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 fastcall __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 fastcall 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_pgdat(pgdat) {
370 zones = pgdat->node_zonelists[gfp_zone(GFP_NOFS)].zones;
372 try_to_free_pages(zones, 0, GFP_NOFS);
377 * I/O completion handler for block_read_full_page() - pages
378 * which come unlocked at the end of I/O.
380 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
383 struct buffer_head *first;
384 struct buffer_head *tmp;
386 int page_uptodate = 1;
388 BUG_ON(!buffer_async_read(bh));
392 set_buffer_uptodate(bh);
394 clear_buffer_uptodate(bh);
395 if (printk_ratelimit())
401 * Be _very_ careful from here on. Bad things can happen if
402 * two buffer heads end IO at almost the same time and both
403 * decide that the page is now completely done.
405 first = page_buffers(page);
406 local_irq_save(flags);
407 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
408 clear_buffer_async_read(bh);
412 if (!buffer_uptodate(tmp))
414 if (buffer_async_read(tmp)) {
415 BUG_ON(!buffer_locked(tmp));
418 tmp = tmp->b_this_page;
420 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
421 local_irq_restore(flags);
424 * If none of the buffers had errors and they are all
425 * uptodate then we can set the page uptodate.
427 if (page_uptodate && !PageError(page))
428 SetPageUptodate(page);
433 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
434 local_irq_restore(flags);
439 * Completion handler for block_write_full_page() - pages which are unlocked
440 * during I/O, and which have PageWriteback cleared upon I/O completion.
442 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
444 char b[BDEVNAME_SIZE];
446 struct buffer_head *first;
447 struct buffer_head *tmp;
450 BUG_ON(!buffer_async_write(bh));
454 set_buffer_uptodate(bh);
456 if (printk_ratelimit()) {
458 printk(KERN_WARNING "lost page write due to "
460 bdevname(bh->b_bdev, b));
462 set_bit(AS_EIO, &page->mapping->flags);
463 set_buffer_write_io_error(bh);
464 clear_buffer_uptodate(bh);
468 first = page_buffers(page);
469 local_irq_save(flags);
470 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
472 clear_buffer_async_write(bh);
474 tmp = bh->b_this_page;
476 if (buffer_async_write(tmp)) {
477 BUG_ON(!buffer_locked(tmp));
480 tmp = tmp->b_this_page;
482 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
483 local_irq_restore(flags);
484 end_page_writeback(page);
488 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
489 local_irq_restore(flags);
494 * If a page's buffers are under async readin (end_buffer_async_read
495 * completion) then there is a possibility that another thread of
496 * control could lock one of the buffers after it has completed
497 * but while some of the other buffers have not completed. This
498 * locked buffer would confuse end_buffer_async_read() into not unlocking
499 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
500 * that this buffer is not under async I/O.
502 * The page comes unlocked when it has no locked buffer_async buffers
505 * PageLocked prevents anyone starting new async I/O reads any of
508 * PageWriteback is used to prevent simultaneous writeout of the same
511 * PageLocked prevents anyone from starting writeback of a page which is
512 * under read I/O (PageWriteback is only ever set against a locked page).
514 static void mark_buffer_async_read(struct buffer_head *bh)
516 bh->b_end_io = end_buffer_async_read;
517 set_buffer_async_read(bh);
520 void mark_buffer_async_write(struct buffer_head *bh)
522 bh->b_end_io = end_buffer_async_write;
523 set_buffer_async_write(bh);
525 EXPORT_SYMBOL(mark_buffer_async_write);
529 * fs/buffer.c contains helper functions for buffer-backed address space's
530 * fsync functions. A common requirement for buffer-based filesystems is
531 * that certain data from the backing blockdev needs to be written out for
532 * a successful fsync(). For example, ext2 indirect blocks need to be
533 * written back and waited upon before fsync() returns.
535 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
536 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
537 * management of a list of dependent buffers at ->i_mapping->private_list.
539 * Locking is a little subtle: try_to_free_buffers() will remove buffers
540 * from their controlling inode's queue when they are being freed. But
541 * try_to_free_buffers() will be operating against the *blockdev* mapping
542 * at the time, not against the S_ISREG file which depends on those buffers.
543 * So the locking for private_list is via the private_lock in the address_space
544 * which backs the buffers. Which is different from the address_space
545 * against which the buffers are listed. So for a particular address_space,
546 * mapping->private_lock does *not* protect mapping->private_list! In fact,
547 * mapping->private_list will always be protected by the backing blockdev's
550 * Which introduces a requirement: all buffers on an address_space's
551 * ->private_list must be from the same address_space: the blockdev's.
553 * address_spaces which do not place buffers at ->private_list via these
554 * utility functions are free to use private_lock and private_list for
555 * whatever they want. The only requirement is that list_empty(private_list)
556 * be true at clear_inode() time.
558 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
559 * filesystems should do that. invalidate_inode_buffers() should just go
560 * BUG_ON(!list_empty).
562 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
563 * take an address_space, not an inode. And it should be called
564 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
567 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
568 * list if it is already on a list. Because if the buffer is on a list,
569 * it *must* already be on the right one. If not, the filesystem is being
570 * silly. This will save a ton of locking. But first we have to ensure
571 * that buffers are taken *off* the old inode's list when they are freed
572 * (presumably in truncate). That requires careful auditing of all
573 * filesystems (do it inside bforget()). It could also be done by bringing
578 * The buffer's backing address_space's private_lock must be held
580 static inline void __remove_assoc_queue(struct buffer_head *bh)
582 list_del_init(&bh->b_assoc_buffers);
583 WARN_ON(!bh->b_assoc_map);
584 if (buffer_write_io_error(bh))
585 set_bit(AS_EIO, &bh->b_assoc_map->flags);
586 bh->b_assoc_map = NULL;
589 int inode_has_buffers(struct inode *inode)
591 return !list_empty(&inode->i_data.private_list);
595 * osync is designed to support O_SYNC io. It waits synchronously for
596 * all already-submitted IO to complete, but does not queue any new
597 * writes to the disk.
599 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
600 * you dirty the buffers, and then use osync_inode_buffers to wait for
601 * completion. Any other dirty buffers which are not yet queued for
602 * write will not be flushed to disk by the osync.
604 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
606 struct buffer_head *bh;
612 list_for_each_prev(p, list) {
614 if (buffer_locked(bh)) {
618 if (!buffer_uptodate(bh))
630 * sync_mapping_buffers - write out and wait upon a mapping's "associated"
632 * @mapping: the mapping which wants those buffers written
634 * Starts I/O against the buffers at mapping->private_list, and waits upon
637 * Basically, this is a convenience function for fsync().
638 * @mapping is a file or directory which needs those buffers to be written for
639 * a successful fsync().
641 int sync_mapping_buffers(struct address_space *mapping)
643 struct address_space *buffer_mapping = mapping->assoc_mapping;
645 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
648 return fsync_buffers_list(&buffer_mapping->private_lock,
649 &mapping->private_list);
651 EXPORT_SYMBOL(sync_mapping_buffers);
654 * Called when we've recently written block `bblock', and it is known that
655 * `bblock' was for a buffer_boundary() buffer. This means that the block at
656 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
657 * dirty, schedule it for IO. So that indirects merge nicely with their data.
659 void write_boundary_block(struct block_device *bdev,
660 sector_t bblock, unsigned blocksize)
662 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
664 if (buffer_dirty(bh))
665 ll_rw_block(WRITE, 1, &bh);
670 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
672 struct address_space *mapping = inode->i_mapping;
673 struct address_space *buffer_mapping = bh->b_page->mapping;
675 mark_buffer_dirty(bh);
676 if (!mapping->assoc_mapping) {
677 mapping->assoc_mapping = buffer_mapping;
679 BUG_ON(mapping->assoc_mapping != buffer_mapping);
681 if (list_empty(&bh->b_assoc_buffers)) {
682 spin_lock(&buffer_mapping->private_lock);
683 list_move_tail(&bh->b_assoc_buffers,
684 &mapping->private_list);
685 bh->b_assoc_map = mapping;
686 spin_unlock(&buffer_mapping->private_lock);
689 EXPORT_SYMBOL(mark_buffer_dirty_inode);
692 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
695 * If warn is true, then emit a warning if the page is not uptodate and has
696 * not been truncated.
