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>
44 #include <linux/cleancache.h>
46 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
48 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
51 init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
53 bh->b_end_io = handler;
54 bh->b_private = private;
56 EXPORT_SYMBOL(init_buffer);
58 static int sleep_on_buffer(void *word)
64 void __lock_buffer(struct buffer_head *bh)
66 wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
67 TASK_UNINTERRUPTIBLE);
69 EXPORT_SYMBOL(__lock_buffer);
71 void unlock_buffer(struct buffer_head *bh)
73 clear_bit_unlock(BH_Lock, &bh->b_state);
74 smp_mb__after_clear_bit();
75 wake_up_bit(&bh->b_state, BH_Lock);
77 EXPORT_SYMBOL(unlock_buffer);
80 * Block until a buffer comes unlocked. This doesn't stop it
81 * from becoming locked again - you have to lock it yourself
82 * if you want to preserve its state.
84 void __wait_on_buffer(struct buffer_head * bh)
86 wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
88 EXPORT_SYMBOL(__wait_on_buffer);
91 __clear_page_buffers(struct page *page)
93 ClearPagePrivate(page);
94 set_page_private(page, 0);
95 page_cache_release(page);
99 static int quiet_error(struct buffer_head *bh)
101 if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
107 static void buffer_io_error(struct buffer_head *bh)
109 char b[BDEVNAME_SIZE];
110 printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
111 bdevname(bh->b_bdev, b),
112 (unsigned long long)bh->b_blocknr);
116 * End-of-IO handler helper function which does not touch the bh after
118 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
119 * a race there is benign: unlock_buffer() only use the bh's address for
120 * hashing after unlocking the buffer, so it doesn't actually touch the bh
123 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
126 set_buffer_uptodate(bh);
128 /* This happens, due to failed READA attempts. */
129 clear_buffer_uptodate(bh);
135 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
136 * unlock the buffer. This is what ll_rw_block uses too.
138 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
140 __end_buffer_read_notouch(bh, uptodate);
143 EXPORT_SYMBOL(end_buffer_read_sync);
145 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
147 char b[BDEVNAME_SIZE];
150 set_buffer_uptodate(bh);
152 if (!quiet_error(bh)) {
154 printk(KERN_WARNING "lost page write due to "
156 bdevname(bh->b_bdev, b));
158 set_buffer_write_io_error(bh);
159 clear_buffer_uptodate(bh);
164 EXPORT_SYMBOL(end_buffer_write_sync);
167 * Various filesystems appear to want __find_get_block to be non-blocking.
168 * But it's the page lock which protects the buffers. To get around this,
169 * we get exclusion from try_to_free_buffers with the blockdev mapping's
172 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
173 * may be quite high. This code could TryLock the page, and if that
174 * succeeds, there is no need to take private_lock. (But if
175 * private_lock is contended then so is mapping->tree_lock).
177 static struct buffer_head *
178 __find_get_block_slow(struct block_device *bdev, sector_t block)
180 struct inode *bd_inode = bdev->bd_inode;
181 struct address_space *bd_mapping = bd_inode->i_mapping;
182 struct buffer_head *ret = NULL;
184 struct buffer_head *bh;
185 struct buffer_head *head;
189 index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
190 page = find_get_page(bd_mapping, index);
194 spin_lock(&bd_mapping->private_lock);
195 if (!page_has_buffers(page))
197 head = page_buffers(page);
200 if (!buffer_mapped(bh))
202 else if (bh->b_blocknr == block) {
207 bh = bh->b_this_page;
208 } while (bh != head);
210 /* we might be here because some of the buffers on this page are
211 * not mapped. This is due to various races between
212 * file io on the block device and getblk. It gets dealt with
213 * elsewhere, don't buffer_error if we had some unmapped buffers
216 char b[BDEVNAME_SIZE];
218 printk("__find_get_block_slow() failed. "
219 "block=%llu, b_blocknr=%llu\n",
220 (unsigned long long)block,
221 (unsigned long long)bh->b_blocknr);
222 printk("b_state=0x%08lx, b_size=%zu\n",
223 bh->b_state, bh->b_size);
224 printk("device %s blocksize: %d\n", bdevname(bdev, b),
225 1 << bd_inode->i_blkbits);
228 spin_unlock(&bd_mapping->private_lock);
229 page_cache_release(page);
234 /* If invalidate_buffers() will trash dirty buffers, it means some kind
235 of fs corruption is going on. Trashing dirty data always imply losing
236 information that was supposed to be just stored on the physical layer
239 Thus invalidate_buffers in general usage is not allwowed to trash
240 dirty buffers. For example ioctl(FLSBLKBUF) expects dirty data to
241 be preserved. These buffers are simply skipped.
243 We also skip buffers which are still in use. For example this can
244 happen if a userspace program is reading the block device.
246 NOTE: In the case where the user removed a removable-media-disk even if
247 there's still dirty data not synced on disk (due a bug in the device driver
248 or due an error of the user), by not destroying the dirty buffers we could
249 generate corruption also on the next media inserted, thus a parameter is
250 necessary to handle this case in the most safe way possible (trying
251 to not corrupt also the new disk inserted with the data belonging to
252 the old now corrupted disk). Also for the ramdisk the natural thing
253 to do in order to release the ramdisk memory is to destroy dirty buffers.
255 These are two special cases. Normal usage imply the device driver
256 to issue a sync on the device (without waiting I/O completion) and
257 then an invalidate_buffers call that doesn't trash dirty buffers.
259 For handling cache coherency with the blkdev pagecache the 'update' case
260 is been introduced. It is needed to re-read from disk any pinned
261 buffer. NOTE: re-reading from disk is destructive so we can do it only
262 when we assume nobody is changing the buffercache under our I/O and when
263 we think the disk contains more recent information than the buffercache.
264 The update == 1 pass marks the buffers we need to update, the update == 2
265 pass does the actual I/O. */
266 void invalidate_bdev(struct block_device *bdev)
268 struct address_space *mapping = bdev->bd_inode->i_mapping;
270 if (mapping->nrpages == 0)
273 invalidate_bh_lrus();
274 lru_add_drain_all(); /* make sure all lru add caches are flushed */
275 invalidate_mapping_pages(mapping, 0, -1);
276 /* 99% of the time, we don't need to flush the cleancache on the bdev.
277 * But, for the strange corners, lets be cautious
279 cleancache_flush_inode(mapping);
281 EXPORT_SYMBOL(invalidate_bdev);
284 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
286 static void free_more_memory(void)
291 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
294 for_each_online_node(nid) {
295 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
296 gfp_zone(GFP_NOFS), NULL,
299 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
305 * I/O completion handler for block_read_full_page() - pages
306 * which come unlocked at the end of I/O.
308 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
311 struct buffer_head *first;
312 struct buffer_head *tmp;
314 int page_uptodate = 1;
316 BUG_ON(!buffer_async_read(bh));
320 set_buffer_uptodate(bh);
322 clear_buffer_uptodate(bh);
323 if (!quiet_error(bh))
329 * Be _very_ careful from here on. Bad things can happen if
330 * two buffer heads end IO at almost the same time and both
331 * decide that the page is now completely done.
333 first = page_buffers(page);
334 local_irq_save(flags);
335 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
336 clear_buffer_async_read(bh);
340 if (!buffer_uptodate(tmp))
342 if (buffer_async_read(tmp)) {
343 BUG_ON(!buffer_locked(tmp));
346 tmp = tmp->b_this_page;
348 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
349 local_irq_restore(flags);
352 * If none of the buffers had errors and they are all
353 * uptodate then we can set the page uptodate.
355 if (page_uptodate && !PageError(page))
356 SetPageUptodate(page);
361 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
362 local_irq_restore(flags);
367 * Completion handler for block_write_full_page() - pages which are unlocked
368 * during I/O, and which have PageWriteback cleared upon I/O completion.
370 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
372 char b[BDEVNAME_SIZE];
374 struct buffer_head *first;
375 struct buffer_head *tmp;
378 BUG_ON(!buffer_async_write(bh));
382 set_buffer_uptodate(bh);
384 if (!quiet_error(bh)) {
386 printk(KERN_WARNING "lost page write due to "
388 bdevname(bh->b_bdev, b));
390 set_bit(AS_EIO, &page->mapping->flags);
391 set_buffer_write_io_error(bh);
392 clear_buffer_uptodate(bh);
396 first = page_buffers(page);
397 local_irq_save(flags);
398 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
400 clear_buffer_async_write(bh);
402 tmp = bh->b_this_page;
404 if (buffer_async_write(tmp)) {
405 BUG_ON(!buffer_locked(tmp));
408 tmp = tmp->b_this_page;
410 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
411 local_irq_restore(flags);
412 end_page_writeback(page);
416 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
417 local_irq_restore(flags);
420 EXPORT_SYMBOL(end_buffer_async_write);
423 * If a page's buffers are under async readin (end_buffer_async_read
424 * completion) then there is a possibility that another thread of
425 * control could lock one of the buffers after it has completed
426 * but while some of the other buffers have not completed. This
427 * locked buffer would confuse end_buffer_async_read() into not unlocking
428 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
429 * that this buffer is not under async I/O.
431 * The page comes unlocked when it has no locked buffer_async buffers
434 * PageLocked prevents anyone starting new async I/O reads any of
437 * PageWriteback is used to prevent simultaneous writeout of the same
440 * PageLocked prevents anyone from starting writeback of a page which is
441 * under read I/O (PageWriteback is only ever set against a locked page).
