4 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
8 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
10 * Removed a lot of unnecessary code and simplified things now that
11 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
13 * Speed up hash, lru, and free list operations. Use gfp() for allocating
14 * hash table, use SLAB cache for buffer heads. SMP threading. -DaveM
16 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
18 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
21 #include <linux/kernel.h>
22 #include <linux/syscalls.h>
25 #include <linux/percpu.h>
26 #include <linux/slab.h>
27 #include <linux/capability.h>
28 #include <linux/blkdev.h>
29 #include <linux/file.h>
30 #include <linux/quotaops.h>
31 #include <linux/highmem.h>
32 #include <linux/export.h>
33 #include <linux/writeback.h>
34 #include <linux/hash.h>
35 #include <linux/suspend.h>
36 #include <linux/buffer_head.h>
37 #include <linux/task_io_accounting_ops.h>
38 #include <linux/bio.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/bitops.h>
42 #include <linux/mpage.h>
43 #include <linux/bit_spinlock.h>
44 #include <trace/events/block.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)
50 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
52 bh->b_end_io = handler;
53 bh->b_private = private;
55 EXPORT_SYMBOL(init_buffer);
57 inline void touch_buffer(struct buffer_head *bh)
59 trace_block_touch_buffer(bh);
60 mark_page_accessed(bh->b_page);
62 EXPORT_SYMBOL(touch_buffer);
64 static int sleep_on_buffer(void *word)
70 void __lock_buffer(struct buffer_head *bh)
72 wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
73 TASK_UNINTERRUPTIBLE);
75 EXPORT_SYMBOL(__lock_buffer);
77 void unlock_buffer(struct buffer_head *bh)
79 clear_bit_unlock(BH_Lock, &bh->b_state);
80 smp_mb__after_clear_bit();
81 wake_up_bit(&bh->b_state, BH_Lock);
83 EXPORT_SYMBOL(unlock_buffer);
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, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
94 EXPORT_SYMBOL(__wait_on_buffer);
97 __clear_page_buffers(struct page *page)
99 ClearPagePrivate(page);
100 set_page_private(page, 0);
101 page_cache_release(page);
105 static int quiet_error(struct buffer_head *bh)
107 if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
113 static void buffer_io_error(struct buffer_head *bh)
115 char b[BDEVNAME_SIZE];
116 printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
117 bdevname(bh->b_bdev, b),
118 (unsigned long long)bh->b_blocknr);
122 * End-of-IO handler helper function which does not touch the bh after
124 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
125 * a race there is benign: unlock_buffer() only use the bh's address for
126 * hashing after unlocking the buffer, so it doesn't actually touch the bh
129 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
132 set_buffer_uptodate(bh);
134 /* This happens, due to failed READA attempts. */
135 clear_buffer_uptodate(bh);
141 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
142 * unlock the buffer. This is what ll_rw_block uses too.
144 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
146 __end_buffer_read_notouch(bh, uptodate);
149 EXPORT_SYMBOL(end_buffer_read_sync);
151 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
153 char b[BDEVNAME_SIZE];
156 set_buffer_uptodate(bh);
158 if (!quiet_error(bh)) {
160 printk(KERN_WARNING "lost page write due to "
162 bdevname(bh->b_bdev, b));
164 set_buffer_write_io_error(bh);
165 clear_buffer_uptodate(bh);
170 EXPORT_SYMBOL(end_buffer_write_sync);
173 * Various filesystems appear to want __find_get_block to be non-blocking.
174 * But it's the page lock which protects the buffers. To get around this,
175 * we get exclusion from try_to_free_buffers with the blockdev mapping's
178 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
179 * may be quite high. This code could TryLock the page, and if that
180 * succeeds, there is no need to take private_lock. (But if
181 * private_lock is contended then so is mapping->tree_lock).
183 static struct buffer_head *
184 __find_get_block_slow(struct block_device *bdev, sector_t block)
186 struct inode *bd_inode = bdev->bd_inode;
187 struct address_space *bd_mapping = bd_inode->i_mapping;
188 struct buffer_head *ret = NULL;
190 struct buffer_head *bh;
191 struct buffer_head *head;
195 index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
196 page = find_get_page(bd_mapping, index);
200 spin_lock(&bd_mapping->private_lock);
201 if (!page_has_buffers(page))
203 head = page_buffers(page);
206 if (!buffer_mapped(bh))
208 else if (bh->b_blocknr == block) {
213 bh = bh->b_this_page;
214 } while (bh != head);
216 /* we might be here because some of the buffers on this page are
217 * not mapped. This is due to various races between
218 * file io on the block device and getblk. It gets dealt with
219 * elsewhere, don't buffer_error if we had some unmapped buffers
222 char b[BDEVNAME_SIZE];
224 printk("__find_get_block_slow() failed. "
225 "block=%llu, b_blocknr=%llu\n",
226 (unsigned long long)block,
227 (unsigned long long)bh->b_blocknr);
228 printk("b_state=0x%08lx, b_size=%zu\n",
229 bh->b_state, bh->b_size);
230 printk("device %s blocksize: %d\n", bdevname(bdev, b),
231 1 << bd_inode->i_blkbits);
234 spin_unlock(&bd_mapping->private_lock);
235 page_cache_release(page);
241 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
243 static void free_more_memory(void)
248 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
251 for_each_online_node(nid) {
252 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
253 gfp_zone(GFP_NOFS), NULL,
256 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
262 * I/O completion handler for block_read_full_page() - pages
263 * which come unlocked at the end of I/O.
265 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
268 struct buffer_head *first;
269 struct buffer_head *tmp;
271 int page_uptodate = 1;
273 BUG_ON(!buffer_async_read(bh));
277 set_buffer_uptodate(bh);
279 clear_buffer_uptodate(bh);
280 if (!quiet_error(bh))
286 * Be _very_ careful from here on. Bad things can happen if
287 * two buffer heads end IO at almost the same time and both
288 * decide that the page is now completely done.
290 first = page_buffers(page);
291 local_irq_save(flags);
292 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
293 clear_buffer_async_read(bh);
297 if (!buffer_uptodate(tmp))
299 if (buffer_async_read(tmp)) {
300 BUG_ON(!buffer_locked(tmp));
303 tmp = tmp->b_this_page;
305 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
306 local_irq_restore(flags);
309 * If none of the buffers had errors and they are all
310 * uptodate then we can set the page uptodate.
312 if (page_uptodate && !PageError(page))
313 SetPageUptodate(page);
318 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
319 local_irq_restore(flags);
324 * Completion handler for block_write_full_page() - pages which are unlocked
325 * during I/O, and which have PageWriteback cleared upon I/O completion.
327 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
329 char b[BDEVNAME_SIZE];
331 struct buffer_head *first;
332 struct buffer_head *tmp;
335 BUG_ON(!buffer_async_write(bh));
339 set_buffer_uptodate(bh);
341 if (!quiet_error(bh)) {
343 printk(KERN_WARNING "lost page write due to "
345 bdevname(bh->b_bdev, b));
347 set_bit(AS_EIO, &page->mapping->flags);
348 set_buffer_write_io_error(bh);
349 clear_buffer_uptodate(bh);
353 first = page_buffers(page);
354 local_irq_save(flags);
355 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
357 clear_buffer_async_write(bh);
359 tmp = bh->b_this_page;
361 if (buffer_async_write(tmp)) {
362 BUG_ON(!buffer_locked(tmp));
365 tmp = tmp->b_this_page;
367 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
368 local_irq_restore(flags);
369 end_page_writeback(page);
373 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
374 local_irq_restore(flags);
377 EXPORT_SYMBOL(end_buffer_async_write);
380 * If a page's buffers are under async readin (end_buffer_async_read
381 * completion) then there is a possibility that another thread of
382 * control could lock one of the buffers after it has completed
383 * but while some of the other buffers have not completed. This
384 * locked buffer would confuse end_buffer_async_read() into not unlocking
385 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
386 * that this buffer is not under async I/O.
388 * The page comes unlocked when it has no locked buffer_async buffers
391 * PageLocked prevents anyone starting new async I/O reads any of
394 * PageWriteback is used to prevent simultaneous writeout of the same
397 * PageLocked prevents anyone from starting writeback of a page which is
398 * under read I/O (PageWriteback is only ever set against a locked page).
400 static void mark_buffer_async_read(struct buffer_head *bh)
402 bh->b_end_io = end_buffer_async_read;
403 set_buffer_async_read(bh);
406 static void mark_buffer_async_write_endio(struct buffer_head *bh,
407 bh_end_io_t *handler)
409 bh->b_end_io = handler;
410 set_buffer_async_write(bh);
413 void mark_buffer_async_write(struct buffer_head *bh)
415 mark_buffer_async_write_endio(bh, end_buffer_async_write);
417 EXPORT_SYMBOL(mark_buffer_async_write);
421 * fs/buffer.c contains helper functions for buffer-backed address space's
422 * fsync functions. A common requirement for buffer-based filesystems is
423 * that certain data from the backing blockdev needs to be written out for
424 * a successful fsync(). For example, ext2 indirect blocks need to be
425 * written back and waited upon before fsync() returns.
427 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
428 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
429 * management of a list of dependent buffers at ->i_mapping->private_list.
431 * Locking is a little subtle: try_to_free_buffers() will remove buffers
432 * from their controlling inode's queue when they are being freed. But
433 * try_to_free_buffers() will be operating against the *blockdev* mapping
434 * at the time, not against the S_ISREG file which depends on those buffers.
435 * So the locking for private_list is via the private_lock in the address_space
436 * which backs the buffers. Which is different from the address_space
437 * against which the buffers are listed. So for a particular address_space,
438 * mapping->private_lock does *not* protect mapping->private_list! In fact,
439 * mapping->private_list will always be protected by the backing blockdev's
442 * Which introduces a requirement: all buffers on an address_space's
443 * ->private_list must be from the same address_space: the blockdev's.