698 static int __set_page_dirty(struct page *page,
699 struct address_space *mapping, int warn)
701 if (unlikely(!mapping))
702 return !TestSetPageDirty(page);
704 if (TestSetPageDirty(page))
707 write_lock_irq(&mapping->tree_lock);
708 if (page->mapping) { /* Race with truncate? */
709 WARN_ON_ONCE(warn && !PageUptodate(page));
711 if (mapping_cap_account_dirty(mapping)) {
712 __inc_zone_page_state(page, NR_FILE_DIRTY);
713 task_io_account_write(PAGE_CACHE_SIZE);
715 radix_tree_tag_set(&mapping->page_tree,
716 page_index(page), PAGECACHE_TAG_DIRTY);
718 write_unlock_irq(&mapping->tree_lock);
719 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
725 * Add a page to the dirty page list.
727 * It is a sad fact of life that this function is called from several places
728 * deeply under spinlocking. It may not sleep.
730 * If the page has buffers, the uptodate buffers are set dirty, to preserve
731 * dirty-state coherency between the page and the buffers. It the page does
732 * not have buffers then when they are later attached they will all be set
735 * The buffers are dirtied before the page is dirtied. There's a small race
736 * window in which a writepage caller may see the page cleanness but not the
737 * buffer dirtiness. That's fine. If this code were to set the page dirty
738 * before the buffers, a concurrent writepage caller could clear the page dirty
739 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
740 * page on the dirty page list.
742 * We use private_lock to lock against try_to_free_buffers while using the
743 * page's buffer list. Also use this to protect against clean buffers being
744 * added to the page after it was set dirty.
746 * FIXME: may need to call ->reservepage here as well. That's rather up to the
747 * address_space though.
749 int __set_page_dirty_buffers(struct page *page)
751 struct address_space *mapping = page_mapping(page);
753 if (unlikely(!mapping))
754 return !TestSetPageDirty(page);
756 spin_lock(&mapping->private_lock);
757 if (page_has_buffers(page)) {
758 struct buffer_head *head = page_buffers(page);
759 struct buffer_head *bh = head;
762 set_buffer_dirty(bh);
763 bh = bh->b_this_page;
764 } while (bh != head);
766 spin_unlock(&mapping->private_lock);
768 return __set_page_dirty(page, mapping, 1);
770 EXPORT_SYMBOL(__set_page_dirty_buffers);
773 * Write out and wait upon a list of buffers.
775 * We have conflicting pressures: we want to make sure that all
776 * initially dirty buffers get waited on, but that any subsequently
777 * dirtied buffers don't. After all, we don't want fsync to last
778 * forever if somebody is actively writing to the file.
780 * Do this in two main stages: first we copy dirty buffers to a
781 * temporary inode list, queueing the writes as we go. Then we clean
782 * up, waiting for those writes to complete.
784 * During this second stage, any subsequent updates to the file may end
785 * up refiling the buffer on the original inode's dirty list again, so
786 * there is a chance we will end up with a buffer queued for write but
787 * not yet completed on that list. So, as a final cleanup we go through
788 * the osync code to catch these locked, dirty buffers without requeuing
789 * any newly dirty buffers for write.
791 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
793 struct buffer_head *bh;
794 struct list_head tmp;
797 INIT_LIST_HEAD(&tmp);
800 while (!list_empty(list)) {
801 bh = BH_ENTRY(list->next);
802 __remove_assoc_queue(bh);
803 if (buffer_dirty(bh) || buffer_locked(bh)) {
804 list_add(&bh->b_assoc_buffers, &tmp);
805 if (buffer_dirty(bh)) {
809 * Ensure any pending I/O completes so that
810 * ll_rw_block() actually writes the current
811 * contents - it is a noop if I/O is still in
812 * flight on potentially older contents.
814 ll_rw_block(SWRITE, 1, &bh);
821 while (!list_empty(&tmp)) {
822 bh = BH_ENTRY(tmp.prev);
823 list_del_init(&bh->b_assoc_buffers);
827 if (!buffer_uptodate(bh))
834 err2 = osync_buffers_list(lock, list);
842 * Invalidate any and all dirty buffers on a given inode. We are
843 * probably unmounting the fs, but that doesn't mean we have already
844 * done a sync(). Just drop the buffers from the inode list.
846 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
847 * assumes that all the buffers are against the blockdev. Not true
850 void invalidate_inode_buffers(struct inode *inode)
852 if (inode_has_buffers(inode)) {
853 struct address_space *mapping = &inode->i_data;
854 struct list_head *list = &mapping->private_list;
855 struct address_space *buffer_mapping = mapping->assoc_mapping;
857 spin_lock(&buffer_mapping->private_lock);
858 while (!list_empty(list))
859 __remove_assoc_queue(BH_ENTRY(list->next));
860 spin_unlock(&buffer_mapping->private_lock);
865 * Remove any clean buffers from the inode's buffer list. This is called
866 * when we're trying to free the inode itself. Those buffers can pin it.
868 * Returns true if all buffers were removed.
870 int remove_inode_buffers(struct inode *inode)
874 if (inode_has_buffers(inode)) {
875 struct address_space *mapping = &inode->i_data;
876 struct list_head *list = &mapping->private_list;
877 struct address_space *buffer_mapping = mapping->assoc_mapping;
879 spin_lock(&buffer_mapping->private_lock);
880 while (!list_empty(list)) {
881 struct buffer_head *bh = BH_ENTRY(list->next);
882 if (buffer_dirty(bh)) {
886 __remove_assoc_queue(bh);
888 spin_unlock(&buffer_mapping->private_lock);
894 * Create the appropriate buffers when given a page for data area and
895 * the size of each buffer.. Use the bh->b_this_page linked list to
896 * follow the buffers created. Return NULL if unable to create more
899 * The retry flag is used to differentiate async IO (paging, swapping)
900 * which may not fail from ordinary buffer allocations.
902 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
905 struct buffer_head *bh, *head;
911 while ((offset -= size) >= 0) {
912 bh = alloc_buffer_head(GFP_NOFS);
917 bh->b_this_page = head;
922 atomic_set(&bh->b_count, 0);
923 bh->b_private = NULL;
926 /* Link the buffer to its page */
927 set_bh_page(bh, page, offset);
929 init_buffer(bh, NULL, NULL);
933 * In case anything failed, we just free everything we got.
939 head = head->b_this_page;
940 free_buffer_head(bh);
945 * Return failure for non-async IO requests. Async IO requests
946 * are not allowed to fail, so we have to wait until buffer heads
947 * become available. But we don't want tasks sleeping with
948 * partially complete buffers, so all were released above.
953 /* We're _really_ low on memory. Now we just
954 * wait for old buffer heads to become free due to
955 * finishing IO. Since this is an async request and
956 * the reserve list is empty, we're sure there are
957 * async buffer heads in use.
962 EXPORT_SYMBOL_GPL(alloc_page_buffers);
965 link_dev_buffers(struct page *page, struct buffer_head *head)
967 struct buffer_head *bh, *tail;
972 bh = bh->b_this_page;
974 tail->b_this_page = head;
975 attach_page_buffers(page, head);
979 * Initialise the state of a blockdev page's buffers.
982 init_page_buffers(struct page *page, struct block_device *bdev,
983 sector_t block, int size)
985 struct buffer_head *head = page_buffers(page);
986 struct buffer_head *bh = head;
987 int uptodate = PageUptodate(page);
990 if (!buffer_mapped(bh)) {
991 init_buffer(bh, NULL, NULL);
993 bh->b_blocknr = block;
995 set_buffer_uptodate(bh);
996 set_buffer_mapped(bh);
999 bh = bh->b_this_page;
1000 } while (bh != head);
1004 * Create the page-cache page that contains the requested block.
1006 * This is user purely for blockdev mappings.
1008 static struct page *
1009 grow_dev_page(struct block_device *bdev, sector_t block,
1010 pgoff_t index, int size)
1012 struct inode *inode = bdev->bd_inode;
1014 struct buffer_head *bh;
1016 page = find_or_create_page(inode->i_mapping, index,
1017 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1021 BUG_ON(!PageLocked(page));
1023 if (page_has_buffers(page)) {
1024 bh = page_buffers(page);
1025 if (bh->b_size == size) {
1026 init_page_buffers(page, bdev, block, size);
1029 if (!try_to_free_buffers(page))
1034 * Allocate some buffers for this page
1036 bh = alloc_page_buffers(page, size, 0);
1041 * Link the page to the buffers and initialise them. Take the
1042 * lock to be atomic wrt __find_get_block(), which does not
1043 * run under the page lock.
