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;
56 static int sync_buffer(void *word)
58 struct block_device *bd;
59 struct buffer_head *bh
60 = container_of(word, struct buffer_head, b_state);
65 blk_run_address_space(bd->bd_inode->i_mapping);
70 void __lock_buffer(struct buffer_head *bh)
72 wait_on_bit_lock(&bh->b_state, BH_Lock, sync_buffer,
73 TASK_UNINTERRUPTIBLE);
75 EXPORT_SYMBOL(__lock_buffer);
77 void unlock_buffer(struct buffer_head *bh)
79 smp_mb__before_clear_bit();
80 clear_buffer_locked(bh);
81 smp_mb__after_clear_bit();
82 wake_up_bit(&bh->b_state, BH_Lock);
86 * Block until a buffer comes unlocked. This doesn't stop it
87 * from becoming locked again - you have to lock it yourself
88 * if you want to preserve its state.
90 void __wait_on_buffer(struct buffer_head * bh)
92 wait_on_bit(&bh->b_state, BH_Lock, sync_buffer, TASK_UNINTERRUPTIBLE);
96 __clear_page_buffers(struct page *page)
98 ClearPagePrivate(page);
99 set_page_private(page, 0);
100 page_cache_release(page);
103 static void buffer_io_error(struct buffer_head *bh)
105 char b[BDEVNAME_SIZE];
107 printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
108 bdevname(bh->b_bdev, b),
109 (unsigned long long)bh->b_blocknr);
113 * End-of-IO handler helper function which does not touch the bh after
115 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
116 * a race there is benign: unlock_buffer() only use the bh's address for
117 * hashing after unlocking the buffer, so it doesn't actually touch the bh
120 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
123 set_buffer_uptodate(bh);
125 /* This happens, due to failed READA attempts. */
126 clear_buffer_uptodate(bh);
132 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
133 * unlock the buffer. This is what ll_rw_block uses too.
135 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
137 __end_buffer_read_notouch(bh, uptodate);
141 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
143 char b[BDEVNAME_SIZE];
146 set_buffer_uptodate(bh);
148 if (!buffer_eopnotsupp(bh) && printk_ratelimit()) {
150 printk(KERN_WARNING "lost page write due to "
152 bdevname(bh->b_bdev, b));
154 set_buffer_write_io_error(bh);
155 clear_buffer_uptodate(bh);
162 * Write out and wait upon all the dirty data associated with a block
163 * device via its mapping. Does not take the superblock lock.
165 int sync_blockdev(struct block_device *bdev)
170 ret = filemap_write_and_wait(bdev->bd_inode->i_mapping);
173 EXPORT_SYMBOL(sync_blockdev);
176 * Write out and wait upon all dirty data associated with this
177 * device. Filesystem data as well as the underlying block
178 * device. Takes the superblock lock.
180 int fsync_bdev(struct block_device *bdev)
182 struct super_block *sb = get_super(bdev);
184 int res = fsync_super(sb);
188 return sync_blockdev(bdev);
192 * freeze_bdev -- lock a filesystem and force it into a consistent state
193 * @bdev: blockdevice to lock
195 * This takes the block device bd_mount_sem to make sure no new mounts
196 * happen on bdev until thaw_bdev() is called.
197 * If a superblock is found on this device, we take the s_umount semaphore
198 * on it to make sure nobody unmounts until the snapshot creation is done.
200 struct super_block *freeze_bdev(struct block_device *bdev)
202 struct super_block *sb;
204 down(&bdev->bd_mount_sem);
205 sb = get_super(bdev);
206 if (sb && !(sb->s_flags & MS_RDONLY)) {
207 sb->s_frozen = SB_FREEZE_WRITE;
212 sb->s_frozen = SB_FREEZE_TRANS;
215 sync_blockdev(sb->s_bdev);
217 if (sb->s_op->write_super_lockfs)
218 sb->s_op->write_super_lockfs(sb);
222 return sb; /* thaw_bdev releases s->s_umount and bd_mount_sem */
224 EXPORT_SYMBOL(freeze_bdev);
227 * thaw_bdev -- unlock filesystem
228 * @bdev: blockdevice to unlock
229 * @sb: associated superblock
231 * Unlocks the filesystem and marks it writeable again after freeze_bdev().
233 void thaw_bdev(struct block_device *bdev, struct super_block *sb)
236 BUG_ON(sb->s_bdev != bdev);
238 if (sb->s_op->unlockfs)
239 sb->s_op->unlockfs(sb);
240 sb->s_frozen = SB_UNFROZEN;
242 wake_up(&sb->s_wait_unfrozen);
246 up(&bdev->bd_mount_sem);
248 EXPORT_SYMBOL(thaw_bdev);
251 * Various filesystems appear to want __find_get_block to be non-blocking.
252 * But it's the page lock which protects the buffers. To get around this,
253 * we get exclusion from try_to_free_buffers with the blockdev mapping's
256 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
257 * may be quite high. This code could TryLock the page, and if that
258 * succeeds, there is no need to take private_lock. (But if
259 * private_lock is contended then so is mapping->tree_lock).
261 static struct buffer_head *
262 __find_get_block_slow(struct block_device *bdev, sector_t block)
264 struct inode *bd_inode = bdev->bd_inode;
265 struct address_space *bd_mapping = bd_inode->i_mapping;
266 struct buffer_head *ret = NULL;
268 struct buffer_head *bh;
269 struct buffer_head *head;
273 index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
274 page = find_get_page(bd_mapping, index);
278 spin_lock(&bd_mapping->private_lock);
279 if (!page_has_buffers(page))
281 head = page_buffers(page);
284 if (bh->b_blocknr == block) {
289 if (!buffer_mapped(bh))
291 bh = bh->b_this_page;
292 } while (bh != head);
294 /* we might be here because some of the buffers on this page are
295 * not mapped. This is due to various races between
296 * file io on the block device and getblk. It gets dealt with
297 * elsewhere, don't buffer_error if we had some unmapped buffers
300 printk("__find_get_block_slow() failed. "
301 "block=%llu, b_blocknr=%llu\n",
302 (unsigned long long)block,
303 (unsigned long long)bh->b_blocknr);
304 printk("b_state=0x%08lx, b_size=%zu\n",
305 bh->b_state, bh->b_size);
306 printk("device blocksize: %d\n", 1 << bd_inode->i_blkbits);
309 spin_unlock(&bd_mapping->private_lock);
310 page_cache_release(page);
315 /* If invalidate_buffers() will trash dirty buffers, it means some kind
316 of fs corruption is going on. Trashing dirty data always imply losing
317 information that was supposed to be just stored on the physical layer
320 Thus invalidate_buffers in general usage is not allwowed to trash
321 dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
322 be preserved. These buffers are simply skipped.
324 We also skip buffers which are still in use. For example this can
325 happen if a userspace program is reading the block device.
327 NOTE: In the case where the user removed a removable-media-disk even if
328 there's still dirty data not synced on disk (due a bug in the device driver
329 or due an error of the user), by not destroying the dirty buffers we could
330 generate corruption also on the next media inserted, thus a parameter is
331 necessary to handle this case in the most safe way possible (trying
332 to not corrupt also the new disk inserted with the data belonging to
333 the old now corrupted disk). Also for the ramdisk the natural thing
334 to do in order to release the ramdisk memory is to destroy dirty buffers.
336 These are two special cases. Normal usage imply the device driver
337 to issue a sync on the device (without waiting I/O completion) and
338 then an invalidate_buffers call that doesn't trash dirty buffers.
340 For handling cache coherency with the blkdev pagecache the 'update' case
341 is been introduced. It is needed to re-read from disk any pinned
342 buffer. NOTE: re-reading from disk is destructive so we can do it only
343 when we assume nobody is changing the buffercache under our I/O and when
344 we think the disk contains more recent information than the buffercache.
345 The update == 1 pass marks the buffers we need to update, the update == 2
346 pass does the actual I/O. */
347 void invalidate_bdev(struct block_device *bdev)
349 struct address_space *mapping = bdev->bd_inode->i_mapping;
351 if (mapping->nrpages == 0)
354 invalidate_bh_lrus();
355 invalidate_mapping_pages(mapping, 0, -1);
359 * Kick pdflush then try to free up some ZONE_NORMAL memory.
361 static void free_more_memory(void)
366 wakeup_pdflush(1024);
369 for_each_online_node(nid) {
370 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
371 gfp_zone(GFP_NOFS), NULL,
374 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
380 * I/O completion handler for block_read_full_page() - pages
381 * which come unlocked at the end of I/O.
383 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
386 struct buffer_head *first;
387 struct buffer_head *tmp;
389 int page_uptodate = 1;
391 BUG_ON(!buffer_async_read(bh));
395 set_buffer_uptodate(bh);
397 clear_buffer_uptodate(bh);
398 if (printk_ratelimit())
404 * Be _very_ careful from here on. Bad things can happen if
405 * two buffer heads end IO at almost the same time and both
406 * decide that the page is now completely done.
408 first = page_buffers(page);
409 local_irq_save(flags);
410 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
411 clear_buffer_async_read(bh);
415 if (!buffer_uptodate(tmp))
417 if (buffer_async_read(tmp)) {
418 BUG_ON(!buffer_locked(tmp));
421 tmp = tmp->b_this_page;
423 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
424 local_irq_restore(flags);
427 * If none of the buffers had errors and they are all
428 * uptodate then we can set the page uptodate.
430 if (page_uptodate && !PageError(page))
431 SetPageUptodate(page);
436 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
437 local_irq_restore(flags);
442 * Completion handler for block_write_full_page() - pages which are unlocked
443 * during I/O, and which have PageWriteback cleared upon I/O completion.
