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 * Returns if the page has dirty or writeback buffers. If all the buffers
87 * are unlocked and clean then the PageDirty information is stale. If
88 * any of the pages are locked, it is assumed they are locked for IO.
90 void buffer_check_dirty_writeback(struct page *page,
91 bool *dirty, bool *writeback)
93 struct buffer_head *head, *bh;
97 BUG_ON(!PageLocked(page));
99 if (!page_has_buffers(page))
102 if (PageWriteback(page))
105 head = page_buffers(page);
108 if (buffer_locked(bh))
111 if (buffer_dirty(bh))
114 bh = bh->b_this_page;
115 } while (bh != head);
117 EXPORT_SYMBOL(buffer_check_dirty_writeback);
120 * Block until a buffer comes unlocked. This doesn't stop it
121 * from becoming locked again - you have to lock it yourself
122 * if you want to preserve its state.
124 void __wait_on_buffer(struct buffer_head * bh)
126 wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
128 EXPORT_SYMBOL(__wait_on_buffer);
131 __clear_page_buffers(struct page *page)
133 ClearPagePrivate(page);
134 set_page_private(page, 0);
135 page_cache_release(page);
139 static int quiet_error(struct buffer_head *bh)
141 if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
147 static void buffer_io_error(struct buffer_head *bh)
149 char b[BDEVNAME_SIZE];
150 printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
151 bdevname(bh->b_bdev, b),
152 (unsigned long long)bh->b_blocknr);
156 * End-of-IO handler helper function which does not touch the bh after
158 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
159 * a race there is benign: unlock_buffer() only use the bh's address for
160 * hashing after unlocking the buffer, so it doesn't actually touch the bh
163 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
166 set_buffer_uptodate(bh);
168 /* This happens, due to failed READA attempts. */
169 clear_buffer_uptodate(bh);
175 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
176 * unlock the buffer. This is what ll_rw_block uses too.
178 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
180 __end_buffer_read_notouch(bh, uptodate);
183 EXPORT_SYMBOL(end_buffer_read_sync);
185 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
187 char b[BDEVNAME_SIZE];
190 set_buffer_uptodate(bh);
192 if (!quiet_error(bh)) {
194 printk(KERN_WARNING "lost page write due to "
196 bdevname(bh->b_bdev, b));
198 set_buffer_write_io_error(bh);
199 clear_buffer_uptodate(bh);
204 EXPORT_SYMBOL(end_buffer_write_sync);
207 * Various filesystems appear to want __find_get_block to be non-blocking.
208 * But it's the page lock which protects the buffers. To get around this,
209 * we get exclusion from try_to_free_buffers with the blockdev mapping's
212 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
213 * may be quite high. This code could TryLock the page, and if that
214 * succeeds, there is no need to take private_lock. (But if
215 * private_lock is contended then so is mapping->tree_lock).
217 static struct buffer_head *
218 __find_get_block_slow(struct block_device *bdev, sector_t block)
220 struct inode *bd_inode = bdev->bd_inode;
221 struct address_space *bd_mapping = bd_inode->i_mapping;
222 struct buffer_head *ret = NULL;
224 struct buffer_head *bh;
225 struct buffer_head *head;
229 index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
230 page = find_get_page(bd_mapping, index);
234 spin_lock(&bd_mapping->private_lock);
235 if (!page_has_buffers(page))
237 head = page_buffers(page);
240 if (!buffer_mapped(bh))
242 else if (bh->b_blocknr == block) {
247 bh = bh->b_this_page;
248 } while (bh != head);
250 /* we might be here because some of the buffers on this page are
251 * not mapped. This is due to various races between
252 * file io on the block device and getblk. It gets dealt with
253 * elsewhere, don't buffer_error if we had some unmapped buffers
256 char b[BDEVNAME_SIZE];
258 printk("__find_get_block_slow() failed. "
259 "block=%llu, b_blocknr=%llu\n",
260 (unsigned long long)block,
261 (unsigned long long)bh->b_blocknr);
262 printk("b_state=0x%08lx, b_size=%zu\n",
263 bh->b_state, bh->b_size);
264 printk("device %s blocksize: %d\n", bdevname(bdev, b),
265 1 << bd_inode->i_blkbits);
268 spin_unlock(&bd_mapping->private_lock);
269 page_cache_release(page);
275 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
277 static void free_more_memory(void)
282 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
285 for_each_online_node(nid) {
286 (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
287 gfp_zone(GFP_NOFS), NULL,
290 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
296 * I/O completion handler for block_read_full_page() - pages
297 * which come unlocked at the end of I/O.
299 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
302 struct buffer_head *first;
303 struct buffer_head *tmp;
305 int page_uptodate = 1;
307 BUG_ON(!buffer_async_read(bh));
311 set_buffer_uptodate(bh);
313 clear_buffer_uptodate(bh);
314 if (!quiet_error(bh))
320 * Be _very_ careful from here on. Bad things can happen if
321 * two buffer heads end IO at almost the same time and both
322 * decide that the page is now completely done.
324 first = page_buffers(page);
325 local_irq_save(flags);
326 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
327 clear_buffer_async_read(bh);
331 if (!buffer_uptodate(tmp))
333 if (buffer_async_read(tmp)) {
334 BUG_ON(!buffer_locked(tmp));
337 tmp = tmp->b_this_page;
339 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
340 local_irq_restore(flags);
343 * If none of the buffers had errors and they are all
344 * uptodate then we can set the page uptodate.
346 if (page_uptodate && !PageError(page))
347 SetPageUptodate(page);
352 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
353 local_irq_restore(flags);
358 * Completion handler for block_write_full_page() - pages which are unlocked
359 * during I/O, and which have PageWriteback cleared upon I/O completion.
361 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
363 char b[BDEVNAME_SIZE];
365 struct buffer_head *first;
366 struct buffer_head *tmp;
369 BUG_ON(!buffer_async_write(bh));
373 set_buffer_uptodate(bh);
375 if (!quiet_error(bh)) {
377 printk(KERN_WARNING "lost page write due to "
379 bdevname(bh->b_bdev, b));
381 set_bit(AS_EIO, &page->mapping->flags);
382 set_buffer_write_io_error(bh);
383 clear_buffer_uptodate(bh);
387 first = page_buffers(page);
388 local_irq_save(flags);
389 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
391 clear_buffer_async_write(bh);
393 tmp = bh->b_this_page;
395 if (buffer_async_write(tmp)) {
396 BUG_ON(!buffer_locked(tmp));
399 tmp = tmp->b_this_page;
401 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
402 local_irq_restore(flags);
403 end_page_writeback(page);
407 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
408 local_irq_restore(flags);
411 EXPORT_SYMBOL(end_buffer_async_write);
414 * If a page's buffers are under async readin (end_buffer_async_read
415 * completion) then there is a possibility that another thread of
416 * control could lock one of the buffers after it has completed
417 * but while some of the other buffers have not completed. This
418 * locked buffer would confuse end_buffer_async_read() into not unlocking
419 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
420 * that this buffer is not under async I/O.
422 * The page comes unlocked when it has no locked buffer_async buffers
425 * PageLocked prevents anyone starting new async I/O reads any of
428 * PageWriteback is used to prevent simultaneous writeout of the same
431 * PageLocked prevents anyone from starting writeback of a page which is
432 * under read I/O (PageWriteback is only ever set against a locked page).