1045 spin_lock(&inode->i_mapping->private_lock);
1046 link_dev_buffers(page, bh);
1047 init_page_buffers(page, bdev, block, size);
1048 spin_unlock(&inode->i_mapping->private_lock);
1054 page_cache_release(page);
1059 * Create buffers for the specified block device block's page. If
1060 * that page was dirty, the buffers are set dirty also.
1063 grow_buffers(struct block_device *bdev, sector_t block, int size)
1072 } while ((size << sizebits) < PAGE_SIZE);
1074 index = block >> sizebits;
1077 * Check for a block which wants to lie outside our maximum possible
1078 * pagecache index. (this comparison is done using sector_t types).
1080 if (unlikely(index != block >> sizebits)) {
1081 char b[BDEVNAME_SIZE];
1083 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1085 __FUNCTION__, (unsigned long long)block,
1089 block = index << sizebits;
1090 /* Create a page with the proper size buffers.. */
1091 page = grow_dev_page(bdev, block, index, size);
1095 page_cache_release(page);
1099 static struct buffer_head *
1100 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1102 /* Size must be multiple of hard sectorsize */
1103 if (unlikely(size & (bdev_hardsect_size(bdev)-1) ||
1104 (size < 512 || size > PAGE_SIZE))) {
1105 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1107 printk(KERN_ERR "hardsect size: %d\n",
1108 bdev_hardsect_size(bdev));
1115 struct buffer_head * bh;
1118 bh = __find_get_block(bdev, block, size);
1122 ret = grow_buffers(bdev, block, size);
1131 * The relationship between dirty buffers and dirty pages:
1133 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1134 * the page is tagged dirty in its radix tree.
1136 * At all times, the dirtiness of the buffers represents the dirtiness of
1137 * subsections of the page. If the page has buffers, the page dirty bit is
1138 * merely a hint about the true dirty state.
1140 * When a page is set dirty in its entirety, all its buffers are marked dirty
1141 * (if the page has buffers).
1143 * When a buffer is marked dirty, its page is dirtied, but the page's other
1146 * Also. When blockdev buffers are explicitly read with bread(), they
1147 * individually become uptodate. But their backing page remains not
1148 * uptodate - even if all of its buffers are uptodate. A subsequent
1149 * block_read_full_page() against that page will discover all the uptodate
1150 * buffers, will set the page uptodate and will perform no I/O.
1154 * mark_buffer_dirty - mark a buffer_head as needing writeout
1155 * @bh: the buffer_head to mark dirty
1157 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1158 * backing page dirty, then tag the page as dirty in its address_space's radix
1159 * tree and then attach the address_space's inode to its superblock's dirty
1162 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1163 * mapping->tree_lock and the global inode_lock.
1165 void fastcall mark_buffer_dirty(struct buffer_head *bh)
1167 WARN_ON_ONCE(!buffer_uptodate(bh));
1168 if (!buffer_dirty(bh) && !test_set_buffer_dirty(bh))
1169 __set_page_dirty(bh->b_page, page_mapping(bh->b_page), 0);
1173 * Decrement a buffer_head's reference count. If all buffers against a page
1174 * have zero reference count, are clean and unlocked, and if the page is clean
1175 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1176 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1177 * a page but it ends up not being freed, and buffers may later be reattached).
1179 void __brelse(struct buffer_head * buf)
1181 if (atomic_read(&buf->b_count)) {
1185 printk(KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1190 * bforget() is like brelse(), except it discards any
1191 * potentially dirty data.
1193 void __bforget(struct buffer_head *bh)
1195 clear_buffer_dirty(bh);
1196 if (!list_empty(&bh->b_assoc_buffers)) {
1197 struct address_space *buffer_mapping = bh->b_page->mapping;
1199 spin_lock(&buffer_mapping->private_lock);
1200 list_del_init(&bh->b_assoc_buffers);
1201 bh->b_assoc_map = NULL;
1202 spin_unlock(&buffer_mapping->private_lock);
1207 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1210 if (buffer_uptodate(bh)) {
1215 bh->b_end_io = end_buffer_read_sync;
1216 submit_bh(READ, bh);
1218 if (buffer_uptodate(bh))
1226 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1227 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1228 * refcount elevated by one when they're in an LRU. A buffer can only appear
1229 * once in a particular CPU's LRU. A single buffer can be present in multiple
1230 * CPU's LRUs at the same time.
1232 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1233 * sb_find_get_block().
1235 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1236 * a local interrupt disable for that.
1239 #define BH_LRU_SIZE 8
1242 struct buffer_head *bhs[BH_LRU_SIZE];
1245 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1248 #define bh_lru_lock() local_irq_disable()
1249 #define bh_lru_unlock() local_irq_enable()
1251 #define bh_lru_lock() preempt_disable()
1252 #define bh_lru_unlock() preempt_enable()
1255 static inline void check_irqs_on(void)
1257 #ifdef irqs_disabled
1258 BUG_ON(irqs_disabled());
1263 * The LRU management algorithm is dopey-but-simple. Sorry.
1265 static void bh_lru_install(struct buffer_head *bh)
1267 struct buffer_head *evictee = NULL;
1272 lru = &__get_cpu_var(bh_lrus);
1273 if (lru->bhs[0] != bh) {
1274 struct buffer_head *bhs[BH_LRU_SIZE];
1280 for (in = 0; in < BH_LRU_SIZE; in++) {
1281 struct buffer_head *bh2 = lru->bhs[in];
1286 if (out >= BH_LRU_SIZE) {
1287 BUG_ON(evictee != NULL);
1294 while (out < BH_LRU_SIZE)
1296 memcpy(lru->bhs, bhs, sizeof(bhs));
1305 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1307 static struct buffer_head *
1308 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1310 struct buffer_head *ret = NULL;
1316 lru = &__get_cpu_var(bh_lrus);
1317 for (i = 0; i < BH_LRU_SIZE; i++) {
1318 struct buffer_head *bh = lru->bhs[i];
1320 if (bh && bh->b_bdev == bdev &&
1321 bh->b_blocknr == block && bh->b_size == size) {
1324 lru->bhs[i] = lru->bhs[i - 1];
1339 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1340 * it in the LRU and mark it as accessed. If it is not present then return
1343 struct buffer_head *
1344 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1346 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1349 bh = __find_get_block_slow(bdev, block);
1357 EXPORT_SYMBOL(__find_get_block);
1360 * __getblk will locate (and, if necessary, create) the buffer_head
1361 * which corresponds to the passed block_device, block and size. The
1362 * returned buffer has its reference count incremented.
1364 * __getblk() cannot fail - it just keeps trying. If you pass it an
1365 * illegal block number, __getblk() will happily return a buffer_head
1366 * which represents the non-existent block. Very weird.
1368 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1369 * attempt is failing. FIXME, perhaps?
1371 struct buffer_head *
1372 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1374 struct buffer_head *bh = __find_get_block(bdev, block, size);
1378 bh = __getblk_slow(bdev, block, size);
1381 EXPORT_SYMBOL(__getblk);
1384 * Do async read-ahead on a buffer..
1386 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1388 struct buffer_head *bh = __getblk(bdev, block, size);
1390 ll_rw_block(READA, 1, &bh);
1394 EXPORT_SYMBOL(__breadahead);
1397 * __bread() - reads a specified block and returns the bh
1398 * @bdev: the block_device to read from
1399 * @block: number of block
1400 * @size: size (in bytes) to read
1402 * Reads a specified block, and returns buffer head that contains it.
1403 * It returns NULL if the block was unreadable.
1405 struct buffer_head *
1406 __bread(struct block_device *bdev, sector_t block, unsigned size)
1408 struct buffer_head *bh = __getblk(bdev, block, size);
1410 if (likely(bh) && !buffer_uptodate(bh))
1411 bh = __bread_slow(bh);
1414 EXPORT_SYMBOL(__bread);
1417 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1418 * This doesn't race because it runs in each cpu either in irq
1419 * or with preempt disabled.
1421 static void invalidate_bh_lru(void *arg)
1423 struct bh_lru *b = &get_cpu_var(bh_lrus);
1426 for (i = 0; i < BH_LRU_SIZE; i++) {
1430 put_cpu_var(bh_lrus);
1433 void invalidate_bh_lrus(void)
1435 on_each_cpu(invalidate_bh_lru, NULL, 1, 1);
1438 void set_bh_page(struct buffer_head *bh,
1439 struct page *page, unsigned long offset)
1442 BUG_ON(offset >= PAGE_SIZE);
1443 if (PageHighMem(page))
1445 * This catches illegal uses and preserves the offset:
1447 bh->b_data = (char *)(0 + offset);
1449 bh->b_data = page_address(page) + offset;
1451 EXPORT_SYMBOL(set_bh_page);
1454 * Called when truncating a buffer on a page completely.
