bh);
1785 submit_bh(write_op, bh);
1789 } while (bh != head);
1795 * If a page has any new buffers, zero them out here, and mark them uptodate
1796 * and dirty so they'll be written out (in order to prevent uninitialised
1797 * block data from leaking). And clear the new bit.
1799 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1801 unsigned int block_start, block_end;
1802 struct buffer_head *head, *bh;
1804 BUG_ON(!PageLocked(page));
1805 if (!page_has_buffers(page))
1808 bh = head = page_buffers(page);
1811 block_end = block_start + bh->b_size;
1813 if (buffer_new(bh)) {
1814 if (block_end > from && block_start < to) {
1815 if (!PageUptodate(page)) {
1816 unsigned start, size;
1818 start = max(from, block_start);
1819 size = min(to, block_end) - start;
1821 zero_user(page, start, size);
1822 set_buffer_uptodate(bh);
1825 clear_buffer_new(bh);
1826 mark_buffer_dirty(bh);
1830 block_start = block_end;
1831 bh = bh->b_this_page;
1832 } while (bh != head);
1834 EXPORT_SYMBOL(page_zero_new_buffers);
1836 int block_prepare_write(struct page *page, unsigned from, unsigned to,
1837 get_block_t *get_block)
1839 struct inode *inode = page->mapping->host;
1840 unsigned block_start, block_end;
1843 unsigned blocksize, bbits;
1844 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1846 BUG_ON(!PageLocked(page));
1847 BUG_ON(from > PAGE_CACHE_SIZE);
1848 BUG_ON(to > PAGE_CACHE_SIZE);
1851 blocksize = 1 << inode->i_blkbits;
1852 if (!page_has_buffers(page))
1853 create_empty_buffers(page, blocksize, 0);
1854 head = page_buffers(page);
1856 bbits = inode->i_blkbits;
1857 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1859 for(bh = head, block_start = 0; bh != head || !block_start;
1860 block++, block_start=block_end, bh = bh->b_this_page) {
1861 block_end = block_start + blocksize;
1862 if (block_end <= from || block_start >= to) {
1863 if (PageUptodate(page)) {
1864 if (!buffer_uptodate(bh))
1865 set_buffer_uptodate(bh);
1870 clear_buffer_new(bh);
1871 if (!buffer_mapped(bh)) {
1872 WARN_ON(bh->b_size != blocksize);
1873 err = get_block(inode, block, bh, 1);
1876 if (buffer_new(bh)) {
1877 unmap_underlying_metadata(bh->b_bdev,
1879 if (PageUptodate(page)) {
1880 clear_buffer_new(bh);
1881 set_buffer_uptodate(bh);
1882 mark_buffer_dirty(bh);
1885 if (block_end > to || block_start < from)
1886 zero_user_segments(page,
1892 if (PageUptodate(page)) {
1893 if (!buffer_uptodate(bh))
1894 set_buffer_uptodate(bh);
1897 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1898 !buffer_unwritten(bh) &&
1899 (block_start < from || block_end > to)) {
1900 ll_rw_block(READ, 1, &bh);
1905 * If we issued read requests - let them complete.
1907 while(wait_bh > wait) {
1908 wait_on_buffer(*--wait_bh);
1909 if (!buffer_uptodate(*wait_bh))
1912 if (unlikely(err)) {
1913 page_zero_new_buffers(page, from, to);
1914 ClearPageUptodate(page);
1918 EXPORT_SYMBOL(block_prepare_write);
1920 static int __block_commit_write(struct inode *inode, struct page *page,
1921 unsigned from, unsigned to)
1923 unsigned block_start, block_end;
1926 struct buffer_head *bh, *head;
1928 blocksize = 1 << inode->i_blkbits;
1930 for(bh = head = page_buffers(page), block_start = 0;
1931 bh != head || !block_start;
1932 block_start=block_end, bh = bh->b_this_page) {
1933 block_end = block_start + blocksize;
1934 if (block_end <= from || block_start >= to) {
1935 if (!buffer_uptodate(bh))
1938 set_buffer_uptodate(bh);
1939 mark_buffer_dirty(bh);
1941 clear_buffer_new(bh);
1945 * If this is a partial write which happened to make all buffers
1946 * uptodate then we can optimize away a bogus readpage() for
1947 * the next read(). Here we 'discover' whether the page went
1948 * uptodate as a result of this (potentially partial) write.
1951 SetPageUptodate(page);
1955 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1956 get_block_t *get_block)
1958 unsigned start = pos & (PAGE_CACHE_SIZE - 1);
1960 return block_prepare_write(page, start, start + len, get_block);
1962 EXPORT_SYMBOL(__block_write_begin);
1965 * Filesystems implementing the new truncate sequence should use the
1966 * _newtrunc postfix variant which won't incorrectly call vmtruncate.
1967 * The filesystem needs to handle block truncation upon failure.