434 static void mark_buffer_async_read(struct buffer_head *bh)
436 bh->b_end_io = end_buffer_async_read;
437 set_buffer_async_read(bh);
440 static void mark_buffer_async_write_endio(struct buffer_head *bh,
441 bh_end_io_t *handler)
443 bh->b_end_io = handler;
444 set_buffer_async_write(bh);
447 void mark_buffer_async_write(struct buffer_head *bh)
449 mark_buffer_async_write_endio(bh, end_buffer_async_write);
451 EXPORT_SYMBOL(mark_buffer_async_write);
455 * fs/buffer.c contains helper functions for buffer-backed address space's
456 * fsync functions. A common requirement for buffer-based filesystems is
457 * that certain data from the backing blockdev needs to be written out for
458 * a successful fsync(). For example, ext2 indirect blocks need to be
459 * written back and waited upon before fsync() returns.
461 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
462 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
463 * management of a list of dependent buffers at ->i_mapping->private_list.
465 * Locking is a little subtle: try_to_free_buffers() will remove buffers
466 * from their controlling inode's queue when they are being freed. But
467 * try_to_free_buffers() will be operating against the *blockdev* mapping
468 * at the time, not against the S_ISREG file which depends on those buffers.
469 * So the locking for private_list is via the private_lock in the address_space
470 * which backs the buffers. Which is different from the address_space
471 * against which the buffers are listed. So for a particular address_space,
472 * mapping->private_lock does *not* protect mapping->private_list! In fact,
473 * mapping->private_list will always be protected by the backing blockdev's
476 * Which introduces a requirement: all buffers on an address_space's
477 * ->private_list must be from the same address_space: the blockdev's.
479 * address_spaces which do not place buffers at ->private_list via these
480 * utility functions are free to use private_lock and private_list for
481 * whatever they want. The only requirement is that list_empty(private_list)
482 * be true at clear_inode() time.
484 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
485 * filesystems should do that. invalidate_inode_buffers() should just go
486 * BUG_ON(!list_empty).
488 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
489 * take an address_space, not an inode. And it should be called
490 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
493 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
494 * list if it is already on a list. Because if the buffer is on a list,
495 * it *must* already be on the right one. If not, the filesystem is being
496 * silly. This will save a ton of locking. But first we have to ensure
497 * that buffers are taken *off* the old inode's list when they are freed
498 * (presumably in truncate). That requires careful auditing of all
499 * filesystems (do it inside bforget()). It could also be done by bringing
504 * The buffer's backing address_space's private_lock must be held
506 static void __remove_assoc_queue(struct buffer_head *bh)
508 list_del_init(&bh->b_assoc_buffers);
509 WARN_ON(!bh->b_assoc_map);
510 if (buffer_write_io_error(bh))
511 set_bit(AS_EIO, &bh->b_assoc_map->flags);
512 bh->b_assoc_map = NULL;
515 int inode_has_buffers(struct inode *inode)
517 return !list_empty(&inode->i_data.private_list);
521 * osync is designed to support O_SYNC io. It waits synchronously for
522 * all already-submitted IO to complete, but does not queue any new
523 * writes to the disk.
525 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
526 * you dirty the buffers, and then use osync_inode_buffers to wait for
527 * completion. Any other dirty buffers which are not yet queued for
528 * write will not be flushed to disk by the osync.
530 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
532 struct buffer_head *bh;
538 list_for_each_prev(p, list) {
540 if (buffer_locked(bh)) {
544 if (!buffer_uptodate(bh))
555 static void do_thaw_one(struct super_block *sb, void *unused)
557 char b[BDEVNAME_SIZE];
558 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
559 printk(KERN_WARNING "Emergency Thaw on %s\n",
560 bdevname(sb->s_bdev, b));
563 static void do_thaw_all(struct work_struct *work)
565 iterate_supers(do_thaw_one, NULL);
567 printk(KERN_WARNING "Emergency Thaw complete\n");
571 * emergency_thaw_all -- forcibly thaw every frozen filesystem
573 * Used for emergency unfreeze of all filesystems via SysRq
575 void emergency_thaw_all(void)
577 struct work_struct *work;
579 work = kmalloc(sizeof(*work), GFP_ATOMIC);
581 INIT_WORK(work, do_thaw_all);
587 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
588 * @mapping: the mapping which wants those buffers written
590 * Starts I/O against the buffers at mapping->private_list, and waits upon
593 * Basically, this is a convenience function for fsync().
594 * @mapping is a file or directory which needs those buffers to be written for
595 * a successful fsync().
597 int sync_mapping_buffers(struct address_space *mapping)
599 struct address_space *buffer_mapping = mapping->private_data;
601 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
604 return fsync_buffers_list(&buffer_mapping->private_lock,
605 &mapping->private_list);
607 EXPORT_SYMBOL(sync_mapping_buffers);
610 * Called when we've recently written block `bblock', and it is known that
611 * `bblock' was for a buffer_boundary() buffer. This means that the block at
612 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
613 * dirty, schedule it for IO. So that indirects merge nicely with their data.
615 void write_boundary_block(struct block_device *bdev,
616 sector_t bblock, unsigned blocksize)
618 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
620 if (buffer_dirty(bh))
621 ll_rw_block(WRITE, 1, &bh);
626 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
628 struct address_space *mapping = inode->i_mapping;
629 struct address_space *buffer_mapping = bh->b_page->mapping;
631 mark_buffer_dirty(bh);
632 if (!mapping->private_data) {
633 mapping->private_data = buffer_mapping;
635 BUG_ON(mapping->private_data != buffer_mapping);
637 if (!bh->b_assoc_map) {
638 spin_lock(&buffer_mapping->private_lock);
639 list_move_tail(&bh->b_assoc_buffers,
640 &mapping->private_list);
641 bh->b_assoc_map = mapping;
642 spin_unlock(&buffer_mapping->private_lock);
645 EXPORT_SYMBOL(mark_buffer_dirty_inode);
648 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
651 * If warn is true, then emit a warning if the page is not uptodate and has
652 * not been truncated.
654 static void __set_page_dirty(struct page *page,
655 struct address_space *mapping, int warn)
659 spin_lock_irqsave(&mapping->tree_lock, flags);
660 if (page->mapping) { /* Race with truncate? */
661 WARN_ON_ONCE(warn && !PageUptodate(page));
662 account_page_dirtied(page, mapping);
663 radix_tree_tag_set(&mapping->page_tree,
664 page_index(page), PAGECACHE_TAG_DIRTY);
666 spin_unlock_irqrestore(&mapping->tree_lock, flags);
667 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
671 * Add a page to the dirty page list.
673 * It is a sad fact of life that this function is called from several places
674 * deeply under spinlocking. It may not sleep.
676 * If the page has buffers, the uptodate buffers are set dirty, to preserve
677 * dirty-state coherency between the page and the buffers. It the page does
678 * not have buffers then when they are later attached they will all be set
681 * The buffers are dirtied before the page is dirtied. There's a small race
682 * window in which a writepage caller may see the page cleanness but not the
683 * buffer dirtiness. That's fine. If this code were to set the page dirty
684 * before the buffers, a concurrent writepage caller could clear the page dirty
685 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
686 * page on the dirty page list.
688 * We use private_lock to lock against try_to_free_buffers while using the
689 * page's buffer list. Also use this to protect against clean buffers being
690 * added to the page after it was set dirty.
692 * FIXME: may need to call ->reservepage here as well. That's rather up to the
693 * address_space though.