443 static void mark_buffer_async_read(struct buffer_head *bh)
445 bh->b_end_io = end_buffer_async_read;
446 set_buffer_async_read(bh);
449 static void mark_buffer_async_write_endio(struct buffer_head *bh,
450 bh_end_io_t *handler)
452 bh->b_end_io = handler;
453 set_buffer_async_write(bh);
456 void mark_buffer_async_write(struct buffer_head *bh)
458 mark_buffer_async_write_endio(bh, end_buffer_async_write);
460 EXPORT_SYMBOL(mark_buffer_async_write);
464 * fs/buffer.c contains helper functions for buffer-backed address space's
465 * fsync functions. A common requirement for buffer-based filesystems is
466 * that certain data from the backing blockdev needs to be written out for
467 * a successful fsync(). For example, ext2 indirect blocks need to be
468 * written back and waited upon before fsync() returns.
470 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
471 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
472 * management of a list of dependent buffers at ->i_mapping->private_list.
474 * Locking is a little subtle: try_to_free_buffers() will remove buffers
475 * from their controlling inode's queue when they are being freed. But
476 * try_to_free_buffers() will be operating against the *blockdev* mapping
477 * at the time, not against the S_ISREG file which depends on those buffers.
478 * So the locking for private_list is via the private_lock in the address_space
479 * which backs the buffers. Which is different from the address_space
480 * against which the buffers are listed. So for a particular address_space,
481 * mapping->private_lock does *not* protect mapping->private_list! In fact,
482 * mapping->private_list will always be protected by the backing blockdev's
485 * Which introduces a requirement: all buffers on an address_space's
486 * ->private_list must be from the same address_space: the blockdev's.
488 * address_spaces which do not place buffers at ->private_list via these
489 * utility functions are free to use private_lock and private_list for
490 * whatever they want. The only requirement is that list_empty(private_list)
491 * be true at clear_inode() time.
493 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
494 * filesystems should do that. invalidate_inode_buffers() should just go
495 * BUG_ON(!list_empty).
497 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
498 * take an address_space, not an inode. And it should be called
499 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
502 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
503 * list if it is already on a list. Because if the buffer is on a list,
504 * it *must* already be on the right one. If not, the filesystem is being
505 * silly. This will save a ton of locking. But first we have to ensure
506 * that buffers are taken *off* the old inode's list when they are freed
507 * (presumably in truncate). That requires careful auditing of all
508 * filesystems (do it inside bforget()). It could also be done by bringing
513 * The buffer's backing address_space's private_lock must be held
515 static void __remove_assoc_queue(struct buffer_head *bh)
517 list_del_init(&bh->b_assoc_buffers);
518 WARN_ON(!bh->b_assoc_map);
519 if (buffer_write_io_error(bh))
520 set_bit(AS_EIO, &bh->b_assoc_map->flags);
521 bh->b_assoc_map = NULL;
524 int inode_has_buffers(struct inode *inode)
526 return !list_empty(&inode->i_data.private_list);
530 * osync is designed to support O_SYNC io. It waits synchronously for
531 * all already-submitted IO to complete, but does not queue any new
532 * writes to the disk.
534 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
535 * you dirty the buffers, and then use osync_inode_buffers to wait for
536 * completion. Any other dirty buffers which are not yet queued for
537 * write will not be flushed to disk by the osync.
539 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
541 struct buffer_head *bh;
547 list_for_each_prev(p, list) {
549 if (buffer_locked(bh)) {
553 if (!buffer_uptodate(bh))
564 static void do_thaw_one(struct super_block *sb, void *unused)
566 char b[BDEVNAME_SIZE];
567 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
568 printk(KERN_WARNING "Emergency Thaw on %s\n",
569 bdevname(sb->s_bdev, b));
572 static void do_thaw_all(struct work_struct *work)
574 iterate_supers(do_thaw_one, NULL);
576 printk(KERN_WARNING "Emergency Thaw complete\n");
580 * emergency_thaw_all -- forcibly thaw every frozen filesystem
582 * Used for emergency unfreeze of all filesystems via SysRq
584 void emergency_thaw_all(void)
586 struct work_struct *work;
588 work = kmalloc(sizeof(*work), GFP_ATOMIC);
590 INIT_WORK(work, do_thaw_all);
596 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
597 * @mapping: the mapping which wants those buffers written
599 * Starts I/O against the buffers at mapping->private_list, and waits upon
602 * Basically, this is a convenience function for fsync().
603 * @mapping is a file or directory which needs those buffers to be written for
604 * a successful fsync().
606 int sync_mapping_buffers(struct address_space *mapping)
608 struct address_space *buffer_mapping = mapping->assoc_mapping;
610 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
613 return fsync_buffers_list(&buffer_mapping->private_lock,
614 &mapping->private_list);
616 EXPORT_SYMBOL(sync_mapping_buffers);
619 * Called when we've recently written block `bblock', and it is known that
620 * `bblock' was for a buffer_boundary() buffer. This means that the block at
621 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
622 * dirty, schedule it for IO. So that indirects merge nicely with their data.
624 void write_boundary_block(struct block_device *bdev,
625 sector_t bblock, unsigned blocksize)
627 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
629 if (buffer_dirty(bh))
630 ll_rw_block(WRITE, 1, &bh);
635 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
637 struct address_space *mapping = inode->i_mapping;
638 struct address_space *buffer_mapping = bh->b_page->mapping;
640 mark_buffer_dirty(bh);
641 if (!mapping->assoc_mapping) {
642 mapping->assoc_mapping = buffer_mapping;
644 BUG_ON(mapping->assoc_mapping != buffer_mapping);
646 if (!bh->b_assoc_map) {
647 spin_lock(&buffer_mapping->private_lock);
648 list_move_tail(&bh->b_assoc_buffers,
649 &mapping->private_list);
650 bh->b_assoc_map = mapping;
651 spin_unlock(&buffer_mapping->private_lock);
654 EXPORT_SYMBOL(mark_buffer_dirty_inode);
657 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
660 * If warn is true, then emit a warning if the page is not uptodate and has
661 * not been truncated.
663 static void __set_page_dirty(struct page *page,
664 struct address_space *mapping, int warn)
668 spin_lock_irqsave(&mapping->tree_lock, flags);
669 if (page->mapping) { /* Race with truncate? */
670 WARN_ON_ONCE(warn && !PageUptodate(page));
671 account_page_dirtied(page, mapping);
672 radix_tree_tag_set(&mapping->page_tree,
673 page_index(page), PAGECACHE_TAG_DIRTY);
675 spin_unlock_irqrestore(&mapping->tree_lock, flags);
676 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
680 * Add a page to the dirty page list.
682 * It is a sad fact of life that this function is called from several places
683 * deeply under spinlocking. It may not sleep.
685 * If the page has buffers, the uptodate buffers are set dirty, to preserve
686 * dirty-state coherency between the page and the buffers. It the page does
687 * not have buffers then when they are later attached they will all be set
690 * The buffers are dirtied before the page is dirtied. There's a small race
691 * window in which a writepage caller may see the page cleanness but not the
692 * buffer dirtiness. That's fine. If this code were to set the page dirty
693 * before the buffers, a concurrent writepage caller could clear the page dirty
694 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
695 * page on the dirty page list.
697 * We use private_lock to lock against try_to_free_buffers while using the
698 * page's buffer list. Also use this to protect against clean buffers being
699 * added to the page after it was set dirty.
701 * FIXME: may need to call ->reservepage here as well. That's rather up to the
702 * address_space though.
704 int __set_page_dirty_buffers(struct page *page)
707 struct address_space *mapping = page_mapping(page);
709 if (unlikely(!mapping))
710 return !TestSetPageDirty(page);
712 spin_lock(&mapping->private_lock);
713 if (page_has_buffers(page)) {
714 struct buffer_head *head = page_buffers(page);
715 struct buffer_head *bh = head;
718 set_buffer_dirty(bh);
719 bh = bh->b_this_page;
720 } while (bh != head);
722 newly_dirty = !TestSetPageDirty(page);
723 spin_unlock(&mapping->private_lock);
726 __set_page_dirty(page, mapping, 1);
729 EXPORT_SYMBOL(__set_page_dirty_buffers);
732 * Write out and wait upon a list of buffers.
734 * We have conflicting pressures: we want to make sure that all
735 * initially dirty buffers get waited on, but that any subsequently
736 * dirtied buffers don't. After all, we don't want fsync to last
737 * forever if somebody is actively writing to the file.
739 * Do this in two main stages: first we copy dirty buffers to a
740 * temporary inode list, queueing the writes as we go. Then we clean
741 * up, waiting for those writes to complete.
743 * During this second stage, any subsequent updates to the file may end
744 * up refiling the buffer on the original inode's dirty list again, so
745 * there is a chance we will end up with a buffer queued for write but
746 * not yet completed on that list. So, as a final cleanup we go through
747 * the osync code to catch these locked, dirty buffers without requeuing
748 * any newly dirty buffers for write.
750 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
752 struct buffer_head *bh;
753 struct list_head tmp;
754 struct address_space *mapping;
756 struct blk_plug plug;
758 INIT_LIST_HEAD(&tmp);
759 blk_start_plug(&plug);
762 while (!list_empty(list)) {
763 bh = BH_ENTRY(list->next);
764 mapping = bh->b_assoc_map;
765 __remove_assoc_queue(bh);
766 /* Avoid race with mark_buffer_dirty_inode() which does
767 * a lockless check and we rely on seeing the dirty bit */
769 if (buffer_dirty(bh) || buffer_locked(bh)) {
770 list_add(&bh->b_assoc_buffers, &tmp);
771 bh->b_assoc_map = mapping;
772 if (buffer_dirty(bh)) {
776 * Ensure any pending I/O completes so that
777 * write_dirty_buffer() actually writes the
778 * current contents - it is a noop if I/O is
779 * still in flight on potentially older
782 write_dirty_buffer(bh, WRITE_SYNC);
785 * Kick off IO for the previous mapping. Note
786 * that we will not run the very last mapping,
787 * wait_on_buffer() will do that for us
788 * through sync_buffer().