1969 int block_write_begin_newtrunc(struct file *file, struct address_space *mapping,
1970 loff_t pos, unsigned len, unsigned flags,
1971 struct page **pagep, void **fsdata,
1972 get_block_t *get_block)
1974 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1978 page = grab_cache_page_write_begin(mapping, index, flags);
1982 status = __block_write_begin(page, pos, len, get_block);
1983 if (unlikely(status)) {
1985 page_cache_release(page);
1992 EXPORT_SYMBOL(block_write_begin_newtrunc);
1995 * block_write_begin takes care of the basic task of block allocation and
1996 * bringing partial write blocks uptodate first.
1998 * If *pagep is not NULL, then block_write_begin uses the locked page
1999 * at *pagep rather than allocating its own. In this case, the page will
2000 * not be unlocked or deallocated on failure.
2002 int block_write_begin(struct file *file, struct address_space *mapping,
2003 loff_t pos, unsigned len, unsigned flags,
2004 struct page **pagep, void **fsdata,
2005 get_block_t *get_block)
2009 ret = block_write_begin_newtrunc(file, mapping, pos, len, flags,
2010 pagep, fsdata, get_block);
2013 * prepare_write() may have instantiated a few blocks
2014 * outside i_size. Trim these off again. Don't need
2015 * i_size_read because we hold i_mutex.
2017 * Filesystems which pass down their own page also cannot
2018 * call into vmtruncate here because it would lead to lock
2019 * inversion problems (*pagep is locked). This is a further
2020 * example of where the old truncate sequence is inadequate.
2022 if (unlikely(ret) && *pagep == NULL) {
2023 loff_t isize = mapping->host->i_size;
2024 if (pos + len > isize)
2025 vmtruncate(mapping->host, isize);
2030 EXPORT_SYMBOL(block_write_begin);
2032 int block_write_end(struct file *file, struct address_space *mapping,
2033 loff_t pos, unsigned len, unsigned copied,
2034 struct page *page, void *fsdata)
2036 struct inode *inode = mapping->host;
2039 start = pos & (PAGE_CACHE_SIZE - 1);
2041 if (unlikely(copied < len)) {
2043 * The buffers that were written will now be uptodate, so we
2044 * don't have to worry about a readpage reading them and
2045 * overwriting a partial write. However if we have encountered
2046 * a short write and only partially written into a buffer, it
2047 * will not be marked uptodate, so a readpage might come in and
2048 * destroy our partial write.
2050 * Do the simplest thing, and just treat any short write to a
2051 * non uptodate page as a zero-length write, and force the
2052 * caller to redo the whole thing.
2054 if (!PageUptodate(page))
2057 page_zero_new_buffers(page, start+copied, start+len);
2059 flush_dcache_page(page);
2061 /* This could be a short (even 0-length) commit */
2062 __block_commit_write(inode, page, start, start+copied);
2066 EXPORT_SYMBOL(block_write_end);
2068 int generic_write_end(struct file *file, struct address_space *mapping,
2069 loff_t pos, unsigned len, unsigned copied,
2070 struct page *page, void *fsdata)
2072 struct inode *inode = mapping->host;
2073 int i_size_changed = 0;
2075 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2078 * No need to use i_size_read() here, the i_size
2079 * cannot change under us because we hold i_mutex.
2081 * But it's important to update i_size while still holding page lock:
2082 * page writeout could otherwise come in and zero beyond i_size.
2084 if (pos+copied > inode->i_size) {
2085 i_size_write(inode, pos+copied);
2090 page_cache_release(page);
2093 * Don't mark the inode dirty under page lock. First, it unnecessarily
2094 * makes the holding time of page lock longer. Second, it forces lock
2095 * ordering of page lock and transaction start for journaling
2099 mark_inode_dirty(inode);
2103 EXPORT_SYMBOL(generic_write_end);
2106 * block_is_partially_uptodate checks whether buffers within a page are
2109 * Returns true if all buffers which correspond to a file portion
2110 * we want to read are uptodate.
2112 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2115 struct inode *inode = page->mapping->host;
2116 unsigned block_start, block_end, blocksize;
2118 struct buffer_head *bh, *head;
2121 if (!page_has_buffers(page))
2124 blocksize = 1 << inode->i_blkbits;
2125 to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2127 if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2130 head = page_buffers(page);
2134 block_end = block_start + blocksize;
2135 if (block_end > from && block_start < to) {
2136 if (!buffer_uptodate(bh)) {
2140 if (block_end >= to)
2143 block_start = block_end;
2144 bh = bh->b_this_page;
2145 } while (bh != head);
2149 EXPORT_SYMBOL(block_is_partially_uptodate);
2152 * Generic "read page" function for block devices that have the normal
2153 * get_block functionality. This is most of the block device filesystems.