695 int __set_page_dirty_buffers(struct page *page)
698 struct address_space *mapping = page_mapping(page);
700 if (unlikely(!mapping))
701 return !TestSetPageDirty(page);
703 spin_lock(&mapping->private_lock);
704 if (page_has_buffers(page)) {
705 struct buffer_head *head = page_buffers(page);
706 struct buffer_head *bh = head;
709 set_buffer_dirty(bh);
710 bh = bh->b_this_page;
711 } while (bh != head);
713 newly_dirty = !TestSetPageDirty(page);
714 spin_unlock(&mapping->private_lock);
717 __set_page_dirty(page, mapping, 1);
720 EXPORT_SYMBOL(__set_page_dirty_buffers);
723 * Write out and wait upon a list of buffers.
725 * We have conflicting pressures: we want to make sure that all
726 * initially dirty buffers get waited on, but that any subsequently
727 * dirtied buffers don't. After all, we don't want fsync to last
728 * forever if somebody is actively writing to the file.
730 * Do this in two main stages: first we copy dirty buffers to a
731 * temporary inode list, queueing the writes as we go. Then we clean
732 * up, waiting for those writes to complete.
734 * During this second stage, any subsequent updates to the file may end
735 * up refiling the buffer on the original inode's dirty list again, so
736 * there is a chance we will end up with a buffer queued for write but
737 * not yet completed on that list. So, as a final cleanup we go through
738 * the osync code to catch these locked, dirty buffers without requeuing
739 * any newly dirty buffers for write.
741 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
743 struct buffer_head *bh;
744 struct list_head tmp;
745 struct address_space *mapping;
747 struct blk_plug plug;
749 INIT_LIST_HEAD(&tmp);
750 blk_start_plug(&plug);
753 while (!list_empty(list)) {
754 bh = BH_ENTRY(list->next);
755 mapping = bh->b_assoc_map;
756 __remove_assoc_queue(bh);
757 /* Avoid race with mark_buffer_dirty_inode() which does
758 * a lockless check and we rely on seeing the dirty bit */
760 if (buffer_dirty(bh) || buffer_locked(bh)) {
761 list_add(&bh->b_assoc_buffers, &tmp);
762 bh->b_assoc_map = mapping;
763 if (buffer_dirty(bh)) {
767 * Ensure any pending I/O completes so that
768 * write_dirty_buffer() actually writes the
769 * current contents - it is a noop if I/O is
770 * still in flight on potentially older
773 write_dirty_buffer(bh, WRITE_SYNC);
776 * Kick off IO for the previous mapping. Note
777 * that we will not run the very last mapping,
778 * wait_on_buffer() will do that for us
779 * through sync_buffer().
788 blk_finish_plug(&plug);
791 while (!list_empty(&tmp)) {
792 bh = BH_ENTRY(tmp.prev);
794 mapping = bh->b_assoc_map;
795 __remove_assoc_queue(bh);
796 /* Avoid race with mark_buffer_dirty_inode() which does
797 * a lockless check and we rely on seeing the dirty bit */
799 if (buffer_dirty(bh)) {
800 list_add(&bh->b_assoc_buffers,
801 &mapping->private_list);
802 bh->b_assoc_map = mapping;
806 if (!buffer_uptodate(bh))
813 err2 = osync_buffers_list(lock, list);
821 * Invalidate any and all dirty buffers on a given inode. We are
822 * probably unmounting the fs, but that doesn't mean we have already
823 * done a sync(). Just drop the buffers from the inode list.
825 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
826 * assumes that all the buffers are against the blockdev. Not true
829 void invalidate_inode_buffers(struct inode *inode)
831 if (inode_has_buffers(inode)) {
832 struct address_space *mapping = &inode->i_data;
833 struct list_head *list = &mapping->private_list;
834 struct address_space *buffer_mapping = mapping->private_data;
836 spin_lock(&buffer_mapping->private_lock);
837 while (!list_empty(list))
838 __remove_assoc_queue(BH_ENTRY(list->next));
839 spin_unlock(&buffer_mapping->private_lock);
842 EXPORT_SYMBOL(invalidate_inode_buffers);
845 * Remove any clean buffers from the inode's buffer list. This is called
846 * when we're trying to free the inode itself. Those buffers can pin it.
848 * Returns true if all buffers were removed.
850 int remove_inode_buffers(struct inode *inode)
854 if (inode_has_buffers(inode)) {
855 struct address_space *mapping = &inode->i_data;
856 struct list_head *list = &mapping->private_list;
857 struct address_space *buffer_mapping = mapping->private_data;
859 spin_lock(&buffer_mapping->private_lock);
860 while (!list_empty(list)) {
861 struct buffer_head *bh = BH_ENTRY(list->next);
862 if (buffer_dirty(bh)) {
866 __remove_assoc_queue(bh);
868 spin_unlock(&buffer_mapping->private_lock);
874 * Create the appropriate buffers when given a page for data area and
875 * the size of each buffer.. Use the bh->b_this_page linked list to
876 * follow the buffers created. Return NULL if unable to create more
879 * The retry flag is used to differentiate async IO (paging, swapping)
880 * which may not fail from ordinary buffer allocations.
882 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
885 struct buffer_head *bh, *head;
891 while ((offset -= size) >= 0) {
892 bh = alloc_buffer_head(GFP_NOFS);
896 bh->b_this_page = head;
902 /* Link the buffer to its page */
903 set_bh_page(bh, page, offset);
907 * In case anything failed, we just free everything we got.
913 head = head->b_this_page;
914 free_buffer_head(bh);
919 * Return failure for non-async IO requests. Async IO requests
920 * are not allowed to fail, so we have to wait until buffer heads
921 * become available. But we don't want tasks sleeping with
922 * partially complete buffers, so all were released above.
927 /* We're _really_ low on memory. Now we just
928 * wait for old buffer heads to become free due to
929 * finishing IO. Since this is an async request and
930 * the reserve list is empty, we're sure there are
931 * async buffer heads in use.
936 EXPORT_SYMBOL_GPL(alloc_page_buffers);
939 link_dev_buffers(struct page *page, struct buffer_head *head)
941 struct buffer_head *bh, *tail;
946 bh = bh->b_this_page;
948 tail->b_this_page = head;
949 attach_page_buffers(page, head);
952 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
954 sector_t retval = ~((sector_t)0);
955 loff_t sz = i_size_read(bdev->bd_inode);
958 unsigned int sizebits = blksize_bits(size);
959 retval = (sz >> sizebits);
965 * Initialise the state of a blockdev page's buffers.
968 init_page_buffers(struct page *page, struct block_device *bdev,
969 sector_t block, int size)
971 struct buffer_head *head = page_buffers(page);
972 struct buffer_head *bh = head;
973 int uptodate = PageUptodate(page);
974 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
977 if (!buffer_mapped(bh)) {
978 init_buffer(bh, NULL, NULL);
980 bh->b_blocknr = block;
982 set_buffer_uptodate(bh);
983 if (block < end_block)
984 set_buffer_mapped(bh);
987 bh = bh->b_this_page;
988 } while (bh != head);
991 * Caller needs to validate requested block against end of device.
997 * Create the page-cache page that contains the requested block.
999 * This is used purely for blockdev mappings.