445 * address_spaces which do not place buffers at ->private_list via these
446 * utility functions are free to use private_lock and private_list for
447 * whatever they want. The only requirement is that list_empty(private_list)
448 * be true at clear_inode() time.
450 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
451 * filesystems should do that. invalidate_inode_buffers() should just go
452 * BUG_ON(!list_empty).
454 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
455 * take an address_space, not an inode. And it should be called
456 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
459 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
460 * list if it is already on a list. Because if the buffer is on a list,
461 * it *must* already be on the right one. If not, the filesystem is being
462 * silly. This will save a ton of locking. But first we have to ensure
463 * that buffers are taken *off* the old inode's list when they are freed
464 * (presumably in truncate). That requires careful auditing of all
465 * filesystems (do it inside bforget()). It could also be done by bringing
470 * The buffer's backing address_space's private_lock must be held
472 static void __remove_assoc_queue(struct buffer_head *bh)
474 list_del_init(&bh->b_assoc_buffers);
475 WARN_ON(!bh->b_assoc_map);
476 if (buffer_write_io_error(bh))
477 set_bit(AS_EIO, &bh->b_assoc_map->flags);
478 bh->b_assoc_map = NULL;
481 int inode_has_buffers(struct inode *inode)
483 return !list_empty(&inode->i_data.private_list);
487 * osync is designed to support O_SYNC io. It waits synchronously for
488 * all already-submitted IO to complete, but does not queue any new
489 * writes to the disk.
491 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
492 * you dirty the buffers, and then use osync_inode_buffers to wait for
493 * completion. Any other dirty buffers which are not yet queued for
494 * write will not be flushed to disk by the osync.
496 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
498 struct buffer_head *bh;
504 list_for_each_prev(p, list) {
506 if (buffer_locked(bh)) {
510 if (!buffer_uptodate(bh))
521 static void do_thaw_one(struct super_block *sb, void *unused)
523 char b[BDEVNAME_SIZE];
524 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
525 printk(KERN_WARNING "Emergency Thaw on %s\n",
526 bdevname(sb->s_bdev, b));
529 static void do_thaw_all(struct work_struct *work)
531 iterate_supers(do_thaw_one, NULL);
533 printk(KERN_WARNING "Emergency Thaw complete\n");
537 * emergency_thaw_all -- forcibly thaw every frozen filesystem
539 * Used for emergency unfreeze of all filesystems via SysRq
541 void emergency_thaw_all(void)
543 struct work_struct *work;
545 work = kmalloc(sizeof(*work), GFP_ATOMIC);
547 INIT_WORK(work, do_thaw_all);
553 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
554 * @mapping: the mapping which wants those buffers written
556 * Starts I/O against the buffers at mapping->private_list, and waits upon
559 * Basically, this is a convenience function for fsync().
560 * @mapping is a file or directory which needs those buffers to be written for
561 * a successful fsync().
563 int sync_mapping_buffers(struct address_space *mapping)
565 struct address_space *buffer_mapping = mapping->private_data;
567 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
570 return fsync_buffers_list(&buffer_mapping->private_lock,
571 &mapping->private_list);
573 EXPORT_SYMBOL(sync_mapping_buffers);
576 * Called when we've recently written block `bblock', and it is known that
577 * `bblock' was for a buffer_boundary() buffer. This means that the block at
578 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
579 * dirty, schedule it for IO. So that indirects merge nicely with their data.
581 void write_boundary_block(struct block_device *bdev,
582 sector_t bblock, unsigned blocksize)
584 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
586 if (buffer_dirty(bh))
587 ll_rw_block(WRITE, 1, &bh);
592 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
594 struct address_space *mapping = inode->i_mapping;
595 struct address_space *buffer_mapping = bh->b_page->mapping;
597 mark_buffer_dirty(bh);
598 if (!mapping->private_data) {
599 mapping->private_data = buffer_mapping;
601 BUG_ON(mapping->private_data != buffer_mapping);
603 if (!bh->b_assoc_map) {
604 spin_lock(&buffer_mapping->private_lock);
605 list_move_tail(&bh->b_assoc_buffers,
606 &mapping->private_list);
607 bh->b_assoc_map = mapping;
608 spin_unlock(&buffer_mapping->private_lock);
611 EXPORT_SYMBOL(mark_buffer_dirty_inode);
614 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
617 * If warn is true, then emit a warning if the page is not uptodate and has
618 * not been truncated.
620 static void __set_page_dirty(struct page *page,
621 struct address_space *mapping, int warn)
625 spin_lock_irqsave(&mapping->tree_lock, flags);
626 if (page->mapping) { /* Race with truncate? */
627 WARN_ON_ONCE(warn && !PageUptodate(page));
628 account_page_dirtied(page, mapping);
629 radix_tree_tag_set(&mapping->page_tree,
630 page_index(page), PAGECACHE_TAG_DIRTY);
632 spin_unlock_irqrestore(&mapping->tree_lock, flags);
633 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
637 * Add a page to the dirty page list.
639 * It is a sad fact of life that this function is called from several places
640 * deeply under spinlocking. It may not sleep.
642 * If the page has buffers, the uptodate buffers are set dirty, to preserve
643 * dirty-state coherency between the page and the buffers. It the page does
644 * not have buffers then when they are later attached they will all be set
647 * The buffers are dirtied before the page is dirtied. There's a small race
648 * window in which a writepage caller may see the page cleanness but not the
649 * buffer dirtiness. That's fine. If this code were to set the page dirty
650 * before the buffers, a concurrent writepage caller could clear the page dirty
651 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
652 * page on the dirty page list.
654 * We use private_lock to lock against try_to_free_buffers while using the
655 * page's buffer list. Also use this to protect against clean buffers being
656 * added to the page after it was set dirty.
658 * FIXME: may need to call ->reservepage here as well. That's rather up to the
659 * address_space though.
661 int __set_page_dirty_buffers(struct page *page)
664 struct address_space *mapping = page_mapping(page);
666 if (unlikely(!mapping))
667 return !TestSetPageDirty(page);
669 spin_lock(&mapping->private_lock);
670 if (page_has_buffers(page)) {
671 struct buffer_head *head = page_buffers(page);
672 struct buffer_head *bh = head;
675 set_buffer_dirty(bh);
676 bh = bh->b_this_page;
677 } while (bh != head);
679 newly_dirty = !TestSetPageDirty(page);
680 spin_unlock(&mapping->private_lock);
683 __set_page_dirty(page, mapping, 1);
686 EXPORT_SYMBOL(__set_page_dirty_buffers);
689 * Write out and wait upon a list of buffers.
691 * We have conflicting pressures: we want to make sure that all
692 * initially dirty buffers get waited on, but that any subsequently
693 * dirtied buffers don't. After all, we don't want fsync to last
694 * forever if somebody is actively writing to the file.
696 * Do this in two main stages: first we copy dirty buffers to a
697 * temporary inode list, queueing the writes as we go. Then we clean
698 * up, waiting for those writes to complete.
700 * During this second stage, any subsequent updates to the file may end
701 * up refiling the buffer on the original inode's dirty list again, so
702 * there is a chance we will end up with a buffer queued for write but
703 * not yet completed on that list. So, as a final cleanup we go through
704 * the osync code to catch these locked, dirty buffers without requeuing
705 * any newly dirty buffers for write.
707 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
709 struct buffer_head *bh;
710 struct list_head tmp;
711 struct address_space *mapping;
713 struct blk_plug plug;
715 INIT_LIST_HEAD(&tmp);
716 blk_start_plug(&plug);
719 while (!list_empty(list)) {
720 bh = BH_ENTRY(list->next);
721 mapping = bh->b_assoc_map;
722 __remove_assoc_queue(bh);
723 /* Avoid race with mark_buffer_dirty_inode() which does
724 * a lockless check and we rely on seeing the dirty bit */
726 if (buffer_dirty(bh) || buffer_locked(bh)) {
727 list_add(&bh->b_assoc_buffers, &tmp);
728 bh->b_assoc_map = mapping;
729 if (buffer_dirty(bh)) {
733 * Ensure any pending I/O completes so that
734 * write_dirty_buffer() actually writes the
735 * current contents - it is a noop if I/O is
736 * still in flight on potentially older
739 write_dirty_buffer(bh, WRITE_SYNC);
742 * Kick off IO for the previous mapping. Note
743 * that we will not run the very last mapping,
744 * wait_on_buffer() will do that for us
745 * through sync_buffer().
754 blk_finish_plug(&plug);
757 while (!list_empty(&tmp)) {
758 bh = BH_ENTRY(tmp.prev);
760 mapping = bh->b_assoc_map;
761 __remove_assoc_queue(bh);
762 /* Avoid race with mark_buffer_dirty_inode() which does
763 * a lockless check and we rely on seeing the dirty bit */
765 if (buffer_dirty(bh)) {
766 list_add(&bh->b_assoc_buffers,
767 &mapping->private_list);
768 bh->b_assoc_map = mapping;
772 if (!buffer_uptodate(bh))
779 err2 = osync_buffers_list(lock, list);
787 * Invalidate any and all dirty buffers on a given inode. We are
788 * probably unmounting the fs, but that doesn't mean we have already
789 * done a sync(). Just drop the buffers from the inode list.
791 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
792 * assumes that all the buffers are against the blockdev. Not true
795 void invalidate_inode_buffers(struct inode *inode)
797 if (inode_has_buffers(inode)) {
798 struct address_space *mapping = &inode->i_data;
799 struct list_head *list = &mapping->private_list;
800 struct address_space *buffer_mapping = mapping->private_data;
802 spin_lock(&buffer_mapping->private_lock);
803 while (!list_empty(list))
804 __remove_assoc_queue(BH_ENTRY(list->next));
805 spin_unlock(&buffer_mapping->private_lock);
808 EXPORT_SYMBOL(invalidate_inode_buffers);
811 * Remove any clean buffers from the inode's buffer list. This is called
812 * when we're trying to free the inode itself. Those buffers can pin it.