2154 * Reads the page asynchronously --- the unlock_buffer() and
2155 * set/clear_buffer_uptodate() functions propagate buffer state into the
2156 * page struct once IO has completed.
2158 int block_read_full_page(struct page *page, get_block_t *get_block)
2160 struct inode *inode = page->mapping->host;
2161 sector_t iblock, lblock;
2162 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2163 unsigned int blocksize;
2165 int fully_mapped = 1;
2167 BUG_ON(!PageLocked(page));
2168 blocksize = 1 << inode->i_blkbits;
2169 if (!page_has_buffers(page))
2170 create_empty_buffers(page, blocksize, 0);
2171 head = page_buffers(page);
2173 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2174 lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2180 if (buffer_uptodate(bh))
2183 if (!buffer_mapped(bh)) {
2187 if (iblock < lblock) {
2188 WARN_ON(bh->b_size != blocksize);
2189 err = get_block(inode, iblock, bh, 0);
2193 if (!buffer_mapped(bh)) {
2194 zero_user(page, i * blocksize, blocksize);
2196 set_buffer_uptodate(bh);
2200 * get_block() might have updated the buffer
2203 if (buffer_uptodate(bh))
2207 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2210 SetPageMappedToDisk(page);
2214 * All buffers are uptodate - we can set the page uptodate
2215 * as well. But not if get_block() returned an error.
2217 if (!PageError(page))
2218 SetPageUptodate(page);
2223 /* Stage two: lock the buffers */
2224 for (i = 0; i < nr; i++) {
2227 mark_buffer_async_read(bh);
2231 * Stage 3: start the IO. Check for uptodateness
2232 * inside the buffer lock in case another process reading
2233 * the underlying blockdev brought it uptodate (the sct fix).
2235 for (i = 0; i < nr; i++) {
2237 if (buffer_uptodate(bh))
2238 end_buffer_async_read(bh, 1);
2240 submit_bh(READ, bh);
2244 EXPORT_SYMBOL(block_read_full_page);
2246 /* utility function for filesystems that need to do work on expanding
2247 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2248 * deal with the hole.
2250 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2252 struct address_space *mapping = inode->i_mapping;
2257 err = inode_newsize_ok(inode, size);
2261 err = pagecache_write_begin(NULL, mapping, size, 0,
2262 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2267 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2273 EXPORT_SYMBOL(generic_cont_expand_simple);
2275 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2276 loff_t pos, loff_t *bytes)
2278 struct inode *inode = mapping->host;
2279 unsigned blocksize = 1 << inode->i_blkbits;
2282 pgoff_t index, curidx;
2284 unsigned zerofrom, offset, len;
2287 index = pos >> PAGE_CACHE_SHIFT;
2288 offset = pos & ~PAGE_CACHE_MASK;
2290 while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2291 zerofrom = curpos & ~PAGE_CACHE_MASK;
2292 if (zerofrom & (blocksize-1)) {
2293 *bytes |= (blocksize-1);
2296 len = PAGE_CACHE_SIZE - zerofrom;
2298 err = pagecache_write_begin(file, mapping, curpos, len,
2299 AOP_FLAG_UNINTERRUPTIBLE,
2303 zero_user(page, zerofrom, len);
2304 err = pagecache_write_end(file, mapping, curpos, len, len,
2311 balance_dirty_pages_ratelimited(mapping);
2314 /* page covers the boundary, find the boundary offset */
2315 if (index == curidx) {
2316 zerofrom = curpos & ~PAGE_CACHE_MASK;
2317 /* if we will expand the thing last block will be filled */
2318 if (offset <= zerofrom) {
2321 if (zerofrom & (blocksize-1)) {
2322 *bytes |= (blocksize-1);
2325 len = offset - zerofrom;
2327 err = pagecache_write_begin(file, mapping, curpos, len,
2328 AOP_FLAG_UNINTERRUPTIBLE,
2332 zero_user(page, zerofrom, len);
2333 err = pagecache_write_end(file, mapping, curpos, len, len,
2345 * For moronic filesystems that do not allow holes in file.
2346 * We may have to extend the file.
2348 int cont_write_begin(struct file *file, struct address_space *mapping,
2349 loff_t pos, unsigned len, unsigned flags,
2350 struct page **pagep, void **fsdata,
2351 get_block_t *get_block, loff_t *bytes)
2353 struct inode *inode = mapping->host;
2354 unsigned blocksize = 1 << inode->i_blkbits;
2358 err = cont_expand_zero(file, mapping, pos, bytes);
2362 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2363 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2364 *bytes |= (blocksize-1);
2369 err = block_write_begin_newtrunc(file, mapping, pos, len,
2370 flags, pagep, fsdata, get_block);
2374 EXPORT_SYMBOL(cont_write_begin);
2376 int block_commit_write(struct page *page, unsigned from, unsigned to)
2378 struct inode *inode = page->mapping->host;
2379 __block_commit_write(inode,page,from,to);
2382 EXPORT_SYMBOL(block_commit_write);
2385 * block_page_mkwrite() is not allowed to change the file size as it gets
2386 * called from a page fault handler when a page is first dirtied. Hence we must
2387 * be careful to check for EOF conditions here. We set the page up correctly
2388 * for a written page which means we get ENOSPC checking when writing into
2389 * holes and correct delalloc and unwritten extent mapping on filesystems that
2390 * support these features.