1456 static void discard_buffer(struct buffer_head * bh)
1459 clear_buffer_dirty(bh);
1461 clear_buffer_mapped(bh);
1462 clear_buffer_req(bh);
1463 clear_buffer_new(bh);
1464 clear_buffer_delay(bh);
1465 clear_buffer_unwritten(bh);
1470 * block_invalidatepage - invalidate part of all of a buffer-backed page
1472 * @page: the page which is affected
1473 * @offset: the index of the truncation point
1475 * block_invalidatepage() is called when all or part of the page has become
1476 * invalidatedby a truncate operation.
1478 * block_invalidatepage() does not have to release all buffers, but it must
1479 * ensure that no dirty buffer is left outside @offset and that no I/O
1480 * is underway against any of the blocks which are outside the truncation
1481 * point. Because the caller is about to free (and possibly reuse) those
1484 void block_invalidatepage(struct page *page, unsigned long offset)
1486 struct buffer_head *head, *bh, *next;
1487 unsigned int curr_off = 0;
1489 BUG_ON(!PageLocked(page));
1490 if (!page_has_buffers(page))
1493 head = page_buffers(page);
1496 unsigned int next_off = curr_off + bh->b_size;
1497 next = bh->b_this_page;
1500 * is this block fully invalidated?
1502 if (offset <= curr_off)
1504 curr_off = next_off;
1506 } while (bh != head);
1509 * We release buffers only if the entire page is being invalidated.
1510 * The get_block cached value has been unconditionally invalidated,
1511 * so real IO is not possible anymore.
1514 try_to_release_page(page, 0);
1518 EXPORT_SYMBOL(block_invalidatepage);
1521 * We attach and possibly dirty the buffers atomically wrt
1522 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1523 * is already excluded via the page lock.
1525 void create_empty_buffers(struct page *page,
1526 unsigned long blocksize, unsigned long b_state)
1528 struct buffer_head *bh, *head, *tail;
1530 head = alloc_page_buffers(page, blocksize, 1);
1533 bh->b_state |= b_state;
1535 bh = bh->b_this_page;
1537 tail->b_this_page = head;
1539 spin_lock(&page->mapping->private_lock);
1540 if (PageUptodate(page) || PageDirty(page)) {
1543 if (PageDirty(page))
1544 set_buffer_dirty(bh);
1545 if (PageUptodate(page))
1546 set_buffer_uptodate(bh);
1547 bh = bh->b_this_page;
1548 } while (bh != head);
1550 attach_page_buffers(page, head);
1551 spin_unlock(&page->mapping->private_lock);
1553 EXPORT_SYMBOL(create_empty_buffers);
1556 * We are taking a block for data and we don't want any output from any
1557 * buffer-cache aliases starting from return from that function and
1558 * until the moment when something will explicitly mark the buffer
1559 * dirty (hopefully that will not happen until we will free that block ;-)
1560 * We don't even need to mark it not-uptodate - nobody can expect
1561 * anything from a newly allocated buffer anyway. We used to used
1562 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1563 * don't want to mark the alias unmapped, for example - it would confuse
1564 * anyone who might pick it with bread() afterwards...
1566 * Also.. Note that bforget() doesn't lock the buffer. So there can
1567 * be writeout I/O going on against recently-freed buffers. We don't
1568 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1569 * only if we really need to. That happens here.
1571 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1573 struct buffer_head *old_bh;
1577 old_bh = __find_get_block_slow(bdev, block);
1579 clear_buffer_dirty(old_bh);
1580 wait_on_buffer(old_bh);
1581 clear_buffer_req(old_bh);
1585 EXPORT_SYMBOL(unmap_underlying_metadata);
1588 * NOTE! All mapped/uptodate combinations are valid:
1590 * Mapped Uptodate Meaning
1592 * No No "unknown" - must do get_block()
1593 * No Yes "hole" - zero-filled
1594 * Yes No "allocated" - allocated on disk, not read in
1595 * Yes Yes "valid" - allocated and up-to-date in memory.
1597 * "Dirty" is valid only with the last case (mapped+uptodate).
1601 * While block_write_full_page is writing back the dirty buffers under
1602 * the page lock, whoever dirtied the buffers may decide to clean them
1603 * again at any time. We handle that by only looking at the buffer
1604 * state inside lock_buffer().
1606 * If block_write_full_page() is called for regular writeback
1607 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1608 * locked buffer. This only can happen if someone has written the buffer
1609 * directly, with submit_bh(). At the address_space level PageWriteback
1610 * prevents this contention from occurring.
1612 static int __block_write_full_page(struct inode *inode, struct page *page,
1613 get_block_t *get_block, struct writeback_control *wbc)
1617 sector_t last_block;
1618 struct buffer_head *bh, *head;
1619 const unsigned blocksize = 1 << inode->i_blkbits;
1620 int nr_underway = 0;
1622 BUG_ON(!PageLocked(page));
1624 last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1626 if (!page_has_buffers(page)) {
1627 create_empty_buffers(page, blocksize,
1628 (1 << BH_Dirty)|(1 << BH_Uptodate));
1632 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1633 * here, and the (potentially unmapped) buffers may become dirty at
1634 * any time. If a buffer becomes dirty here after we've inspected it
1635 * then we just miss that fact, and the page stays dirty.
1637 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1638 * handle that here by just cleaning them.
1641 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1642 head = page_buffers(page);
1646 * Get all the dirty buffers mapped to disk addresses and
1647 * handle any aliases from the underlying blockdev's mapping.
1650 if (block > last_block) {
1652 * mapped buffers outside i_size will occur, because
1653 * this page can be outside i_size when there is a
1654 * truncate in progress.
1657 * The buffer was zeroed by block_write_full_page()
1659 clear_buffer_dirty(bh);
1660 set_buffer_uptodate(bh);
1661 } else if (!buffer_mapped(bh) && buffer_dirty(bh)) {
1662 WARN_ON(bh->b_size != blocksize);
1663 err = get_block(inode, block, bh, 1);
1666 if (buffer_new(bh)) {
1667 /* blockdev mappings never come here */
1668 clear_buffer_new(bh);
1669 unmap_underlying_metadata(bh->b_bdev,
1673 bh = bh->b_this_page;
1675 } while (bh != head);
1678 if (!buffer_mapped(bh))
1681 * If it's a fully non-blocking write attempt and we cannot
1682 * lock the buffer then redirty the page. Note that this can
1683 * potentially cause a busy-wait loop from pdflush and kswapd
1684 * activity, but those code paths have their own higher-level
1687 if (wbc->sync_mode != WB_SYNC_NONE || !wbc->nonblocking) {
1689 } else if (test_set_buffer_locked(bh)) {
1690 redirty_page_for_writepage(wbc, page);
1693 if (test_clear_buffer_dirty(bh)) {
1694 mark_buffer_async_write(bh);
1698 } while ((bh = bh->b_this_page) != head);
1701 * The page and its buffers are protected by PageWriteback(), so we can
1702 * drop the bh refcounts early.
1704 BUG_ON(PageWriteback(page));
1705 set_page_writeback(page);
1708 struct buffer_head *next = bh->b_this_page;
1709 if (buffer_async_write(bh)) {
1710 submit_bh(WRITE, bh);
1714 } while (bh != head);
1719 if (nr_underway == 0) {
1721 * The page was marked dirty, but the buffers were
1722 * clean. Someone wrote them back by hand with
1723 * ll_rw_block/submit_bh. A rare case.
1725 end_page_writeback(page);
1728 * The page and buffer_heads can be released at any time from
1731 wbc->pages_skipped++; /* We didn't write this page */
1737 * ENOSPC, or some other error. We may already have added some
1738 * blocks to the file, so we need to write these out to avoid
1739 * exposing stale data.
1740 * The page is currently locked and not marked for writeback
1743 /* Recovery: lock and submit the mapped buffers */
1745 if (buffer_mapped(bh) && buffer_dirty(bh)) {
1747 mark_buffer_async_write(bh);
1750 * The buffer may have been set dirty during
1751 * attachment to a dirty page.