445 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
447 char b[BDEVNAME_SIZE];
449 struct buffer_head *first;
450 struct buffer_head *tmp;
453 BUG_ON(!buffer_async_write(bh));
457 set_buffer_uptodate(bh);
459 if (printk_ratelimit()) {
461 printk(KERN_WARNING "lost page write due to "
463 bdevname(bh->b_bdev, b));
465 set_bit(AS_EIO, &page->mapping->flags);
466 set_buffer_write_io_error(bh);
467 clear_buffer_uptodate(bh);
471 first = page_buffers(page);
472 local_irq_save(flags);
473 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
475 clear_buffer_async_write(bh);
477 tmp = bh->b_this_page;
479 if (buffer_async_write(tmp)) {
480 BUG_ON(!buffer_locked(tmp));
483 tmp = tmp->b_this_page;
485 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
486 local_irq_restore(flags);
487 end_page_writeback(page);
491 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
492 local_irq_restore(flags);
497 * If a page's buffers are under async readin (end_buffer_async_read
498 * completion) then there is a possibility that another thread of
499 * control could lock one of the buffers after it has completed
500 * but while some of the other buffers have not completed. This
501 * locked buffer would confuse end_buffer_async_read() into not unlocking
502 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
503 * that this buffer is not under async I/O.
505 * The page comes unlocked when it has no locked buffer_async buffers
508 * PageLocked prevents anyone starting new async I/O reads any of
511 * PageWriteback is used to prevent simultaneous writeout of the same
514 * PageLocked prevents anyone from starting writeback of a page which is
515 * under read I/O (PageWriteback is only ever set against a locked page).
517 static void mark_buffer_async_read(struct buffer_head *bh)
519 bh->b_end_io = end_buffer_async_read;
520 set_buffer_async_read(bh);
523 void mark_buffer_async_write(struct buffer_head *bh)
525 bh->b_end_io = end_buffer_async_write;
526 set_buffer_async_write(bh);
528 EXPORT_SYMBOL(mark_buffer_async_write);
532 * fs/buffer.c contains helper functions for buffer-backed address space's
533 * fsync functions. A common requirement for buffer-based filesystems is
534 * that certain data from the backing blockdev needs to be written out for
535 * a successful fsync(). For example, ext2 indirect blocks need to be
536 * written back and waited upon before fsync() returns.
538 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
539 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
540 * management of a list of dependent buffers at ->i_mapping->private_list.
542 * Locking is a little subtle: try_to_free_buffers() will remove buffers
543 * from their controlling inode's queue when they are being freed. But
544 * try_to_free_buffers() will be operating against the *blockdev* mapping
545 * at the time, not against the S_ISREG file which depends on those buffers.
546 * So the locking for private_list is via the private_lock in the address_space
547 * which backs the buffers. Which is different from the address_space
548 * against which the buffers are listed. So for a particular address_space,
549 * mapping->private_lock does *not* protect mapping->private_list! In fact,
550 * mapping->private_list will always be protected by the backing blockdev's
553 * Which introduces a requirement: all buffers on an address_space's
554 * ->private_list must be from the same address_space: the blockdev's.
556 * address_spaces which do not place buffers at ->private_list via these
557 * utility functions are free to use private_lock and private_list for
558 * whatever they want. The only requirement is that list_empty(private_list)
559 * be true at clear_inode() time.
561 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
562 * filesystems should do that. invalidate_inode_buffers() should just go
563 * BUG_ON(!list_empty).
565 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
566 * take an address_space, not an inode. And it should be called
567 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
570 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
571 * list if it is already on a list. Because if the buffer is on a list,
572 * it *must* already be on the right one. If not, the filesystem is being
573 * silly. This will save a ton of locking. But first we have to ensure
574 * that buffers are taken *off* the old inode's list when they are freed
575 * (presumably in truncate). That requires careful auditing of all
576 * filesystems (do it inside bforget()). It could also be done by bringing
581 * The buffer's backing address_space's private_lock must be held
583 static inline void __remove_assoc_queue(struct buffer_head *bh)
585 list_del_init(&bh->b_assoc_buffers);
586 WARN_ON(!bh->b_assoc_map);
587 if (buffer_write_io_error(bh))
588 set_bit(AS_EIO, &bh->b_assoc_map->flags);
589 bh->b_assoc_map = NULL;
592 int inode_has_buffers(struct inode *inode)
594 return !list_empty(&inode->i_data.private_list);
598 * osync is designed to support O_SYNC io. It waits synchronously for
599 * all already-submitted IO to complete, but does not queue any new
600 * writes to the disk.
602 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
603 * you dirty the buffers, and then use osync_inode_buffers to wait for
604 * completion. Any other dirty buffers which are not yet queued for
605 * write will not be flushed to disk by the osync.
607 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
609 struct buffer_head *bh;
615 list_for_each_prev(p, list) {
617 if (buffer_locked(bh)) {
621 if (!buffer_uptodate(bh))
633 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
634 * @mapping: the mapping which wants those buffers written
636 * Starts I/O against the buffers at mapping->private_list, and waits upon
639 * Basically, this is a convenience function for fsync().
640 * @mapping is a file or directory which needs those buffers to be written for
641 * a successful fsync().
643 int sync_mapping_buffers(struct address_space *mapping)
645 struct address_space *buffer_mapping = mapping->assoc_mapping;
647 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
650 return fsync_buffers_list(&buffer_mapping->private_lock,
651 &mapping->private_list);
653 EXPORT_SYMBOL(sync_mapping_buffers);
656 * Called when we've recently written block `bblock', and it is known that
657 * `bblock' was for a buffer_boundary() buffer. This means that the block at
658 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
659 * dirty, schedule it for IO. So that indirects merge nicely with their data.
661 void write_boundary_block(struct block_device *bdev,
662 sector_t bblock, unsigned blocksize)
664 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
666 if (buffer_dirty(bh))
667 ll_rw_block(WRITE, 1, &bh);
672 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
674 struct address_space *mapping = inode->i_mapping;
675 struct address_space *buffer_mapping = bh->b_page->mapping;
677 mark_buffer_dirty(bh);
678 if (!mapping->assoc_mapping) {
679 mapping->assoc_mapping = buffer_mapping;
681 BUG_ON(mapping->assoc_mapping != buffer_mapping);
683 if (!bh->b_assoc_map) {
684 spin_lock(&buffer_mapping->private_lock);
685 list_move_tail(&bh->b_assoc_buffers,
686 &mapping->private_list);
687 bh->b_assoc_map = mapping;
688 spin_unlock(&buffer_mapping->private_lock);
691 EXPORT_SYMBOL(mark_buffer_dirty_inode);
694 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
697 * If warn is true, then emit a warning if the page is not uptodate and has
698 * not been truncated.
700 static int __set_page_dirty(struct page *page,
701 struct address_space *mapping, int warn)
703 if (unlikely(!mapping))
704 return !TestSetPageDirty(page);
706 if (TestSetPageDirty(page))
709 write_lock_irq(&mapping->tree_lock);
710 if (page->mapping) { /* Race with truncate? */
711 WARN_ON_ONCE(warn && !PageUptodate(page));
713 if (mapping_cap_account_dirty(mapping)) {
714 __inc_zone_page_state(page, NR_FILE_DIRTY);
715 __inc_bdi_stat(mapping->backing_dev_info,
717 task_io_account_write(PAGE_CACHE_SIZE);
719 radix_tree_tag_set(&mapping->page_tree,
720 page_index(page), PAGECACHE_TAG_DIRTY);
722 write_unlock_irq(&mapping->tree_lock);
723 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
729 * Add a page to the dirty page list.
731 * It is a sad fact of life that this function is called from several places
732 * deeply under spinlocking. It may not sleep.
734 * If the page has buffers, the uptodate buffers are set dirty, to preserve
735 * dirty-state coherency between the page and the buffers. It the page does
736 * not have buffers then when they are later attached they will all be set
739 * The buffers are dirtied before the page is dirtied. There's a small race
740 * window in which a writepage caller may see the page cleanness but not the
741 * buffer dirtiness. That's fine. If this code were to set the page dirty
742 * before the buffers, a concurrent writepage caller could clear the page dirty
743 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
744 * page on the dirty page list.
746 * We use private_lock to lock against try_to_free_buffers while using the
747 * page's buffer list. Also use this to protect against clean buffers being
748 * added to the page after it was set dirty.
750 * FIXME: may need to call ->reservepage here as well. That's rather up to the
751 * address_space though.
753 int __set_page_dirty_buffers(struct page *page)
755 struct address_space *mapping = page_mapping(page);
757 if (unlikely(!mapping))
758 return !TestSetPageDirty(page);
760 spin_lock(&mapping->private_lock);
761 if (page_has_buffers(page)) {
762 struct buffer_head *head = page_buffers(page);
763 struct buffer_head *bh = head;
766 set_buffer_dirty(bh);
767 bh = bh->b_this_page;
768 } while (bh != head);
770 spin_unlock(&mapping->private_lock);
772 return __set_page_dirty(page, mapping, 1);
774 EXPORT_SYMBOL(__set_page_dirty_buffers);
777 * Write out and wait upon a list of buffers.
779 * We have conflicting pressures: we want to make sure that all
780 * initially dirty buffers get waited on, but that any subsequently
781 * dirtied buffers don't. After all, we don't want fsync to last
782 * forever if somebody is actively writing to the file.
784 * Do this in two main stages: first we copy dirty buffers to a
785 * temporary inode list, queueing the writes as we go. Then we clean
786 * up, waiting for those writes to complete.
788 * During this second stage, any subsequent updates to the file may end
789 * up refiling the buffer on the original inode's dirty list again, so
790 * there is a chance we will end up with a buffer queued for write but
791 * not yet completed on that list. So, as a final cleanup we go through
792 * the osync code to catch these locked, dirty buffers without requeuing
793 * any newly dirty buffers for write.