814 * Returns true if all buffers were removed.
816 int remove_inode_buffers(struct inode *inode)
820 if (inode_has_buffers(inode)) {
821 struct address_space *mapping = &inode->i_data;
822 struct list_head *list = &mapping->private_list;
823 struct address_space *buffer_mapping = mapping->private_data;
825 spin_lock(&buffer_mapping->private_lock);
826 while (!list_empty(list)) {
827 struct buffer_head *bh = BH_ENTRY(list->next);
828 if (buffer_dirty(bh)) {
832 __remove_assoc_queue(bh);
834 spin_unlock(&buffer_mapping->private_lock);
840 * Create the appropriate buffers when given a page for data area and
841 * the size of each buffer.. Use the bh->b_this_page linked list to
842 * follow the buffers created. Return NULL if unable to create more
845 * The retry flag is used to differentiate async IO (paging, swapping)
846 * which may not fail from ordinary buffer allocations.
848 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
851 struct buffer_head *bh, *head;
857 while ((offset -= size) >= 0) {
858 bh = alloc_buffer_head(GFP_NOFS);
862 bh->b_this_page = head;
868 /* Link the buffer to its page */
869 set_bh_page(bh, page, offset);
873 * In case anything failed, we just free everything we got.
879 head = head->b_this_page;
880 free_buffer_head(bh);
885 * Return failure for non-async IO requests. Async IO requests
886 * are not allowed to fail, so we have to wait until buffer heads
887 * become available. But we don't want tasks sleeping with
888 * partially complete buffers, so all were released above.
893 /* We're _really_ low on memory. Now we just
894 * wait for old buffer heads to become free due to
895 * finishing IO. Since this is an async request and
896 * the reserve list is empty, we're sure there are
897 * async buffer heads in use.
902 EXPORT_SYMBOL_GPL(alloc_page_buffers);
905 link_dev_buffers(struct page *page, struct buffer_head *head)
907 struct buffer_head *bh, *tail;
912 bh = bh->b_this_page;
914 tail->b_this_page = head;
915 attach_page_buffers(page, head);
918 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
920 sector_t retval = ~((sector_t)0);
921 loff_t sz = i_size_read(bdev->bd_inode);
924 unsigned int sizebits = blksize_bits(size);
925 retval = (sz >> sizebits);
931 * Initialise the state of a blockdev page's buffers.
934 init_page_buffers(struct page *page, struct block_device *bdev,
935 sector_t block, int size)
937 struct buffer_head *head = page_buffers(page);
938 struct buffer_head *bh = head;
939 int uptodate = PageUptodate(page);
940 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
943 if (!buffer_mapped(bh)) {
944 init_buffer(bh, NULL, NULL);
946 bh->b_blocknr = block;
948 set_buffer_uptodate(bh);
949 if (block < end_block)
950 set_buffer_mapped(bh);
953 bh = bh->b_this_page;
954 } while (bh != head);
957 * Caller needs to validate requested block against end of device.
963 * Create the page-cache page that contains the requested block.
965 * This is used purely for blockdev mappings.
968 grow_dev_page(struct block_device *bdev, sector_t block,
969 pgoff_t index, int size, int sizebits)
971 struct inode *inode = bdev->bd_inode;
973 struct buffer_head *bh;
975 int ret = 0; /* Will call free_more_memory() */
977 page = find_or_create_page(inode->i_mapping, index,
978 (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
982 BUG_ON(!PageLocked(page));
984 if (page_has_buffers(page)) {
985 bh = page_buffers(page);
986 if (bh->b_size == size) {
987 end_block = init_page_buffers(page, bdev,
988 (sector_t)index << sizebits,
992 if (!try_to_free_buffers(page))
997 * Allocate some buffers for this page
999 bh = alloc_page_buffers(page, size, 0);
1004 * Link the page to the buffers and initialise them. Take the
1005 * lock to be atomic wrt __find_get_block(), which does not
1006 * run under the page lock.
1008 spin_lock(&inode->i_mapping->private_lock);
1009 link_dev_buffers(page, bh);
1010 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1012 spin_unlock(&inode->i_mapping->private_lock);
1014 ret = (block < end_block) ? 1 : -ENXIO;
1017 page_cache_release(page);
1022 * Create buffers for the specified block device block's page. If
1023 * that page was dirty, the buffers are set dirty also.
1026 grow_buffers(struct block_device *bdev, sector_t block, int size)
1034 } while ((size << sizebits) < PAGE_SIZE);
1036 index = block >> sizebits;
1039 * Check for a block which wants to lie outside our maximum possible
1040 * pagecache index. (this comparison is done using sector_t types).
1042 if (unlikely(index != block >> sizebits)) {
1043 char b[BDEVNAME_SIZE];
1045 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1047 __func__, (unsigned long long)block,
1052 /* Create a page with the proper size buffers.. */
1053 return grow_dev_page(bdev, block, index, size, sizebits);
1056 static struct buffer_head *
1057 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1059 /* Size must be multiple of hard sectorsize */
1060 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1061 (size < 512 || size > PAGE_SIZE))) {
1062 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1064 printk(KERN_ERR "logical block size: %d\n",
1065 bdev_logical_block_size(bdev));
1072 struct buffer_head *bh;
1075 bh = __find_get_block(bdev, block, size);
1079 ret = grow_buffers(bdev, block, size);
1088 * The relationship between dirty buffers and dirty pages:
1090 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1091 * the page is tagged dirty in its radix tree.
1093 * At all times, the dirtiness of the buffers represents the dirtiness of
1094 * subsections of the page. If the page has buffers, the page dirty bit is
1095 * merely a hint about the true dirty state.
1097 * When a page is set dirty in its entirety, all its buffers are marked dirty
1098 * (if the page has buffers).
1100 * When a buffer is marked dirty, its page is dirtied, but the page's other
1103 * Also. When blockdev buffers are explicitly read with bread(), they
1104 * individually become uptodate. But their backing page remains not
1105 * uptodate - even if all of its buffers are uptodate. A subsequent
1106 * block_read_full_page() against that page will discover all the uptodate
1107 * buffers, will set the page uptodate and will perform no I/O.
1111 * mark_buffer_dirty - mark a buffer_head as needing writeout
1112 * @bh: the buffer_head to mark dirty
1114 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1115 * backing page dirty, then tag the page as dirty in its address_space's radix
1116 * tree and then attach the address_space's inode to its superblock's dirty
1119 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1120 * mapping->tree_lock and mapping->host->i_lock.
1122 void mark_buffer_dirty(struct buffer_head *bh)
1124 WARN_ON_ONCE(!buffer_uptodate(bh));
1126 trace_block_dirty_buffer(bh);
1129 * Very *carefully* optimize the it-is-already-dirty case.
1131 * Don't let the final "is it dirty" escape to before we
1132 * perhaps modified the buffer.
1134 if (buffer_dirty(bh)) {
1136 if (buffer_dirty(bh))
1140 if (!test_set_buffer_dirty(bh)) {
1141 struct page *page = bh->b_page;
1142 if (!TestSetPageDirty(page)) {
1143 struct address_space *mapping = page_mapping(page);
1145 __set_page_dirty(page, mapping, 0);
1149 EXPORT_SYMBOL(mark_buffer_dirty);
1152 * Decrement a buffer_head's reference count. If all buffers against a page
1153 * have zero reference count, are clean and unlocked, and if the page is clean
1154 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1155 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1156 * a page but it ends up not being freed, and buffers may later be reattached).
1158 void __brelse(struct buffer_head * buf)
1160 if (atomic_read(&buf->b_count)) {
1164 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1166 EXPORT_SYMBOL(__brelse);
1169 * bforget() is like brelse(), except it discards any
1170 * potentially dirty data.
1172 void __bforget(struct buffer_head *bh)
1174 clear_buffer_dirty(bh);
1175 if (bh->b_assoc_map) {
1176 struct address_space *buffer_mapping = bh->b_page->mapping;
1178 spin_lock(&buffer_mapping->private_lock);
1179 list_del_init(&bh->b_assoc_buffers);
1180 bh->b_assoc_map = NULL;
1181 spin_unlock(&buffer_mapping->private_lock);
1185 EXPORT_SYMBOL(__bforget);
1187 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1190 if (buffer_uptodate(bh)) {
1195 bh->b_end_io = end_buffer_read_sync;
1196 submit_bh(READ, bh);
1198 if (buffer_uptodate(bh))
1206 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1207 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1208 * refcount elevated by one when they're in an LRU. A buffer can only appear
1209 * once in a particular CPU's LRU. A single buffer can be present in multiple
1210 * CPU's LRUs at the same time.
1212 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1213 * sb_find_get_block().
1215 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1216 * a local interrupt disable for that.
1219 #define BH_LRU_SIZE 8
1222 struct buffer_head *bhs[BH_LRU_SIZE];
1225 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1228 #define bh_lru_lock() local_irq_disable()
1229 #define bh_lru_unlock() local_irq_enable()
1231 #define bh_lru_lock() preempt_disable()
1232 #define bh_lru_unlock() preempt_enable()
1235 static inline void check_irqs_on(void)
1237 #ifdef irqs_disabled
1238 BUG_ON(irqs_disabled());
1243 * The LRU management algorithm is dopey-but-simple. Sorry.