797 blk_finish_plug(&plug);
800 while (!list_empty(&tmp)) {
801 bh = BH_ENTRY(tmp.prev);
803 mapping = bh->b_assoc_map;
804 __remove_assoc_queue(bh);
805 /* Avoid race with mark_buffer_dirty_inode() which does
806 * a lockless check and we rely on seeing the dirty bit */
808 if (buffer_dirty(bh)) {
809 list_add(&bh->b_assoc_buffers,
810 &mapping->private_list);
811 bh->b_assoc_map = mapping;
815 if (!buffer_uptodate(bh))
822 err2 = osync_buffers_list(lock, list);
830 * Invalidate any and all dirty buffers on a given inode. We are
831 * probably unmounting the fs, but that doesn't mean we have already
832 * done a sync(). Just drop the buffers from the inode list.
834 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
835 * assumes that all the buffers are against the blockdev. Not true
838 void invalidate_inode_buffers(struct inode *inode)
840 if (inode_has_buffers(inode)) {
841 struct address_space *mapping = &inode->i_data;
842 struct list_head *list = &mapping->private_list;
843 struct address_space *buffer_mapping = mapping->assoc_mapping;
845 spin_lock(&buffer_mapping->private_lock);
846 while (!list_empty(list))
847 __remove_assoc_queue(BH_ENTRY(list->next));
848 spin_unlock(&buffer_mapping->private_lock);
851 EXPORT_SYMBOL(invalidate_inode_buffers);
854 * Remove any clean buffers from the inode's buffer list. This is called
855 * when we're trying to free the inode itself. Those buffers can pin it.
857 * Returns true if all buffers were removed.
859 int remove_inode_buffers(struct inode *inode)
863 if (inode_has_buffers(inode)) {
864 struct address_space *mapping = &inode->i_data;
865 struct list_head *list = &mapping->private_list;
866 struct address_space *buffer_mapping = mapping->assoc_mapping;
868 spin_lock(&buffer_mapping->private_lock);
869 while (!list_empty(list)) {
870 struct buffer_head *bh = BH_ENTRY(list->next);
871 if (buffer_dirty(bh)) {
875 __remove_assoc_queue(bh);
877 spin_unlock(&buffer_mapping->private_lock);
883 * Create the appropriate buffers when given a page for data area and
884 * the size of each buffer.. Use the bh->b_this_page linked list to
885 * follow the buffers created. Return NULL if unable to create more
888 * The retry flag is used to differentiate async IO (paging, swapping)
889 * which may not fail from ordinary buffer allocations.
891 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
894 struct buffer_head *bh, *head;
900 while ((offset -= size) >= 0) {
901 bh = alloc_buffer_head(GFP_NOFS);
906 bh->b_this_page = head;
911 atomic_set(&bh->b_count, 0);
914 /* Link the buffer to its page */
915 set_bh_page(bh, page, offset);
917 init_buffer(bh, NULL, NULL);
921 * In case anything failed, we just free everything we got.
927 head = head->b_this_page;
928 free_buffer_head(bh);
933 * Return failure for non-async IO requests. Async IO requests
934 * are not allowed to fail, so we have to wait until buffer heads
935 * become available. But we don't want tasks sleeping with
936 * partially complete buffers, so all were released above.
941 /* We're _really_ low on memory. Now we just
942 * wait for old buffer heads to become free due to
943 * finishing IO. Since this is an async request and
944 * the reserve list is empty, we're sure there are
945 * async buffer heads in use.
950 EXPORT_SYMBOL_GPL(alloc_page_buffers);
953 link_dev_buffers(struct page *page, struct buffer_head *head)
955 struct buffer_head *bh, *tail;
960 bh = bh->b_this_page;
962 tail->b_this_page = head;
963 attach_page_buffers(page, head);
967 * Initialise the state of a blockdev page's buffers.
970 init_page_buffers(struct page *page, struct block_device *bdev,
971 sector_t block, int size)
973 struct buffer_head *head = page_buffers(page);
974 struct buffer_head *bh = head;
975 int uptodate = PageUptodate(page);
976 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode));
979 if (!buffer_mapped(bh)) {
980 init_buffer(bh, NULL, NULL);
982 bh->b_blocknr = block;
984 set_buffer_uptodate(bh);
985 if (block < end_block)
986 set_buffer_mapped(bh);
989 bh = bh->b_this_page;
990 } while (bh != head);
993 * Caller needs to validate requested block against end of device.
999 * Create the page-cache page that contains the requested block.
1001 * This is used purely for blockdev mappings.
1004 grow_dev_page(struct block_device *bdev, sector_t block,
1005 pgoff_t index, int size, int sizebits)
1007 struct inode *inode = bdev->bd_inode;
1009 struct buffer_head *bh;
1011 int ret = 0; /* Will call free_more_memory() */
1013 page = find_or_create_page(inode->i_mapping, index,
1014 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
1018 BUG_ON(!PageLocked(page));
1020 if (page_has_buffers(page)) {
1021 bh = page_buffers(page);
1022 if (bh->b_size == size) {
1023 end_block = init_page_buffers(page, bdev,
1024 (sector_t)index << sizebits,
1028 if (!try_to_free_buffers(page))
1033 * Allocate some buffers for this page
1035 bh = alloc_page_buffers(page, size, 0);
1040 * Link the page to the buffers and initialise them. Take the
1041 * lock to be atomic wrt __find_get_block(), which does not
1042 * run under the page lock.
1044 spin_lock(&inode->i_mapping->private_lock);
1045 link_dev_buffers(page, bh);
1046 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1048 spin_unlock(&inode->i_mapping->private_lock);
1050 ret = (block < end_block) ? 1 : -ENXIO;
1053 page_cache_release(page);
1058 * Create buffers for the specified block device block's page. If
1059 * that page was dirty, the buffers are set dirty also.
1062 grow_buffers(struct block_device *bdev, sector_t block, int size)
1070 } while ((size << sizebits) < PAGE_SIZE);
1072 index = block >> sizebits;
1075 * Check for a block which wants to lie outside our maximum possible
1076 * pagecache index. (this comparison is done using sector_t types).
1078 if (unlikely(index != block >> sizebits)) {
1079 char b[BDEVNAME_SIZE];
1081 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1083 __func__, (unsigned long long)block,
1088 /* Create a page with the proper size buffers.. */
1089 return grow_dev_page(bdev, block, index, size, sizebits);
1092 static struct buffer_head *
1093 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1095 /* Size must be multiple of hard sectorsize */
1096 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1097 (size < 512 || size > PAGE_SIZE))) {
1098 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1100 printk(KERN_ERR "logical block size: %d\n",
1101 bdev_logical_block_size(bdev));
1108 struct buffer_head *bh;
1111 bh = __find_get_block(bdev, block, size);
1115 ret = grow_buffers(bdev, block, size);
1124 * The relationship between dirty buffers and dirty pages:
1126 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1127 * the page is tagged dirty in its radix tree.
1129 * At all times, the dirtiness of the buffers represents the dirtiness of
1130 * subsections of the page. If the page has buffers, the page dirty bit is
1131 * merely a hint about the true dirty state.
1133 * When a page is set dirty in its entirety, all its buffers are marked dirty
1134 * (if the page has buffers).
1136 * When a buffer is marked dirty, its page is dirtied, but the page's other
1139 * Also. When blockdev buffers are explicitly read with bread(), they
1140 * individually become uptodate. But their backing page remains not
1141 * uptodate - even if all of its buffers are uptodate. A subsequent
1142 * block_read_full_page() against that page will discover all the uptodate
1143 * buffers, will set the page uptodate and will perform no I/O.
1147 * mark_buffer_dirty - mark a buffer_head as needing writeout
1148 * @bh: the buffer_head to mark dirty
1150 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1151 * backing page dirty, then tag the page as dirty in its address_space's radix
1152 * tree and then attach the address_space's inode to its superblock's dirty
1155 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1156 * mapping->tree_lock and mapping->host->i_lock.
1158 void mark_buffer_dirty(struct buffer_head *bh)
1160 WARN_ON_ONCE(!buffer_uptodate(bh));
1163 * Very *carefully* optimize the it-is-already-dirty case.
1165 * Don't let the final "is it dirty" escape to before we
1166 * perhaps modified the buffer.
1168 if (buffer_dirty(bh)) {
1170 if (buffer_dirty(bh))
1174 if (!test_set_buffer_dirty(bh)) {
1175 struct page *page = bh->b_page;
1176 if (!TestSetPageDirty(page)) {
1177 struct address_space *mapping = page_mapping(page);
1179 __set_page_dirty(page, mapping, 0);
1183 EXPORT_SYMBOL(mark_buffer_dirty);
1186 * Decrement a buffer_head's reference count. If all buffers against a page
1187 * have zero reference count, are clean and unlocked, and if the page is clean
1188 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1189 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1190 * a page but it ends up not being freed, and buffers may later be reattached).
1192 void __brelse(struct buffer_head * buf)
1194 if (atomic_read(&buf->b_count)) {
1198 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1200 EXPORT_SYMBOL(__brelse);
1203 * bforget() is like brelse(), except it discards any
1204 * potentially dirty data.
1206 void __bforget(struct buffer_head *bh)
1208 clear_buffer_dirty(bh);
1209 if (bh->b_assoc_map) {
1210 struct address_space *buffer_mapping = bh->b_page->mapping;
1212 spin_lock(&buffer_mapping->private_lock);
1213 list_del_init(&bh->b_assoc_buffers);
1214 bh->b_assoc_map = NULL;
1215 spin_unlock(&buffer_mapping->private_lock);
1219 EXPORT_SYMBOL(__bforget);
1221 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1224 if (buffer_uptodate(bh)) {
1229 bh->b_end_io = end_buffer_read_sync;
1230 submit_bh(READ, bh);
1232 if (buffer_uptodate(bh))
1240 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1241 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1242 * refcount elevated by one when they're in an LRU. A buffer can only appear
1243 * once in a particular CPU's LRU. A single buffer can be present in multiple
1244 * CPU's LRUs at the same time.