1002 grow_dev_page(struct block_device *bdev, sector_t block,
1003 pgoff_t index, int size, int sizebits)
1005 struct inode *inode = bdev->bd_inode;
1007 struct buffer_head *bh;
1009 int ret = 0; /* Will call free_more_memory() */
1012 gfp_mask = mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS;
1013 gfp_mask |= __GFP_MOVABLE;
1015 * XXX: __getblk_slow() can not really deal with failure and
1016 * will endlessly loop on improvised global reclaim. Prefer
1017 * looping in the allocator rather than here, at least that
1018 * code knows what it's doing.
1020 gfp_mask |= __GFP_NOFAIL;
1022 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1026 BUG_ON(!PageLocked(page));
1028 if (page_has_buffers(page)) {
1029 bh = page_buffers(page);
1030 if (bh->b_size == size) {
1031 end_block = init_page_buffers(page, bdev,
1032 (sector_t)index << sizebits,
1036 if (!try_to_free_buffers(page))
1041 * Allocate some buffers for this page
1043 bh = alloc_page_buffers(page, size, 0);
1048 * Link the page to the buffers and initialise them. Take the
1049 * lock to be atomic wrt __find_get_block(), which does not
1050 * run under the page lock.
1052 spin_lock(&inode->i_mapping->private_lock);
1053 link_dev_buffers(page, bh);
1054 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1056 spin_unlock(&inode->i_mapping->private_lock);
1058 ret = (block < end_block) ? 1 : -ENXIO;
1061 page_cache_release(page);
1066 * Create buffers for the specified block device block's page. If
1067 * that page was dirty, the buffers are set dirty also.
1070 grow_buffers(struct block_device *bdev, sector_t block, int size)
1078 } while ((size << sizebits) < PAGE_SIZE);
1080 index = block >> sizebits;
1083 * Check for a block which wants to lie outside our maximum possible
1084 * pagecache index. (this comparison is done using sector_t types).
1086 if (unlikely(index != block >> sizebits)) {
1087 char b[BDEVNAME_SIZE];
1089 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1091 __func__, (unsigned long long)block,
1096 /* Create a page with the proper size buffers.. */
1097 return grow_dev_page(bdev, block, index, size, sizebits);
1100 static struct buffer_head *
1101 __getblk_slow(struct block_device *bdev, sector_t block, int size)
1103 /* Size must be multiple of hard sectorsize */
1104 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1105 (size < 512 || size > PAGE_SIZE))) {
1106 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1108 printk(KERN_ERR "logical block size: %d\n",
1109 bdev_logical_block_size(bdev));
1116 struct buffer_head *bh;
1119 bh = __find_get_block(bdev, block, size);
1123 ret = grow_buffers(bdev, block, size);
1132 * The relationship between dirty buffers and dirty pages:
1134 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1135 * the page is tagged dirty in its radix tree.
1137 * At all times, the dirtiness of the buffers represents the dirtiness of
1138 * subsections of the page. If the page has buffers, the page dirty bit is
1139 * merely a hint about the true dirty state.
1141 * When a page is set dirty in its entirety, all its buffers are marked dirty
1142 * (if the page has buffers).
1144 * When a buffer is marked dirty, its page is dirtied, but the page's other
1147 * Also. When blockdev buffers are explicitly read with bread(), they
1148 * individually become uptodate. But their backing page remains not
1149 * uptodate - even if all of its buffers are uptodate. A subsequent
1150 * block_read_full_page() against that page will discover all the uptodate
1151 * buffers, will set the page uptodate and will perform no I/O.
1155 * mark_buffer_dirty - mark a buffer_head as needing writeout
1156 * @bh: the buffer_head to mark dirty
1158 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1159 * backing page dirty, then tag the page as dirty in its address_space's radix
1160 * tree and then attach the address_space's inode to its superblock's dirty
1163 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1164 * mapping->tree_lock and mapping->host->i_lock.
1166 void mark_buffer_dirty(struct buffer_head *bh)
1168 WARN_ON_ONCE(!buffer_uptodate(bh));
1170 trace_block_dirty_buffer(bh);
1173 * Very *carefully* optimize the it-is-already-dirty case.
1175 * Don't let the final "is it dirty" escape to before we
1176 * perhaps modified the buffer.
1178 if (buffer_dirty(bh)) {
1180 if (buffer_dirty(bh))
1184 if (!test_set_buffer_dirty(bh)) {
1185 struct page *page = bh->b_page;
1186 if (!TestSetPageDirty(page)) {
1187 struct address_space *mapping = page_mapping(page);
1189 __set_page_dirty(page, mapping, 0);
1193 EXPORT_SYMBOL(mark_buffer_dirty);
1196 * Decrement a buffer_head's reference count. If all buffers against a page
1197 * have zero reference count, are clean and unlocked, and if the page is clean
1198 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1199 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1200 * a page but it ends up not being freed, and buffers may later be reattached).
1202 void __brelse(struct buffer_head * buf)
1204 if (atomic_read(&buf->b_count)) {
1208 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1210 EXPORT_SYMBOL(__brelse);
1213 * bforget() is like brelse(), except it discards any
1214 * potentially dirty data.
1216 void __bforget(struct buffer_head *bh)
1218 clear_buffer_dirty(bh);
1219 if (bh->b_assoc_map) {
1220 struct address_space *buffer_mapping = bh->b_page->mapping;
1222 spin_lock(&buffer_mapping->private_lock);
1223 list_del_init(&bh->b_assoc_buffers);
1224 bh->b_assoc_map = NULL;
1225 spin_unlock(&buffer_mapping->private_lock);
1229 EXPORT_SYMBOL(__bforget);
1231 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1234 if (buffer_uptodate(bh)) {
1239 bh->b_end_io = end_buffer_read_sync;
1240 submit_bh(READ, bh);
1242 if (buffer_uptodate(bh))
1250 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1251 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1252 * refcount elevated by one when they're in an LRU. A buffer can only appear
1253 * once in a particular CPU's LRU. A single buffer can be present in multiple
1254 * CPU's LRUs at the same time.
1256 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1257 * sb_find_get_block().
1259 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1260 * a local interrupt disable for that.
1263 #define BH_LRU_SIZE 8
1266 struct buffer_head *bhs[BH_LRU_SIZE];
1269 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1272 #define bh_lru_lock() local_irq_disable()
1273 #define bh_lru_unlock() local_irq_enable()
1275 #define bh_lru_lock() preempt_disable()
1276 #define bh_lru_unlock() preempt_enable()
1279 static inline void check_irqs_on(void)
1281 #ifdef irqs_disabled
1282 BUG_ON(irqs_disabled());
1287 * The LRU management algorithm is dopey-but-simple. Sorry.
1289 static void bh_lru_install(struct buffer_head *bh)
1291 struct buffer_head *evictee = NULL;
1295 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1296 struct buffer_head *bhs[BH_LRU_SIZE];
1302 for (in = 0; in < BH_LRU_SIZE; in++) {
1303 struct buffer_head *bh2 =
1304 __this_cpu_read(bh_lrus.bhs[in]);
1309 if (out >= BH_LRU_SIZE) {
1310 BUG_ON(evictee != NULL);
1317 while (out < BH_LRU_SIZE)
1319 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1328 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1330 static struct buffer_head *
1331 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1333 struct buffer_head *ret = NULL;
1338 for (i = 0; i < BH_LRU_SIZE; i++) {
1339 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1341 if (bh && bh->b_bdev == bdev &&
1342 bh->b_blocknr == block && bh->b_size == size) {
1345 __this_cpu_write(bh_lrus.bhs[i],
1346 __this_cpu_read(bh_lrus.bhs[i - 1]));
1349 __this_cpu_write(bh_lrus.bhs[0], bh);
1361 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1362 * it in the LRU and mark it as accessed. If it is not present then return
1365 struct buffer_head *
1366 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1368 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1371 bh = __find_get_block_slow(bdev, block);
1379 EXPORT_SYMBOL(__find_get_block);
1382 * __getblk will locate (and, if necessary, create) the buffer_head
1383 * which corresponds to the passed block_device, block and size. The
1384 * returned buffer has its reference count incremented.