1245 static void bh_lru_install(struct buffer_head *bh)
1247 struct buffer_head *evictee = NULL;
1251 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1252 struct buffer_head *bhs[BH_LRU_SIZE];
1258 for (in = 0; in < BH_LRU_SIZE; in++) {
1259 struct buffer_head *bh2 =
1260 __this_cpu_read(bh_lrus.bhs[in]);
1265 if (out >= BH_LRU_SIZE) {
1266 BUG_ON(evictee != NULL);
1273 while (out < BH_LRU_SIZE)
1275 memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1284 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1286 static struct buffer_head *
1287 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1289 struct buffer_head *ret = NULL;
1294 for (i = 0; i < BH_LRU_SIZE; i++) {
1295 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1297 if (bh && bh->b_bdev == bdev &&
1298 bh->b_blocknr == block && bh->b_size == size) {
1301 __this_cpu_write(bh_lrus.bhs[i],
1302 __this_cpu_read(bh_lrus.bhs[i - 1]));
1305 __this_cpu_write(bh_lrus.bhs[0], bh);
1317 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1318 * it in the LRU and mark it as accessed. If it is not present then return
1321 struct buffer_head *
1322 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1324 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1327 bh = __find_get_block_slow(bdev, block);
1335 EXPORT_SYMBOL(__find_get_block);
1338 * __getblk will locate (and, if necessary, create) the buffer_head
1339 * which corresponds to the passed block_device, block and size. The
1340 * returned buffer has its reference count incremented.
1342 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1343 * attempt is failing. FIXME, perhaps?
1345 struct buffer_head *
1346 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1348 struct buffer_head *bh = __find_get_block(bdev, block, size);
1352 bh = __getblk_slow(bdev, block, size);
1355 EXPORT_SYMBOL(__getblk);
1358 * Do async read-ahead on a buffer..
1360 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1362 struct buffer_head *bh = __getblk(bdev, block, size);
1364 ll_rw_block(READA, 1, &bh);
1368 EXPORT_SYMBOL(__breadahead);
1371 * __bread() - reads a specified block and returns the bh
1372 * @bdev: the block_device to read from
1373 * @block: number of block
1374 * @size: size (in bytes) to read
1376 * Reads a specified block, and returns buffer head that contains it.
1377 * It returns NULL if the block was unreadable.
1379 struct buffer_head *
1380 __bread(struct block_device *bdev, sector_t block, unsigned size)
1382 struct buffer_head *bh = __getblk(bdev, block, size);
1384 if (likely(bh) && !buffer_uptodate(bh))
1385 bh = __bread_slow(bh);
1388 EXPORT_SYMBOL(__bread);
1391 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1392 * This doesn't race because it runs in each cpu either in irq
1393 * or with preempt disabled.
1395 static void invalidate_bh_lru(void *arg)
1397 struct bh_lru *b = &get_cpu_var(bh_lrus);
1400 for (i = 0; i < BH_LRU_SIZE; i++) {
1404 put_cpu_var(bh_lrus);
1407 static bool has_bh_in_lru(int cpu, void *dummy)
1409 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1412 for (i = 0; i < BH_LRU_SIZE; i++) {
1420 void invalidate_bh_lrus(void)
1422 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1424 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1426 void set_bh_page(struct buffer_head *bh,
1427 struct page *page, unsigned long offset)
1430 BUG_ON(offset >= PAGE_SIZE);
1431 if (PageHighMem(page))
1433 * This catches illegal uses and preserves the offset:
1435 bh->b_data = (char *)(0 + offset);
1437 bh->b_data = page_address(page) + offset;
1439 EXPORT_SYMBOL(set_bh_page);
1442 * Called when truncating a buffer on a page completely.
1444 static void discard_buffer(struct buffer_head * bh)
1447 clear_buffer_dirty(bh);
1449 clear_buffer_mapped(bh);
1450 clear_buffer_req(bh);
1451 clear_buffer_new(bh);
1452 clear_buffer_delay(bh);
1453 clear_buffer_unwritten(bh);
1458 * block_invalidatepage - invalidate part or all of a buffer-backed page
1460 * @page: the page which is affected
1461 * @offset: the index of the truncation point
1463 * block_invalidatepage() is called when all or part of the page has become
1464 * invalidated by a truncate operation.
1466 * block_invalidatepage() does not have to release all buffers, but it must
1467 * ensure that no dirty buffer is left outside @offset and that no I/O
1468 * is underway against any of the blocks which are outside the truncation
1469 * point. Because the caller is about to free (and possibly reuse) those
1472 void block_invalidatepage(struct page *page, unsigned long offset)
1474 struct buffer_head *head, *bh, *next;
1475 unsigned int curr_off = 0;
1477 BUG_ON(!PageLocked(page));
1478 if (!page_has_buffers(page))
1481 head = page_buffers(page);
1484 unsigned int next_off = curr_off + bh->b_size;
1485 next = bh->b_this_page;
1488 * is this block fully invalidated?
1490 if (offset <= curr_off)
1492 curr_off = next_off;
1494 } while (bh != head);
1497 * We release buffers only if the entire page is being invalidated.
1498 * The get_block cached value has been unconditionally invalidated,
1499 * so real IO is not possible anymore.
1502 try_to_release_page(page, 0);
1506 EXPORT_SYMBOL(block_invalidatepage);
1509 * We attach and possibly dirty the buffers atomically wrt
1510 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1511 * is already excluded via the page lock.
1513 void create_empty_buffers(struct page *page,
1514 unsigned long blocksize, unsigned long b_state)
1516 struct buffer_head *bh, *head, *tail;
1518 head = alloc_page_buffers(page, blocksize, 1);
1521 bh->b_state |= b_state;
1523 bh = bh->b_this_page;
1525 tail->b_this_page = head;
1527 spin_lock(&page->mapping->private_lock);
1528 if (PageUptodate(page) || PageDirty(page)) {
1531 if (PageDirty(page))
1532 set_buffer_dirty(bh);
1533 if (PageUptodate(page))
1534 set_buffer_uptodate(bh);
1535 bh = bh->b_this_page;
1536 } while (bh != head);
1538 attach_page_buffers(page, head);
1539 spin_unlock(&page->mapping->private_lock);
1541 EXPORT_SYMBOL(create_empty_buffers);
1544 * We are taking a block for data and we don't want any output from any
1545 * buffer-cache aliases starting from return from that function and
1546 * until the moment when something will explicitly mark the buffer
1547 * dirty (hopefully that will not happen until we will free that block ;-)
1548 * We don't even need to mark it not-uptodate - nobody can expect
1549 * anything from a newly allocated buffer anyway. We used to used
1550 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1551 * don't want to mark the alias unmapped, for example - it would confuse
1552 * anyone who might pick it with bread() afterwards...
1554 * Also.. Note that bforget() doesn't lock the buffer. So there can
1555 * be writeout I/O going on against recently-freed buffers. We don't
1556 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1557 * only if we really need to. That happens here.
1559 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1561 struct buffer_head *old_bh;
1565 old_bh = __find_get_block_slow(bdev, block);
1567 clear_buffer_dirty(old_bh);
1568 wait_on_buffer(old_bh);
1569 clear_buffer_req(old_bh);
1573 EXPORT_SYMBOL(unmap_underlying_metadata);
1576 * Size is a power-of-two in the range 512..PAGE_SIZE,
1577 * and the case we care about most is PAGE_SIZE.
1579 * So this *could* possibly be written with those
1580 * constraints in mind (relevant mostly if some
1581 * architecture has a slow bit-scan instruction)
1583 static inline int block_size_bits(unsigned int blocksize)
1585 return ilog2(blocksize);
1588 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1590 BUG_ON(!PageLocked(page));
1592 if (!page_has_buffers(page))
1593 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1594 return page_buffers(page);
1598 * NOTE! All mapped/uptodate combinations are valid:
1600 * Mapped Uptodate Meaning
1602 * No No "unknown" - must do get_block()
1603 * No Yes "hole" - zero-filled
1604 * Yes No "allocated" - allocated on disk, not read in
1605 * Yes Yes "valid" - allocated and up-to-date in memory.
1607 * "Dirty" is valid only with the last case (mapped+uptodate).
1611 * While block_write_full_page is writing back the dirty buffers under
1612 * the page lock, whoever dirtied the buffers may decide to clean them
1613 * again at any time. We handle that by only looking at the buffer
1614 * state inside lock_buffer().
1616 * If block_write_full_page() is called for regular writeback
1617 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1618 * locked buffer. This only can happen if someone has written the buffer
1619 * directly, with submit_bh(). At the address_space level PageWriteback
1620 * prevents this contention from occurring.
1622 * If block_write_full_page() is called with wbc->sync_mode ==
1623 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1624 * causes the writes to be flagged as synchronous writes.
1626 static int __block_write_full_page(struct inode *inode, struct page *page,
1627 get_block_t *get_block, struct writeback_control *wbc,
1628 bh_end_io_t *handler)
1632 sector_t last_block;
1633 struct buffer_head *bh, *head;
1634 unsigned int blocksize, bbits;
1635 int nr_underway = 0;
1636 int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1637 WRITE_SYNC : WRITE);
1639 head = create_page_buffers(page, inode,
1640 (1 << BH_Dirty)|(1 << BH_Uptodate));
1643 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1644 * here, and the (potentially unmapped) buffers may become dirty at
1645 * any time. If a buffer becomes dirty here after we've inspected it
1646 * then we just miss that fact, and the page stays dirty.
1648 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1649 * handle that here by just cleaning them.
1653 blocksize = bh->b_size;
1654 bbits = block_size_bits(blocksize);
1656 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1657 last_block = (i_size_read(inode) - 1) >> bbits;
1660 * Get all the dirty buffers mapped to disk addresses and
1661 * handle any aliases from the underlying blockdev's mapping.
1664 if (block > last_block) {
1666 * mapped buffers outside i_size will occur, because
1667 * this page can be outside i_size when there is a
1668 * truncate in progress.
1671 * The buffer was zeroed by block_write_full_page()
1673 clear_buffer_dirty(bh);
1674 set_buffer_uptodate(bh);
1675 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1677 WARN_ON(bh->b_size != blocksize);
1678 err = get_block(inode, block, bh, 1);
1681 clear_buffer_delay(bh);
1682 if (buffer_new(bh)) {
1683 /* blockdev mappings never come here */
1684 clear_buffer_new(bh);
1685 unmap_underlying_metadata(bh->b_bdev,
1689 bh = bh->b_this_page;
1691 } while (bh != head);
1694 if (!buffer_mapped(bh))
1697 * If it's a fully non-blocking write attempt and we cannot
1698 * lock the buffer then redirty the page. Note that this can
1699 * potentially cause a busy-wait loop from writeback threads
1700 * and kswapd activity, but those code paths have their own
1701 * higher-level throttling.