2392 * We are not allowed to take the i_mutex here so we have to play games to
2393 * protect against truncate races as the page could now be beyond EOF. Because
2394 * truncate writes the inode size before removing pages, once we have the
2395 * page lock we can determine safely if the page is beyond EOF. If it is not
2396 * beyond EOF, then the page is guaranteed safe against truncation until we
2400 block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2401 get_block_t get_block)
2403 struct page *page = vmf->page;
2404 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2407 int ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
2410 size = i_size_read(inode);
2411 if ((page->mapping != inode->i_mapping) ||
2412 (page_offset(page) > size)) {
2413 /* page got truncated out from underneath us */
2418 /* page is wholly or partially inside EOF */
2419 if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2420 end = size & ~PAGE_CACHE_MASK;
2422 end = PAGE_CACHE_SIZE;
2424 ret = block_prepare_write(page, 0, end, get_block);
2426 ret = block_commit_write(page, 0, end);
2428 if (unlikely(ret)) {
2432 else /* -ENOSPC, -EIO, etc */
2433 ret = VM_FAULT_SIGBUS;
2435 ret = VM_FAULT_LOCKED;
2440 EXPORT_SYMBOL(block_page_mkwrite);
2443 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2444 * immediately, while under the page lock. So it needs a special end_io
2445 * handler which does not touch the bh after unlocking it.
2447 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2449 __end_buffer_read_notouch(bh, uptodate);
2453 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2454 * the page (converting it to circular linked list and taking care of page
2457 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2459 struct buffer_head *bh;
2461 BUG_ON(!PageLocked(page));
2463 spin_lock(&page->mapping->private_lock);
2466 if (PageDirty(page))
2467 set_buffer_dirty(bh);
2468 if (!bh->b_this_page)
2469 bh->b_this_page = head;
2470 bh = bh->b_this_page;
2471 } while (bh != head);
2472 attach_page_buffers(page, head);
2473 spin_unlock(&page->mapping->private_lock);
2477 * On entry, the page is fully not uptodate.
2478 * On exit the page is fully uptodate in the areas outside (from,to)
2479 * The filesystem needs to handle block truncation upon failure.
2481 int nobh_write_begin(struct address_space *mapping,
2482 loff_t pos, unsigned len, unsigned flags,
2483 struct page **pagep, void **fsdata,
2484 get_block_t *get_block)
2486 struct inode *inode = mapping->host;
2487 const unsigned blkbits = inode->i_blkbits;
2488 const unsigned blocksize = 1 << blkbits;
2489 struct buffer_head *head, *bh;
2493 unsigned block_in_page;
2494 unsigned block_start, block_end;
2495 sector_t block_in_file;
2498 int is_mapped_to_disk = 1;
2500 index = pos >> PAGE_CACHE_SHIFT;
2501 from = pos & (PAGE_CACHE_SIZE - 1);
2504 page = grab_cache_page_write_begin(mapping, index, flags);
2510 if (page_has_buffers(page)) {
2512 page_cache_release(page);
2514 return block_write_begin_newtrunc(NULL, mapping, pos, len,
2515 flags, pagep, fsdata, get_block);
2518 if (PageMappedToDisk(page))
2522 * Allocate buffers so that we can keep track of state, and potentially
2523 * attach them to the page if an error occurs. In the common case of
2524 * no error, they will just be freed again without ever being attached
2525 * to the page (which is all OK, because we're under the page lock).
2527 * Be careful: the buffer linked list is a NULL terminated one, rather
2528 * than the circular one we're used to.
2530 head = alloc_page_buffers(page, blocksize, 0);
2536 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2539 * We loop across all blocks in the page, whether or not they are
2540 * part of the affected region. This is so we can discover if the
2541 * page is fully mapped-to-disk.
2543 for (block_start = 0, block_in_page = 0, bh = head;
2544 block_start < PAGE_CACHE_SIZE;
2545 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2548 block_end = block_start + blocksize;
2551 if (block_start >= to)
2553 ret = get_block(inode, block_in_file + block_in_page,
2557 if (!buffer_mapped(bh))
2558 is_mapped_to_disk = 0;
2560 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2561 if (PageUptodate(page)) {
2562 set_buffer_uptodate(bh);
2565 if (buffer_new(bh) || !buffer_mapped(bh)) {
2566 zero_user_segments(page, block_start, from,
2570 if (buffer_uptodate(bh))
2571 continue; /* reiserfs does this */
2572 if (block_start < from || block_end > to) {
2574 bh->b_end_io = end_buffer_read_nobh;
2575 submit_bh(READ, bh);
2582 * The page is locked, so these buffers are protected from
2583 * any VM or truncate activity. Hence we don't need to care
2584 * for the buffer_head refcounts.