795 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
797 struct buffer_head *bh;
798 struct list_head tmp;
799 struct address_space *mapping;
802 INIT_LIST_HEAD(&tmp);
805 while (!list_empty(list)) {
806 bh = BH_ENTRY(list->next);
807 mapping = bh->b_assoc_map;
808 __remove_assoc_queue(bh);
809 /* Avoid race with mark_buffer_dirty_inode() which does
810 * a lockless check and we rely on seeing the dirty bit */
812 if (buffer_dirty(bh) || buffer_locked(bh)) {
813 list_add(&bh->b_assoc_buffers, &tmp);
814 bh->b_assoc_map = mapping;
815 if (buffer_dirty(bh)) {
819 * Ensure any pending I/O completes so that
820 * ll_rw_block() actually writes the current
821 * contents - it is a noop if I/O is still in
822 * flight on potentially older contents.
824 ll_rw_block(SWRITE_SYNC, 1, &bh);
831 while (!list_empty(&tmp)) {
832 bh = BH_ENTRY(tmp.prev);
834 mapping = bh->b_assoc_map;
835 __remove_assoc_queue(bh);
836 /* Avoid race with mark_buffer_dirty_inode() which does
837 * a lockless check and we rely on seeing the dirty bit */
839 if (buffer_dirty(bh)) {
840 list_add(&bh->b_assoc_buffers,
841 &mapping->private_list);
842 bh->b_assoc_map = mapping;
846 if (!buffer_uptodate(bh))
853 err2 = osync_buffers_list(lock, list);
861 * Invalidate any and all dirty buffers on a given inode. We are
862 * probably unmounting the fs, but that doesn't mean we have already
863 * done a sync(). Just drop the buffers from the inode list.
865 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
866 * assumes that all the buffers are against the blockdev. Not true
869 void invalidate_inode_buffers(struct inode *inode)
871 if (inode_has_buffers(inode)) {
872 struct address_space *mapping = &inode->i_data;
873 struct list_head *list = &mapping->private_list;
874 struct address_space *buffer_mapping = mapping->assoc_mapping;
876 spin_lock(&buffer_mapping->private_lock);
877 while (!list_empty(list))
878 __remove_assoc_queue(BH_ENTRY(list->next));
879 spin_unlock(&buffer_mapping->private_lock);
884 * Remove any clean buffers from the inode's buffer list. This is called
885 * when we're trying to free the inode itself. Those buffers can pin it.
887 * Returns true if all buffers were removed.
889 int remove_inode_buffers(struct inode *inode)
893 if (inode_has_buffers(inode)) {
894 struct address_space *mapping = &inode->i_data;
895 struct list_head *list = &mapping->private_list;
896 struct address_space *buffer_mapping = mapping->assoc_mapping;
898 spin_lock(&buffer_mapping->private_lock);
899 while (!list_empty(list)) {
900 struct buffer_head *bh = BH_ENTRY(list->next);
901 if (buffer_dirty(bh)) {
905 __remove_assoc_queue(bh);
907 spin_unlock(&buffer_mapping->private_lock);
913 * Create the appropriate buffers when given a page for data area and
914 * the size of each buffer.. Use the bh->b_this_page linked list to
915 * follow the buffers created. Return NULL if unable to create more
918 * The retry flag is used to differentiate async IO (paging, swapping)
919 * which may not fail from ordinary buffer allocations.
921 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
924 struct buffer_head *bh, *head;
930 while ((offset -= size) >= 0) {
931 bh = alloc_buffer_head(GFP_NOFS);
936 bh->b_this_page = head;
941 atomic_set(&bh->b_count, 0);
942 bh->b_private = NULL;
945 /* Link the buffer to its page */
946 set_bh_page(bh, page, offset);
948 init_buffer(bh, NULL, NULL);
952 * In case anything failed, we just free everything we got.
958 head = head->b_this_page;
959 free_buffer_head(bh);
964 * Return failure for non-async IO requests. Async IO requests
965 * are not allowed to fail, so we have to wait until buffer heads
966 * become available. But we don't want tasks sleeping with
967 * partially complete buffers, so all were released above.
972 /* We're _really_ low on memory. Now we just
973 * wait for old buffer heads to become free due to
974 * finishing IO. Since this is an async request and
975 * the reserve list is empty, we're sure there are
976 * async buffer heads in use.
981 EXPORT_SYMBOL_GPL(alloc_page_buffers);
984 link_dev_buffers(struct page *page, struct buffer_head *head)
986 struct buffer_head *bh, *tail;
991 bh = bh->b_this_page;
993 tail->b_this_page = head;
994 attach_page_buffers(page, head);
998 * Initialise the state of a blockdev page's buffers.
1001 init_page_buffers(struct page *page, struct block_device *bdev,
1002 sector_t block, int size)
1004 struct buffer_head *head = page_buffers(page);
1005 struct buffer_head *bh = head;
1006 int uptodate = PageUptodate(page);
1009 if (!buffer_mapped(bh)) {
1010 init_buffer(bh, NULL, NULL);
1012 bh->b_blocknr = block;
1014 set_buffer_uptodate(bh);
1015 set_buffer_mapped(bh);
1018 bh = bh->b_this_page;
1019 } while (bh != head);
1023 * Create the page-cache page that contains the requested block.
1025 * This is user purely for blockdev mappings.
1027 static struct page *
1028 grow_dev_page(struct block_device *bdev, sector_t block,
1029 pgoff_t index, int size)
1031 struct inode *inode = bdev->bd_inode;
1033 struct buffer_head *bh;
1035 page = find_or_create_page(inode->i_mapping, index,
1036 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1040 BUG_ON(!PageLocked(page));
1042 if (page_has_buffers(page)) {
1043 bh = page_buffers(page);
1044 if (bh->b_size == size) {
1045 init_page_buffers(page, bdev, block, size);
1048 if (!try_to_free_buffers(page))
1053 * Allocate some buffers for this page
1055 bh = alloc_page_buffers(page, size, 0);
1060 * Link the page to the buffers and initialise them. Take the
1061 * lock to be atomic wrt __find_get_block(), which does not
1062 * run under the page lock.
1064 spin_lock(&inode->i_mapping->private_lock);
1065 link_dev_buffers(page, bh);
1066 init_page_buffers(page, bdev, block, size);
1067 spin_unlock(&inode->i_mapping->private_lock);
1073 page_cache_release(page);
1078 * Create buffers for the specified block device block's page. If
1079 * that page was dirty, the buffers are set dirty also.
1082 grow_buffers(struct block_device *bdev, sector_t block, int size)
1091 } while ((size << sizebits) < PAGE_SIZE);
1093 index = block >> sizebits;
1096 * Check for a block which wants to lie outside our maximum possible
1097 * pagecache index. (this comparison is done using sector_t types).
1099 if (unlikely(index != block >> sizebits)) {
1100 char b[BDEVNAME_SIZE];
1102 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1104 __func__, (unsigned long long)block,
1108 block = index << sizebits;
1109 /* Create a page with the proper size buffers.. */
1110 page = grow_dev_page(bdev, block, index, size);
1114 page_cache_release(page);
1118 static struct buffer_head *
1119 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1121 /* Size must be multiple of hard sectorsize */
1122 if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1123 (size < 512 || size > PAGE_SIZE))) {
1124 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1126 printk(KERN_ERR "hardsect size: %d\n",
1127 bdev_hardsect_size(bdev));
1134 struct buffer_head * bh;
1137 bh = __find_get_block(bdev, block, size);
1141 ret = grow_buffers(bdev, block, size);
1150 * The relationship between dirty buffers and dirty pages:
1152 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1153 * the page is tagged dirty in its radix tree.
1155 * At all times, the dirtiness of the buffers represents the dirtiness of
1156 * subsections of the page. If the page has buffers, the page dirty bit is
1157 * merely a hint about the true dirty state.
1159 * When a page is set dirty in its entirety, all its buffers are marked dirty
1160 * (if the page has buffers).
1162 * When a buffer is marked dirty, its page is dirtied, but the page's other
1165 * Also. When blockdev buffers are explicitly read with bread(), they
1166 * individually become uptodate. But their backing page remains not
1167 * uptodate - even if all of its buffers are uptodate. A subsequent
1168 * block_read_full_page() against that page will discover all the uptodate
1169 * buffers, will set the page uptodate and will perform no I/O.
1173 * mark_buffer_dirty - mark a buffer_head as needing writeout
1174 * @bh: the buffer_head to mark dirty
1176 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1177 * backing page dirty, then tag the page as dirty in its address_space's radix
1178 * tree and then attach the address_space's inode to its superblock's dirty
1181 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1182 * mapping->tree_lock and the global inode_lock.
1184 void mark_buffer_dirty(struct buffer_head *bh)
1186 WARN_ON_ONCE(!buffer_uptodate(bh));
1189 * Very *carefully* optimize the it-is-already-dirty case.
1191 * Don't let the final "is it dirty" escape to before we
1192 * perhaps modified the buffer.
1194 if (buffer_dirty(bh)) {
1196 if (buffer_dirty(bh))
1200 if (!test_set_buffer_dirty(bh))
1201 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1205 * Decrement a buffer_head's reference count. If all buffers against a page
1206 * have zero reference count, are clean and unlocked, and if the page is clean
1207 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1208 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1209 * a page but it ends up not being freed, and buffers may later be reattached).
1211 void __brelse(struct buffer_head * buf)
1213 if (atomic_read(&buf->b_count)) {
1217 printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1222 * bforget() is like brelse(), except it discards any
1223 * potentially dirty data.
1225 void __bforget(struct buffer_head *bh)
1227 clear_buffer_dirty(bh);
1228 if (bh->b_assoc_map) {
1229 struct address_space *buffer_mapping = bh->b_page->mapping;
1231 spin_lock(&buffer_mapping->private_lock);
1232 list_del_init(&bh->b_assoc_buffers);
1233 bh->b_assoc_map = NULL;
1234 spin_unlock(&buffer_mapping->private_lock);
1239 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1242 if (buffer_uptodate(bh)) {
1247 bh->b_end_io = end_buffer_read_sync;
1248 submit_bh(READ, bh);
1250 if (buffer_uptodate(bh))
1258 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1259 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1260 * refcount elevated by one when they're in an LRU. A buffer can only appear
1261 * once in a particular CPU's LRU. A single buffer can be present in multiple
1262 * CPU's LRUs at the same time.