1703 if (wbc->sync_mode != WB_SYNC_NONE) {
1705 } else if (!trylock_buffer(bh)) {
1706 redirty_page_for_writepage(wbc, page);
1709 if (test_clear_buffer_dirty(bh)) {
1710 mark_buffer_async_write_endio(bh, handler);
1714 } while ((bh = bh->b_this_page) != head);
1717 * The page and its buffers are protected by PageWriteback(), so we can
1718 * drop the bh refcounts early.
1720 BUG_ON(PageWriteback(page));
1721 set_page_writeback(page);
1724 struct buffer_head *next = bh->b_this_page;
1725 if (buffer_async_write(bh)) {
1726 submit_bh(write_op, bh);
1730 } while (bh != head);
1735 if (nr_underway == 0) {
1737 * The page was marked dirty, but the buffers were
1738 * clean. Someone wrote them back by hand with
1739 * ll_rw_block/submit_bh. A rare case.
1741 end_page_writeback(page);
1744 * The page and buffer_heads can be released at any time from
1752 * ENOSPC, or some other error. We may already have added some
1753 * blocks to the file, so we need to write these out to avoid
1754 * exposing stale data.
1755 * The page is currently locked and not marked for writeback
1758 /* Recovery: lock and submit the mapped buffers */
1760 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1761 !buffer_delay(bh)) {
1763 mark_buffer_async_write_endio(bh, handler);
1766 * The buffer may have been set dirty during
1767 * attachment to a dirty page.
1769 clear_buffer_dirty(bh);
1771 } while ((bh = bh->b_this_page) != head);
1773 BUG_ON(PageWriteback(page));
1774 mapping_set_error(page->mapping, err);
1775 set_page_writeback(page);
1777 struct buffer_head *next = bh->b_this_page;
1778 if (buffer_async_write(bh)) {
1779 clear_buffer_dirty(bh);
1780 submit_bh(write_op, bh);
1784 } while (bh != head);
1790 * If a page has any new buffers, zero them out here, and mark them uptodate
1791 * and dirty so they'll be written out (in order to prevent uninitialised
1792 * block data from leaking). And clear the new bit.
1794 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1796 unsigned int block_start, block_end;
1797 struct buffer_head *head, *bh;
1799 BUG_ON(!PageLocked(page));
1800 if (!page_has_buffers(page))
1803 bh = head = page_buffers(page);
1806 block_end = block_start + bh->b_size;
1808 if (buffer_new(bh)) {
1809 if (block_end > from && block_start < to) {
1810 if (!PageUptodate(page)) {
1811 unsigned start, size;
1813 start = max(from, block_start);
1814 size = min(to, block_end) - start;
1816 zero_user(page, start, size);
1817 set_buffer_uptodate(bh);
1820 clear_buffer_new(bh);
1821 mark_buffer_dirty(bh);
1825 block_start = block_end;
1826 bh = bh->b_this_page;
1827 } while (bh != head);
1829 EXPORT_SYMBOL(page_zero_new_buffers);
1831 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1832 get_block_t *get_block)
1834 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1835 unsigned to = from + len;
1836 struct inode *inode = page->mapping->host;
1837 unsigned block_start, block_end;
1840 unsigned blocksize, bbits;
1841 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1843 BUG_ON(!PageLocked(page));
1844 BUG_ON(from > PAGE_CACHE_SIZE);
1845 BUG_ON(to > PAGE_CACHE_SIZE);
1848 head = create_page_buffers(page, inode, 0);
1849 blocksize = head->b_size;
1850 bbits = block_size_bits(blocksize);
1852 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1854 for(bh = head, block_start = 0; bh != head || !block_start;
1855 block++, block_start=block_end, bh = bh->b_this_page) {
1856 block_end = block_start + blocksize;
1857 if (block_end <= from || block_start >= to) {
1858 if (PageUptodate(page)) {
1859 if (!buffer_uptodate(bh))
1860 set_buffer_uptodate(bh);
1865 clear_buffer_new(bh);
1866 if (!buffer_mapped(bh)) {
1867 WARN_ON(bh->b_size != blocksize);
1868 err = get_block(inode, block, bh, 1);
1871 if (buffer_new(bh)) {
1872 unmap_underlying_metadata(bh->b_bdev,
1874 if (PageUptodate(page)) {
1875 clear_buffer_new(bh);
1876 set_buffer_uptodate(bh);
1877 mark_buffer_dirty(bh);
1880 if (block_end > to || block_start < from)
1881 zero_user_segments(page,
1887 if (PageUptodate(page)) {
1888 if (!buffer_uptodate(bh))
1889 set_buffer_uptodate(bh);
1892 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1893 !buffer_unwritten(bh) &&
1894 (block_start < from || block_end > to)) {
1895 ll_rw_block(READ, 1, &bh);
1900 * If we issued read requests - let them complete.
1902 while(wait_bh > wait) {
1903 wait_on_buffer(*--wait_bh);
1904 if (!buffer_uptodate(*wait_bh))
1908 page_zero_new_buffers(page, from, to);
1911 EXPORT_SYMBOL(__block_write_begin);
1913 static int __block_commit_write(struct inode *inode, struct page *page,
1914 unsigned from, unsigned to)
1916 unsigned block_start, block_end;
1919 struct buffer_head *bh, *head;
1921 bh = head = page_buffers(page);
1922 blocksize = bh->b_size;
1926 block_end = block_start + blocksize;
1927 if (block_end <= from || block_start >= to) {
1928 if (!buffer_uptodate(bh))
1931 set_buffer_uptodate(bh);
1932 mark_buffer_dirty(bh);
1934 clear_buffer_new(bh);
1936 block_start = block_end;
1937 bh = bh->b_this_page;
1938 } while (bh != head);
1941 * If this is a partial write which happened to make all buffers
1942 * uptodate then we can optimize away a bogus readpage() for
1943 * the next read(). Here we 'discover' whether the page went
1944 * uptodate as a result of this (potentially partial) write.
1947 SetPageUptodate(page);
1952 * block_write_begin takes care of the basic task of block allocation and
1953 * bringing partial write blocks uptodate first.
1955 * The filesystem needs to handle block truncation upon failure.
1957 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1958 unsigned flags, struct page **pagep, get_block_t *get_block)
1960 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1964 page = grab_cache_page_write_begin(mapping, index, flags);
1968 status = __block_write_begin(page, pos, len, get_block);
1969 if (unlikely(status)) {
1971 page_cache_release(page);
1978 EXPORT_SYMBOL(block_write_begin);
1980 int block_write_end(struct file *file, struct address_space *mapping,
1981 loff_t pos, unsigned len, unsigned copied,
1982 struct page *page, void *fsdata)
1984 struct inode *inode = mapping->host;
1987 start = pos & (PAGE_CACHE_SIZE - 1);
1989 if (unlikely(copied < len)) {
1991 * The buffers that were written will now be uptodate, so we
1992 * don't have to worry about a readpage reading them and
1993 * overwriting a partial write. However if we have encountered
1994 * a short write and only partially written into a buffer, it
1995 * will not be marked uptodate, so a readpage might come in and
1996 * destroy our partial write.
1998 * Do the simplest thing, and just treat any short write to a
1999 * non uptodate page as a zero-length write, and force the
2000 * caller to redo the whole thing.
2002 if (!PageUptodate(page))
2005 page_zero_new_buffers(page, start+copied, start+len);
2007 flush_dcache_page(page);
2009 /* This could be a short (even 0-length) commit */
2010 __block_commit_write(inode, page, start, start+copied);
2014 EXPORT_SYMBOL(block_write_end);
2016 int generic_write_end(struct file *file, struct address_space *mapping,
2017 loff_t pos, unsigned len, unsigned copied,
2018 struct page *page, void *fsdata)
2020 struct inode *inode = mapping->host;
2021 loff_t old_size = inode->i_size;
2022 int i_size_changed = 0;
2024 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2027 * No need to use i_size_read() here, the i_size
2028 * cannot change under us because we hold i_mutex.
2030 * But it's important to update i_size while still holding page lock:
2031 * page writeout could otherwise come in and zero beyond i_size.
2033 if (pos+copied > inode->i_size) {
2034 i_size_write(inode, pos+copied);
2039 page_cache_release(page);
2042 pagecache_isize_extended(inode, old_size, pos);
2044 * Don't mark the inode dirty under page lock. First, it unnecessarily
2045 * makes the holding time of page lock longer. Second, it forces lock
2046 * ordering of page lock and transaction start for journaling
2050 mark_inode_dirty(inode);
2054 EXPORT_SYMBOL(generic_write_end);
2057 * block_is_partially_uptodate checks whether buffers within a page are
2060 * Returns true if all buffers which correspond to a file portion
2061 * we want to read are uptodate.
2063 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2066 unsigned block_start, block_end, blocksize;
2068 struct buffer_head *bh, *head;
2071 if (!page_has_buffers(page))
2074 head = page_buffers(page);
2075 blocksize = head->b_size;
2076 to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2078 if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2084 block_end = block_start + blocksize;
2085 if (block_end > from && block_start < to) {
2086 if (!buffer_uptodate(bh)) {
2090 if (block_end >= to)
2093 block_start = block_end;
2094 bh = bh->b_this_page;
2095 } while (bh != head);
2099 EXPORT_SYMBOL(block_is_partially_uptodate);
2102 * Generic "read page" function for block devices that have the normal
2103 * get_block functionality. This is most of the block device filesystems.
2104 * Reads the page asynchronously --- the unlock_buffer() and
2105 * set/clear_buffer_uptodate() functions propagate buffer state into the
2106 * page struct once IO has completed.