1386 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1387 * attempt is failing. FIXME, perhaps?
1389 struct buffer_head *
1390 __getblk(struct block_device *bdev, sector_t block, unsigned size)
1392 struct buffer_head *bh = __find_get_block(bdev, block, size);
1396 bh = __getblk_slow(bdev, block, size);
1399 EXPORT_SYMBOL(__getblk);
1402 * Do async read-ahead on a buffer..
1404 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1406 struct buffer_head *bh = __getblk(bdev, block, size);
1408 ll_rw_block(READA, 1, &bh);
1412 EXPORT_SYMBOL(__breadahead);
1415 * __bread() - reads a specified block and returns the bh
1416 * @bdev: the block_device to read from
1417 * @block: number of block
1418 * @size: size (in bytes) to read
1420 * Reads a specified block, and returns buffer head that contains it.
1421 * It returns NULL if the block was unreadable.
1423 struct buffer_head *
1424 __bread(struct block_device *bdev, sector_t block, unsigned size)
1426 struct buffer_head *bh = __getblk(bdev, block, size);
1428 if (likely(bh) && !buffer_uptodate(bh))
1429 bh = __bread_slow(bh);
1432 EXPORT_SYMBOL(__bread);
1435 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1436 * This doesn't race because it runs in each cpu either in irq
1437 * or with preempt disabled.
1439 static void invalidate_bh_lru(void *arg)
1441 struct bh_lru *b = &get_cpu_var(bh_lrus);
1444 for (i = 0; i < BH_LRU_SIZE; i++) {
1448 put_cpu_var(bh_lrus);
1451 static bool has_bh_in_lru(int cpu, void *dummy)
1453 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1456 for (i = 0; i < BH_LRU_SIZE; i++) {
1464 void invalidate_bh_lrus(void)
1466 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1468 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1470 void set_bh_page(struct buffer_head *bh,
1471 struct page *page, unsigned long offset)
1474 BUG_ON(offset >= PAGE_SIZE);
1475 if (PageHighMem(page))
1477 * This catches illegal uses and preserves the offset:
1479 bh->b_data = (char *)(0 + offset);
1481 bh->b_data = page_address(page) + offset;
1483 EXPORT_SYMBOL(set_bh_page);
1486 * Called when truncating a buffer on a page completely.
1488 static void discard_buffer(struct buffer_head * bh)
1491 clear_buffer_dirty(bh);
1493 clear_buffer_mapped(bh);
1494 clear_buffer_req(bh);
1495 clear_buffer_new(bh);
1496 clear_buffer_delay(bh);
1497 clear_buffer_unwritten(bh);
1502 * block_invalidatepage - invalidate part or all of a buffer-backed page
1504 * @page: the page which is affected
1505 * @offset: start of the range to invalidate
1506 * @length: length of the range to invalidate
1508 * block_invalidatepage() is called when all or part of the page has become
1509 * invalidated by a truncate operation.
1511 * block_invalidatepage() does not have to release all buffers, but it must
1512 * ensure that no dirty buffer is left outside @offset and that no I/O
1513 * is underway against any of the blocks which are outside the truncation
1514 * point. Because the caller is about to free (and possibly reuse) those
1517 void block_invalidatepage(struct page *page, unsigned int offset,
1518 unsigned int length)
1520 struct buffer_head *head, *bh, *next;
1521 unsigned int curr_off = 0;
1522 unsigned int stop = length + offset;
1524 BUG_ON(!PageLocked(page));
1525 if (!page_has_buffers(page))
1529 * Check for overflow
1531 BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
1533 head = page_buffers(page);
1536 unsigned int next_off = curr_off + bh->b_size;
1537 next = bh->b_this_page;
1540 * Are we still fully in range ?
1542 if (next_off > stop)
1546 * is this block fully invalidated?
1548 if (offset <= curr_off)
1550 curr_off = next_off;
1552 } while (bh != head);
1555 * We release buffers only if the entire page is being invalidated.
1556 * The get_block cached value has been unconditionally invalidated,
1557 * so real IO is not possible anymore.
1560 try_to_release_page(page, 0);
1564 EXPORT_SYMBOL(block_invalidatepage);
1568 * We attach and possibly dirty the buffers atomically wrt
1569 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1570 * is already excluded via the page lock.
1572 void create_empty_buffers(struct page *page,
1573 unsigned long blocksize, unsigned long b_state)
1575 struct buffer_head *bh, *head, *tail;
1577 head = alloc_page_buffers(page, blocksize, 1);
1580 bh->b_state |= b_state;
1582 bh = bh->b_this_page;
1584 tail->b_this_page = head;
1586 spin_lock(&page->mapping->private_lock);
1587 if (PageUptodate(page) || PageDirty(page)) {
1590 if (PageDirty(page))
1591 set_buffer_dirty(bh);
1592 if (PageUptodate(page))
1593 set_buffer_uptodate(bh);
1594 bh = bh->b_this_page;
1595 } while (bh != head);
1597 attach_page_buffers(page, head);
1598 spin_unlock(&page->mapping->private_lock);
1600 EXPORT_SYMBOL(create_empty_buffers);
1603 * We are taking a block for data and we don't want any output from any
1604 * buffer-cache aliases starting from return from that function and
1605 * until the moment when something will explicitly mark the buffer
1606 * dirty (hopefully that will not happen until we will free that block ;-)
1607 * We don't even need to mark it not-uptodate - nobody can expect
1608 * anything from a newly allocated buffer anyway. We used to used
1609 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1610 * don't want to mark the alias unmapped, for example - it would confuse
1611 * anyone who might pick it with bread() afterwards...
1613 * Also.. Note that bforget() doesn't lock the buffer. So there can
1614 * be writeout I/O going on against recently-freed buffers. We don't
1615 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1616 * only if we really need to. That happens here.
1618 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1620 struct buffer_head *old_bh;
1624 old_bh = __find_get_block_slow(bdev, block);
1626 clear_buffer_dirty(old_bh);
1627 wait_on_buffer(old_bh);
1628 clear_buffer_req(old_bh);
1632 EXPORT_SYMBOL(unmap_underlying_metadata);
1635 * Size is a power-of-two in the range 512..PAGE_SIZE,
1636 * and the case we care about most is PAGE_SIZE.
1638 * So this *could* possibly be written with those
1639 * constraints in mind (relevant mostly if some
1640 * architecture has a slow bit-scan instruction)
1642 static inline int block_size_bits(unsigned int blocksize)
1644 return ilog2(blocksize);
1647 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1649 BUG_ON(!PageLocked(page));
1651 if (!page_has_buffers(page))
1652 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1653 return page_buffers(page);
1657 * NOTE! All mapped/uptodate combinations are valid:
1659 * Mapped Uptodate Meaning
1661 * No No "unknown" - must do get_block()
1662 * No Yes "hole" - zero-filled
1663 * Yes No "allocated" - allocated on disk, not read in
1664 * Yes Yes "valid" - allocated and up-to-date in memory.