1264 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1265 * sb_find_get_block().
1267 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1268 * a local interrupt disable for that.
1271 #define BH_LRU_SIZE 8
1274 struct buffer_head *bhs[BH_LRU_SIZE];
1277 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1280 #define bh_lru_lock() local_irq_disable()
1281 #define bh_lru_unlock() local_irq_enable()
1283 #define bh_lru_lock() preempt_disable()
1284 #define bh_lru_unlock() preempt_enable()
1287 static inline void check_irqs_on(void)
1289 #ifdef irqs_disabled
1290 BUG_ON(irqs_disabled());
1295 * The LRU management algorithm is dopey-but-simple. Sorry.
1297 static void bh_lru_install(struct buffer_head *bh)
1299 struct buffer_head *evictee = NULL;
1304 lru = &__get_cpu_var(bh_lrus);
1305 if (lru->bhs[0] != bh) {
1306 struct buffer_head *bhs[BH_LRU_SIZE];
1312 for (in = 0; in < BH_LRU_SIZE; in++) {
1313 struct buffer_head *bh2 = lru->bhs[in];
1318 if (out >= BH_LRU_SIZE) {
1319 BUG_ON(evictee != NULL);
1326 while (out < BH_LRU_SIZE)
1328 memcpy(lru->bhs, bhs, sizeof(bhs));
1337 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1339 static struct buffer_head *
1340 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1342 struct buffer_head *ret = NULL;
1348 lru = &__get_cpu_var(bh_lrus);
1349 for (i = 0; i < BH_LRU_SIZE; i++) {
1350 struct buffer_head *bh = lru->bhs[i];
1352 if (bh && bh->b_bdev == bdev &&
1353 bh->b_blocknr == block && bh->b_size == size) {
1356 lru->bhs[i] = lru->bhs[i - 1];
1371 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1372 * it in the LRU and mark it as accessed. If it is not present then return
1375 struct buffer_head *
1376 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1378 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1381 bh = __find_get_block_slow(bdev, block);
1389 EXPORT_SYMBOL(__find_get_block);
1392 * __getblk will locate (and, if necessary, create) the buffer_head
1393 * which corresponds to the passed block_device, block and size. The
1394 * returned buffer has its reference count incremented.
1396 * __getblk() cannot fail - it just keeps trying. If you pass it an
1397 * illegal block number, __getblk() will happily return a buffer_head
1398 * which represents the non-existent block. Very weird.
1400 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1401 * attempt is failing. FIXME, perhaps?
1403 struct buffer_head *
1404 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1406 struct buffer_head *bh = __find_get_block(bdev, block, size);
1410 bh = __getblk_slow(bdev, block, size);
1413 EXPORT_SYMBOL(__getblk);
1416 * Do async read-ahead on a buffer..
1418 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1420 struct buffer_head *bh = __getblk(bdev, block, size);
1422 ll_rw_block(READA, 1, &bh);
1426 EXPORT_SYMBOL(__breadahead);
1429 * __bread() - reads a specified block and returns the bh
1430 * @bdev: the block_device to read from
1431 * @block: number of block
1432 * @size: size (in bytes) to read
1434 * Reads a specified block, and returns buffer head that contains it.
1435 * It returns NULL if the block was unreadable.
1437 struct buffer_head *
1438 __bread(struct block_device *bdev, sector_t block, unsigned size)
1440 struct buffer_head *bh = __getblk(bdev, block, size);
1442 if (likely(bh) && !buffer_uptodate(bh))
1443 bh = __bread_slow(bh);
1446 EXPORT_SYMBOL(__bread);
1449 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1450 * This doesn't race because it runs in each cpu either in irq
1451 * or with preempt disabled.
1453 static void invalidate_bh_lru(void *arg)
1455 struct bh_lru *b = &get_cpu_var(bh_lrus);
1458 for (i = 0; i < BH_LRU_SIZE; i++) {
1462 put_cpu_var(bh_lrus);
1465 void invalidate_bh_lrus(void)
1467 on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1469 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1471 void set_bh_page(struct buffer_head *bh,
1472 struct page *page, unsigned long offset)
1475 BUG_ON(offset >= PAGE_SIZE);
1476 if (PageHighMem(page))
1478 * This catches illegal uses and preserves the offset:
1480 bh->b_data = (char *)(0 + offset);
1482 bh->b_data = page_address(page) + offset;
1484 EXPORT_SYMBOL(set_bh_page);
1487 * Called when truncating a buffer on a page completely.
1489 static void discard_buffer(struct buffer_head * bh)
1492 clear_buffer_dirty(bh);
1494 clear_buffer_mapped(bh);
1495 clear_buffer_req(bh);
1496 clear_buffer_new(bh);
1497 clear_buffer_delay(bh);
1498 clear_buffer_unwritten(bh);
1503 * block_invalidatepage - invalidate part of all of a buffer-backed page
1505 * @page: the page which is affected
1506 * @offset: the index of the truncation point
1508 * block_invalidatepage() is called when all or part of the page has become
1509 * invalidatedby a truncate operation.
1511 * block_invalidatepage() does not have to release all buffers, but it must
1512 * ensure that no dirty buffer is left outside @offset and that no I/O
1513 * is underway against any of the blocks which are outside the truncation
1514 * point. Because the caller is about to free (and possibly reuse) those
1517 void block_invalidatepage(struct page *page, unsigned long offset)
1519 struct buffer_head *head, *bh, *next;
1520 unsigned int curr_off = 0;
1522 BUG_ON(!PageLocked(page));
1523 if (!page_has_buffers(page))
1526 head = page_buffers(page);
1529 unsigned int next_off = curr_off + bh->b_size;
1530 next = bh->b_this_page;
1533 * is this block fully invalidated?
1535 if (offset <= curr_off)
1537 curr_off = next_off;
1539 } while (bh != head);
1542 * We release buffers only if the entire page is being invalidated.
1543 * The get_block cached value has been unconditionally invalidated,
1544 * so real IO is not possible anymore.
1547 try_to_release_page(page, 0);
1551 EXPORT_SYMBOL(block_invalidatepage);
1554 * We attach and possibly dirty the buffers atomically wrt
1555 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1556 * is already excluded via the page lock.
1558 void create_empty_buffers(struct page *page,
1559 unsigned long blocksize, unsigned long b_state)
1561 struct buffer_head *bh, *head, *tail;
1563 head = alloc_page_buffers(page, blocksize, 1);
1566 bh->b_state |= b_state;
1568 bh = bh->b_this_page;
1570 tail->b_this_page = head;
1572 spin_lock(&page->mapping->private_lock);
1573 if (PageUptodate(page) || PageDirty(page)) {
1576 if (PageDirty(page))
1577 set_buffer_dirty(bh);
1578 if (PageUptodate(page))
1579 set_buffer_uptodate(bh);
1580 bh = bh->b_this_page;
1581 } while (bh != head);
1583 attach_page_buffers(page, head);
1584 spin_unlock(&page->mapping->private_lock);
1586 EXPORT_SYMBOL(create_empty_buffers);
1589 * We are taking a block for data and we don't want any output from any
1590 * buffer-cache aliases starting from return from that function and
1591 * until the moment when something will explicitly mark the buffer
1592 * dirty (hopefully that will not happen until we will free that block ;-)
1593 * We don't even need to mark it not-uptodate - nobody can expect
1594 * anything from a newly allocated buffer anyway. We used to used
1595 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1596 * don't want to mark the alias unmapped, for example - it would confuse
1597 * anyone who might pick it with bread() afterwards...
1599 * Also.. Note that bforget() doesn't lock the buffer. So there can
1600 * be writeout I/O going on against recently-freed buffers. We don't
1601 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1602 * only if we really need to. That happens here.
1604 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1606 struct buffer_head *old_bh;
1610 old_bh = __find_get_block_slow(bdev, block);
1612 clear_buffer_dirty(old_bh);
1613 wait_on_buffer(old_bh);
1614 clear_buffer_req(old_bh);
1618 EXPORT_SYMBOL(unmap_underlying_metadata);
1621 * NOTE! All mapped/uptodate combinations are valid:
1623 * Mapped Uptodate Meaning
1625 * No No "unknown" - must do get_block()
1626 * No Yes "hole" - zero-filled
1627 * Yes No "allocated" - allocated on disk, not read in
1628 * Yes Yes "valid" - allocated and up-to-date in memory.
1630 * "Dirty" is valid only with the last case (mapped+uptodate).
1634 * While block_write_full_page is writing back the dirty buffers under
1635 * the page lock, whoever dirtied the buffers may decide to clean them
1636 * again at any time. We handle that by only looking at the buffer
1637 * state inside lock_buffer().
1639 * If block_write_full_page() is called for regular writeback
1640 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1641 * locked buffer. This only can happen if someone has written the buffer
1642 * directly, with submit_bh(). At the address_space level PageWriteback
1643 * prevents this contention from occurring.
1645 static int __block_write_full_page(struct inode *inode, struct page *page,
1646 get_block_t *get_block, struct writeback_control *wbc)
1650 sector_t last_block;
1651 struct buffer_head *bh, *head;
1652 const unsigned blocksize = 1 << inode->i_blkbits;
1653 int nr_underway = 0;
1655 BUG_ON(!PageLocked(page));
1657 last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1659 if (!page_has_buffers(page)) {
1660 create_empty_buffers(page, blocksize,
1661 (1 << BH_Dirty)|(1 << BH_Uptodate));
1665 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1666 * here, and the (potentially unmapped) buffers may become dirty at
1667 * any time. If a buffer becomes dirty here after we've inspected it
1668 * then we just miss that fact, and the page stays dirty.
1670 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1671 * handle that here by just cleaning them.