1753 clear_buffer_dirty(bh);
1755 } while ((bh = bh->b_this_page) != head);
1757 BUG_ON(PageWriteback(page));
1758 mapping_set_error(page->mapping, err);
1759 set_page_writeback(page);
1761 struct buffer_head *next = bh->b_this_page;
1762 if (buffer_async_write(bh)) {
1763 clear_buffer_dirty(bh);
1764 submit_bh(WRITE, bh);
1768 } while (bh != head);
1773 static int __block_prepare_write(struct inode *inode, struct page *page,
1774 unsigned from, unsigned to, get_block_t *get_block)
1776 unsigned block_start, block_end;
1779 unsigned blocksize, bbits;
1780 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1782 BUG_ON(!PageLocked(page));
1783 BUG_ON(from > PAGE_CACHE_SIZE);
1784 BUG_ON(to > PAGE_CACHE_SIZE);
1787 blocksize = 1 << inode->i_blkbits;
1788 if (!page_has_buffers(page))
1789 create_empty_buffers(page, blocksize, 0);
1790 head = page_buffers(page);
1792 bbits = inode->i_blkbits;
1793 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1795 for(bh = head, block_start = 0; bh != head || !block_start;
1796 block++, block_start=block_end, bh = bh->b_this_page) {
1797 block_end = block_start + blocksize;
1798 if (block_end <= from || block_start >= to) {
1799 if (PageUptodate(page)) {
1800 if (!buffer_uptodate(bh))
1801 set_buffer_uptodate(bh);
1806 clear_buffer_new(bh);
1807 if (!buffer_mapped(bh)) {
1808 WARN_ON(bh->b_size != blocksize);
1809 err = get_block(inode, block, bh, 1);
1812 if (buffer_new(bh)) {
1813 unmap_underlying_metadata(bh->b_bdev,
1815 if (PageUptodate(page)) {
1816 set_buffer_uptodate(bh);
1819 if (block_end > to || block_start < from) {
1822 kaddr = kmap_atomic(page, KM_USER0);
1826 if (block_start < from)
1827 memset(kaddr+block_start,
1828 0, from-block_start);
1829 flush_dcache_page(page);
1830 kunmap_atomic(kaddr, KM_USER0);
1835 if (PageUptodate(page)) {
1836 if (!buffer_uptodate(bh))
1837 set_buffer_uptodate(bh);
1840 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1841 !buffer_unwritten(bh) &&
1842 (block_start < from || block_end > to)) {
1843 ll_rw_block(READ, 1, &bh);
1848 * If we issued read requests - let them complete.
1850 while(wait_bh > wait) {
1851 wait_on_buffer(*--wait_bh);
1852 if (!buffer_uptodate(*wait_bh))
1859 clear_buffer_new(bh);
1860 } while ((bh = bh->b_this_page) != head);
1865 * Zero out any newly allocated blocks to avoid exposing stale
1866 * data. If BH_New is set, we know that the block was newly
1867 * allocated in the above loop.
1872 block_end = block_start+blocksize;
1873 if (block_end <= from)
1875 if (block_start >= to)
1877 if (buffer_new(bh)) {
1878 clear_buffer_new(bh);
1879 zero_user_page(page, block_start, bh->b_size, KM_USER0);
1880 set_buffer_uptodate(bh);
1881 mark_buffer_dirty(bh);
1884 block_start = block_end;
1885 bh = bh->b_this_page;
1886 } while (bh != head);
1890 static int __block_commit_write(struct inode *inode, struct page *page,
1891 unsigned from, unsigned to)
1893 unsigned block_start, block_end;
1896 struct buffer_head *bh, *head;
1898 blocksize = 1 << inode->i_blkbits;
1900 for(bh = head = page_buffers(page), block_start = 0;
1901 bh != head || !block_start;
1902 block_start=block_end, bh = bh->b_this_page) {
1903 block_end = block_start + blocksize;
1904 if (block_end <= from || block_start >= to) {
1905 if (!buffer_uptodate(bh))
1908 set_buffer_uptodate(bh);
1909 mark_buffer_dirty(bh);
1914 * If this is a partial write which happened to make all buffers
1915 * uptodate then we can optimize away a bogus readpage() for
1916 * the next read(). Here we 'discover' whether the page went
1917 * uptodate as a result of this (potentially partial) write.
1920 SetPageUptodate(page);
1925 * Generic "read page" function for block devices that have the normal
1926 * get_block functionality. This is most of the block device filesystems.
1927 * Reads the page asynchronously --- the unlock_buffer() and
1928 * set/clear_buffer_uptodate() functions propagate buffer state into the
1929 * page struct once IO has completed.
1931 int block_read_full_page(struct page *page, get_block_t *get_block)
1933 struct inode *inode = page->mapping->host;
1934 sector_t iblock, lblock;
1935 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
1936 unsigned int blocksize;
1938 int fully_mapped = 1;
1940 BUG_ON(!PageLocked(page));
1941 blocksize = 1 << inode->i_blkbits;
1942 if (!page_has_buffers(page))
1943 create_empty_buffers(page, blocksize, 0);
1944 head = page_buffers(page);
1946 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1947 lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
1953 if (buffer_uptodate(bh))
1956 if (!buffer_mapped(bh)) {
1960 if (iblock < lblock) {
1961 WARN_ON(bh->b_size != blocksize);
1962 err = get_block(inode, iblock, bh, 0);
1966 if (!buffer_mapped(bh)) {
1967 zero_user_page(page, i * blocksize, blocksize,
1970 set_buffer_uptodate(bh);
1974 * get_block() might have updated the buffer
1977 if (buffer_uptodate(bh))
1981 } while (i++, iblock++, (bh = bh->b_this_page) != head);
1984 SetPageMappedToDisk(page);
1988 * All buffers are uptodate - we can set the page uptodate
1989 * as well. But not if get_block() returned an error.
1991 if (!PageError(page))
1992 SetPageUptodate(page);
1997 /* Stage two: lock the buffers */
1998 for (i = 0; i < nr; i++) {
2001 mark_buffer_async_read(bh);
2005 * Stage 3: start the IO. Check for uptodateness
2006 * inside the buffer lock in case another process reading
2007 * the underlying blockdev brought it uptodate (the sct fix).
2009 for (i = 0; i < nr; i++) {
2011 if (buffer_uptodate(bh))
2012 end_buffer_async_read(bh, 1);
2014 submit_bh(READ, bh);
2019 /* utility function for filesystems that need to do work on expanding
2020 * truncates. Uses prepare/commit_write to allow the filesystem to
2021 * deal with the hole.
2023 static int __generic_cont_expand(struct inode *inode, loff_t size,
2024 pgoff_t index, unsigned int offset)
2026 struct address_space *mapping = inode->i_mapping;
2028 unsigned long limit;
2032 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2033 if (limit != RLIM_INFINITY && size > (loff_t)limit) {
2034 send_sig(SIGXFSZ, current, 0);
2037 if (size > inode->i_sb->s_maxbytes)
2041 page = grab_cache_page(mapping, index);
2044 err = mapping->a_ops->prepare_write(NULL, page, offset, offset);
2047 * ->prepare_write() may have instantiated a few blocks
2048 * outside i_size. Trim these off again.
2051 page_cache_release(page);
2052 vmtruncate(inode, inode->i_size);
2056 err = mapping->a_ops->commit_write(NULL, page, offset, offset);
2059 page_cache_release(page);
2066 int generic_cont_expand(struct inode *inode, loff_t size)
2069 unsigned int offset;
2071 offset = (size & (PAGE_CACHE_SIZE - 1)); /* Within page */
2073 /* ugh. in prepare/commit_write, if from==to==start of block, we
2074 ** skip the prepare. make sure we never send an offset for the start
2077 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
2078 /* caller must handle this extra byte. */
2081 index = size >> PAGE_CACHE_SHIFT;
2083 return __generic_cont_expand(inode, size, index, offset);
2086 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2088 loff_t pos = size - 1;
2089 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2090 unsigned int offset = (pos & (PAGE_CACHE_SIZE - 1)) + 1;
2092 /* prepare/commit_write can handle even if from==to==start of block. */
2093 return __generic_cont_expand(inode, size, index, offset);
2097 * For moronic filesystems that do not allow holes in file.
2098 * We may have to extend the file.