2108 int block_read_full_page(struct page *page, get_block_t *get_block)
2110 struct inode *inode = page->mapping->host;
2111 sector_t iblock, lblock;
2112 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2113 unsigned int blocksize, bbits;
2115 int fully_mapped = 1;
2117 head = create_page_buffers(page, inode, 0);
2118 blocksize = head->b_size;
2119 bbits = block_size_bits(blocksize);
2121 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2122 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2128 if (buffer_uptodate(bh))
2131 if (!buffer_mapped(bh)) {
2135 if (iblock < lblock) {
2136 WARN_ON(bh->b_size != blocksize);
2137 err = get_block(inode, iblock, bh, 0);
2141 if (!buffer_mapped(bh)) {
2142 zero_user(page, i * blocksize, blocksize);
2144 set_buffer_uptodate(bh);
2148 * get_block() might have updated the buffer
2151 if (buffer_uptodate(bh))
2155 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2158 SetPageMappedToDisk(page);
2162 * All buffers are uptodate - we can set the page uptodate
2163 * as well. But not if get_block() returned an error.
2165 if (!PageError(page))
2166 SetPageUptodate(page);
2171 /* Stage two: lock the buffers */
2172 for (i = 0; i < nr; i++) {
2175 mark_buffer_async_read(bh);
2179 * Stage 3: start the IO. Check for uptodateness
2180 * inside the buffer lock in case another process reading
2181 * the underlying blockdev brought it uptodate (the sct fix).
2183 for (i = 0; i < nr; i++) {
2185 if (buffer_uptodate(bh))
2186 end_buffer_async_read(bh, 1);
2188 submit_bh(READ, bh);
2192 EXPORT_SYMBOL(block_read_full_page);
2194 /* utility function for filesystems that need to do work on expanding
2195 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2196 * deal with the hole.
2198 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2200 struct address_space *mapping = inode->i_mapping;
2205 err = inode_newsize_ok(inode, size);
2209 err = pagecache_write_begin(NULL, mapping, size, 0,
2210 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2215 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2221 EXPORT_SYMBOL(generic_cont_expand_simple);
2223 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2224 loff_t pos, loff_t *bytes)
2226 struct inode *inode = mapping->host;
2227 unsigned blocksize = 1 << inode->i_blkbits;
2230 pgoff_t index, curidx;
2232 unsigned zerofrom, offset, len;
2235 index = pos >> PAGE_CACHE_SHIFT;
2236 offset = pos & ~PAGE_CACHE_MASK;
2238 while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2239 zerofrom = curpos & ~PAGE_CACHE_MASK;
2240 if (zerofrom & (blocksize-1)) {
2241 *bytes |= (blocksize-1);
2244 len = PAGE_CACHE_SIZE - zerofrom;
2246 err = pagecache_write_begin(file, mapping, curpos, len,
2247 AOP_FLAG_UNINTERRUPTIBLE,
2251 zero_user(page, zerofrom, len);
2252 err = pagecache_write_end(file, mapping, curpos, len, len,
2259 balance_dirty_pages_ratelimited(mapping);
2261 if (unlikely(fatal_signal_pending(current))) {
2267 /* page covers the boundary, find the boundary offset */
2268 if (index == curidx) {
2269 zerofrom = curpos & ~PAGE_CACHE_MASK;
2270 /* if we will expand the thing last block will be filled */
2271 if (offset <= zerofrom) {
2274 if (zerofrom & (blocksize-1)) {
2275 *bytes |= (blocksize-1);
2278 len = offset - zerofrom;
2280 err = pagecache_write_begin(file, mapping, curpos, len,
2281 AOP_FLAG_UNINTERRUPTIBLE,
2285 zero_user(page, zerofrom, len);
2286 err = pagecache_write_end(file, mapping, curpos, len, len,
2298 * For moronic filesystems that do not allow holes in file.
2299 * We may have to extend the file.
2301 int cont_write_begin(struct file *file, struct address_space *mapping,
2302 loff_t pos, unsigned len, unsigned flags,
2303 struct page **pagep, void **fsdata,
2304 get_block_t *get_block, loff_t *bytes)
2306 struct inode *inode = mapping->host;
2307 unsigned blocksize = 1 << inode->i_blkbits;
2311 err = cont_expand_zero(file, mapping, pos, bytes);
2315 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2316 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2317 *bytes |= (blocksize-1);
2321 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2323 EXPORT_SYMBOL(cont_write_begin);
2325 int block_commit_write(struct page *page, unsigned from, unsigned to)
2327 struct inode *inode = page->mapping->host;
2328 __block_commit_write(inode,page,from,to);
2331 EXPORT_SYMBOL(block_commit_write);
2334 * block_page_mkwrite() is not allowed to change the file size as it gets
2335 * called from a page fault handler when a page is first dirtied. Hence we must
2336 * be careful to check for EOF conditions here. We set the page up correctly
2337 * for a written page which means we get ENOSPC checking when writing into
2338 * holes and correct delalloc and unwritten extent mapping on filesystems that
2339 * support these features.
2341 * We are not allowed to take the i_mutex here so we have to play games to
2342 * protect against truncate races as the page could now be beyond EOF. Because
2343 * truncate writes the inode size before removing pages, once we have the
2344 * page lock we can determine safely if the page is beyond EOF. If it is not
2345 * beyond EOF, then the page is guaranteed safe against truncation until we
2348 * Direct callers of this function should protect against filesystem freezing
2349 * using sb_start_write() - sb_end_write() functions.
2351 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2352 get_block_t get_block)
2354 struct page *page = vmf->page;
2355 struct inode *inode = file_inode(vma->vm_file);
2361 size = i_size_read(inode);
2362 if ((page->mapping != inode->i_mapping) ||
2363 (page_offset(page) > size)) {
2364 /* We overload EFAULT to mean page got truncated */
2369 /* page is wholly or partially inside EOF */
2370 if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2371 end = size & ~PAGE_CACHE_MASK;
2373 end = PAGE_CACHE_SIZE;
2375 ret = __block_write_begin(page, 0, end, get_block);
2377 ret = block_commit_write(page, 0, end);
2379 if (unlikely(ret < 0))
2381 set_page_dirty(page);
2382 wait_for_stable_page(page);
2388 EXPORT_SYMBOL(__block_page_mkwrite);
2390 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2391 get_block_t get_block)
2394 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
2396 sb_start_pagefault(sb);
2399 * Update file times before taking page lock. We may end up failing the
2400 * fault so this update may be superfluous but who really cares...
2402 file_update_time(vma->vm_file);
2404 ret = __block_page_mkwrite(vma, vmf, get_block);
2405 sb_end_pagefault(sb);
2406 return block_page_mkwrite_return(ret);
2408 EXPORT_SYMBOL(block_page_mkwrite);
2411 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2412 * immediately, while under the page lock. So it needs a special end_io
2413 * handler which does not touch the bh after unlocking it.
2415 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2417 __end_buffer_read_notouch(bh, uptodate);
2421 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2422 * the page (converting it to circular linked list and taking care of page
2425 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2427 struct buffer_head *bh;
2429 BUG_ON(!PageLocked(page));
2431 spin_lock(&page->mapping->private_lock);
2434 if (PageDirty(page))
2435 set_buffer_dirty(bh);
2436 if (!bh->b_this_page)
2437 bh->b_this_page = head;
2438 bh = bh->b_this_page;
2439 } while (bh != head);
2440 attach_page_buffers(page, head);
2441 spin_unlock(&page->mapping->private_lock);
2445 * On entry, the page is fully not uptodate.
2446 * On exit the page is fully uptodate in the areas outside (from,to)
2447 * The filesystem needs to handle block truncation upon failure.
2449 int nobh_write_begin(struct address_space *mapping,
2450 loff_t pos, unsigned len, unsigned flags,
2451 struct page **pagep, void **fsdata,
2452 get_block_t *get_block)
2454 struct inode *inode = mapping->host;
2455 const unsigned blkbits = inode->i_blkbits;
2456 const unsigned blocksize = 1 << blkbits;
2457 struct buffer_head *head, *bh;
2461 unsigned block_in_page;
2462 unsigned block_start, block_end;
2463 sector_t block_in_file;
2466 int is_mapped_to_disk = 1;
2468 index = pos >> PAGE_CACHE_SHIFT;
2469 from = pos & (PAGE_CACHE_SIZE - 1);
2472 page = grab_cache_page_write_begin(mapping, index, flags);
2478 if (page_has_buffers(page)) {
2479 ret = __block_write_begin(page, pos, len, get_block);
2485 if (PageMappedToDisk(page))
2489 * Allocate buffers so that we can keep track of state, and potentially
2490 * attach them to the page if an error occurs. In the common case of
2491 * no error, they will just be freed again without ever being attached
2492 * to the page (which is all OK, because we're under the page lock).
2494 * Be careful: the buffer linked list is a NULL terminated one, rather
2495 * than the circular one we're used to.
2497 head = alloc_page_buffers(page, blocksize, 0);
2503 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2506 * We loop across all blocks in the page, whether or not they are
2507 * part of the affected region. This is so we can discover if the
2508 * page is fully mapped-to-disk.
2510 for (block_start = 0, block_in_page = 0, bh = head;
2511 block_start < PAGE_CACHE_SIZE;
2512 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2515 block_end = block_start + blocksize;
2518 if (block_start >= to)
2520 ret = get_block(inode, block_in_file + block_in_page,
2524 if (!buffer_mapped(bh))
2525 is_mapped_to_disk = 0;
2527 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2528 if (PageUptodate(page)) {
2529 set_buffer_uptodate(bh);
2532 if (buffer_new(bh) || !buffer_mapped(bh)) {
2533 zero_user_segments(page, block_start, from,
2537 if (buffer_uptodate(bh))
2538 continue; /* reiserfs does this */
2539 if (block_start < from || block_end > to) {
2541 bh->b_end_io = end_buffer_read_nobh;
2542 submit_bh(READ, bh);
2549 * The page is locked, so these buffers are protected from
2550 * any VM or truncate activity. Hence we don't need to care
2551 * for the buffer_head refcounts.