1246 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1247 * sb_find_get_block().
1249 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1250 * a local interrupt disable for that.
1253 #define BH_LRU_SIZE 8
1256 struct buffer_head *bhs[BH_LRU_SIZE];
1259 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1262 #define bh_lru_lock() local_irq_disable()
1263 #define bh_lru_unlock() local_irq_enable()
1265 #define bh_lru_lock() preempt_disable()
1266 #define bh_lru_unlock() preempt_enable()
1269 static inline void check_irqs_on(void)
1271 #ifdef irqs_disabled
1272 BUG_ON(irqs_disabled());
1277 * The LRU management algorithm is dopey-but-simple. Sorry.
1279 static void bh_lru_install(struct buffer_head *bh)
1281 struct buffer_head *evictee = NULL;
1285 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1286 struct buffer_head *bhs[BH_LRU_SIZE];
1292 for (in = 0; in < BH_LRU_SIZE; in++) {
1293 struct buffer_head *bh2 =
1294 __this_cpu_read(bh_lrus.bhs[in]);
1299 if (out >= BH_LRU_SIZE) {
1300 BUG_ON(evictee != NULL);
1307 while (out < BH_LRU_SIZE)
1309 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1318 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1320 static struct buffer_head *
1321 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1323 struct buffer_head *ret = NULL;
1328 for (i = 0; i < BH_LRU_SIZE; i++) {
1329 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1331 if (bh && bh->b_bdev == bdev &&
1332 bh->b_blocknr == block && bh->b_size == size) {
1335 __this_cpu_write(bh_lrus.bhs[i],
1336 __this_cpu_read(bh_lrus.bhs[i - 1]));
1339 __this_cpu_write(bh_lrus.bhs[0], bh);
1351 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1352 * it in the LRU and mark it as accessed. If it is not present then return
1355 struct buffer_head *
1356 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1358 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1361 bh = __find_get_block_slow(bdev, block);
1369 EXPORT_SYMBOL(__find_get_block);
1372 * __getblk will locate (and, if necessary, create) the buffer_head
1373 * which corresponds to the passed block_device, block and size. The
1374 * returned buffer has its reference count incremented.
1376 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1377 * attempt is failing. FIXME, perhaps?
1379 struct buffer_head *
1380 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1382 struct buffer_head *bh = __find_get_block(bdev, block, size);
1386 bh = __getblk_slow(bdev, block, size);
1389 EXPORT_SYMBOL(__getblk);
1392 * Do async read-ahead on a buffer..
1394 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1396 struct buffer_head *bh = __getblk(bdev, block, size);
1398 ll_rw_block(READA, 1, &bh);
1402 EXPORT_SYMBOL(__breadahead);
1405 * __bread() - reads a specified block and returns the bh
1406 * @bdev: the block_device to read from
1407 * @block: number of block
1408 * @size: size (in bytes) to read
1410 * Reads a specified block, and returns buffer head that contains it.
1411 * It returns NULL if the block was unreadable.
1413 struct buffer_head *
1414 __bread(struct block_device *bdev, sector_t block, unsigned size)
1416 struct buffer_head *bh = __getblk(bdev, block, size);
1418 if (likely(bh) && !buffer_uptodate(bh))
1419 bh = __bread_slow(bh);
1422 EXPORT_SYMBOL(__bread);
1425 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1426 * This doesn't race because it runs in each cpu either in irq
1427 * or with preempt disabled.
1429 static void invalidate_bh_lru(void *arg)
1431 struct bh_lru *b = &get_cpu_var(bh_lrus);
1434 for (i = 0; i < BH_LRU_SIZE; i++) {
1438 put_cpu_var(bh_lrus);
1441 void invalidate_bh_lrus(void)
1443 on_each_cpu(invalidate_bh_lru, NULL, 1);
1445 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1447 void set_bh_page(struct buffer_head *bh,
1448 struct page *page, unsigned long offset)
1451 BUG_ON(offset >= PAGE_SIZE);
1452 if (PageHighMem(page))
1454 * This catches illegal uses and preserves the offset:
1456 bh->b_data = (char *)(0 + offset);
1458 bh->b_data = page_address(page) + offset;
1460 EXPORT_SYMBOL(set_bh_page);
1463 * Called when truncating a buffer on a page completely.
1465 static void discard_buffer(struct buffer_head * bh)
1468 clear_buffer_dirty(bh);
1470 clear_buffer_mapped(bh);
1471 clear_buffer_req(bh);
1472 clear_buffer_new(bh);
1473 clear_buffer_delay(bh);
1474 clear_buffer_unwritten(bh);
1479 * block_invalidatepage - invalidate part or all of a buffer-backed page
1481 * @page: the page which is affected
1482 * @offset: the index of the truncation point
1484 * block_invalidatepage() is called when all or part of the page has become
1485 * invalidated by a truncate operation.
1487 * block_invalidatepage() does not have to release all buffers, but it must
1488 * ensure that no dirty buffer is left outside @offset and that no I/O
1489 * is underway against any of the blocks which are outside the truncation
1490 * point. Because the caller is about to free (and possibly reuse) those
1493 void block_invalidatepage(struct page *page, unsigned long offset)
1495 struct buffer_head *head, *bh, *next;
1496 unsigned int curr_off = 0;
1498 BUG_ON(!PageLocked(page));
1499 if (!page_has_buffers(page))
1502 head = page_buffers(page);
1505 unsigned int next_off = curr_off + bh->b_size;
1506 next = bh->b_this_page;
1509 * is this block fully invalidated?
1511 if (offset <= curr_off)
1513 curr_off = next_off;
1515 } while (bh != head);
1518 * We release buffers only if the entire page is being invalidated.
1519 * The get_block cached value has been unconditionally invalidated,
1520 * so real IO is not possible anymore.
1523 try_to_release_page(page, 0);
1527 EXPORT_SYMBOL(block_invalidatepage);
1530 * We attach and possibly dirty the buffers atomically wrt
1531 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1532 * is already excluded via the page lock.
1534 void create_empty_buffers(struct page *page,
1535 unsigned long blocksize, unsigned long b_state)
1537 struct buffer_head *bh, *head, *tail;
1539 head = alloc_page_buffers(page, blocksize, 1);
1542 bh->b_state |= b_state;
1544 bh = bh->b_this_page;
1546 tail->b_this_page = head;
1548 spin_lock(&page->mapping->private_lock);
1549 if (PageUptodate(page) || PageDirty(page)) {
1552 if (PageDirty(page))
1553 set_buffer_dirty(bh);
1554 if (PageUptodate(page))
1555 set_buffer_uptodate(bh);
1556 bh = bh->b_this_page;
1557 } while (bh != head);
1559 attach_page_buffers(page, head);
1560 spin_unlock(&page->mapping->private_lock);
1562 EXPORT_SYMBOL(create_empty_buffers);
1565 * We are taking a block for data and we don't want any output from any
1566 * buffer-cache aliases starting from return from that function and
1567 * until the moment when something will explicitly mark the buffer
1568 * dirty (hopefully that will not happen until we will free that block ;-)
1569 * We don't even need to mark it not-uptodate - nobody can expect
1570 * anything from a newly allocated buffer anyway. We used to used
1571 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1572 * don't want to mark the alias unmapped, for example - it would confuse
1573 * anyone who might pick it with bread() afterwards...
1575 * Also.. Note that bforget() doesn't lock the buffer. So there can
1576 * be writeout I/O going on against recently-freed buffers. We don't
1577 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1578 * only if we really need to. That happens here.
1580 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1582 struct buffer_head *old_bh;
1586 old_bh = __find_get_block_slow(bdev, block);
1588 clear_buffer_dirty(old_bh);
1589 wait_on_buffer(old_bh);
1590 clear_buffer_req(old_bh);
1594 EXPORT_SYMBOL(unmap_underlying_metadata);
1597 * NOTE! All mapped/uptodate combinations are valid:
1599 * Mapped Uptodate Meaning
1601 * No No "unknown" - must do get_block()
1602 * No Yes "hole" - zero-filled
1603 * Yes No "allocated" - allocated on disk, not read in
1604 * Yes Yes "valid" - allocated and up-to-date in memory.
1606 * "Dirty" is valid only with the last case (mapped+uptodate).
1610 * While block_write_full_page is writing back the dirty buffers under
1611 * the page lock, whoever dirtied the buffers may decide to clean them
1612 * again at any time. We handle that by only looking at the buffer
1613 * state inside lock_buffer().
1615 * If block_write_full_page() is called for regular writeback
1616 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1617 * locked buffer. This only can happen if someone has written the buffer
1618 * directly, with submit_bh(). At the address_space level PageWriteback
1619 * prevents this contention from occurring.
1621 * If block_write_full_page() is called with wbc->sync_mode ==
1622 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1623 * causes the writes to be flagged as synchronous writes.
1625 static int __block_write_full_page(struct inode *inode, struct page *page,
1626 get_block_t *get_block, struct writeback_control *wbc,
1627 bh_end_io_t *handler)
1631 sector_t last_block;
1632 struct buffer_head *bh, *head;
1633 const unsigned blocksize = 1 << inode->i_blkbits;
1634 int nr_underway = 0;
1635 int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1636 WRITE_SYNC : WRITE);
1638 BUG_ON(!PageLocked(page));
1640 last_block = (i_size_read(inode) - 1) >> inode->i_blkbits;
1642 if (!page_has_buffers(page)) {
1643 create_empty_buffers(page, blocksize,
1644 (1 << BH_Dirty)|(1 << BH_Uptodate));
1648 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1649 * here, and the (potentially unmapped) buffers may become dirty at
1650 * any time. If a buffer becomes dirty here after we've inspected it
1651 * then we just miss that fact, and the page stays dirty.
1653 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1654 * handle that here by just cleaning them.
1657 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1658 head = page_buffers(page);
1662 * Get all the dirty buffers mapped to disk addresses and
1663 * handle any aliases from the underlying blockdev's mapping.