2101 int cont_prepare_write(struct page *page, unsigned offset,
2102 unsigned to, get_block_t *get_block, loff_t *bytes)
2104 struct address_space *mapping = page->mapping;
2105 struct inode *inode = mapping->host;
2106 struct page *new_page;
2110 unsigned blocksize = 1 << inode->i_blkbits;
2112 while(page->index > (pgpos = *bytes>>PAGE_CACHE_SHIFT)) {
2114 new_page = grab_cache_page(mapping, pgpos);
2117 /* we might sleep */
2118 if (*bytes>>PAGE_CACHE_SHIFT != pgpos) {
2119 unlock_page(new_page);
2120 page_cache_release(new_page);
2123 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2124 if (zerofrom & (blocksize-1)) {
2125 *bytes |= (blocksize-1);
2128 status = __block_prepare_write(inode, new_page, zerofrom,
2129 PAGE_CACHE_SIZE, get_block);
2132 zero_user_page(new_page, zerofrom, PAGE_CACHE_SIZE - zerofrom,
2134 generic_commit_write(NULL, new_page, zerofrom, PAGE_CACHE_SIZE);
2135 unlock_page(new_page);
2136 page_cache_release(new_page);
2139 if (page->index < pgpos) {
2140 /* completely inside the area */
2143 /* page covers the boundary, find the boundary offset */
2144 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2146 /* if we will expand the thing last block will be filled */
2147 if (to > zerofrom && (zerofrom & (blocksize-1))) {
2148 *bytes |= (blocksize-1);
2152 /* starting below the boundary? Nothing to zero out */
2153 if (offset <= zerofrom)
2156 status = __block_prepare_write(inode, page, zerofrom, to, get_block);
2159 if (zerofrom < offset) {
2160 zero_user_page(page, zerofrom, offset - zerofrom, KM_USER0);
2161 __block_commit_write(inode, page, zerofrom, offset);
2165 ClearPageUptodate(page);
2169 ClearPageUptodate(new_page);
2170 unlock_page(new_page);
2171 page_cache_release(new_page);
2176 int block_prepare_write(struct page *page, unsigned from, unsigned to,
2177 get_block_t *get_block)
2179 struct inode *inode = page->mapping->host;
2180 int err = __block_prepare_write(inode, page, from, to, get_block);
2182 ClearPageUptodate(page);
2186 int block_commit_write(struct page *page, unsigned from, unsigned to)
2188 struct inode *inode = page->mapping->host;
2189 __block_commit_write(inode,page,from,to);
2193 int generic_commit_write(struct file *file, struct page *page,
2194 unsigned from, unsigned to)
2196 struct inode *inode = page->mapping->host;
2197 loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2198 __block_commit_write(inode,page,from,to);
2200 * No need to use i_size_read() here, the i_size
2201 * cannot change under us because we hold i_mutex.
2203 if (pos > inode->i_size) {
2204 i_size_write(inode, pos);
2205 mark_inode_dirty(inode);
2211 * block_page_mkwrite() is not allowed to change the file size as it gets
2212 * called from a page fault handler when a page is first dirtied. Hence we must
2213 * be careful to check for EOF conditions here. We set the page up correctly
2214 * for a written page which means we get ENOSPC checking when writing into
2215 * holes and correct delalloc and unwritten extent mapping on filesystems that
2216 * support these features.
2218 * We are not allowed to take the i_mutex here so we have to play games to
2219 * protect against truncate races as the page could now be beyond EOF. Because
2220 * vmtruncate() writes the inode size before removing pages, once we have the
2221 * page lock we can determine safely if the page is beyond EOF. If it is not
2222 * beyond EOF, then the page is guaranteed safe against truncation until we
2226 block_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2227 get_block_t get_block)
2229 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2235 size = i_size_read(inode);
2236 if ((page->mapping != inode->i_mapping) ||
2237 (page_offset(page) > size)) {
2238 /* page got truncated out from underneath us */
2242 /* page is wholly or partially inside EOF */
2243 if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2244 end = size & ~PAGE_CACHE_MASK;
2246 end = PAGE_CACHE_SIZE;
2248 ret = block_prepare_write(page, 0, end, get_block);
2250 ret = block_commit_write(page, 0, end);
2258 * nobh_prepare_write()'s prereads are special: the buffer_heads are freed
2259 * immediately, while under the page lock. So it needs a special end_io
2260 * handler which does not touch the bh after unlocking it.
2262 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2264 __end_buffer_read_notouch(bh, uptodate);
2268 * On entry, the page is fully not uptodate.
2269 * On exit the page is fully uptodate in the areas outside (from,to)
2271 int nobh_prepare_write(struct page *page, unsigned from, unsigned to,
2272 get_block_t *get_block)
2274 struct inode *inode = page->mapping->host;
2275 const unsigned blkbits = inode->i_blkbits;
2276 const unsigned blocksize = 1 << blkbits;
2277 struct buffer_head map_bh;
2278 struct buffer_head *read_bh[MAX_BUF_PER_PAGE];
2279 unsigned block_in_page;
2280 unsigned block_start;
2281 sector_t block_in_file;
2286 int is_mapped_to_disk = 1;
2288 if (PageMappedToDisk(page))
2291 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2292 map_bh.b_page = page;
2295 * We loop across all blocks in the page, whether or not they are
2296 * part of the affected region. This is so we can discover if the
2297 * page is fully mapped-to-disk.
2299 for (block_start = 0, block_in_page = 0;
2300 block_start < PAGE_CACHE_SIZE;
2301 block_in_page++, block_start += blocksize) {
2302 unsigned block_end = block_start + blocksize;
2307 if (block_start >= to)
2309 map_bh.b_size = blocksize;
2310 ret = get_block(inode, block_in_file + block_in_page,
2314 if (!buffer_mapped(&map_bh))
2315 is_mapped_to_disk = 0;
2316 if (buffer_new(&map_bh))
2317 unmap_underlying_metadata(map_bh.b_bdev,
2319 if (PageUptodate(page))
2321 if (buffer_new(&map_bh) || !buffer_mapped(&map_bh)) {
2322 kaddr = kmap_atomic(page, KM_USER0);
2323 if (block_start < from)
2324 memset(kaddr+block_start, 0, from-block_start);
2326 memset(kaddr + to, 0, block_end - to);
2327 flush_dcache_page(page);
2328 kunmap_atomic(kaddr, KM_USER0);
2331 if (buffer_uptodate(&map_bh))
2332 continue; /* reiserfs does this */
2333 if (block_start < from || block_end > to) {
2334 struct buffer_head *bh = alloc_buffer_head(GFP_NOFS);
2340 bh->b_state = map_bh.b_state;
2341 atomic_set(&bh->b_count, 0);
2342 bh->b_this_page = NULL;
2344 bh->b_blocknr = map_bh.b_blocknr;
2345 bh->b_size = blocksize;
2346 bh->b_data = (char *)(long)block_start;
2347 bh->b_bdev = map_bh.b_bdev;
2348 bh->b_private = NULL;
2349 read_bh[nr_reads++] = bh;
2354 struct buffer_head *bh;
2357 * The page is locked, so these buffers are protected from
2358 * any VM or truncate activity. Hence we don't need to care
2359 * for the buffer_head refcounts.
2361 for (i = 0; i < nr_reads; i++) {
2364 bh->b_end_io = end_buffer_read_nobh;
2365 submit_bh(READ, bh);
2367 for (i = 0; i < nr_reads; i++) {
2370 if (!buffer_uptodate(bh))
2372 free_buffer_head(bh);
2379 if (is_mapped_to_disk)
2380 SetPageMappedToDisk(page);
2385 for (i = 0; i < nr_reads; i++) {
2387 free_buffer_head(read_bh[i]);
2391 * Error recovery is pretty slack. Clear the page and mark it dirty
2392 * so we'll later zero out any blocks which _were_ allocated.
2394 zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
2395 SetPageUptodate(page);
2396 set_page_dirty(page);
2399 EXPORT_SYMBOL(nobh_prepare_write);
2402 * Make sure any changes to nobh_commit_write() are reflected in
2403 * nobh_truncate_page(), since it doesn't call commit_write().