2553 for (bh = head; bh; bh = bh->b_this_page) {
2555 if (!buffer_uptodate(bh))
2562 if (is_mapped_to_disk)
2563 SetPageMappedToDisk(page);
2565 *fsdata = head; /* to be released by nobh_write_end */
2572 * Error recovery is a bit difficult. We need to zero out blocks that
2573 * were newly allocated, and dirty them to ensure they get written out.
2574 * Buffers need to be attached to the page at this point, otherwise
2575 * the handling of potential IO errors during writeout would be hard
2576 * (could try doing synchronous writeout, but what if that fails too?)
2578 attach_nobh_buffers(page, head);
2579 page_zero_new_buffers(page, from, to);
2583 page_cache_release(page);
2588 EXPORT_SYMBOL(nobh_write_begin);
2590 int nobh_write_end(struct file *file, struct address_space *mapping,
2591 loff_t pos, unsigned len, unsigned copied,
2592 struct page *page, void *fsdata)
2594 struct inode *inode = page->mapping->host;
2595 struct buffer_head *head = fsdata;
2596 struct buffer_head *bh;
2597 BUG_ON(fsdata != NULL && page_has_buffers(page));
2599 if (unlikely(copied < len) && head)
2600 attach_nobh_buffers(page, head);
2601 if (page_has_buffers(page))
2602 return generic_write_end(file, mapping, pos, len,
2603 copied, page, fsdata);
2605 SetPageUptodate(page);
2606 set_page_dirty(page);
2607 if (pos+copied > inode->i_size) {
2608 i_size_write(inode, pos+copied);
2609 mark_inode_dirty(inode);
2613 page_cache_release(page);
2617 head = head->b_this_page;
2618 free_buffer_head(bh);
2623 EXPORT_SYMBOL(nobh_write_end);
2626 * nobh_writepage() - based on block_full_write_page() except
2627 * that it tries to operate without attaching bufferheads to
2630 int nobh_writepage(struct page *page, get_block_t *get_block,
2631 struct writeback_control *wbc)
2633 struct inode * const inode = page->mapping->host;
2634 loff_t i_size = i_size_read(inode);
2635 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2639 /* Is the page fully inside i_size? */
2640 if (page->index < end_index)
2643 /* Is the page fully outside i_size? (truncate in progress) */
2644 offset = i_size & (PAGE_CACHE_SIZE-1);
2645 if (page->index >= end_index+1 || !offset) {
2647 * The page may have dirty, unmapped buffers. For example,
2648 * they may have been added in ext3_writepage(). Make them
2649 * freeable here, so the page does not leak.
2652 /* Not really sure about this - do we need this ? */
2653 if (page->mapping->a_ops->invalidatepage)
2654 page->mapping->a_ops->invalidatepage(page, offset);
2657 return 0; /* don't care */
2661 * The page straddles i_size. It must be zeroed out on each and every
2662 * writepage invocation because it may be mmapped. "A file is mapped
2663 * in multiples of the page size. For a file that is not a multiple of
2664 * the page size, the remaining memory is zeroed when mapped, and
2665 * writes to that region are not written out to the file."
2667 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2669 ret = mpage_writepage(page, get_block, wbc);
2671 ret = __block_write_full_page(inode, page, get_block, wbc,
2672 end_buffer_async_write);
2675 EXPORT_SYMBOL(nobh_writepage);
2677 int nobh_truncate_page(struct address_space *mapping,
2678 loff_t from, get_block_t *get_block)
2680 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2681 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2684 unsigned length, pos;
2685 struct inode *inode = mapping->host;
2687 struct buffer_head map_bh;
2690 blocksize = 1 << inode->i_blkbits;
2691 length = offset & (blocksize - 1);
2693 /* Block boundary? Nothing to do */
2697 length = blocksize - length;
2698 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2700 page = grab_cache_page(mapping, index);
2705 if (page_has_buffers(page)) {
2708 page_cache_release(page);
2709 return block_truncate_page(mapping, from, get_block);
2712 /* Find the buffer that contains "offset" */
2714 while (offset >= pos) {
2719 map_bh.b_size = blocksize;
2721 err = get_block(inode, iblock, &map_bh, 0);
2724 /* unmapped? It's a hole - nothing to do */
2725 if (!buffer_mapped(&map_bh))
2728 /* Ok, it's mapped. Make sure it's up-to-date */
2729 if (!PageUptodate(page)) {
2730 err = mapping->a_ops->readpage(NULL, page);
2732 page_cache_release(page);
2736 if (!PageUptodate(page)) {
2740 if (page_has_buffers(page))
2743 zero_user(page, offset, length);
2744 set_page_dirty(page);
2749 page_cache_release(page);
2753 EXPORT_SYMBOL(nobh_truncate_page);
2755 int block_truncate_page(struct address_space *mapping,
2756 loff_t from, get_block_t *get_block)
2758 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2759 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2762 unsigned length, pos;
2763 struct inode *inode = mapping->host;
2765 struct buffer_head *bh;
2768 blocksize = 1 << inode->i_blkbits;
2769 length = offset & (blocksize - 1);
2771 /* Block boundary? Nothing to do */
2775 length = blocksize - length;
2776 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2778 page = grab_cache_page(mapping, index);
2783 if (!page_has_buffers(page))
2784 create_empty_buffers(page, blocksize, 0);
2786 /* Find the buffer that contains "offset" */
2787 bh = page_buffers(page);
2789 while (offset >= pos) {
2790 bh = bh->b_this_page;
2796 if (!buffer_mapped(bh)) {
2797 WARN_ON(bh->b_size != blocksize);
2798 err = get_block(inode, iblock, bh, 0);
2801 /* unmapped? It's a hole - nothing to do */
2802 if (!buffer_mapped(bh))
2806 /* Ok, it's mapped. Make sure it's up-to-date */
2807 if (PageUptodate(page))
2808 set_buffer_uptodate(bh);
2810 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2812 ll_rw_block(READ, 1, &bh);
2814 /* Uhhuh. Read error. Complain and punt. */
2815 if (!buffer_uptodate(bh))
2819 zero_user(page, offset, length);
2820 mark_buffer_dirty(bh);
2825 page_cache_release(page);
2829 EXPORT_SYMBOL(block_truncate_page);
2832 * The generic ->writepage function for buffer-backed address_spaces
2833 * this form passes in the end_io handler used to finish the IO.
2835 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2836 struct writeback_control *wbc, bh_end_io_t *handler)
2838 struct inode * const inode = page->mapping->host;
2839 loff_t i_size = i_size_read(inode);
2840 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2843 /* Is the page fully inside i_size? */
2844 if (page->index < end_index)
2845 return __block_write_full_page(inode, page, get_block, wbc,
2848 /* Is the page fully outside i_size? (truncate in progress) */
2849 offset = i_size & (PAGE_CACHE_SIZE-1);
2850 if (page->index >= end_index+1 || !offset) {
2852 * The page may have dirty, unmapped buffers. For example,
2853 * they may have been added in ext3_writepage(). Make them
2854 * freeable here, so the page does not leak.
2856 do_invalidatepage(page, 0);
2858 return 0; /* don't care */
2862 * The page straddles i_size. It must be zeroed out on each and every
2863 * writepage invocation because it may be mmapped. "A file is mapped
2864 * in multiples of the page size. For a file that is not a multiple of
2865 * the page size, the remaining memory is zeroed when mapped, and
2866 * writes to that region are not written out to the file."
2868 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2869 return __block_write_full_page(inode, page, get_block, wbc, handler);
2871 EXPORT_SYMBOL(block_write_full_page_endio);
2874 * The generic ->writepage function for buffer-backed address_spaces
2876 int block_write_full_page(struct page *page, get_block_t *get_block,
2877 struct writeback_control *wbc)
2879 return block_write_full_page_endio(page, get_block, wbc,
2880 end_buffer_async_write);
2882 EXPORT_SYMBOL(block_write_full_page);
2884 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2885 get_block_t *get_block)
2887 struct buffer_head tmp;
2888 struct inode *inode = mapping->host;
2891 tmp.b_size = 1 << inode->i_blkbits;
2892 get_block(inode, block, &tmp, 0);
2893 return tmp.b_blocknr;
2895 EXPORT_SYMBOL(generic_block_bmap);
2897 static void end_bio_bh_io_sync(struct bio *bio, int err)
2899 struct buffer_head *bh = bio->bi_private;
2901 if (err == -EOPNOTSUPP) {
2902 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2905 if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2906 set_bit(BH_Quiet, &bh->b_state);
2908 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2913 * This allows us to do IO even on the odd last sectors
2914 * of a device, even if the bh block size is some multiple
2915 * of the physical sector size.
2917 * We'll just truncate the bio to the size of the device,
2918 * and clear the end of the buffer head manually.
2920 * Truly out-of-range accesses will turn into actual IO
2921 * errors, this only handles the "we need to be able to
2922 * do IO at the final sector" case.