1666 if (block > last_block) {
1668 * mapped buffers outside i_size will occur, because
1669 * this page can be outside i_size when there is a
1670 * truncate in progress.
1673 * The buffer was zeroed by block_write_full_page()
1675 clear_buffer_dirty(bh);
1676 set_buffer_uptodate(bh);
1677 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1679 WARN_ON(bh->b_size != blocksize);
1680 err = get_block(inode, block, bh, 1);
1683 clear_buffer_delay(bh);
1684 if (buffer_new(bh)) {
1685 /* blockdev mappings never come here */
1686 clear_buffer_new(bh);
1687 unmap_underlying_metadata(bh->b_bdev,
1691 bh = bh->b_this_page;
1693 } while (bh != head);
1696 if (!buffer_mapped(bh))
1699 * If it's a fully non-blocking write attempt and we cannot
1700 * lock the buffer then redirty the page. Note that this can
1701 * potentially cause a busy-wait loop from writeback threads
1702 * and kswapd activity, but those code paths have their own
1703 * higher-level throttling.
1705 if (wbc->sync_mode != WB_SYNC_NONE) {
1707 } else if (!trylock_buffer(bh)) {
1708 redirty_page_for_writepage(wbc, page);
1711 if (test_clear_buffer_dirty(bh)) {
1712 mark_buffer_async_write_endio(bh, handler);
1716 } while ((bh = bh->b_this_page) != head);
1719 * The page and its buffers are protected by PageWriteback(), so we can
1720 * drop the bh refcounts early.
1722 BUG_ON(PageWriteback(page));
1723 set_page_writeback(page);
1726 struct buffer_head *next = bh->b_this_page;
1727 if (buffer_async_write(bh)) {
1728 submit_bh(write_op, bh);
1732 } while (bh != head);
1737 if (nr_underway == 0) {
1739 * The page was marked dirty, but the buffers were
1740 * clean. Someone wrote them back by hand with
1741 * ll_rw_block/submit_bh. A rare case.
1743 end_page_writeback(page);
1746 * The page and buffer_heads can be released at any time from
1754 * ENOSPC, or some other error. We may already have added some
1755 * blocks to the file, so we need to write these out to avoid
1756 * exposing stale data.
1757 * The page is currently locked and not marked for writeback
1760 /* Recovery: lock and submit the mapped buffers */
1762 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1763 !buffer_delay(bh)) {
1765 mark_buffer_async_write_endio(bh, handler);
1768 * The buffer may have been set dirty during
1769 * attachment to a dirty page.
1771 clear_buffer_dirty(bh);
1773 } while ((bh = bh->b_this_page) != head);
1775 BUG_ON(PageWriteback(page));
1776 mapping_set_error(page->mapping, err);
1777 set_page_writeback(page);
1779 struct buffer_head *next = bh->b_this_page;
1780 if (buffer_async_write(bh)) {
1781 clear_buffer_dirty(bh);
1782 submit_bh(write_op, bh);
1786 } while (bh != head);
1792 * If a page has any new buffers, zero them out here, and mark them uptodate
1793 * and dirty so they'll be written out (in order to prevent uninitialised
1794 * block data from leaking). And clear the new bit.
1796 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1798 unsigned int block_start, block_end;
1799 struct buffer_head *head, *bh;
1801 BUG_ON(!PageLocked(page));
1802 if (!page_has_buffers(page))
1805 bh = head = page_buffers(page);
1808 block_end = block_start + bh->b_size;
1810 if (buffer_new(bh)) {
1811 if (block_end > from && block_start < to) {
1812 if (!PageUptodate(page)) {
1813 unsigned start, size;
1815 start = max(from, block_start);
1816 size = min(to, block_end) - start;
1818 zero_user(page, start, size);
1819 set_buffer_uptodate(bh);
1822 clear_buffer_new(bh);
1823 mark_buffer_dirty(bh);
1827 block_start = block_end;
1828 bh = bh->b_this_page;
1829 } while (bh != head);
1831 EXPORT_SYMBOL(page_zero_new_buffers);
1833 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1834 get_block_t *get_block)
1836 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1837 unsigned to = from + len;
1838 struct inode *inode = page->mapping->host;
1839 unsigned block_start, block_end;
1842 unsigned blocksize, bbits;
1843 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1845 BUG_ON(!PageLocked(page));
1846 BUG_ON(from > PAGE_CACHE_SIZE);
1847 BUG_ON(to > PAGE_CACHE_SIZE);
1850 blocksize = 1 << inode->i_blkbits;
1851 if (!page_has_buffers(page))
1852 create_empty_buffers(page, blocksize, 0);
1853 head = page_buffers(page);
1855 bbits = inode->i_blkbits;
1856 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1858 for(bh = head, block_start = 0; bh != head || !block_start;
1859 block++, block_start=block_end, bh = bh->b_this_page) {
1860 block_end = block_start + blocksize;
1861 if (block_end <= from || block_start >= to) {
1862 if (PageUptodate(page)) {
1863 if (!buffer_uptodate(bh))
1864 set_buffer_uptodate(bh);
1869 clear_buffer_new(bh);
1870 if (!buffer_mapped(bh)) {
1871 WARN_ON(bh->b_size != blocksize);
1872 err = get_block(inode, block, bh, 1);
1875 if (buffer_new(bh)) {
1876 unmap_underlying_metadata(bh->b_bdev,
1878 if (PageUptodate(page)) {
1879 clear_buffer_new(bh);
1880 set_buffer_uptodate(bh);
1881 mark_buffer_dirty(bh);
1884 if (block_end > to || block_start < from)
1885 zero_user_segments(page,
1891 if (PageUptodate(page)) {
1892 if (!buffer_uptodate(bh))
1893 set_buffer_uptodate(bh);
1896 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1897 !buffer_unwritten(bh) &&
1898 (block_start < from || block_end > to)) {
1899 ll_rw_block(READ, 1, &bh);
1904 * If we issued read requests - let them complete.
1906 while(wait_bh > wait) {
1907 wait_on_buffer(*--wait_bh);
1908 if (!buffer_uptodate(*wait_bh))
1912 page_zero_new_buffers(page, from, to);
1915 EXPORT_SYMBOL(__block_write_begin);
1917 static int __block_commit_write(struct inode *inode, struct page *page,
1918 unsigned from, unsigned to)
1920 unsigned block_start, block_end;
1923 struct buffer_head *bh, *head;
1925 blocksize = 1 << inode->i_blkbits;
1927 for(bh = head = page_buffers(page), block_start = 0;
1928 bh != head || !block_start;
1929 block_start=block_end, bh = bh->b_this_page) {
1930 block_end = block_start + blocksize;
1931 if (block_end <= from || block_start >= to) {
1932 if (!buffer_uptodate(bh))
1935 set_buffer_uptodate(bh);
1936 mark_buffer_dirty(bh);
1938 clear_buffer_new(bh);
1942 * If this is a partial write which happened to make all buffers
1943 * uptodate then we can optimize away a bogus readpage() for
1944 * the next read(). Here we 'discover' whether the page went
1945 * uptodate as a result of this (potentially partial) write.
1948 SetPageUptodate(page);
1953 * block_write_begin takes care of the basic task of block allocation and
1954 * bringing partial write blocks uptodate first.
1956 * The filesystem needs to handle block truncation upon failure.
1958 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1959 unsigned flags, struct page **pagep, get_block_t *get_block)
1961 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1965 page = grab_cache_page_write_begin(mapping, index, flags);
1969 status = __block_write_begin(page, pos, len, get_block);
1970 if (unlikely(status)) {
1972 page_cache_release(page);
1979 EXPORT_SYMBOL(block_write_begin);
1981 int block_write_end(struct file *file, struct address_space *mapping,
1982 loff_t pos, unsigned len, unsigned copied,
1983 struct page *page, void *fsdata)
1985 struct inode *inode = mapping->host;
1988 start = pos & (PAGE_CACHE_SIZE - 1);
1990 if (unlikely(copied < len)) {
1992 * The buffers that were written will now be uptodate, so we
1993 * don't have to worry about a readpage reading them and
1994 * overwriting a partial write. However if we have encountered
1995 * a short write and only partially written into a buffer, it
1996 * will not be marked uptodate, so a readpage might come in and
1997 * destroy our partial write.
1999 * Do the simplest thing, and just treat any short write to a
2000 * non uptodate page as a zero-length write, and force the
2001 * caller to redo the whole thing.
2003 if (!PageUptodate(page))
2006 page_zero_new_buffers(page, start+copied, start+len);
2008 flush_dcache_page(page);
2010 /* This could be a short (even 0-length) commit */
2011 __block_commit_write(inode, page, start, start+copied);
2015 EXPORT_SYMBOL(block_write_end);
2017 int generic_write_end(struct file *file, struct address_space *mapping,
2018 loff_t pos, unsigned len, unsigned copied,
2019 struct page *page, void *fsdata)
2021 struct inode *inode = mapping->host;
2022 loff_t old_size = inode->i_size;
2023 int i_size_changed = 0;
2025 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2028 * No need to use i_size_read() here, the i_size
2029 * cannot change under us because we hold i_mutex.
2031 * But it's important to update i_size while still holding page lock:
2032 * page writeout could otherwise come in and zero beyond i_size.
2034 if (pos+copied > inode->i_size) {
2035 i_size_write(inode, pos+copied);
2040 page_cache_release(page);
2043 pagecache_isize_extended(inode, old_size, pos);
2045 * Don't mark the inode dirty under page lock. First, it unnecessarily
2046 * makes the holding time of page lock longer. Second, it forces lock
2047 * ordering of page lock and transaction start for journaling
2051 mark_inode_dirty(inode);
2055 EXPORT_SYMBOL(generic_write_end);
2058 * block_is_partially_uptodate checks whether buffers within a page are
2061 * Returns true if all buffers which correspond to a file portion
2062 * we want to read are uptodate.