2586 for (bh = head; bh; bh = bh->b_this_page) {
2588 if (!buffer_uptodate(bh))
2595 if (is_mapped_to_disk)
2596 SetPageMappedToDisk(page);
2598 *fsdata = head; /* to be released by nobh_write_end */
2605 * Error recovery is a bit difficult. We need to zero out blocks that
2606 * were newly allocated, and dirty them to ensure they get written out.
2607 * Buffers need to be attached to the page at this point, otherwise
2608 * the handling of potential IO errors during writeout would be hard
2609 * (could try doing synchronous writeout, but what if that fails too?)
2611 attach_nobh_buffers(page, head);
2612 page_zero_new_buffers(page, from, to);
2616 page_cache_release(page);
2621 EXPORT_SYMBOL(nobh_write_begin);
2623 int nobh_write_end(struct file *file, struct address_space *mapping,
2624 loff_t pos, unsigned len, unsigned copied,
2625 struct page *page, void *fsdata)
2627 struct inode *inode = page->mapping->host;
2628 struct buffer_head *head = fsdata;
2629 struct buffer_head *bh;
2630 BUG_ON(fsdata != NULL && page_has_buffers(page));
2632 if (unlikely(copied < len) && head)
2633 attach_nobh_buffers(page, head);
2634 if (page_has_buffers(page))
2635 return generic_write_end(file, mapping, pos, len,
2636 copied, page, fsdata);
2638 SetPageUptodate(page);
2639 set_page_dirty(page);
2640 if (pos+copied > inode->i_size) {
2641 i_size_write(inode, pos+copied);
2642 mark_inode_dirty(inode);
2646 page_cache_release(page);
2650 head = head->b_this_page;
2651 free_buffer_head(bh);
2656 EXPORT_SYMBOL(nobh_write_end);
2659 * nobh_writepage() - based on block_full_write_page() except
2660 * that it tries to operate without attaching bufferheads to
2663 int nobh_writepage(struct page *page, get_block_t *get_block,
2664 struct writeback_control *wbc)
2666 struct inode * const inode = page->mapping->host;
2667 loff_t i_size = i_size_read(inode);
2668 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2672 /* Is the page fully inside i_size? */
2673 if (page->index < end_index)
2676 /* Is the page fully outside i_size? (truncate in progress) */
2677 offset = i_size & (PAGE_CACHE_SIZE-1);
2678 if (page->index >= end_index+1 || !offset) {
2680 * The page may have dirty, unmapped buffers. For example,
2681 * they may have been added in ext3_writepage(). Make them
2682 * freeable here, so the page does not leak.
2685 /* Not really sure about this - do we need this ? */
2686 if (page->mapping->a_ops->invalidatepage)
2687 page->mapping->a_ops->invalidatepage(page, offset);
2690 return 0; /* don't care */
2694 * The page straddles i_size. It must be zeroed out on each and every
2695 * writepage invocation because it may be mmapped. "A file is mapped
2696 * in multiples of the page size. For a file that is not a multiple of
2697 * the page size, the remaining memory is zeroed when mapped, and
2698 * writes to that region are not written out to the file."