2405 int nobh_commit_write(struct file *file, struct page *page,
2406 unsigned from, unsigned to)
2408 struct inode *inode = page->mapping->host;
2409 loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
2411 SetPageUptodate(page);
2412 set_page_dirty(page);
2413 if (pos > inode->i_size) {
2414 i_size_write(inode, pos);
2415 mark_inode_dirty(inode);
2419 EXPORT_SYMBOL(nobh_commit_write);
2422 * nobh_writepage() - based on block_full_write_page() except
2423 * that it tries to operate without attaching bufferheads to
2426 int nobh_writepage(struct page *page, get_block_t *get_block,
2427 struct writeback_control *wbc)
2429 struct inode * const inode = page->mapping->host;
2430 loff_t i_size = i_size_read(inode);
2431 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2435 /* Is the page fully inside i_size? */
2436 if (page->index < end_index)
2439 /* Is the page fully outside i_size? (truncate in progress) */
2440 offset = i_size & (PAGE_CACHE_SIZE-1);
2441 if (page->index >= end_index+1 || !offset) {
2443 * The page may have dirty, unmapped buffers. For example,
2444 * they may have been added in ext3_writepage(). Make them
2445 * freeable here, so the page does not leak.
2448 /* Not really sure about this - do we need this ? */
2449 if (page->mapping->a_ops->invalidatepage)
2450 page->mapping->a_ops->invalidatepage(page, offset);
2453 return 0; /* don't care */
2457 * The page straddles i_size. It must be zeroed out on each and every
2458 * writepage invocation because it may be mmapped. "A file is mapped
2459 * in multiples of the page size. For a file that is not a multiple of
2460 * the page size, the remaining memory is zeroed when mapped, and
2461 * writes to that region are not written out to the file."
2463 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset, KM_USER0);
2465 ret = mpage_writepage(page, get_block, wbc);
2467 ret = __block_write_full_page(inode, page, get_block, wbc);
2470 EXPORT_SYMBOL(nobh_writepage);
2473 * This function assumes that ->prepare_write() uses nobh_prepare_write().
2475 int nobh_truncate_page(struct address_space *mapping, loff_t from)
2477 struct inode *inode = mapping->host;
2478 unsigned blocksize = 1 << inode->i_blkbits;
2479 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2480 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2483 const struct address_space_operations *a_ops = mapping->a_ops;
2486 if ((offset & (blocksize - 1)) == 0)
2490 page = grab_cache_page(mapping, index);
2494 to = (offset + blocksize) & ~(blocksize - 1);
2495 ret = a_ops->prepare_write(NULL, page, offset, to);
2497 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset,
2500 * It would be more correct to call aops->commit_write()
2501 * here, but this is more efficient.
2503 SetPageUptodate(page);
2504 set_page_dirty(page);
2507 page_cache_release(page);
2511 EXPORT_SYMBOL(nobh_truncate_page);
2513 int block_truncate_page(struct address_space *mapping,
2514 loff_t from, get_block_t *get_block)
2516 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2517 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2520 unsigned length, pos;
2521 struct inode *inode = mapping->host;
2523 struct buffer_head *bh;
2526 blocksize = 1 << inode->i_blkbits;
2527 length = offset & (blocksize - 1);
2529 /* Block boundary? Nothing to do */
2533 length = blocksize - length;
2534 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2536 page = grab_cache_page(mapping, index);
2541 if (!page_has_buffers(page))
2542 create_empty_buffers(page, blocksize, 0);
2544 /* Find the buffer that contains "offset" */
2545 bh = page_buffers(page);
2547 while (offset >= pos) {
2548 bh = bh->b_this_page;
2554 if (!buffer_mapped(bh)) {
2555 WARN_ON(bh->b_size != blocksize);
2556 err = get_block(inode, iblock, bh, 0);
2559 /* unmapped? It's a hole - nothing to do */
2560 if (!buffer_mapped(bh))
2564 /* Ok, it's mapped. Make sure it's up-to-date */
2565 if (PageUptodate(page))
2566 set_buffer_uptodate(bh);
2568 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2570 ll_rw_block(READ, 1, &bh);
2572 /* Uhhuh. Read error. Complain and punt. */
2573 if (!buffer_uptodate(bh))
2577 zero_user_page(page, offset, length, KM_USER0);
2578 mark_buffer_dirty(bh);
2583 page_cache_release(page);
2589 * The generic ->writepage function for buffer-backed address_spaces
2591 int block_write_full_page(struct page *page, get_block_t *get_block,
2592 struct writeback_control *wbc)
2594 struct inode * const inode = page->mapping->host;
2595 loff_t i_size = i_size_read(inode);
2596 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2599 /* Is the page fully inside i_size? */
2600 if (page->index < end_index)
2601 return __block_write_full_page(inode, page, get_block, wbc);
2603 /* Is the page fully outside i_size? (truncate in progress) */
2604 offset = i_size & (PAGE_CACHE_SIZE-1);
2605 if (page->index >= end_index+1 || !offset) {
2607 * The page may have dirty, unmapped buffers. For example,
2608 * they may have been added in ext3_writepage(). Make them
2609 * freeable here, so the page does not leak.
2611 do_invalidatepage(page, 0);
2613 return 0; /* don't care */
2617 * The page straddles i_size. It must be zeroed out on each and every
2618 * writepage invokation because it may be mmapped. "A file is mapped
2619 * in multiples of the page size. For a file that is not a multiple of
2620 * the page size, the remaining memory is zeroed when mapped, and
2621 * writes to that region are not written out to the file."
2623 zero_user_page(page, offset, PAGE_CACHE_SIZE - offset, KM_USER0);
2624 return __block_write_full_page(inode, page, get_block, wbc);
2627 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2628 get_block_t *get_block)
2630 struct buffer_head tmp;
2631 struct inode *inode = mapping->host;
2634 tmp.b_size = 1 << inode->i_blkbits;
2635 get_block(inode, block, &tmp, 0);
2636 return tmp.b_blocknr;
2639 static void end_bio_bh_io_sync(struct bio *bio, int err)
2641 struct buffer_head *bh = bio->bi_private;
2643 if (err == -EOPNOTSUPP) {
2644 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2645 set_bit(BH_Eopnotsupp, &bh->b_state);
2648 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2652 int submit_bh(int rw, struct buffer_head * bh)
2657 BUG_ON(!buffer_locked(bh));
2658 BUG_ON(!buffer_mapped(bh));
2659 BUG_ON(!bh->b_end_io);
2661 if (buffer_ordered(bh) && (rw == WRITE))
2665 * Only clear out a write error when rewriting, should this
2666 * include WRITE_SYNC as well?
2668 if (test_set_buffer_req(bh) && (rw == WRITE || rw == WRITE_BARRIER))
2669 clear_buffer_write_io_error(bh);
2672 * from here on down, it's all bio -- do the initial mapping,
2673 * submit_bio -> generic_make_request may further map this bio around
2675 bio = bio_alloc(GFP_NOIO, 1);
2677 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2678 bio->bi_bdev = bh->b_bdev;
2679 bio->bi_io_vec[0].bv_page = bh->b_page;
2680 bio->bi_io_vec[0].bv_len = bh->b_size;
2681 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2685 bio->bi_size = bh->b_size;
2687 bio->bi_end_io = end_bio_bh_io_sync;
2688 bio->bi_private = bh;
2691 submit_bio(rw, bio);
2693 if (bio_flagged(bio, BIO_EOPNOTSUPP))
2701 * ll_rw_block: low-level access to block devices (DEPRECATED)
2702 * @rw: whether to %READ or %WRITE or %SWRITE or maybe %READA (readahead)
2703 * @nr: number of &struct buffer_heads in the array
2704 * @bhs: array of pointers to &struct buffer_head
2706 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2707 * requests an I/O operation on them, either a %READ or a %WRITE. The third
2708 * %SWRITE is like %WRITE only we make sure that the *current* data in buffers
2709 * are sent to disk. The fourth %READA option is described in the documentation
2710 * for generic_make_request() which ll_rw_block() calls.
2712 * This function drops any buffer that it cannot get a lock on (with the
2713 * BH_Lock state bit) unless SWRITE is required, any buffer that appears to be
2714 * clean when doing a write request, and any buffer that appears to be
2715 * up-to-date when doing read request. Further it marks as clean buffers that
2716 * are processed for writing (the buffer cache won't assume that they are
2717 * actually clean until the buffer gets unlocked).
2719 * ll_rw_block sets b_end_io to simple completion handler that marks
2720 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2723 * All of the buffers must be for the same device, and must also be a
2724 * multiple of the current approved size for the device.