2064 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2067 struct inode *inode = page->mapping->host;
2068 unsigned block_start, block_end, blocksize;
2070 struct buffer_head *bh, *head;
2073 if (!page_has_buffers(page))
2076 blocksize = 1 << inode->i_blkbits;
2077 to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2079 if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2082 head = page_buffers(page);
2086 block_end = block_start + blocksize;
2087 if (block_end > from && block_start < to) {
2088 if (!buffer_uptodate(bh)) {
2092 if (block_end >= to)
2095 block_start = block_end;
2096 bh = bh->b_this_page;
2097 } while (bh != head);
2101 EXPORT_SYMBOL(block_is_partially_uptodate);
2104 * Generic "read page" function for block devices that have the normal
2105 * get_block functionality. This is most of the block device filesystems.
2106 * Reads the page asynchronously --- the unlock_buffer() and
2107 * set/clear_buffer_uptodate() functions propagate buffer state into the
2108 * page struct once IO has completed.
2110 int block_read_full_page(struct page *page, get_block_t *get_block)
2112 struct inode *inode = page->mapping->host;
2113 sector_t iblock, lblock;
2114 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2115 unsigned int blocksize;
2117 int fully_mapped = 1;
2119 BUG_ON(!PageLocked(page));
2120 blocksize = 1 << inode->i_blkbits;
2121 if (!page_has_buffers(page))
2122 create_empty_buffers(page, blocksize, 0);
2123 head = page_buffers(page);
2125 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2126 lblock = (i_size_read(inode)+blocksize-1) >> inode->i_blkbits;
2132 if (buffer_uptodate(bh))
2135 if (!buffer_mapped(bh)) {
2139 if (iblock < lblock) {
2140 WARN_ON(bh->b_size != blocksize);
2141 err = get_block(inode, iblock, bh, 0);
2145 if (!buffer_mapped(bh)) {
2146 zero_user(page, i * blocksize, blocksize);
2148 set_buffer_uptodate(bh);
2152 * get_block() might have updated the buffer
2155 if (buffer_uptodate(bh))
2159 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2162 SetPageMappedToDisk(page);
2166 * All buffers are uptodate - we can set the page uptodate
2167 * as well. But not if get_block() returned an error.
2169 if (!PageError(page))
2170 SetPageUptodate(page);
2175 /* Stage two: lock the buffers */
2176 for (i = 0; i < nr; i++) {
2179 mark_buffer_async_read(bh);
2183 * Stage 3: start the IO. Check for uptodateness
2184 * inside the buffer lock in case another process reading
2185 * the underlying blockdev brought it uptodate (the sct fix).
2187 for (i = 0; i < nr; i++) {
2189 if (buffer_uptodate(bh))
2190 end_buffer_async_read(bh, 1);
2192 submit_bh(READ, bh);
2196 EXPORT_SYMBOL(block_read_full_page);
2198 /* utility function for filesystems that need to do work on expanding
2199 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2200 * deal with the hole.
2202 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2204 struct address_space *mapping = inode->i_mapping;
2209 err = inode_newsize_ok(inode, size);
2213 err = pagecache_write_begin(NULL, mapping, size, 0,
2214 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2219 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2225 EXPORT_SYMBOL(generic_cont_expand_simple);
2227 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2228 loff_t pos, loff_t *bytes)
2230 struct inode *inode = mapping->host;
2231 unsigned blocksize = 1 << inode->i_blkbits;
2234 pgoff_t index, curidx;
2236 unsigned zerofrom, offset, len;
2239 index = pos >> PAGE_CACHE_SHIFT;
2240 offset = pos & ~PAGE_CACHE_MASK;
2242 while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2243 zerofrom = curpos & ~PAGE_CACHE_MASK;
2244 if (zerofrom & (blocksize-1)) {
2245 *bytes |= (blocksize-1);
2248 len = PAGE_CACHE_SIZE - zerofrom;
2250 err = pagecache_write_begin(file, mapping, curpos, len,
2251 AOP_FLAG_UNINTERRUPTIBLE,
2255 zero_user(page, zerofrom, len);
2256 err = pagecache_write_end(file, mapping, curpos, len, len,
2263 balance_dirty_pages_ratelimited(mapping);
2265 if (unlikely(fatal_signal_pending(current))) {
2271 /* page covers the boundary, find the boundary offset */
2272 if (index == curidx) {
2273 zerofrom = curpos & ~PAGE_CACHE_MASK;
2274 /* if we will expand the thing last block will be filled */
2275 if (offset <= zerofrom) {
2278 if (zerofrom & (blocksize-1)) {
2279 *bytes |= (blocksize-1);
2282 len = offset - zerofrom;
2284 err = pagecache_write_begin(file, mapping, curpos, len,
2285 AOP_FLAG_UNINTERRUPTIBLE,
2289 zero_user(page, zerofrom, len);
2290 err = pagecache_write_end(file, mapping, curpos, len, len,
2302 * For moronic filesystems that do not allow holes in file.
2303 * We may have to extend the file.
2305 int cont_write_begin(struct file *file, struct address_space *mapping,
2306 loff_t pos, unsigned len, unsigned flags,
2307 struct page **pagep, void **fsdata,
2308 get_block_t *get_block, loff_t *bytes)
2310 struct inode *inode = mapping->host;
2311 unsigned blocksize = 1 << inode->i_blkbits;
2315 err = cont_expand_zero(file, mapping, pos, bytes);
2319 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2320 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2321 *bytes |= (blocksize-1);
2325 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2327 EXPORT_SYMBOL(cont_write_begin);
2329 int block_commit_write(struct page *page, unsigned from, unsigned to)
2331 struct inode *inode = page->mapping->host;
2332 __block_commit_write(inode,page,from,to);
2335 EXPORT_SYMBOL(block_commit_write);
2338 * block_page_mkwrite() is not allowed to change the file size as it gets
2339 * called from a page fault handler when a page is first dirtied. Hence we must
2340 * be careful to check for EOF conditions here. We set the page up correctly
2341 * for a written page which means we get ENOSPC checking when writing into
2342 * holes and correct delalloc and unwritten extent mapping on filesystems that
2343 * support these features.
2345 * We are not allowed to take the i_mutex here so we have to play games to
2346 * protect against truncate races as the page could now be beyond EOF. Because
2347 * truncate writes the inode size before removing pages, once we have the
2348 * page lock we can determine safely if the page is beyond EOF. If it is not
2349 * beyond EOF, then the page is guaranteed safe against truncation until we
2352 * Direct callers of this function should call vfs_check_frozen() so that page
2353 * fault does not busyloop until the fs is thawed.
2355 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2356 get_block_t get_block)
2358 struct page *page = vmf->page;
2359 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2365 size = i_size_read(inode);
2366 if ((page->mapping != inode->i_mapping) ||
2367 (page_offset(page) > size)) {
2368 /* We overload EFAULT to mean page got truncated */
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_write_begin(page, 0, end, get_block);
2381 ret = block_commit_write(page, 0, end);
2383 if (unlikely(ret < 0))
2386 * Freezing in progress? We check after the page is marked dirty and
2387 * with page lock held so if the test here fails, we are sure freezing
2388 * code will wait during syncing until the page fault is done - at that
2389 * point page will be dirty and unlocked so freezing code will write it
2390 * and writeprotect it again.
2392 set_page_dirty(page);
2393 if (inode->i_sb->s_frozen != SB_UNFROZEN) {
2397 wait_on_page_writeback(page);
2403 EXPORT_SYMBOL(__block_page_mkwrite);
2405 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2406 get_block_t get_block)
2409 struct super_block *sb = vma->vm_file->f_path.dentry->d_inode->i_sb;
2412 * This check is racy but catches the common case. The check in
2413 * __block_page_mkwrite() is reliable.
2415 vfs_check_frozen(sb, SB_FREEZE_WRITE);
2416 ret = __block_page_mkwrite(vma, vmf, get_block);
2417 return block_page_mkwrite_return(ret);
2419 EXPORT_SYMBOL(block_page_mkwrite);
2422 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2423 * immediately, while under the page lock. So it needs a special end_io
2424 * handler which does not touch the bh after unlocking it.
2426 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2428 __end_buffer_read_notouch(bh, uptodate);
2432 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2433 * the page (converting it to circular linked list and taking care of page
2436 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2438 struct buffer_head *bh;
2440 BUG_ON(!PageLocked(page));
2442 spin_lock(&page->mapping->private_lock);
2445 if (PageDirty(page))
2446 set_buffer_dirty(bh);
2447 if (!bh->b_this_page)
2448 bh->b_this_page = head;
2449 bh = bh->b_this_page;
2450 } while (bh != head);
2451 attach_page_buffers(page, head);
2452 spin_unlock(&page->mapping->private_lock);
2456 * On entry, the page is fully not uptodate.
2457 * On exit the page is fully uptodate in the areas outside (from,to)
2458 * The filesystem needs to handle block truncation upon failure.
2460 int nobh_write_begin(struct address_space *mapping,
2461 loff_t pos, unsigned len, unsigned flags,
2462 struct page **pagep, void **fsdata,
2463 get_block_t *get_block)
2465 struct inode *inode = mapping->host;
2466 const unsigned blkbits = inode->i_blkbits;
2467 const unsigned blocksize = 1 << blkbits;
2468 struct buffer_head *head, *bh;
2472 unsigned block_in_page;
2473 unsigned block_start, block_end;
2474 sector_t block_in_file;
2477 int is_mapped_to_disk = 1;
2479 index = pos >> PAGE_CACHE_SHIFT;
2480 from = pos & (PAGE_CACHE_SIZE - 1);
2483 page = grab_cache_page_write_begin(mapping, index, flags);
2489 if (page_has_buffers(page)) {
2490 ret = __block_write_begin(page, pos, len, get_block);
2496 if (PageMappedToDisk(page))
2500 * Allocate buffers so that we can keep track of state, and potentially
2501 * attach them to the page if an error occurs. In the common case of
2502 * no error, they will just be freed again without ever being attached
2503 * to the page (which is all OK, because we're under the page lock).