2700 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2702 ret = mpage_writepage(page, get_block, wbc);
2704 ret = __block_write_full_page(inode, page, get_block, wbc,
2705 end_buffer_async_write);
2708 EXPORT_SYMBOL(nobh_writepage);
2710 int nobh_truncate_page(struct address_space *mapping,
2711 loff_t from, get_block_t *get_block)
2713 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2714 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2717 unsigned length, pos;
2718 struct inode *inode = mapping->host;
2720 struct buffer_head map_bh;
2723 blocksize = 1 << inode->i_blkbits;
2724 length = offset & (blocksize - 1);
2726 /* Block boundary? Nothing to do */
2730 length = blocksize - length;
2731 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2733 page = grab_cache_page(mapping, index);
2738 if (page_has_buffers(page)) {
2741 page_cache_release(page);
2742 return block_truncate_page(mapping, from, get_block);
2745 /* Find the buffer that contains "offset" */
2747 while (offset >= pos) {
2752 map_bh.b_size = blocksize;
2754 err = get_block(inode, iblock, &map_bh, 0);
2757 /* unmapped? It's a hole - nothing to do */
2758 if (!buffer_mapped(&map_bh))
2761 /* Ok, it's mapped. Make sure it's up-to-date */
2762 if (!PageUptodate(page)) {
2763 err = mapping->a_ops->readpage(NULL, page);
2765 page_cache_release(page);
2769 if (!PageUptodate(page)) {
2773 if (page_has_buffers(page))
2776 zero_user(page, offset, length);
2777 set_page_dirty(page);
2782 page_cache_release(page);
2786 EXPORT_SYMBOL(nobh_truncate_page);
2788 int block_truncate_page(struct address_space *mapping,
2789 loff_t from, get_block_t *get_block)
2791 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2792 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2795 unsigned length, pos;
2796 struct inode *inode = mapping->host;
2798 struct buffer_head *bh;
2801 blocksize = 1 << inode->i_blkbits;
2802 length = offset & (blocksize - 1);
2804 /* Block boundary? Nothing to do */
2808 length = blocksize - length;
2809 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2811 page = grab_cache_page(mapping, index);
2816 if (!page_has_buffers(page))
2817 create_empty_buffers(page, blocksize, 0);
2819 /* Find the buffer that contains "offset" */
2820 bh = page_buffers(page);
2822 while (offset >= pos) {
2823 bh = bh->b_this_page;
2829 if (!buffer_mapped(bh)) {
2830 WARN_ON(bh->b_size != blocksize);
2831 err = get_block(inode, iblock, bh, 0);
2834 /* unmapped? It's a hole - nothing to do */
2835 if (!buffer_mapped(bh))
2839 /* Ok, it's mapped. Make sure it's up-to-date */
2840 if (PageUptodate(page))
2841 set_buffer_uptodate(bh);
2843 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2845 ll_rw_block(READ, 1, &bh);
2847 /* Uhhuh. Read error. Complain and punt. */
2848 if (!buffer_uptodate(bh))
2852 zero_user(page, offset, length);
2853 mark_buffer_dirty(bh);
2858 page_cache_release(page);
2862 EXPORT_SYMBOL(block_truncate_page);
2865 * The generic ->writepage function for buffer-backed address_spaces
2866 * this form passes in the end_io handler used to finish the IO.
2868 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2869 struct writeback_control *wbc, bh_end_io_t *handler)
2871 struct inode * const inode = page->mapping->host;
2872 loff_t i_size = i_size_read(inode);
2873 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2876 /* Is the page fully inside i_size? */
2877 if (page->index < end_index)
2878 return __block_write_full_page(inode, page, get_block, wbc,
2881 /* Is the page fully outside i_size? (truncate in progress) */
2882 offset = i_size & (PAGE_CACHE_SIZE-1);
2883 if (page->index >= end_index+1 || !offset) {
2885 * The page may have dirty, unmapped buffers. For example,
2886 * they may have been added in ext3_writepage(). Make them
2887 * freeable here, so the page does not leak.
2889 do_invalidatepage(page, 0);
2891 return 0; /* don't care */
2895 * The page straddles i_size. It must be zeroed out on each and every
2896 * writepage invocation because it may be mmapped. "A file is mapped
2897 * in multiples of the page size. For a file that is not a multiple of
2898 * the page size, the remaining memory is zeroed when mapped, and
2899 * writes to that region are not written out to the file."
2901 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2902 return __block_write_full_page(inode, page, get_block, wbc, handler);
2904 EXPORT_SYMBOL(block_write_full_page_endio);
2907 * The generic ->writepage function for buffer-backed address_spaces
2909 int block_write_full_page(struct page *page, get_block_t *get_block,
2910 struct writeback_control *wbc)
2912 return block_write_full_page_endio(page, get_block, wbc,
2913 end_buffer_async_write);
2915 EXPORT_SYMBOL(block_write_full_page);
2917 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2918 get_block_t *get_block)
2920 struct buffer_head tmp;
2921 struct inode *inode = mapping->host;
2924 tmp.b_size = 1 << inode->i_blkbits;
2925 get_block(inode, block, &tmp, 0);
2926 return tmp.b_blocknr;
2928 EXPORT_SYMBOL(generic_block_bmap);
2930 static void end_bio_bh_io_sync(struct bio *bio, int err)
2932 struct buffer_head *bh = bio->bi_private;
2934 if (err == -EOPNOTSUPP) {
2935 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2936 set_bit(BH_Eopnotsupp, &bh->b_state);
2939 if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2940 set_bit(BH_Quiet, &bh->b_state);
2942 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2946 int submit_bh(int rw, struct buffer_head * bh)
2951 BUG_ON(!buffer_locked(bh));
2952 BUG_ON(!buffer_mapped(bh));
2953 BUG_ON(!bh->b_end_io);
2954 BUG_ON(buffer_delay(bh));
2955 BUG_ON(buffer_unwritten(bh));
2958 * Mask in barrier bit for a write (could be either a WRITE or a
2961 if (buffer_ordered(bh) && (rw & WRITE))
2962 rw |= WRITE_BARRIER;
2965 * Only clear out a write error when rewriting
2967 if (test_set_buffer_req(bh) && (rw & WRITE))
2968 clear_buffer_write_io_error(bh);
2971 * from here on down, it's all bio -- do the initial mapping,
2972 * submit_bio -> generic_make_request may further map this bio around
2974 bio = bio_alloc(GFP_NOIO, 1);
2976 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2977 bio->bi_bdev = bh->b_bdev;
2978 bio->bi_io_vec[0].bv_page = bh->b_page;
2979 bio->bi_io_vec[0].bv_len = bh->b_size;
2980 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2984 bio->bi_size = bh->b_size;
2986 bio->bi_end_io = end_bio_bh_io_sync;
2987 bio->bi_private = bh;
2990 submit_bio(rw, bio);
2992 if (bio_flagged(bio, BIO_EOPNOTSUPP))
2998 EXPORT_SYMBOL(submit_bh);
3001 * ll_rw_block: low-level access to block devices (DEPRECATED)
3002 * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
3003 * @nr: number of &struct buffer_heads in the array
3004 * @bhs: array of pointers to &struct buffer_head
3006 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3007 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3008 * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
3009 * are sent to disk. The fourth %READA option is described in the documentation
3010 * for generic_make_request() which ll_rw_block() calls.