1674 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1675 head = page_buffers(page);
1679 * Get all the dirty buffers mapped to disk addresses and
1680 * handle any aliases from the underlying blockdev's mapping.
1683 if (block > last_block) {
1685 * mapped buffers outside i_size will occur, because
1686 * this page can be outside i_size when there is a
1687 * truncate in progress.
1690 * The buffer was zeroed by block_write_full_page()
1692 clear_buffer_dirty(bh);
1693 set_buffer_uptodate(bh);
1694 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1695 WARN_ON(bh->b_size != blocksize);
1696 err = get_block(inode, block, bh, 1);
1699 if (buffer_new(bh)) {
1700 /* blockdev mappings never come here */
1701 clear_buffer_new(bh);
1702 unmap_underlying_metadata(bh->b_bdev,
1706 bh = bh->b_this_page;
1708 } while (bh != head);
1711 if (!buffer_mapped(bh))
1714 * If it's a fully non-blocking write attempt and we cannot
1715 * lock the buffer then redirty the page. Note that this can
1716 * potentially cause a busy-wait loop from pdflush and kswapd
1717 * activity, but those code paths have their own higher-level
1720 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1722 } else if (test_set_buffer_locked(bh)) {
1723 redirty_page_for_writepage(wbc, page);
1726 if (test_clear_buffer_dirty(bh)) {
1727 mark_buffer_async_write(bh);
1731 } while ((bh = bh->b_this_page) != head);
1734 * The page and its buffers are protected by PageWriteback(), so we can
1735 * drop the bh refcounts early.
1737 BUG_ON(PageWriteback(page));
1738 set_page_writeback(page);
1741 struct buffer_head *next = bh->b_this_page;
1742 if (buffer_async_write(bh)) {
1743 submit_bh(WRITE, bh);
1747 } while (bh != head);
1752 if (nr_underway == 0) {
1754 * The page was marked dirty, but the buffers were
1755 * clean. Someone wrote them back by hand with
1756 * ll_rw_block/submit_bh. A rare case.
1758 end_page_writeback(page);
1761 * The page and buffer_heads can be released at any time from
1769 * ENOSPC, or some other error. We may already have added some
1770 * blocks to the file, so we need to write these out to avoid
1771 * exposing stale data.
1772 * The page is currently locked and not marked for writeback
1775 /* Recovery: lock and submit the mapped buffers */
1777 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1779 mark_buffer_async_write(bh);
1782 * The buffer may have been set dirty during
1783 * attachment to a dirty page.
1785 clear_buffer_dirty(bh);
1787 } while ((bh = bh->b_this_page) != head);
1789 BUG_ON(PageWriteback(page));
1790 mapping_set_error(page->mapping, err);
1791 set_page_writeback(page);
1793 struct buffer_head *next = bh->b_this_page;
1794 if (buffer_async_write(bh)) {
1795 clear_buffer_dirty(bh);
1796 submit_bh(WRITE, bh);
1800 } while (bh != head);
1806 * If a page has any new buffers, zero them out here, and mark them uptodate
1807 * and dirty so they'll be written out (in order to prevent uninitialised
1808 * block data from leaking). And clear the new bit.
1810 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1812 unsigned int block_start, block_end;
1813 struct buffer_head *head, *bh;
1815 BUG_ON(!PageLocked(page));
1816 if (!page_has_buffers(page))
1819 bh = head = page_buffers(page);
1822 block_end = block_start + bh->b_size;
1824 if (buffer_new(bh)) {
1825 if (block_end > from && block_start < to) {
1826 if (!PageUptodate(page)) {
1827 unsigned start, size;
1829 start = max(from, block_start);
1830 size = min(to, block_end) - start;
1832 zero_user(page, start, size);
1833 set_buffer_uptodate(bh);
1836 clear_buffer_new(bh);
1837 mark_buffer_dirty(bh);
1841 block_start = block_end;
1842 bh = bh->b_this_page;
1843 } while (bh != head);
1845 EXPORT_SYMBOL(page_zero_new_buffers);
1847 static int __block_prepare_write(struct inode *inode, struct page *page,
1848 unsigned from, unsigned to, get_block_t *get_block)
1850 unsigned block_start, block_end;
1853 unsigned blocksize, bbits;
1854 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1856 BUG_ON(!PageLocked(page));
1857 BUG_ON(from > PAGE_CACHE_SIZE);
1858 BUG_ON(to > PAGE_CACHE_SIZE);
1861 blocksize = 1 << inode->i_blkbits;
1862 if (!page_has_buffers(page))
1863 create_empty_buffers(page, blocksize, 0);
1864 head = page_buffers(page);
1866 bbits = inode->i_blkbits;
1867 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1869 for(bh = head, block_start = 0; bh != head || !block_start;
1870 block++, block_start=block_end, bh = bh->b_this_page) {
1871 block_end = block_start + blocksize;
1872 if (block_end <= from || block_start >= to) {
1873 if (PageUptodate(page)) {
1874 if (!buffer_uptodate(bh))
1875 set_buffer_uptodate(bh);
1880 clear_buffer_new(bh);
1881 if (!buffer_mapped(bh)) {
1882 WARN_ON(bh->b_size != blocksize);
1883 err = get_block(inode, block, bh, 1);
1886 if (buffer_new(bh)) {
1887 unmap_underlying_metadata(bh->b_bdev,
1889 if (PageUptodate(page)) {
1890 clear_buffer_new(bh);
1891 set_buffer_uptodate(bh);
1892 mark_buffer_dirty(bh);
1895 if (block_end > to || block_start < from)
1896 zero_user_segments(page,
1902 if (PageUptodate(page)) {
1903 if (!buffer_uptodate(bh))
1904 set_buffer_uptodate(bh);
1907 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1908 !buffer_unwritten(bh) &&
1909 (block_start < from || block_end > to)) {
1910 ll_rw_block(READ, 1, &bh);
1915 * If we issued read requests - let them complete.
1917 while(wait_bh > wait) {
1918 wait_on_buffer(*--wait_bh);
1919 if (!buffer_uptodate(*wait_bh))
1923 page_zero_new_buffers(page, from, to);
1927 static int __block_commit_write(struct inode *inode, struct page *page,
1928 unsigned from, unsigned to)
1930 unsigned block_start, block_end;
1933 struct buffer_head *bh, *head;
1935 blocksize = 1 << inode->i_blkbits;
1937 for(bh = head = page_buffers(page), block_start = 0;
1938 bh != head || !block_start;
1939 block_start=block_end, bh = bh->b_this_page) {
1940 block_end = block_start + blocksize;
1941 if (block_end <= from || block_start >= to) {
1942 if (!buffer_uptodate(bh))
1945 set_buffer_uptodate(bh);
1946 mark_buffer_dirty(bh);
1948 clear_buffer_new(bh);
1952 * If this is a partial write which happened to make all buffers
1953 * uptodate then we can optimize away a bogus readpage() for
1954 * the next read(). Here we 'discover' whether the page went
1955 * uptodate as a result of this (potentially partial) write.
1958 SetPageUptodate(page);
1963 * block_write_begin takes care of the basic task of block allocation and
1964 * bringing partial write blocks uptodate first.
1966 * If *pagep is not NULL, then block_write_begin uses the locked page
1967 * at *pagep rather than allocating its own. In this case, the page will
1968 * not be unlocked or deallocated on failure.
1970 int block_write_begin(struct file *file, struct address_space *mapping,
1971 loff_t pos, unsigned len, unsigned flags,
1972 struct page **pagep, void **fsdata,
1973 get_block_t *get_block)
1975 struct inode *inode = mapping->host;
1979 unsigned start, end;
1982 index = pos >> PAGE_CACHE_SHIFT;
1983 start = pos & (PAGE_CACHE_SIZE - 1);
1989 page = __grab_cache_page(mapping, index);
1996 BUG_ON(!PageLocked(page));
1998 status = __block_prepare_write(inode, page, start, end, get_block);
1999 if (unlikely(status)) {
2000 ClearPageUptodate(page);
2004 page_cache_release(page);
2008 * prepare_write() may have instantiated a few blocks
2009 * outside i_size. Trim these off again. Don't need
2010 * i_size_read because we hold i_mutex.
2012 if (pos + len > inode->i_size)
2013 vmtruncate(inode, inode->i_size);
2021 EXPORT_SYMBOL(block_write_begin);
2023 int block_write_end(struct file *file, struct address_space *mapping,
2024 loff_t pos, unsigned len, unsigned copied,
2025 struct page *page, void *fsdata)
2027 struct inode *inode = mapping->host;
2030 start = pos & (PAGE_CACHE_SIZE - 1);
2032 if (unlikely(copied < len)) {
2034 * The buffers that were written will now be uptodate, so we
2035 * don't have to worry about a readpage reading them and
2036 * overwriting a partial write. However if we have encountered
2037 * a short write and only partially written into a buffer, it
2038 * will not be marked uptodate, so a readpage might come in and
2039 * destroy our partial write.
2041 * Do the simplest thing, and just treat any short write to a
2042 * non uptodate page as a zero-length write, and force the
2043 * caller to redo the whole thing.
2045 if (!PageUptodate(page))
2048 page_zero_new_buffers(page, start+copied, start+len);
2050 flush_dcache_page(page);
2052 /* This could be a short (even 0-length) commit */
2053 __block_commit_write(inode, page, start, start+copied);
2057 EXPORT_SYMBOL(block_write_end);
2059 int generic_write_end(struct file *file, struct address_space *mapping,
2060 loff_t pos, unsigned len, unsigned copied,
2061 struct page *page, void *fsdata)
2063 struct inode *inode = mapping->host;
2064 int i_size_changed = 0;
2066 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2069 * No need to use i_size_read() here, the i_size
2070 * cannot change under us because we hold i_mutex.
2072 * But it's important to update i_size while still holding page lock:
2073 * page writeout could otherwise come in and zero beyond i_size.