1666 * "Dirty" is valid only with the last case (mapped+uptodate).
1670 * While block_write_full_page is writing back the dirty buffers under
1671 * the page lock, whoever dirtied the buffers may decide to clean them
1672 * again at any time. We handle that by only looking at the buffer
1673 * state inside lock_buffer().
1675 * If block_write_full_page() is called for regular writeback
1676 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1677 * locked buffer. This only can happen if someone has written the buffer
1678 * directly, with submit_bh(). At the address_space level PageWriteback
1679 * prevents this contention from occurring.
1681 * If block_write_full_page() is called with wbc->sync_mode ==
1682 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1683 * causes the writes to be flagged as synchronous writes.
1685 static int __block_write_full_page(struct inode *inode, struct page *page,
1686 get_block_t *get_block, struct writeback_control *wbc,
1687 bh_end_io_t *handler)
1691 sector_t last_block;
1692 struct buffer_head *bh, *head;
1693 unsigned int blocksize, bbits;
1694 int nr_underway = 0;
1695 int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1696 WRITE_SYNC : WRITE);
1698 head = create_page_buffers(page, inode,
1699 (1 << BH_Dirty)|(1 << BH_Uptodate));
1702 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1703 * here, and the (potentially unmapped) buffers may become dirty at
1704 * any time. If a buffer becomes dirty here after we've inspected it
1705 * then we just miss that fact, and the page stays dirty.
1707 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1708 * handle that here by just cleaning them.
1712 blocksize = bh->b_size;
1713 bbits = block_size_bits(blocksize);
1715 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1716 last_block = (i_size_read(inode) - 1) >> bbits;
1719 * Get all the dirty buffers mapped to disk addresses and
1720 * handle any aliases from the underlying blockdev's mapping.
1723 if (block > last_block) {
1725 * mapped buffers outside i_size will occur, because
1726 * this page can be outside i_size when there is a
1727 * truncate in progress.
1730 * The buffer was zeroed by block_write_full_page()
1732 clear_buffer_dirty(bh);
1733 set_buffer_uptodate(bh);
1734 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1736 WARN_ON(bh->b_size != blocksize);
1737 err = get_block(inode, block, bh, 1);
1740 clear_buffer_delay(bh);
1741 if (buffer_new(bh)) {
1742 /* blockdev mappings never come here */
1743 clear_buffer_new(bh);
1744 unmap_underlying_metadata(bh->b_bdev,
1748 bh = bh->b_this_page;
1750 } while (bh != head);
1753 if (!buffer_mapped(bh))
1756 * If it's a fully non-blocking write attempt and we cannot
1757 * lock the buffer then redirty the page. Note that this can
1758 * potentially cause a busy-wait loop from writeback threads
1759 * and kswapd activity, but those code paths have their own
1760 * higher-level throttling.
1762 if (wbc->sync_mode != WB_SYNC_NONE) {
1764 } else if (!trylock_buffer(bh)) {
1765 redirty_page_for_writepage(wbc, page);
1768 if (test_clear_buffer_dirty(bh)) {
1769 mark_buffer_async_write_endio(bh, handler);
1773 } while ((bh = bh->b_this_page) != head);
1776 * The page and its buffers are protected by PageWriteback(), so we can
1777 * drop the bh refcounts early.
1779 BUG_ON(PageWriteback(page));
1780 set_page_writeback(page);
1783 struct buffer_head *next = bh->b_this_page;
1784 if (buffer_async_write(bh)) {
1785 submit_bh(write_op, bh);
1789 } while (bh != head);
1794 if (nr_underway == 0) {
1796 * The page was marked dirty, but the buffers were
1797 * clean. Someone wrote them back by hand with
1798 * ll_rw_block/submit_bh. A rare case.
1800 end_page_writeback(page);
1803 * The page and buffer_heads can be released at any time from
1811 * ENOSPC, or some other error. We may already have added some
1812 * blocks to the file, so we need to write these out to avoid
1813 * exposing stale data.
1814 * The page is currently locked and not marked for writeback
1817 /* Recovery: lock and submit the mapped buffers */
1819 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1820 !buffer_delay(bh)) {
1822 mark_buffer_async_write_endio(bh, handler);
1825 * The buffer may have been set dirty during
1826 * attachment to a dirty page.
1828 clear_buffer_dirty(bh);
1830 } while ((bh = bh->b_this_page) != head);
1832 BUG_ON(PageWriteback(page));
1833 mapping_set_error(page->mapping, err);
1834 set_page_writeback(page);
1836 struct buffer_head *next = bh->b_this_page;
1837 if (buffer_async_write(bh)) {
1838 clear_buffer_dirty(bh);
1839 submit_bh(write_op, bh);
1843 } while (bh != head);
1849 * If a page has any new buffers, zero them out here, and mark them uptodate
1850 * and dirty so they'll be written out (in order to prevent uninitialised
1851 * block data from leaking). And clear the new bit.
1853 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1855 unsigned int block_start, block_end;
1856 struct buffer_head *head, *bh;
1858 BUG_ON(!PageLocked(page));
1859 if (!page_has_buffers(page))
1862 bh = head = page_buffers(page);
1865 block_end = block_start + bh->b_size;
1867 if (buffer_new(bh)) {
1868 if (block_end > from && block_start < to) {
1869 if (!PageUptodate(page)) {
1870 unsigned start, size;
1872 start = max(from, block_start);
1873 size = min(to, block_end) - start;
1875 zero_user(page, start, size);
1876 set_buffer_uptodate(bh);
1879 clear_buffer_new(bh);
1880 mark_buffer_dirty(bh);
1884 block_start = block_end;
1885 bh = bh->b_this_page;
1886 } while (bh != head);
1888 EXPORT_SYMBOL(page_zero_new_buffers);
1890 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1891 get_block_t *get_block)
1893 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1894 unsigned to = from + len;
1895 struct inode *inode = page->mapping->host;
1896 unsigned block_start, block_end;
1899 unsigned blocksize, bbits;
1900 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1902 BUG_ON(!PageLocked(page));
1903 BUG_ON(from > PAGE_CACHE_SIZE);
1904 BUG_ON(to > PAGE_CACHE_SIZE);
1907 head = create_page_buffers(page, inode, 0);
1908 blocksize = head->b_size;
1909 bbits = block_size_bits(blocksize);
1911 block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1913 for(bh = head, block_start = 0; bh != head || !block_start;
1914 block++, block_start=block_end, bh = bh->b_this_page) {
1915 block_end = block_start + blocksize;
1916 if (block_end <= from || block_start >= to) {
1917 if (PageUptodate(page)) {
1918 if (!buffer_uptodate(bh))
1919 set_buffer_uptodate(bh);
1924 clear_buffer_new(bh);
1925 if (!buffer_mapped(bh)) {
1926 WARN_ON(bh->b_size != blocksize);
1927 err = get_block(inode, block, bh, 1);
1930 if (buffer_new(bh)) {
1931 unmap_underlying_metadata(bh->b_bdev,
1933 if (PageUptodate(page)) {
1934 clear_buffer_new(bh);
1935 set_buffer_uptodate(bh);
1936 mark_buffer_dirty(bh);
1939 if (block_end > to || block_start < from)
1940 zero_user_segments(page,
1946 if (PageUptodate(page)) {
1947 if (!buffer_uptodate(bh))
1948 set_buffer_uptodate(bh);
1951 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1952 !buffer_unwritten(bh) &&
1953 (block_start < from || block_end > to)) {
1954 ll_rw_block(READ, 1, &bh);
1959 * If we issued read requests - let them complete.