3012 * This function drops any buffer that it cannot get a lock on (with the
3013 * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
3014 * clean when doing a write request, and any buffer that appears to be
3015 * up-to-date when doing read request. Further it marks as clean buffers that
3016 * are processed for writing (the buffer cache won't assume that they are
3017 * actually clean until the buffer gets unlocked).
3019 * ll_rw_block sets b_end_io to simple completion handler that marks
3020 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3023 * All of the buffers must be for the same device, and must also be a
3024 * multiple of the current approved size for the device.
3026 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3030 for (i = 0; i < nr; i++) {
3031 struct buffer_head *bh = bhs[i];
3033 if (rw == SWRITE || rw == SWRITE_SYNC || rw == SWRITE_SYNC_PLUG)
3035 else if (!trylock_buffer(bh))
3038 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC ||
3039 rw == SWRITE_SYNC_PLUG) {
3040 if (test_clear_buffer_dirty(bh)) {
3041 bh->b_end_io = end_buffer_write_sync;
3043 if (rw == SWRITE_SYNC)
3044 submit_bh(WRITE_SYNC, bh);
3046 submit_bh(WRITE, bh);
3050 if (!buffer_uptodate(bh)) {
3051 bh->b_end_io = end_buffer_read_sync;
3060 EXPORT_SYMBOL(ll_rw_block);
3063 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3064 * and then start new I/O and then wait upon it. The caller must have a ref on
3067 int sync_dirty_buffer(struct buffer_head *bh)
3071 WARN_ON(atomic_read(&bh->b_count) < 1);
3073 if (test_clear_buffer_dirty(bh)) {
3075 bh->b_end_io = end_buffer_write_sync;
3076 ret = submit_bh(WRITE_SYNC, bh);
3078 if (buffer_eopnotsupp(bh)) {
3079 clear_buffer_eopnotsupp(bh);
3082 if (!ret && !buffer_uptodate(bh))
3089 EXPORT_SYMBOL(sync_dirty_buffer);
3092 * try_to_free_buffers() checks if all the buffers on this particular page
3093 * are unused, and releases them if so.
3095 * Exclusion against try_to_free_buffers may be obtained by either
3096 * locking the page or by holding its mapping's private_lock.
3098 * If the page is dirty but all the buffers are clean then we need to
3099 * be sure to mark the page clean as well. This is because the page
3100 * may be against a block device, and a later reattachment of buffers
3101 * to a dirty page will set *all* buffers dirty. Which would corrupt
3102 * filesystem data on the same device.
3104 * The same applies to regular filesystem pages: if all the buffers are
3105 * clean then we set the page clean and proceed. To do that, we require
3106 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3109 * try_to_free_buffers() is non-blocking.
3111 static inline int buffer_busy(struct buffer_head *bh)
3113 return atomic_read(&bh->b_count) |
3114 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3118 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3120 struct buffer_head *head = page_buffers(page);
3121 struct buffer_head *bh;
3125 if (buffer_write_io_error(bh) && page->mapping)
3126 set_bit(AS_EIO, &page->mapping->flags);
3127 if (buffer_busy(bh))
3129 bh = bh->b_this_page;
3130 } while (bh != head);
3133 struct buffer_head *next = bh->b_this_page;
3135 if (bh->b_assoc_map)
3136 __remove_assoc_queue(bh);
3138 } while (bh != head);
3139 *buffers_to_free = head;
3140 __clear_page_buffers(page);
3146 int try_to_free_buffers(struct page *page)
3148 struct address_space * const mapping = page->mapping;
3149 struct buffer_head *buffers_to_free = NULL;
3152 BUG_ON(!PageLocked(page));
3153 if (PageWriteback(page))
3156 if (mapping == NULL) { /* can this still happen? */
3157 ret = drop_buffers(page, &buffers_to_free);
3161 spin_lock(&mapping->private_lock);
3162 ret = drop_buffers(page, &buffers_to_free);
3165 * If the filesystem writes its buffers by hand (eg ext3)
3166 * then we can have clean buffers against a dirty page. We
3167 * clean the page here; otherwise the VM will never notice
3168 * that the filesystem did any IO at all.