2075 if (pos+copied > inode->i_size) {
2076 i_size_write(inode, pos+copied);
2081 page_cache_release(page);
2084 * Don't mark the inode dirty under page lock. First, it unnecessarily
2085 * makes the holding time of page lock longer. Second, it forces lock
2086 * ordering of page lock and transaction start for journaling
2090 mark_inode_dirty(inode);
2094 EXPORT_SYMBOL(generic_write_end);
2097 * Generic "read page" function for block devices that have the normal
2098 * get_block functionality. This is most of the block device filesystems.
2099 * Reads the page asynchronously --- the unlock_buffer() and
2100 * set/clear_buffer_uptodate() functions propagate buffer state into the
2101 * page struct once IO has completed.
2103 int block_read_full_page(struct page *page, get_block_t *get_block)
2105 struct inode *inode = page->mapping->host;
2106 sector_t iblock, lblock;
2107 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2108 unsigned int blocksize;
2110 int fully_mapped = 1;
2112 BUG_ON(!PageLocked(page));
2113 blocksize = 1 << inode->i_blkbits;
2114 if (!page_has_buffers(page))
2115 create_empty_buffers(page, blocksize, 0);
2116 head = page_buffers(page);
2118 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2119 lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2125 if (buffer_uptodate(bh))
2128 if (!buffer_mapped(bh)) {
2132 if (iblock < lblock) {
2133 WARN_ON(bh->b_size != blocksize);
2134 err = get_block(inode, iblock, bh, 0);
2138 if (!buffer_mapped(bh)) {
2139 zero_user(page, i * blocksize, blocksize);
2141 set_buffer_uptodate(bh);
2145 * get_block() might have updated the buffer
2148 if (buffer_uptodate(bh))
2152 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2155 SetPageMappedToDisk(page);
2159 * All buffers are uptodate - we can set the page uptodate
2160 * as well. But not if get_block() returned an error.
2162 if (!PageError(page))
2163 SetPageUptodate(page);
2168 /* Stage two: lock the buffers */
2169 for (i = 0; i < nr; i++) {
2172 mark_buffer_async_read(bh);
2176 * Stage 3: start the IO. Check for uptodateness
2177 * inside the buffer lock in case another process reading
2178 * the underlying blockdev brought it uptodate (the sct fix).
2180 for (i = 0; i < nr; i++) {
2182 if (buffer_uptodate(bh))
2183 end_buffer_async_read(bh, 1);
2185 submit_bh(READ, bh);
2190 /* utility function for filesystems that need to do work on expanding
2191 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2192 * deal with the hole.
2194 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2196 struct address_space *mapping = inode->i_mapping;
2199 unsigned long limit;
2203 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2204 if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2205 send_sig(SIGXFSZ, current, 0);
2208 if (size > inode->i_sb->s_maxbytes)
2211 err = pagecache_write_begin(NULL, mapping, size, 0,
2212 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2217 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2224 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2225 loff_t pos, loff_t *bytes)
2227 struct inode *inode = mapping->host;
2228 unsigned blocksize = 1 << inode->i_blkbits;
2231 pgoff_t index, curidx;
2233 unsigned zerofrom, offset, len;
2236 index = pos >> PAGE_CACHE_SHIFT;
2237 offset = pos & ~PAGE_CACHE_MASK;
2239 while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2240 zerofrom = curpos & ~PAGE_CACHE_MASK;
2241 if (zerofrom & (blocksize-1)) {
2242 *bytes |= (blocksize-1);
2245 len = PAGE_CACHE_SIZE - zerofrom;
2247 err = pagecache_write_begin(file, mapping, curpos, len,
2248 AOP_FLAG_UNINTERRUPTIBLE,
2252 zero_user(page, zerofrom, len);
2253 err = pagecache_write_end(file, mapping, curpos, len, len,
2260 balance_dirty_pages_ratelimited(mapping);
2263 /* page covers the boundary, find the boundary offset */
2264 if (index == curidx) {
2265 zerofrom = curpos & ~PAGE_CACHE_MASK;
2266 /* if we will expand the thing last block will be filled */
2267 if (offset <= zerofrom) {
2270 if (zerofrom & (blocksize-1)) {
2271 *bytes |= (blocksize-1);
2274 len = offset - zerofrom;
2276 err = pagecache_write_begin(file, mapping, curpos, len,
2277 AOP_FLAG_UNINTERRUPTIBLE,
2281 zero_user(page, zerofrom, len);
2282 err = pagecache_write_end(file, mapping, curpos, len, len,
2294 * For moronic filesystems that do not allow holes in file.
2295 * We may have to extend the file.
2297 int cont_write_begin(struct file *file, struct address_space *mapping,
2298 loff_t pos, unsigned len, unsigned flags,
2299 struct page **pagep, void **fsdata,
2300 get_block_t *get_block, loff_t *bytes)
2302 struct inode *inode = mapping->host;
2303 unsigned blocksize = 1 << inode->i_blkbits;
2307 err = cont_expand_zero(file, mapping, pos, bytes);
2311 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2312 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2313 *bytes |= (blocksize-1);
2318 err = block_write_begin(file, mapping, pos, len,
2319 flags, pagep, fsdata, get_block);
2324 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2325 get_block_t *get_block)
2327 struct inode *inode = page->mapping->host;
2328 int err = __block_prepare_write(inode, page, from, to, get_block);
2330 ClearPageUptodate(page);
2334 int block_commit_write(struct page *page, unsigned from, unsigned to)
2336 struct inode *inode = page->mapping->host;
2337 __block_commit_write(inode,page,from,to);
2342 * block_page_mkwrite() is not allowed to change the file size as it gets
2343 * called from a page fault handler when a page is first dirtied. Hence we must
2344 * be careful to check for EOF conditions here. We set the page up correctly
2345 * for a written page which means we get ENOSPC checking when writing into
2346 * holes and correct delalloc and unwritten extent mapping on filesystems that
2347 * support these features.
2349 * We are not allowed to take the i_mutex here so we have to play games to
2350 * protect against truncate races as the page could now be beyond EOF. Because
2351 * vmtruncate() writes the inode size before removing pages, once we have the
2352 * page lock we can determine safely if the page is beyond EOF. If it is not
2353 * beyond EOF, then the page is guaranteed safe against truncation until we
2357 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2358 get_block_t get_block)
2360 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2366 size = i_size_read(inode);
2367 if ((page->mapping != inode->i_mapping) ||
2368 (page_offset(page) > size)) {
2369 /* page got truncated out from underneath us */
2373 /* page is wholly or partially inside EOF */
2374 if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2375 end = size & ~PAGE_CACHE_MASK;
2377 end = PAGE_CACHE_SIZE;
2379 ret = block_prepare_write(page, 0, end, get_block);
2381 ret = block_commit_write(page, 0, end);
2389 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2390 * immediately, while under the page lock. So it needs a special end_io
2391 * handler which does not touch the bh after unlocking it.
2393 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2395 __end_buffer_read_notouch(bh, uptodate);
2399 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2400 * the page (converting it to circular linked list and taking care of page
2403 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2405 struct buffer_head *bh;
2407 BUG_ON(!PageLocked(page));
2409 spin_lock(&page->mapping->private_lock);
2412 if (PageDirty(page))
2413 set_buffer_dirty(bh);
2414 if (!bh->b_this_page)
2415 bh->b_this_page = head;
2416 bh = bh->b_this_page;
2417 } while (bh != head);
2418 attach_page_buffers(page, head);
2419 spin_unlock(&page->mapping->private_lock);
2423 * On entry, the page is fully not uptodate.
2424 * On exit the page is fully uptodate in the areas outside (from,to)
2426 int nobh_write_begin(struct file *file, struct address_space *mapping,
2427 loff_t pos, unsigned len, unsigned flags,
2428 struct page **pagep, void **fsdata,
2429 get_block_t *get_block)
2431 struct inode *inode = mapping->host;
2432 const unsigned blkbits = inode->i_blkbits;
2433 const unsigned blocksize = 1 << blkbits;
2434 struct buffer_head *head, *bh;
2438 unsigned block_in_page;
2439 unsigned block_start, block_end;
2440 sector_t block_in_file;
2443 int is_mapped_to_disk = 1;
2445 index = pos >> PAGE_CACHE_SHIFT;
2446 from = pos & (PAGE_CACHE_SIZE - 1);
2449 page = __grab_cache_page(mapping, index);
2455 if (page_has_buffers(page)) {
2457 page_cache_release(page);
2459 return block_write_begin(file, mapping, pos, len, flags, pagep,
2463 if (PageMappedToDisk(page))
2467 * Allocate buffers so that we can keep track of state, and potentially
2468 * attach them to the page if an error occurs. In the common case of
2469 * no error, they will just be freed again without ever being attached
2470 * to the page (which is all OK, because we're under the page lock).
2472 * Be careful: the buffer linked list is a NULL terminated one, rather
2473 * than the circular one we're used to.
2475 head = alloc_page_buffers(page, blocksize, 0);
2481 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2484 * We loop across all blocks in the page, whether or not they are
2485 * part of the affected region. This is so we can discover if the
2486 * page is fully mapped-to-disk.
2488 for (block_start = 0, block_in_page = 0, bh = head;
2489 block_start < PAGE_CACHE_SIZE;
2490 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2493 block_end = block_start + blocksize;
2496 if (block_start >= to)
2498 ret = get_block(inode, block_in_file + block_in_page,
2502 if (!buffer_mapped(bh))
2503 is_mapped_to_disk = 0;
2505 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2506 if (PageUptodate(page)) {
2507 set_buffer_uptodate(bh);
2510 if (buffer_new(bh) || !buffer_mapped(bh)) {
2511 zero_user_segments(page, block_start, from,
2515 if (buffer_uptodate(bh))
2516 continue; /* reiserfs does this */
2517 if (block_start < from || block_end > to) {
2519 bh->b_end_io = end_buffer_read_nobh;
2520 submit_bh(READ, bh);
2527 * The page is locked, so these buffers are protected from
2528 * any VM or truncate activity. Hence we don't need to care
2529 * for the buffer_head refcounts.