1961 while(wait_bh > wait) {
1962 wait_on_buffer(*--wait_bh);
1963 if (!buffer_uptodate(*wait_bh))
1967 page_zero_new_buffers(page, from, to);
1970 EXPORT_SYMBOL(__block_write_begin);
1972 static int __block_commit_write(struct inode *inode, struct page *page,
1973 unsigned from, unsigned to)
1975 unsigned block_start, block_end;
1978 struct buffer_head *bh, *head;
1980 bh = head = page_buffers(page);
1981 blocksize = bh->b_size;
1985 block_end = block_start + blocksize;
1986 if (block_end <= from || block_start >= to) {
1987 if (!buffer_uptodate(bh))
1990 set_buffer_uptodate(bh);
1991 mark_buffer_dirty(bh);
1993 clear_buffer_new(bh);
1995 block_start = block_end;
1996 bh = bh->b_this_page;
1997 } while (bh != head);
2000 * If this is a partial write which happened to make all buffers
2001 * uptodate then we can optimize away a bogus readpage() for
2002 * the next read(). Here we 'discover' whether the page went
2003 * uptodate as a result of this (potentially partial) write.
2006 SetPageUptodate(page);
2011 * block_write_begin takes care of the basic task of block allocation and
2012 * bringing partial write blocks uptodate first.
2014 * The filesystem needs to handle block truncation upon failure.
2016 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2017 unsigned flags, struct page **pagep, get_block_t *get_block)
2019 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2023 page = grab_cache_page_write_begin(mapping, index, flags);
2027 status = __block_write_begin(page, pos, len, get_block);
2028 if (unlikely(status)) {
2030 page_cache_release(page);
2037 EXPORT_SYMBOL(block_write_begin);
2039 int block_write_end(struct file *file, struct address_space *mapping,
2040 loff_t pos, unsigned len, unsigned copied,
2041 struct page *page, void *fsdata)
2043 struct inode *inode = mapping->host;
2046 start = pos & (PAGE_CACHE_SIZE - 1);
2048 if (unlikely(copied < len)) {
2050 * The buffers that were written will now be uptodate, so we
2051 * don't have to worry about a readpage reading them and
2052 * overwriting a partial write. However if we have encountered
2053 * a short write and only partially written into a buffer, it
2054 * will not be marked uptodate, so a readpage might come in and
2055 * destroy our partial write.
2057 * Do the simplest thing, and just treat any short write to a
2058 * non uptodate page as a zero-length write, and force the
2059 * caller to redo the whole thing.
2061 if (!PageUptodate(page))
2064 page_zero_new_buffers(page, start+copied, start+len);
2066 flush_dcache_page(page);
2068 /* This could be a short (even 0-length) commit */
2069 __block_commit_write(inode, page, start, start+copied);
2073 EXPORT_SYMBOL(block_write_end);
2075 int generic_write_end(struct file *file, struct address_space *mapping,
2076 loff_t pos, unsigned len, unsigned copied,
2077 struct page *page, void *fsdata)
2079 struct inode *inode = mapping->host;
2080 int i_size_changed = 0;
2082 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2085 * No need to use i_size_read() here, the i_size
2086 * cannot change under us because we hold i_mutex.
2088 * But it's important to update i_size while still holding page lock:
2089 * page writeout could otherwise come in and zero beyond i_size.
2091 if (pos+copied > inode->i_size) {
2092 i_size_write(inode, pos+copied);
2097 page_cache_release(page);
2100 * Don't mark the inode dirty under page lock. First, it unnecessarily
2101 * makes the holding time of page lock longer. Second, it forces lock
2102 * ordering of page lock and transaction start for journaling
2106 mark_inode_dirty(inode);
2110 EXPORT_SYMBOL(generic_write_end);
2113 * block_is_partially_uptodate checks whether buffers within a page are
2116 * Returns true if all buffers which correspond to a file portion
2117 * we want to read are uptodate.
2119 int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2122 unsigned block_start, block_end, blocksize;
2124 struct buffer_head *bh, *head;
2127 if (!page_has_buffers(page))
2130 head = page_buffers(page);
2131 blocksize = head->b_size;
2132 to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2134 if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2140 block_end = block_start + blocksize;
2141 if (block_end > from && block_start < to) {
2142 if (!buffer_uptodate(bh)) {
2146 if (block_end >= to)
2149 block_start = block_end;
2150 bh = bh->b_this_page;
2151 } while (bh != head);
2155 EXPORT_SYMBOL(block_is_partially_uptodate);
2158 * Generic "read page" function for block devices that have the normal
2159 * get_block functionality. This is most of the block device filesystems.
2160 * Reads the page asynchronously --- the unlock_buffer() and
2161 * set/clear_buffer_uptodate() functions propagate buffer state into the
2162 * page struct once IO has completed.
2164 int block_read_full_page(struct page *page, get_block_t *get_block)
2166 struct inode *inode = page->mapping->host;
2167 sector_t iblock, lblock;
2168 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2169 unsigned int blocksize, bbits;
2171 int fully_mapped = 1;
2173 head = create_page_buffers(page, inode, 0);
2174 blocksize = head->b_size;
2175 bbits = block_size_bits(blocksize);
2177 iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2178 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2184 if (buffer_uptodate(bh))
2187 if (!buffer_mapped(bh)) {
2191 if (iblock < lblock) {
2192 WARN_ON(bh->b_size != blocksize);
2193 err = get_block(inode, iblock, bh, 0);
2197 if (!buffer_mapped(bh)) {
2198 zero_user(page, i * blocksize, blocksize);
2200 set_buffer_uptodate(bh);
2204 * get_block() might have updated the buffer
2207 if (buffer_uptodate(bh))
2211 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2214 SetPageMappedToDisk(page);
2218 * All buffers are uptodate - we can set the page uptodate
2219 * as well. But not if get_block() returned an error.
2221 if (!PageError(page))
2222 SetPageUptodate(page);
2227 /* Stage two: lock the buffers */
2228 for (i = 0; i < nr; i++) {
2231 mark_buffer_async_read(bh);
2235 * Stage 3: start the IO. Check for uptodateness
2236 * inside the buffer lock in case another process reading
2237 * the underlying blockdev brought it uptodate (the sct fix).
2239 for (i = 0; i < nr; i++) {
2241 if (buffer_uptodate(bh))
2242 end_buffer_async_read(bh, 1);
2244 submit_bh(READ, bh);
2248 EXPORT_SYMBOL(block_read_full_page);
2250 /* utility function for filesystems that need to do work on expanding
2251 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2252 * deal with the hole.