3170 * Also, during truncate, discard_buffer will have marked all
3171 * the page's buffers clean. We discover that here and clean
3174 * private_lock must be held over this entire operation in order
3175 * to synchronise against __set_page_dirty_buffers and prevent the
3176 * dirty bit from being lost.
3179 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3180 spin_unlock(&mapping->private_lock);
3182 if (buffers_to_free) {
3183 struct buffer_head *bh = buffers_to_free;
3186 struct buffer_head *next = bh->b_this_page;
3187 free_buffer_head(bh);
3189 } while (bh != buffers_to_free);
3193 EXPORT_SYMBOL(try_to_free_buffers);
3195 void block_sync_page(struct page *page)
3197 struct address_space *mapping;
3200 mapping = page_mapping(page);
3202 blk_run_backing_dev(mapping->backing_dev_info, page);
3204 EXPORT_SYMBOL(block_sync_page);
3207 * There are no bdflush tunables left. But distributions are
3208 * still running obsolete flush daemons, so we terminate them here.
3210 * Use of bdflush() is deprecated and will be removed in a future kernel.
3211 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3213 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3215 static int msg_count;
3217 if (!capable(CAP_SYS_ADMIN))
3220 if (msg_count < 5) {
3223 "warning: process `%s' used the obsolete bdflush"
3224 " system call\n", current->comm);
3225 printk(KERN_INFO "Fix your initscripts?\n");
3234 * Buffer-head allocation
3236 static struct kmem_cache *bh_cachep;
3239 * Once the number of bh's in the machine exceeds this level, we start
3240 * stripping them in writeback.
3242 static int max_buffer_heads;
3244 int buffer_heads_over_limit;
3246 struct bh_accounting {
3247 int nr; /* Number of live bh's */
3248 int ratelimit; /* Limit cacheline bouncing */
3251 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3253 static void recalc_bh_state(void)
3258 if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3260 __get_cpu_var(bh_accounting).ratelimit = 0;
3261 for_each_online_cpu(i)
3262 tot += per_cpu(bh_accounting, i).nr;
3263 buffer_heads_over_limit = (tot > max_buffer_heads);
3266 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3268 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3270 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3271 get_cpu_var(bh_accounting).nr++;
3273 put_cpu_var(bh_accounting);
3277 EXPORT_SYMBOL(alloc_buffer_head);
3279 void free_buffer_head(struct buffer_head *bh)
3281 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3282 kmem_cache_free(bh_cachep, bh);
3283 get_cpu_var(bh_accounting).nr--;
3285 put_cpu_var(bh_accounting);
3287 EXPORT_SYMBOL(free_buffer_head);
3289 static void buffer_exit_cpu(int cpu)
3292 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3294 for (i = 0; i < BH_LRU_SIZE; i++) {
3298 get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3299 per_cpu(bh_accounting, cpu).nr = 0;
3300 put_cpu_var(bh_accounting);
3303 static int buffer_cpu_notify(struct notifier_block *self,
3304 unsigned long action, void *hcpu)
3306 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3307 buffer_exit_cpu((unsigned long)hcpu);
3312 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3313 * @bh: struct buffer_head
3315 * Return true if the buffer is up-to-date and false,
3316 * with the buffer locked, if not.
3318 int bh_uptodate_or_lock(struct buffer_head *bh)
3320 if (!buffer_uptodate(bh)) {
3322 if (!buffer_uptodate(bh))
3328 EXPORT_SYMBOL(bh_uptodate_or_lock);
3331 * bh_submit_read - Submit a locked buffer for reading
3332 * @bh: struct buffer_head
3334 * Returns zero on success and -EIO on error.
3336 int bh_submit_read(struct buffer_head *bh)
3338 BUG_ON(!buffer_locked(bh));
3340 if (buffer_uptodate(bh)) {
3346 bh->b_end_io = end_buffer_read_sync;
3347 submit_bh(READ, bh);
3349 if (buffer_uptodate(bh))
3353 EXPORT_SYMBOL(bh_submit_read);
3355 void __init buffer_init(void)
3359 bh_cachep = kmem_cache_create("buffer_head",
3360 sizeof(struct buffer_head), 0,
3361 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3366 * Limit the bh occupancy to 10% of ZONE_NORMAL
3368 nrpages = (nr_free_buffer_pages() * 10) / 100;
3369 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3370 hotcpu_notifier(buffer_cpu_notify, 0);