2531 for (bh = head; bh; bh = bh->b_this_page) {
2533 if (!buffer_uptodate(bh))
2540 if (is_mapped_to_disk)
2541 SetPageMappedToDisk(page);
2543 *fsdata = head; /* to be released by nobh_write_end */
2550 * Error recovery is a bit difficult. We need to zero out blocks that
2551 * were newly allocated, and dirty them to ensure they get written out.
2552 * Buffers need to be attached to the page at this point, otherwise
2553 * the handling of potential IO errors during writeout would be hard
2554 * (could try doing synchronous writeout, but what if that fails too?)
2556 attach_nobh_buffers(page, head);
2557 page_zero_new_buffers(page, from, to);
2561 page_cache_release(page);
2564 if (pos + len > inode->i_size)
2565 vmtruncate(inode, inode->i_size);
2569 EXPORT_SYMBOL(nobh_write_begin);
2571 int nobh_write_end(struct file *file, struct address_space *mapping,
2572 loff_t pos, unsigned len, unsigned copied,
2573 struct page *page, void *fsdata)
2575 struct inode *inode = page->mapping->host;
2576 struct buffer_head *head = fsdata;
2577 struct buffer_head *bh;
2578 BUG_ON(fsdata != NULL && page_has_buffers(page));
2580 if (unlikely(copied < len) && !page_has_buffers(page))
2581 attach_nobh_buffers(page, head);
2582 if (page_has_buffers(page))
2583 return generic_write_end(file, mapping, pos, len,
2584 copied, page, fsdata);
2586 SetPageUptodate(page);
2587 set_page_dirty(page);
2588 if (pos+copied > inode->i_size) {
2589 i_size_write(inode, pos+copied);
2590 mark_inode_dirty(inode);
2594 page_cache_release(page);
2598 head = head->b_this_page;
2599 free_buffer_head(bh);
2604 EXPORT_SYMBOL(nobh_write_end);
2607 * nobh_writepage() - based on block_full_write_page() except
2608 * that it tries to operate without attaching bufferheads to
2611 int nobh_writepage(struct page *page, get_block_t *get_block,
2612 struct writeback_control *wbc)
2614 struct inode * const inode = page->mapping->host;
2615 loff_t i_size = i_size_read(inode);
2616 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2620 /* Is the page fully inside i_size? */
2621 if (page->index < end_index)
2624 /* Is the page fully outside i_size? (truncate in progress) */
2625 offset = i_size & (PAGE_CACHE_SIZE-1);
2626 if (page->index >= end_index+1 || !offset) {
2628 * The page may have dirty, unmapped buffers. For example,
2629 * they may have been added in ext3_writepage(). Make them
2630 * freeable here, so the page does not leak.
2633 /* Not really sure about this - do we need this ? */
2634 if (page->mapping->a_ops->invalidatepage)
2635 page->mapping->a_ops->invalidatepage(page, offset);
2638 return 0; /* don't care */
2642 * The page straddles i_size. It must be zeroed out on each and every
2643 * writepage invocation because it may be mmapped. "A file is mapped
2644 * in multiples of the page size. For a file that is not a multiple of
2645 * the page size, the remaining memory is zeroed when mapped, and
2646 * writes to that region are not written out to the file."
2648 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2650 ret = mpage_writepage(page, get_block, wbc);
2652 ret = __block_write_full_page(inode, page, get_block, wbc);
2655 EXPORT_SYMBOL(nobh_writepage);
2657 int nobh_truncate_page(struct address_space *mapping,
2658 loff_t from, get_block_t *get_block)
2660 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2661 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2664 unsigned length, pos;
2665 struct inode *inode = mapping->host;
2667 struct buffer_head map_bh;
2670 blocksize = 1 << inode->i_blkbits;
2671 length = offset & (blocksize - 1);
2673 /* Block boundary? Nothing to do */
2677 length = blocksize - length;
2678 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2680 page = grab_cache_page(mapping, index);
2685 if (page_has_buffers(page)) {
2688 page_cache_release(page);
2689 return block_truncate_page(mapping, from, get_block);
2692 /* Find the buffer that contains "offset" */
2694 while (offset >= pos) {
2699 err = get_block(inode, iblock, &map_bh, 0);
2702 /* unmapped? It's a hole - nothing to do */
2703 if (!buffer_mapped(&map_bh))
2706 /* Ok, it's mapped. Make sure it's up-to-date */
2707 if (!PageUptodate(page)) {
2708 err = mapping->a_ops->readpage(NULL, page);
2710 page_cache_release(page);
2714 if (!PageUptodate(page)) {
2718 if (page_has_buffers(page))
2721 zero_user(page, offset, length);
2722 set_page_dirty(page);
2727 page_cache_release(page);
2731 EXPORT_SYMBOL(nobh_truncate_page);
2733 int block_truncate_page(struct address_space *mapping,
2734 loff_t from, get_block_t *get_block)
2736 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2737 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2740 unsigned length, pos;
2741 struct inode *inode = mapping->host;
2743 struct buffer_head *bh;
2746 blocksize = 1 << inode->i_blkbits;
2747 length = offset & (blocksize - 1);
2749 /* Block boundary? Nothing to do */
2753 length = blocksize - length;
2754 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2756 page = grab_cache_page(mapping, index);
2761 if (!page_has_buffers(page))
2762 create_empty_buffers(page, blocksize, 0);
2764 /* Find the buffer that contains "offset" */
2765 bh = page_buffers(page);
2767 while (offset >= pos) {
2768 bh = bh->b_this_page;
2774 if (!buffer_mapped(bh)) {
2775 WARN_ON(bh->b_size != blocksize);
2776 err = get_block(inode, iblock, bh, 0);
2779 /* unmapped? It's a hole - nothing to do */
2780 if (!buffer_mapped(bh))
2784 /* Ok, it's mapped. Make sure it's up-to-date */
2785 if (PageUptodate(page))
2786 set_buffer_uptodate(bh);
2788 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2790 ll_rw_block(READ, 1, &bh);
2792 /* Uhhuh. Read error. Complain and punt. */
2793 if (!buffer_uptodate(bh))
2797 zero_user(page, offset, length);
2798 mark_buffer_dirty(bh);
2803 page_cache_release(page);
2809 * The generic ->writepage function for buffer-backed address_spaces
2811 int block_write_full_page(struct page *page, get_block_t *get_block,
2812 struct writeback_control *wbc)
2814 struct inode * const inode = page->mapping->host;
2815 loff_t i_size = i_size_read(inode);
2816 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2819 /* Is the page fully inside i_size? */
2820 if (page->index < end_index)
2821 return __block_write_full_page(inode, page, get_block, wbc);
2823 /* Is the page fully outside i_size? (truncate in progress) */
2824 offset = i_size & (PAGE_CACHE_SIZE-1);
2825 if (page->index >= end_index+1 || !offset) {
2827 * The page may have dirty, unmapped buffers. For example,
2828 * they may have been added in ext3_writepage(). Make them
2829 * freeable here, so the page does not leak.
2831 do_invalidatepage(page, 0);
2833 return 0; /* don't care */
2837 * The page straddles i_size. It must be zeroed out on each and every
2838 * writepage invokation because it may be mmapped. "A file is mapped
2839 * in multiples of the page size. For a file that is not a multiple of
2840 * the page size, the remaining memory is zeroed when mapped, and
2841 * writes to that region are not written out to the file."
2843 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2844 return __block_write_full_page(inode, page, get_block, wbc);
2847 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2848 get_block_t *get_block)
2850 struct buffer_head tmp;
2851 struct inode *inode = mapping->host;
2854 tmp.b_size = 1 << inode->i_blkbits;
2855 get_block(inode, block, &tmp, 0);
2856 return tmp.b_blocknr;
2859 static void end_bio_bh_io_sync(struct bio *bio, int err)
2861 struct buffer_head *bh = bio->bi_private;
2863 if (err == -EOPNOTSUPP) {
2864 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2865 set_bit(BH_Eopnotsupp, &bh->b_state);
2868 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2872 int submit_bh(int rw, struct buffer_head * bh)
2877 BUG_ON(!buffer_locked(bh));
2878 BUG_ON(!buffer_mapped(bh));
2879 BUG_ON(!bh->b_end_io);
2881 if (buffer_ordered(bh) && (rw == WRITE))
2885 * Only clear out a write error when rewriting, should this
2886 * include WRITE_SYNC as well?
2888 if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2889 clear_buffer_write_io_error(bh);
2892 * from here on down, it's all bio -- do the initial mapping,
2893 * submit_bio -> generic_make_request may further map this bio around
2895 bio = bio_alloc(GFP_NOIO, 1);
2897 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2898 bio->bi_bdev = bh->b_bdev;
2899 bio->bi_io_vec[0].bv_page = bh->b_page;
2900 bio->bi_io_vec[0].bv_len = bh->b_size;
2901 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2905 bio->bi_size = bh->b_size;
2907 bio->bi_end_io = end_bio_bh_io_sync;
2908 bio->bi_private = bh;
2911 submit_bio(rw, bio);
2913 if (bio_flagged(bio, BIO_EOPNOTSUPP))
2921 * ll_rw_block: low-level access to block devices (DEPRECATED)
2922 * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2923 * @nr: number of &struct buffer_heads in the array
2924 * @bhs: array of pointers to &struct buffer_head
2926 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2927 * requests an I/O operation on them, either a %READ or a %WRITE. The third
2928 * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2929 * are sent to disk. The fourth %READA option is described in the documentation
2930 * for generic_make_request() which ll_rw_block() calls.
2932 * This function drops any buffer that it cannot get a lock on (with the
2933 * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2934 * clean when doing a write request, and any buffer that appears to be
2935 * up-to-date when doing read request. Further it marks as clean buffers that
2936 * are processed for writing (the buffer cache won't assume that they are
2937 * actually clean until the buffer gets unlocked).