2505 * Be careful: the buffer linked list is a NULL terminated one, rather
2506 * than the circular one we're used to.
2508 head = alloc_page_buffers(page, blocksize, 0);
2514 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2517 * We loop across all blocks in the page, whether or not they are
2518 * part of the affected region. This is so we can discover if the
2519 * page is fully mapped-to-disk.
2521 for (block_start = 0, block_in_page = 0, bh = head;
2522 block_start < PAGE_CACHE_SIZE;
2523 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2526 block_end = block_start + blocksize;
2529 if (block_start >= to)
2531 ret = get_block(inode, block_in_file + block_in_page,
2535 if (!buffer_mapped(bh))
2536 is_mapped_to_disk = 0;
2538 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2539 if (PageUptodate(page)) {
2540 set_buffer_uptodate(bh);
2543 if (buffer_new(bh) || !buffer_mapped(bh)) {
2544 zero_user_segments(page, block_start, from,
2548 if (buffer_uptodate(bh))
2549 continue; /* reiserfs does this */
2550 if (block_start < from || block_end > to) {
2552 bh->b_end_io = end_buffer_read_nobh;
2553 submit_bh(READ, bh);
2560 * The page is locked, so these buffers are protected from
2561 * any VM or truncate activity. Hence we don't need to care
2562 * for the buffer_head refcounts.
2564 for (bh = head; bh; bh = bh->b_this_page) {
2566 if (!buffer_uptodate(bh))
2573 if (is_mapped_to_disk)
2574 SetPageMappedToDisk(page);
2576 *fsdata = head; /* to be released by nobh_write_end */
2583 * Error recovery is a bit difficult. We need to zero out blocks that
2584 * were newly allocated, and dirty them to ensure they get written out.
2585 * Buffers need to be attached to the page at this point, otherwise
2586 * the handling of potential IO errors during writeout would be hard
2587 * (could try doing synchronous writeout, but what if that fails too?)
2589 attach_nobh_buffers(page, head);
2590 page_zero_new_buffers(page, from, to);
2594 page_cache_release(page);
2599 EXPORT_SYMBOL(nobh_write_begin);
2601 int nobh_write_end(struct file *file, struct address_space *mapping,
2602 loff_t pos, unsigned len, unsigned copied,
2603 struct page *page, void *fsdata)
2605 struct inode *inode = page->mapping->host;
2606 struct buffer_head *head = fsdata;
2607 struct buffer_head *bh;
2608 BUG_ON(fsdata != NULL && page_has_buffers(page));
2610 if (unlikely(copied < len) && head)
2611 attach_nobh_buffers(page, head);
2612 if (page_has_buffers(page))
2613 return generic_write_end(file, mapping, pos, len,
2614 copied, page, fsdata);
2616 SetPageUptodate(page);
2617 set_page_dirty(page);
2618 if (pos+copied > inode->i_size) {
2619 i_size_write(inode, pos+copied);
2620 mark_inode_dirty(inode);
2624 page_cache_release(page);
2628 head = head->b_this_page;
2629 free_buffer_head(bh);
2634 EXPORT_SYMBOL(nobh_write_end);
2637 * nobh_writepage() - based on block_full_write_page() except
2638 * that it tries to operate without attaching bufferheads to
2641 int nobh_writepage(struct page *page, get_block_t *get_block,
2642 struct writeback_control *wbc)
2644 struct inode * const inode = page->mapping->host;
2645 loff_t i_size = i_size_read(inode);
2646 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2650 /* Is the page fully inside i_size? */
2651 if (page->index < end_index)
2654 /* Is the page fully outside i_size? (truncate in progress) */
2655 offset = i_size & (PAGE_CACHE_SIZE-1);
2656 if (page->index >= end_index+1 || !offset) {
2658 * The page may have dirty, unmapped buffers. For example,
2659 * they may have been added in ext3_writepage(). Make them
2660 * freeable here, so the page does not leak.
2663 /* Not really sure about this - do we need this ? */
2664 if (page->mapping->a_ops->invalidatepage)
2665 page->mapping->a_ops->invalidatepage(page, offset);
2668 return 0; /* don't care */
2672 * The page straddles i_size. It must be zeroed out on each and every
2673 * writepage invocation because it may be mmapped. "A file is mapped
2674 * in multiples of the page size. For a file that is not a multiple of
2675 * the page size, the remaining memory is zeroed when mapped, and
2676 * writes to that region are not written out to the file."
2678 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2680 ret = mpage_writepage(page, get_block, wbc);
2682 ret = __block_write_full_page(inode, page, get_block, wbc,
2683 end_buffer_async_write);
2686 EXPORT_SYMBOL(nobh_writepage);
2688 int nobh_truncate_page(struct address_space *mapping,
2689 loff_t from, get_block_t *get_block)
2691 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2692 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2695 unsigned length, pos;
2696 struct inode *inode = mapping->host;
2698 struct buffer_head map_bh;
2701 blocksize = 1 << inode->i_blkbits;
2702 length = offset & (blocksize - 1);
2704 /* Block boundary? Nothing to do */
2708 length = blocksize - length;
2709 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2711 page = grab_cache_page(mapping, index);
2716 if (page_has_buffers(page)) {
2719 page_cache_release(page);
2720 return block_truncate_page(mapping, from, get_block);
2723 /* Find the buffer that contains "offset" */
2725 while (offset >= pos) {
2730 map_bh.b_size = blocksize;
2732 err = get_block(inode, iblock, &map_bh, 0);
2735 /* unmapped? It's a hole - nothing to do */
2736 if (!buffer_mapped(&map_bh))
2739 /* Ok, it's mapped. Make sure it's up-to-date */
2740 if (!PageUptodate(page)) {
2741 err = mapping->a_ops->readpage(NULL, page);
2743 page_cache_release(page);
2747 if (!PageUptodate(page)) {
2751 if (page_has_buffers(page))
2754 zero_user(page, offset, length);
2755 set_page_dirty(page);
2760 page_cache_release(page);
2764 EXPORT_SYMBOL(nobh_truncate_page);
2766 int block_truncate_page(struct address_space *mapping,
2767 loff_t from, get_block_t *get_block)
2769 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2770 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2773 unsigned length, pos;
2774 struct inode *inode = mapping->host;
2776 struct buffer_head *bh;
2779 blocksize = 1 << inode->i_blkbits;
2780 length = offset & (blocksize - 1);
2782 /* Block boundary? Nothing to do */
2786 length = blocksize - length;
2787 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2789 page = grab_cache_page(mapping, index);
2794 if (!page_has_buffers(page))
2795 create_empty_buffers(page, blocksize, 0);
2797 /* Find the buffer that contains "offset" */
2798 bh = page_buffers(page);
2800 while (offset >= pos) {
2801 bh = bh->b_this_page;
2807 if (!buffer_mapped(bh)) {
2808 WARN_ON(bh->b_size != blocksize);
2809 err = get_block(inode, iblock, bh, 0);
2812 /* unmapped? It's a hole - nothing to do */
2813 if (!buffer_mapped(bh))
2817 /* Ok, it's mapped. Make sure it's up-to-date */
2818 if (PageUptodate(page))
2819 set_buffer_uptodate(bh);
2821 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2823 ll_rw_block(READ, 1, &bh);
2825 /* Uhhuh. Read error. Complain and punt. */
2826 if (!buffer_uptodate(bh))
2830 zero_user(page, offset, length);
2831 mark_buffer_dirty(bh);
2836 page_cache_release(page);
2840 EXPORT_SYMBOL(block_truncate_page);
2843 * The generic ->writepage function for buffer-backed address_spaces
2844 * this form passes in the end_io handler used to finish the IO.
2846 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2847 struct writeback_control *wbc, bh_end_io_t *handler)
2849 struct inode * const inode = page->mapping->host;
2850 loff_t i_size = i_size_read(inode);
2851 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2854 /* Is the page fully inside i_size? */
2855 if (page->index < end_index)
2856 return __block_write_full_page(inode, page, get_block, wbc,
2859 /* Is the page fully outside i_size? (truncate in progress) */
2860 offset = i_size & (PAGE_CACHE_SIZE-1);
2861 if (page->index >= end_index+1 || !offset) {
2863 * The page may have dirty, unmapped buffers. For example,
2864 * they may have been added in ext3_writepage(). Make them
2865 * freeable here, so the page does not leak.
2867 do_invalidatepage(page, 0);
2869 return 0; /* don't care */
2873 * The page straddles i_size. It must be zeroed out on each and every
2874 * writepage invocation because it may be mmapped. "A file is mapped
2875 * in multiples of the page size. For a file that is not a multiple of
2876 * the page size, the remaining memory is zeroed when mapped, and
2877 * writes to that region are not written out to the file."