2726 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2730 for (i = 0; i < nr; i++) {
2731 struct buffer_head *bh = bhs[i];
2735 else if (test_set_buffer_locked(bh))
2738 if (rw == WRITE || rw == SWRITE) {
2739 if (test_clear_buffer_dirty(bh)) {
2740 bh->b_end_io = end_buffer_write_sync;
2742 submit_bh(WRITE, bh);
2746 if (!buffer_uptodate(bh)) {
2747 bh->b_end_io = end_buffer_read_sync;
2758 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2759 * and then start new I/O and then wait upon it. The caller must have a ref on
2762 int sync_dirty_buffer(struct buffer_head *bh)
2766 WARN_ON(atomic_read(&bh->b_count) < 1);
2768 if (test_clear_buffer_dirty(bh)) {
2770 bh->b_end_io = end_buffer_write_sync;
2771 ret = submit_bh(WRITE, bh);
2773 if (buffer_eopnotsupp(bh)) {
2774 clear_buffer_eopnotsupp(bh);
2777 if (!ret && !buffer_uptodate(bh))
2786 * try_to_free_buffers() checks if all the buffers on this particular page
2787 * are unused, and releases them if so.
2789 * Exclusion against try_to_free_buffers may be obtained by either
2790 * locking the page or by holding its mapping's private_lock.
2792 * If the page is dirty but all the buffers are clean then we need to
2793 * be sure to mark the page clean as well. This is because the page
2794 * may be against a block device, and a later reattachment of buffers
2795 * to a dirty page will set *all* buffers dirty. Which would corrupt
2796 * filesystem data on the same device.
2798 * The same applies to regular filesystem pages: if all the buffers are
2799 * clean then we set the page clean and proceed. To do that, we require
2800 * total exclusion from __set_page_dirty_buffers(). That is obtained with
2803 * try_to_free_buffers() is non-blocking.
2805 static inline int buffer_busy(struct buffer_head *bh)
2807 return atomic_read(&bh->b_count) |
2808 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2812 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
2814 struct buffer_head *head = page_buffers(page);
2815 struct buffer_head *bh;
2819 if (buffer_write_io_error(bh) && page->mapping)
2820 set_bit(AS_EIO, &page->mapping->flags);
2821 if (buffer_busy(bh))
2823 bh = bh->b_this_page;
2824 } while (bh != head);
2827 struct buffer_head *next = bh->b_this_page;
2829 if (!list_empty(&bh->b_assoc_buffers))
2830 __remove_assoc_queue(bh);
2832 } while (bh != head);
2833 *buffers_to_free = head;
2834 __clear_page_buffers(page);
2840 int try_to_free_buffers(struct page *page)
2842 struct address_space * const mapping = page->mapping;
2843 struct buffer_head *buffers_to_free = NULL;
2846 BUG_ON(!PageLocked(page));
2847 if (PageWriteback(page))
2850 if (mapping == NULL) { /* can this still happen? */
2851 ret = drop_buffers(page, &buffers_to_free);
2855 spin_lock(&mapping->private_lock);
2856 ret = drop_buffers(page, &buffers_to_free);
2859 * If the filesystem writes its buffers by hand (eg ext3)
2860 * then we can have clean buffers against a dirty page. We
2861 * clean the page here; otherwise the VM will never notice
2862 * that the filesystem did any IO at all.
2864 * Also, during truncate, discard_buffer will have marked all
2865 * the page's buffers clean. We discover that here and clean
2868 * private_lock must be held over this entire operation in order
2869 * to synchronise against __set_page_dirty_buffers and prevent the
2870 * dirty bit from being lost.
2873 cancel_dirty_page(page, PAGE_CACHE_SIZE);
2874 spin_unlock(&mapping->private_lock);
2876 if (buffers_to_free) {
2877 struct buffer_head *bh = buffers_to_free;
2880 struct buffer_head *next = bh->b_this_page;
2881 free_buffer_head(bh);
2883 } while (bh != buffers_to_free);
2887 EXPORT_SYMBOL(try_to_free_buffers);
2889 void block_sync_page(struct page *page)
2891 struct address_space *mapping;
2894 mapping = page_mapping(page);
2896 blk_run_backing_dev(mapping->backing_dev_info, page);
2900 * There are no bdflush tunables left. But distributions are
2901 * still running obsolete flush daemons, so we terminate them here.
2903 * Use of bdflush() is deprecated and will be removed in a future kernel.
2904 * The `pdflush' kernel threads fully replace bdflush daemons and this call.
2906 asmlinkage long sys_bdflush(int func, long data)
2908 static int msg_count;
2910 if (!capable(CAP_SYS_ADMIN))
2913 if (msg_count < 5) {
2916 "warning: process `%s' used the obsolete bdflush"
2917 " system call\n", current->comm);
2918 printk(KERN_INFO "Fix your initscripts?\n");
2927 * Buffer-head allocation
2929 static struct kmem_cache *bh_cachep;
2932 * Once the number of bh's in the machine exceeds this level, we start
2933 * stripping them in writeback.
2935 static int max_buffer_heads;
2937 int buffer_heads_over_limit;
2939 struct bh_accounting {
2940 int nr; /* Number of live bh's */
2941 int ratelimit; /* Limit cacheline bouncing */
2944 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2946 static void recalc_bh_state(void)
2951 if (__get_cpu_var(bh_accounting).ratelimit++ < 4096)
2953 __get_cpu_var(bh_accounting).ratelimit = 0;
2954 for_each_online_cpu(i)
2955 tot += per_cpu(bh_accounting, i).nr;
2956 buffer_heads_over_limit = (tot > max_buffer_heads);
2959 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
2961 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
2963 INIT_LIST_HEAD(&ret->b_assoc_buffers);
2964 get_cpu_var(bh_accounting).nr++;
2966 put_cpu_var(bh_accounting);
2970 EXPORT_SYMBOL(alloc_buffer_head);
2972 void free_buffer_head(struct buffer_head *bh)
2974 BUG_ON(!list_empty(&bh->b_assoc_buffers));
2975 kmem_cache_free(bh_cachep, bh);
2976 get_cpu_var(bh_accounting).nr--;
2978 put_cpu_var(bh_accounting);
2980 EXPORT_SYMBOL(free_buffer_head);
2982 static void buffer_exit_cpu(int cpu)
2985 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
2987 for (i = 0; i < BH_LRU_SIZE; i++) {
2991 get_cpu_var(bh_accounting).nr += per_cpu(bh_accounting, cpu).nr;
2992 per_cpu(bh_accounting, cpu).nr = 0;
2993 put_cpu_var(bh_accounting);
2996 static int buffer_cpu_notify(struct notifier_block *self,
2997 unsigned long action, void *hcpu)
2999 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3000 buffer_exit_cpu((unsigned long)hcpu);
3004 void __init buffer_init(void)
3008 bh_cachep = KMEM_CACHE(buffer_head,
3009 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
3012 * Limit the bh occupancy to 10% of ZONE_NORMAL
3014 nrpages = (nr_free_buffer_pages() * 10) / 100;
3015 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3016 hotcpu_notifier(buffer_cpu_notify, 0);
3019 EXPORT_SYMBOL(__bforget);
3020 EXPORT_SYMBOL(__brelse);
3021 EXPORT_SYMBOL(__wait_on_buffer);
3022 EXPORT_SYMBOL(block_commit_write);
3023 EXPORT_SYMBOL(block_prepare_write);
3024 EXPORT_SYMBOL(block_page_mkwrite);
3025 EXPORT_SYMBOL(block_read_full_page);
3026 EXPORT_SYMBOL(block_sync_page);
3027 EXPORT_SYMBOL(block_truncate_page);
3028 EXPORT_SYMBOL(block_write_full_page);
3029 EXPORT_SYMBOL(cont_prepare_write);
3030 EXPORT_SYMBOL(end_buffer_read_sync);
3031 EXPORT_SYMBOL(end_buffer_write_sync);
3032 EXPORT_SYMBOL(file_fsync);
3033 EXPORT_SYMBOL(fsync_bdev);
3034 EXPORT_SYMBOL(generic_block_bmap);
3035 EXPORT_SYMBOL(generic_commit_write);
3036 EXPORT_SYMBOL(generic_cont_expand);
3037 EXPORT_SYMBOL(generic_cont_expand_simple);
3038 EXPORT_SYMBOL(init_buffer);
3039 EXPORT_SYMBOL(invalidate_bdev);
3040 EXPORT_SYMBOL(ll_rw_block);
3041 EXPORT_SYMBOL(mark_buffer_dirty);
3042 EXPORT_SYMBOL(submit_bh);
3043 EXPORT_SYMBOL(sync_dirty_buffer);
3044 EXPORT_SYMBOL(unlock_buffer);