2939 * ll_rw_block sets b_end_io to simple completion handler that marks
2940 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2943 * All of the buffers must be for the same device, and must also be a
2944 * multiple of the current approved size for the device.
2946 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2950 for (i = 0; i < nr; i++) {
2951 struct buffer_head *bh = bhs[i];
2953 if (rw == SWRITE || rw == SWRITE_SYNC)
2955 else if (test_set_buffer_locked(bh))
2958 if (rw == WRITE || rw == SWRITE || rw == SWRITE_SYNC) {
2959 if (test_clear_buffer_dirty(bh)) {
2960 bh->b_end_io = end_buffer_write_sync;
2962 if (rw == SWRITE_SYNC)
2963 submit_bh(WRITE_SYNC, bh);
2965 submit_bh(WRITE, bh);
2969 if (!buffer_uptodate(bh)) {
2970 bh->b_end_io = end_buffer_read_sync;
2981 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2982 * and then start new I/O and then wait upon it. The caller must have a ref on
2985 int sync_dirty_buffer(struct buffer_head *bh)
2989 WARN_ON(atomic_read(&bh->b_count) < 1);
2991 if (test_clear_buffer_dirty(bh)) {
2993 bh->b_end_io = end_buffer_write_sync;
2994 ret = submit_bh(WRITE_SYNC, bh);
2996 if (buffer_eopnotsupp(bh)) {
2997 clear_buffer_eopnotsupp(bh);
3000 if (!ret && !buffer_uptodate(bh))
3009 * try_to_free_buffers() checks if all the buffers on this particular page
3010 * are unused, and releases them if so.
3012 * Exclusion against try_to_free_buffers may be obtained by either
3013 * locking the page or by holding its mapping's private_lock.
3015 * If the page is dirty but all the buffers are clean then we need to
3016 * be sure to mark the page clean as well. This is because the page
3017 * may be against a block device, and a later reattachment of buffers
3018 * to a dirty page will set *all* buffers dirty. Which would corrupt
3019 * filesystem data on the same device.
3021 * The same applies to regular filesystem pages: if all the buffers are
3022 * clean then we set the page clean and proceed. To do that, we require
3023 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3026 * try_to_free_buffers() is non-blocking.
3028 static inline int buffer_busy(struct buffer_head *bh)
3030 return atomic_read(&bh->b_count) |
3031 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3035 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3037 struct buffer_head *head = page_buffers(page);
3038 struct buffer_head *bh;
3042 if (buffer_write_io_error(bh) && page->mapping)
3043 set_bit(AS_EIO, &page->mapping->flags);
3044 if (buffer_busy(bh))
3046 bh = bh->b_this_page;
3047 } while (bh != head);
3050 struct buffer_head *next = bh->b_this_page;
3052 if (bh->b_assoc_map)
3053 __remove_assoc_queue(bh);
3055 } while (bh != head);
3056 *buffers_to_free = head;
3057 __clear_page_buffers(page);
3063 int try_to_free_buffers(struct page *page)
3065 struct address_space * const mapping = page->mapping;
3066 struct buffer_head *buffers_to_free = NULL;
3069 BUG_ON(!PageLocked(page));
3070 if (PageWriteback(page))
3073 if (mapping == NULL) { /* can this still happen? */
3074 ret = drop_buffers(page, &buffers_to_free);
3078 spin_lock(&mapping->private_lock);
3079 ret = drop_buffers(page, &buffers_to_free);
3082 * If the filesystem writes its buffers by hand (eg ext3)
3083 * then we can have clean buffers against a dirty page. We
3084 * clean the page here; otherwise the VM will never notice
3085 * that the filesystem did any IO at all.
3087 * Also, during truncate, discard_buffer will have marked all
3088 * the page's buffers clean. We discover that here and clean
3091 * private_lock must be held over this entire operation in order
3092 * to synchronise against __set_page_dirty_buffers and prevent the
3093 * dirty bit from being lost.
3096 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3097 spin_unlock(&mapping->private_lock);
3099 if (buffers_to_free) {
3100 struct buffer_head *bh = buffers_to_free;
3103 struct buffer_head *next = bh->b_this_page;
3104 free_buffer_head(bh);
3106 } while (bh != buffers_to_free);
3110 EXPORT_SYMBOL(try_to_free_buffers);
3112 void block_sync_page(struct page *page)
3114 struct address_space *mapping;
3117 mapping = page_mapping(page);
3119 blk_run_backing_dev(mapping->backing_dev_info, page);
3123 * There are no bdflush tunables left. But distributions are
3124 * still running obsolete flush daemons, so we terminate them here.
3126 * Use of bdflush() is deprecated and will be removed in a future kernel.
3127 * The `pdflush' kernel threads fully replace bdflush daemons and this call.
3129 asmlinkage long sys_bdflush(int func, long data)
3131 static int msg_count;
3133 if (!capable(CAP_SYS_ADMIN))
3136 if (msg_count < 5) {
3139 "warning: process `%s' used the obsolete bdflush"
3140 " system call\n", current->comm);
3141 printk(KERN_INFO "Fix your initscripts?\n");
3150 * Buffer-head allocation
3152 static struct kmem_cache *bh_cachep;
3155 * Once the number of bh's in the machine exceeds this level, we start
3156 * stripping them in writeback.
3158 static int max_buffer_heads;
3160 int buffer_heads_over_limit;
3162 struct bh_accounting {
3163 int nr; /* Number of live bh's */
3164 int ratelimit; /* Limit cacheline bouncing */
3167 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3169 static void recalc_bh_state(void)
3174 if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
3176 __get_cpu_var(bh_accounting).ratelimit = 0;
3177 for_each_online_cpu(i)
3178 tot += per_cpu(bh_accounting, i).nr;
3179 buffer_heads_over_limit = (tot > max_buffer_heads);
3182 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3184 struct buffer_head *ret = kmem_cache_alloc(bh_cachep, gfp_flags);
3186 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3187 get_cpu_var(bh_accounting).nr++;
3189 put_cpu_var(bh_accounting);
3193 EXPORT_SYMBOL(alloc_buffer_head);
3195 void free_buffer_head(struct buffer_head *bh)
3197 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3198 kmem_cache_free(bh_cachep, bh);
3199 get_cpu_var(bh_accounting).nr--;
3201 put_cpu_var(bh_accounting);
3203 EXPORT_SYMBOL(free_buffer_head);
3205 static void buffer_exit_cpu(int cpu)
3208 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3210 for (i = 0; i < BH_LRU_SIZE; i++) {
3214 get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
3215 per_cpu(bh_accounting, cpu).nr = 0;
3216 put_cpu_var(bh_accounting);
3219 static int buffer_cpu_notify(struct notifier_block *self,
3220 unsigned long action, void *hcpu)
3222 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3223 buffer_exit_cpu((unsigned long)hcpu);
3228 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3229 * @bh: struct buffer_head
3231 * Return true if the buffer is up-to-date and false,
3232 * with the buffer locked, if not.
3234 int bh_uptodate_or_lock(struct buffer_head *bh)
3236 if (!buffer_uptodate(bh)) {
3238 if (!buffer_uptodate(bh))
3244 EXPORT_SYMBOL(bh_uptodate_or_lock);
3247 * bh_submit_read - Submit a locked buffer for reading
3248 * @bh: struct buffer_head
3250 * Returns zero on success and -EIO on error.
3252 int bh_submit_read(struct buffer_head *bh)
3254 BUG_ON(!buffer_locked(bh));
3256 if (buffer_uptodate(bh)) {
3262 bh->b_end_io = end_buffer_read_sync;
3263 submit_bh(READ, bh);
3265 if (buffer_uptodate(bh))
3269 EXPORT_SYMBOL(bh_submit_read);
3272 init_buffer_head(struct kmem_cache *cachep, void *data)
3274 struct buffer_head *bh = data;
3276 memset(bh, 0, sizeof(*bh));
3277 INIT_LIST_HEAD(&bh->b_assoc_buffers);
3280 void __init buffer_init(void)
3284 bh_cachep = kmem_cache_create("buffer_head",
3285 sizeof(struct buffer_head), 0,
3286 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3291 * Limit the bh occupancy to 10% of ZONE_NORMAL
3293 nrpages = (nr_free_buffer_pages() * 10) / 100;
3294 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3295 hotcpu_notifier(buffer_cpu_notify, 0);
3298 EXPORT_SYMBOL(__bforget);
3299 EXPORT_SYMBOL(__brelse);
3300 EXPORT_SYMBOL(__wait_on_buffer);
3301 EXPORT_SYMBOL(block_commit_write);
3302 EXPORT_SYMBOL(block_prepare_write);
3303 EXPORT_SYMBOL(block_page_mkwrite);
3304 EXPORT_SYMBOL(block_read_full_page);
3305 EXPORT_SYMBOL(block_sync_page);
3306 EXPORT_SYMBOL(block_truncate_page);
3307 EXPORT_SYMBOL(block_write_full_page);
3308 EXPORT_SYMBOL(cont_write_begin);
3309 EXPORT_SYMBOL(end_buffer_read_sync);
3310 EXPORT_SYMBOL(end_buffer_write_sync);
3311 EXPORT_SYMBOL(file_fsync);
3312 EXPORT_SYMBOL(fsync_bdev);
3313 EXPORT_SYMBOL(generic_block_bmap);
3314 EXPORT_SYMBOL(generic_cont_expand_simple);
3315 EXPORT_SYMBOL(init_buffer);
3316 EXPORT_SYMBOL(invalidate_bdev);
3317 EXPORT_SYMBOL(ll_rw_block);
3318 EXPORT_SYMBOL(mark_buffer_dirty);
3319 EXPORT_SYMBOL(submit_bh);
3320 EXPORT_SYMBOL(sync_dirty_buffer);
3321 EXPORT_SYMBOL(unlock_buffer);