2924 static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2929 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2934 * If the *whole* IO is past the end of the device,
2935 * let it through, and the IO layer will turn it into
2938 if (unlikely(bio->bi_sector >= maxsector))
2941 maxsector -= bio->bi_sector;
2942 bytes = bio->bi_size;
2943 if (likely((bytes >> 9) <= maxsector))
2946 /* Uhhuh. We've got a bh that straddles the device size! */
2947 bytes = maxsector << 9;
2949 /* Truncate the bio.. */
2950 bio->bi_size = bytes;
2951 bio->bi_io_vec[0].bv_len = bytes;
2953 /* ..and clear the end of the buffer for reads */
2954 if ((rw & RW_MASK) == READ) {
2955 void *kaddr = kmap_atomic(bh->b_page);
2956 memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
2957 kunmap_atomic(kaddr);
2958 flush_dcache_page(bh->b_page);
2962 int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
2967 BUG_ON(!buffer_locked(bh));
2968 BUG_ON(!buffer_mapped(bh));
2969 BUG_ON(!bh->b_end_io);
2970 BUG_ON(buffer_delay(bh));
2971 BUG_ON(buffer_unwritten(bh));
2974 * Only clear out a write error when rewriting
2976 if (test_set_buffer_req(bh) && (rw & WRITE))
2977 clear_buffer_write_io_error(bh);
2980 * from here on down, it's all bio -- do the initial mapping,
2981 * submit_bio -> generic_make_request may further map this bio around
2983 bio = bio_alloc(GFP_NOIO, 1);
2985 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2986 bio->bi_bdev = bh->b_bdev;
2987 bio->bi_io_vec[0].bv_page = bh->b_page;
2988 bio->bi_io_vec[0].bv_len = bh->b_size;
2989 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2992 bio->bi_size = bh->b_size;
2994 bio->bi_end_io = end_bio_bh_io_sync;
2995 bio->bi_private = bh;
2996 bio->bi_flags |= bio_flags;
2998 /* Take care of bh's that straddle the end of the device */
2999 guard_bh_eod(rw, bio, bh);
3001 if (buffer_meta(bh))
3003 if (buffer_prio(bh))
3007 submit_bio(rw, bio);
3009 if (bio_flagged(bio, BIO_EOPNOTSUPP))
3015 EXPORT_SYMBOL_GPL(_submit_bh);
3017 int submit_bh(int rw, struct buffer_head *bh)
3019 return _submit_bh(rw, bh, 0);
3021 EXPORT_SYMBOL(submit_bh);
3024 * ll_rw_block: low-level access to block devices (DEPRECATED)
3025 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3026 * @nr: number of &struct buffer_heads in the array
3027 * @bhs: array of pointers to &struct buffer_head
3029 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3030 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3031 * %READA option is described in the documentation for generic_make_request()
3032 * which ll_rw_block() calls.
3034 * This function drops any buffer that it cannot get a lock on (with the
3035 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3036 * request, and any buffer that appears to be up-to-date when doing read
3037 * request. Further it marks as clean buffers that are processed for
3038 * writing (the buffer cache won't assume that they are actually clean
3039 * until the buffer gets unlocked).
3041 * ll_rw_block sets b_end_io to simple completion handler that marks
3042 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3045 * All of the buffers must be for the same device, and must also be a
3046 * multiple of the current approved size for the device.
3048 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3052 for (i = 0; i < nr; i++) {
3053 struct buffer_head *bh = bhs[i];
3055 if (!trylock_buffer(bh))
3058 if (test_clear_buffer_dirty(bh)) {
3059 bh->b_end_io = end_buffer_write_sync;
3061 submit_bh(WRITE, bh);
3065 if (!buffer_uptodate(bh)) {
3066 bh->b_end_io = end_buffer_read_sync;
3075 EXPORT_SYMBOL(ll_rw_block);
3077 void write_dirty_buffer(struct buffer_head *bh, int rw)
3080 if (!test_clear_buffer_dirty(bh)) {
3084 bh->b_end_io = end_buffer_write_sync;
3088 EXPORT_SYMBOL(write_dirty_buffer);
3091 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3092 * and then start new I/O and then wait upon it. The caller must have a ref on
3095 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3099 WARN_ON(atomic_read(&bh->b_count) < 1);
3101 if (test_clear_buffer_dirty(bh)) {
3103 bh->b_end_io = end_buffer_write_sync;
3104 ret = submit_bh(rw, bh);
3106 if (!ret && !buffer_uptodate(bh))
3113 EXPORT_SYMBOL(__sync_dirty_buffer);
3115 int sync_dirty_buffer(struct buffer_head *bh)
3117 return __sync_dirty_buffer(bh, WRITE_SYNC);
3119 EXPORT_SYMBOL(sync_dirty_buffer);
3122 * try_to_free_buffers() checks if all the buffers on this particular page
3123 * are unused, and releases them if so.
3125 * Exclusion against try_to_free_buffers may be obtained by either
3126 * locking the page or by holding its mapping's private_lock.
3128 * If the page is dirty but all the buffers are clean then we need to
3129 * be sure to mark the page clean as well. This is because the page
3130 * may be against a block device, and a later reattachment of buffers
3131 * to a dirty page will set *all* buffers dirty. Which would corrupt
3132 * filesystem data on the same device.
3134 * The same applies to regular filesystem pages: if all the buffers are
3135 * clean then we set the page clean and proceed. To do that, we require
3136 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3139 * try_to_free_buffers() is non-blocking.
3141 static inline int buffer_busy(struct buffer_head *bh)
3143 return atomic_read(&bh->b_count) |
3144 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3148 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3150 struct buffer_head *head = page_buffers(page);
3151 struct buffer_head *bh;
3155 if (buffer_write_io_error(bh) && page->mapping)
3156 set_bit(AS_EIO, &page->mapping->flags);
3157 if (buffer_busy(bh))
3159 bh = bh->b_this_page;
3160 } while (bh != head);
3163 struct buffer_head *next = bh->b_this_page;
3165 if (bh->b_assoc_map)
3166 __remove_assoc_queue(bh);
3168 } while (bh != head);
3169 *buffers_to_free = head;
3170 __clear_page_buffers(page);
3176 int try_to_free_buffers(struct page *page)
3178 struct address_space * const mapping = page->mapping;
3179 struct buffer_head *buffers_to_free = NULL;
3182 BUG_ON(!PageLocked(page));
3183 if (PageWriteback(page))
3186 if (mapping == NULL) { /* can this still happen? */
3187 ret = drop_buffers(page, &buffers_to_free);
3191 spin_lock(&mapping->private_lock);
3192 ret = drop_buffers(page, &buffers_to_free);
3195 * If the filesystem writes its buffers by hand (eg ext3)
3196 * then we can have clean buffers against a dirty page. We
3197 * clean the page here; otherwise the VM will never notice
3198 * that the filesystem did any IO at all.
3200 * Also, during truncate, discard_buffer will have marked all
3201 * the page's buffers clean. We discover that here and clean
3204 * private_lock must be held over this entire operation in order
3205 * to synchronise against __set_page_dirty_buffers and prevent the
3206 * dirty bit from being lost.
3209 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3210 spin_unlock(&mapping->private_lock);
3212 if (buffers_to_free) {
3213 struct buffer_head *bh = buffers_to_free;
3216 struct buffer_head *next = bh->b_this_page;
3217 free_buffer_head(bh);
3219 } while (bh != buffers_to_free);
3223 EXPORT_SYMBOL(try_to_free_buffers);
3226 * There are no bdflush tunables left. But distributions are
3227 * still running obsolete flush daemons, so we terminate them here.
3229 * Use of bdflush() is deprecated and will be removed in a future kernel.
3230 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3232 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3234 static int msg_count;
3236 if (!capable(CAP_SYS_ADMIN))
3239 if (msg_count < 5) {
3242 "warning: process `%s' used the obsolete bdflush"
3243 " system call\n", current->comm);
3244 printk(KERN_INFO "Fix your initscripts?\n");
3253 * Buffer-head allocation
3255 static struct kmem_cache *bh_cachep __read_mostly;
3258 * Once the number of bh's in the machine exceeds this level, we start
3259 * stripping them in writeback.
3261 static unsigned long max_buffer_heads;
3263 int buffer_heads_over_limit;
3265 struct bh_accounting {
3266 int nr; /* Number of live bh's */
3267 int ratelimit; /* Limit cacheline bouncing */
3270 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3272 static void recalc_bh_state(void)
3277 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3279 __this_cpu_write(bh_accounting.ratelimit, 0);
3280 for_each_online_cpu(i)
3281 tot += per_cpu(bh_accounting, i).nr;
3282 buffer_heads_over_limit = (tot > max_buffer_heads);
3285 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3287 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3289 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3291 __this_cpu_inc(bh_accounting.nr);
3297 EXPORT_SYMBOL(alloc_buffer_head);
3299 void free_buffer_head(struct buffer_head *bh)
3301 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3302 kmem_cache_free(bh_cachep, bh);
3304 __this_cpu_dec(bh_accounting.nr);
3308 EXPORT_SYMBOL(free_buffer_head);
3310 static void buffer_exit_cpu(int cpu)
3313 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3315 for (i = 0; i < BH_LRU_SIZE; i++) {
3319 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3320 per_cpu(bh_accounting, cpu).nr = 0;
3323 static int buffer_cpu_notify(struct notifier_block *self,
3324 unsigned long action, void *hcpu)
3326 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3327 buffer_exit_cpu((unsigned long)hcpu);
3332 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3333 * @bh: struct buffer_head
3335 * Return true if the buffer is up-to-date and false,
3336 * with the buffer locked, if not.
3338 int bh_uptodate_or_lock(struct buffer_head *bh)
3340 if (!buffer_uptodate(bh)) {
3342 if (!buffer_uptodate(bh))
3348 EXPORT_SYMBOL(bh_uptodate_or_lock);
3351 * bh_submit_read - Submit a locked buffer for reading
3352 * @bh: struct buffer_head
3354 * Returns zero on success and -EIO on error.
3356 int bh_submit_read(struct buffer_head *bh)
3358 BUG_ON(!buffer_locked(bh));
3360 if (buffer_uptodate(bh)) {
3366 bh->b_end_io = end_buffer_read_sync;
3367 submit_bh(READ, bh);
3369 if (buffer_uptodate(bh))
3373 EXPORT_SYMBOL(bh_submit_read);
3375 void __init buffer_init(void)
3377 unsigned long nrpages;
3379 bh_cachep = kmem_cache_create("buffer_head",
3380 sizeof(struct buffer_head), 0,
3381 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3386 * Limit the bh occupancy to 10% of ZONE_NORMAL
3388 nrpages = (nr_free_buffer_pages() * 10) / 100;
3389 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3390 hotcpu_notifier(buffer_cpu_notify, 0);