2879 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2880 return __block_write_full_page(inode, page, get_block, wbc, handler);
2882 EXPORT_SYMBOL(block_write_full_page_endio);
2885 * The generic ->writepage function for buffer-backed address_spaces
2887 int block_write_full_page(struct page *page, get_block_t *get_block,
2888 struct writeback_control *wbc)
2890 return block_write_full_page_endio(page, get_block, wbc,
2891 end_buffer_async_write);
2893 EXPORT_SYMBOL(block_write_full_page);
2895 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2896 get_block_t *get_block)
2898 struct buffer_head tmp;
2899 struct inode *inode = mapping->host;
2902 tmp.b_size = 1 << inode->i_blkbits;
2903 get_block(inode, block, &tmp, 0);
2904 return tmp.b_blocknr;
2906 EXPORT_SYMBOL(generic_block_bmap);
2908 static void end_bio_bh_io_sync(struct bio *bio, int err)
2910 struct buffer_head *bh = bio->bi_private;
2912 if (err == -EOPNOTSUPP) {
2913 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2916 if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2917 set_bit(BH_Quiet, &bh->b_state);
2919 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2923 int submit_bh(int rw, struct buffer_head * bh)
2928 BUG_ON(!buffer_locked(bh));
2929 BUG_ON(!buffer_mapped(bh));
2930 BUG_ON(!bh->b_end_io);
2931 BUG_ON(buffer_delay(bh));
2932 BUG_ON(buffer_unwritten(bh));
2935 * Only clear out a write error when rewriting
2937 if (test_set_buffer_req(bh) && (rw & WRITE))
2938 clear_buffer_write_io_error(bh);
2941 * from here on down, it's all bio -- do the initial mapping,
2942 * submit_bio -> generic_make_request may further map this bio around
2944 bio = bio_alloc(GFP_NOIO, 1);
2946 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2947 bio->bi_bdev = bh->b_bdev;
2948 bio->bi_io_vec[0].bv_page = bh->b_page;
2949 bio->bi_io_vec[0].bv_len = bh->b_size;
2950 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2954 bio->bi_size = bh->b_size;
2956 bio->bi_end_io = end_bio_bh_io_sync;
2957 bio->bi_private = bh;
2960 submit_bio(rw, bio);
2962 if (bio_flagged(bio, BIO_EOPNOTSUPP))
2968 EXPORT_SYMBOL(submit_bh);
2971 * ll_rw_block: low-level access to block devices (DEPRECATED)
2972 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2973 * @nr: number of &struct buffer_heads in the array
2974 * @bhs: array of pointers to &struct buffer_head
2976 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
2977 * requests an I/O operation on them, either a %READ or a %WRITE. The third
2978 * %READA option is described in the documentation for generic_make_request()
2979 * which ll_rw_block() calls.
2981 * This function drops any buffer that it cannot get a lock on (with the
2982 * BH_Lock state bit), any buffer that appears to be clean when doing a write
2983 * request, and any buffer that appears to be up-to-date when doing read
2984 * request. Further it marks as clean buffers that are processed for
2985 * writing (the buffer cache won't assume that they are actually clean
2986 * until the buffer gets unlocked).
2988 * ll_rw_block sets b_end_io to simple completion handler that marks
2989 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
2992 * All of the buffers must be for the same device, and must also be a
2993 * multiple of the current approved size for the device.
2995 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
2999 for (i = 0; i < nr; i++) {
3000 struct buffer_head *bh = bhs[i];
3002 if (!trylock_buffer(bh))
3005 if (test_clear_buffer_dirty(bh)) {
3006 bh->b_end_io = end_buffer_write_sync;
3008 submit_bh(WRITE, bh);
3012 if (!buffer_uptodate(bh)) {
3013 bh->b_end_io = end_buffer_read_sync;
3022 EXPORT_SYMBOL(ll_rw_block);
3024 void write_dirty_buffer(struct buffer_head *bh, int rw)
3027 if (!test_clear_buffer_dirty(bh)) {
3031 bh->b_end_io = end_buffer_write_sync;
3035 EXPORT_SYMBOL(write_dirty_buffer);
3038 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3039 * and then start new I/O and then wait upon it. The caller must have a ref on
3042 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3046 WARN_ON(atomic_read(&bh->b_count) < 1);
3048 if (test_clear_buffer_dirty(bh)) {
3050 bh->b_end_io = end_buffer_write_sync;
3051 ret = submit_bh(rw, bh);
3053 if (!ret && !buffer_uptodate(bh))
3060 EXPORT_SYMBOL(__sync_dirty_buffer);
3062 int sync_dirty_buffer(struct buffer_head *bh)
3064 return __sync_dirty_buffer(bh, WRITE_SYNC);
3066 EXPORT_SYMBOL(sync_dirty_buffer);
3069 * try_to_free_buffers() checks if all the buffers on this particular page
3070 * are unused, and releases them if so.
3072 * Exclusion against try_to_free_buffers may be obtained by either
3073 * locking the page or by holding its mapping's private_lock.
3075 * If the page is dirty but all the buffers are clean then we need to
3076 * be sure to mark the page clean as well. This is because the page
3077 * may be against a block device, and a later reattachment of buffers
3078 * to a dirty page will set *all* buffers dirty. Which would corrupt
3079 * filesystem data on the same device.
3081 * The same applies to regular filesystem pages: if all the buffers are
3082 * clean then we set the page clean and proceed. To do that, we require
3083 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3086 * try_to_free_buffers() is non-blocking.
3088 static inline int buffer_busy(struct buffer_head *bh)
3090 return atomic_read(&bh->b_count) |
3091 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3095 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3097 struct buffer_head *head = page_buffers(page);
3098 struct buffer_head *bh;
3102 if (buffer_write_io_error(bh) && page->mapping)
3103 set_bit(AS_EIO, &page->mapping->flags);
3104 if (buffer_busy(bh))
3106 bh = bh->b_this_page;
3107 } while (bh != head);
3110 struct buffer_head *next = bh->b_this_page;
3112 if (bh->b_assoc_map)
3113 __remove_assoc_queue(bh);
3115 } while (bh != head);
3116 *buffers_to_free = head;
3117 __clear_page_buffers(page);
3123 int try_to_free_buffers(struct page *page)
3125 struct address_space * const mapping = page->mapping;
3126 struct buffer_head *buffers_to_free = NULL;
3129 BUG_ON(!PageLocked(page));
3130 if (PageWriteback(page))
3133 if (mapping == NULL) { /* can this still happen? */
3134 ret = drop_buffers(page, &buffers_to_free);
3138 spin_lock(&mapping->private_lock);
3139 ret = drop_buffers(page, &buffers_to_free);
3142 * If the filesystem writes its buffers by hand (eg ext3)
3143 * then we can have clean buffers against a dirty page. We
3144 * clean the page here; otherwise the VM will never notice
3145 * that the filesystem did any IO at all.
3147 * Also, during truncate, discard_buffer will have marked all
3148 * the page's buffers clean. We discover that here and clean
3151 * private_lock must be held over this entire operation in order
3152 * to synchronise against __set_page_dirty_buffers and prevent the
3153 * dirty bit from being lost.
3156 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3157 spin_unlock(&mapping->private_lock);
3159 if (buffers_to_free) {
3160 struct buffer_head *bh = buffers_to_free;
3163 struct buffer_head *next = bh->b_this_page;
3164 free_buffer_head(bh);
3166 } while (bh != buffers_to_free);
3170 EXPORT_SYMBOL(try_to_free_buffers);
3173 * There are no bdflush tunables left. But distributions are
3174 * still running obsolete flush daemons, so we terminate them here.
3176 * Use of bdflush() is deprecated and will be removed in a future kernel.
3177 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3179 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3181 static int msg_count;
3183 if (!capable(CAP_SYS_ADMIN))
3186 if (msg_count < 5) {
3189 "warning: process `%s' used the obsolete bdflush"
3190 " system call\n", current->comm);
3191 printk(KERN_INFO "Fix your initscripts?\n");
3200 * Buffer-head allocation
3202 static struct kmem_cache *bh_cachep;
3205 * Once the number of bh's in the machine exceeds this level, we start
3206 * stripping them in writeback.
3208 static int max_buffer_heads;
3210 int buffer_heads_over_limit;
3212 struct bh_accounting {
3213 int nr; /* Number of live bh's */
3214 int ratelimit; /* Limit cacheline bouncing */
3217 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3219 static void recalc_bh_state(void)
3224 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3226 __this_cpu_write(bh_accounting.ratelimit, 0);
3227 for_each_online_cpu(i)
3228 tot += per_cpu(bh_accounting, i).nr;
3229 buffer_heads_over_limit = (tot > max_buffer_heads);
3232 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3234 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3236 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3238 __this_cpu_inc(bh_accounting.nr);
3244 EXPORT_SYMBOL(alloc_buffer_head);
3246 void free_buffer_head(struct buffer_head *bh)
3248 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3249 kmem_cache_free(bh_cachep, bh);
3251 __this_cpu_dec(bh_accounting.nr);
3255 EXPORT_SYMBOL(free_buffer_head);
3257 static void buffer_exit_cpu(int cpu)
3260 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3262 for (i = 0; i < BH_LRU_SIZE; i++) {
3266 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3267 per_cpu(bh_accounting, cpu).nr = 0;
3270 static int buffer_cpu_notify(struct notifier_block *self,
3271 unsigned long action, void *hcpu)
3273 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3274 buffer_exit_cpu((unsigned long)hcpu);
3279 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3280 * @bh: struct buffer_head
3282 * Return true if the buffer is up-to-date and false,
3283 * with the buffer locked, if not.
3285 int bh_uptodate_or_lock(struct buffer_head *bh)
3287 if (!buffer_uptodate(bh)) {
3289 if (!buffer_uptodate(bh))
3295 EXPORT_SYMBOL(bh_uptodate_or_lock);
3298 * bh_submit_read - Submit a locked buffer for reading
3299 * @bh: struct buffer_head
3301 * Returns zero on success and -EIO on error.
3303 int bh_submit_read(struct buffer_head *bh)
3305 BUG_ON(!buffer_locked(bh));
3307 if (buffer_uptodate(bh)) {
3313 bh->b_end_io = end_buffer_read_sync;
3314 submit_bh(READ, bh);
3316 if (buffer_uptodate(bh))
3320 EXPORT_SYMBOL(bh_submit_read);
3322 void __init buffer_init(void)
3326 bh_cachep = kmem_cache_create("buffer_head",
3327 sizeof(struct buffer_head), 0,
3328 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3333 * Limit the bh occupancy to 10% of ZONE_NORMAL
3335 nrpages = (nr_free_buffer_pages() * 10) / 100;
3336 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3337 hotcpu_notifier(buffer_cpu_notify, 0);