2254 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2256 struct address_space *mapping = inode->i_mapping;
2261 err = inode_newsize_ok(inode, size);
2265 err = pagecache_write_begin(NULL, mapping, size, 0,
2266 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2271 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2277 EXPORT_SYMBOL(generic_cont_expand_simple);
2279 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2280 loff_t pos, loff_t *bytes)
2282 struct inode *inode = mapping->host;
2283 unsigned blocksize = 1 << inode->i_blkbits;
2286 pgoff_t index, curidx;
2288 unsigned zerofrom, offset, len;
2291 index = pos >> PAGE_CACHE_SHIFT;
2292 offset = pos & ~PAGE_CACHE_MASK;
2294 while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2295 zerofrom = curpos & ~PAGE_CACHE_MASK;
2296 if (zerofrom & (blocksize-1)) {
2297 *bytes |= (blocksize-1);
2300 len = PAGE_CACHE_SIZE - zerofrom;
2302 err = pagecache_write_begin(file, mapping, curpos, len,
2303 AOP_FLAG_UNINTERRUPTIBLE,
2307 zero_user(page, zerofrom, len);
2308 err = pagecache_write_end(file, mapping, curpos, len, len,
2315 balance_dirty_pages_ratelimited(mapping);
2317 if (unlikely(fatal_signal_pending(current))) {
2323 /* page covers the boundary, find the boundary offset */
2324 if (index == curidx) {
2325 zerofrom = curpos & ~PAGE_CACHE_MASK;
2326 /* if we will expand the thing last block will be filled */
2327 if (offset <= zerofrom) {
2330 if (zerofrom & (blocksize-1)) {
2331 *bytes |= (blocksize-1);
2334 len = offset - zerofrom;
2336 err = pagecache_write_begin(file, mapping, curpos, len,
2337 AOP_FLAG_UNINTERRUPTIBLE,
2341 zero_user(page, zerofrom, len);
2342 err = pagecache_write_end(file, mapping, curpos, len, len,
2354 * For moronic filesystems that do not allow holes in file.
2355 * We may have to extend the file.
2357 int cont_write_begin(struct file *file, struct address_space *mapping,
2358 loff_t pos, unsigned len, unsigned flags,
2359 struct page **pagep, void **fsdata,
2360 get_block_t *get_block, loff_t *bytes)
2362 struct inode *inode = mapping->host;
2363 unsigned blocksize = 1 << inode->i_blkbits;
2367 err = cont_expand_zero(file, mapping, pos, bytes);
2371 zerofrom = *bytes & ~PAGE_CACHE_MASK;
2372 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2373 *bytes |= (blocksize-1);
2377 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2379 EXPORT_SYMBOL(cont_write_begin);
2381 int block_commit_write(struct page *page, unsigned from, unsigned to)
2383 struct inode *inode = page->mapping->host;
2384 __block_commit_write(inode,page,from,to);
2387 EXPORT_SYMBOL(block_commit_write);
2390 * block_page_mkwrite() is not allowed to change the file size as it gets
2391 * called from a page fault handler when a page is first dirtied. Hence we must
2392 * be careful to check for EOF conditions here. We set the page up correctly
2393 * for a written page which means we get ENOSPC checking when writing into
2394 * holes and correct delalloc and unwritten extent mapping on filesystems that
2395 * support these features.
2397 * We are not allowed to take the i_mutex here so we have to play games to
2398 * protect against truncate races as the page could now be beyond EOF. Because
2399 * truncate writes the inode size before removing pages, once we have the
2400 * page lock we can determine safely if the page is beyond EOF. If it is not
2401 * beyond EOF, then the page is guaranteed safe against truncation until we
2404 * Direct callers of this function should protect against filesystem freezing
2405 * using sb_start_write() - sb_end_write() functions.
2407 int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2408 get_block_t get_block)
2410 struct page *page = vmf->page;
2411 struct inode *inode = file_inode(vma->vm_file);
2417 size = i_size_read(inode);
2418 if ((page->mapping != inode->i_mapping) ||
2419 (page_offset(page) > size)) {
2420 /* We overload EFAULT to mean page got truncated */
2425 /* page is wholly or partially inside EOF */
2426 if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2427 end = size & ~PAGE_CACHE_MASK;
2429 end = PAGE_CACHE_SIZE;
2431 ret = __block_write_begin(page, 0, end, get_block);
2433 ret = block_commit_write(page, 0, end);
2435 if (unlikely(ret < 0))
2437 set_page_dirty(page);
2438 wait_for_stable_page(page);
2444 EXPORT_SYMBOL(__block_page_mkwrite);
2446 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2447 get_block_t get_block)
2450 struct super_block *sb = file_inode(vma->vm_file)->i_sb;
2452 sb_start_pagefault(sb);
2455 * Update file times before taking page lock. We may end up failing the
2456 * fault so this update may be superfluous but who really cares...
2458 file_update_time(vma->vm_file);
2460 ret = __block_page_mkwrite(vma, vmf, get_block);
2461 sb_end_pagefault(sb);
2462 return block_page_mkwrite_return(ret);
2464 EXPORT_SYMBOL(block_page_mkwrite);
2467 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2468 * immediately, while under the page lock. So it needs a special end_io
2469 * handler which does not touch the bh after unlocking it.
2471 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2473 __end_buffer_read_notouch(bh, uptodate);
2477 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2478 * the page (converting it to circular linked list and taking care of page
2481 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2483 struct buffer_head *bh;
2485 BUG_ON(!PageLocked(page));
2487 spin_lock(&page->mapping->private_lock);
2490 if (PageDirty(page))
2491 set_buffer_dirty(bh);
2492 if (!bh->b_this_page)
2493 bh->b_this_page = head;
2494 bh = bh->b_this_page;
2495 } while (bh != head);
2496 attach_page_buffers(page, head);
2497 spin_unlock(&page->mapping->private_lock);
2501 * On entry, the page is fully not uptodate.
2502 * On exit the page is fully uptodate in the areas outside (from,to)
2503 * The filesystem needs to handle block truncation upon failure.
2505 int nobh_write_begin(struct address_space *mapping,
2506 loff_t pos, unsigned len, unsigned flags,
2507 struct page **pagep, void **fsdata,
2508 get_block_t *get_block)
2510 struct inode *inode = mapping->host;
2511 const unsigned blkbits = inode->i_blkbits;
2512 const unsigned blocksize = 1 << blkbits;
2513 struct buffer_head *head, *bh;
2517 unsigned block_in_page;
2518 unsigned block_start, block_end;
2519 sector_t block_in_file;
2522 int is_mapped_to_disk = 1;
2524 index = pos >> PAGE_CACHE_SHIFT;
2525 from = pos & (PAGE_CACHE_SIZE - 1);
2528 page = grab_cache_page_write_begin(mapping, index, flags);
2534 if (page_has_buffers(page)) {
2535 ret = __block_write_begin(page, pos, len, get_block);
2541 if (PageMappedToDisk(page))
2545 * Allocate buffers so that we can keep track of state, and potentially
2546 * attach them to the page if an error occurs. In the common case of
2547 * no error, they will just be freed again without ever being attached
2548 * to the page (which is all OK, because we're under the page lock).
2550 * Be careful: the buffer linked list is a NULL terminated one, rather
2551 * than the circular one we're used to.
2553 head = alloc_page_buffers(page, blocksize, 0);
2559 block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2562 * We loop across all blocks in the page, whether or not they are
2563 * part of the affected region. This is so we can discover if the
2564 * page is fully mapped-to-disk.
2566 for (block_start = 0, block_in_page = 0, bh = head;
2567 block_start < PAGE_CACHE_SIZE;
2568 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2571 block_end = block_start + blocksize;
2574 if (block_start >= to)
2576 ret = get_block(inode, block_in_file + block_in_page,
2580 if (!buffer_mapped(bh))
2581 is_mapped_to_disk = 0;
2583 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2584 if (PageUptodate(page)) {
2585 set_buffer_uptodate(bh);
2588 if (buffer_new(bh) || !buffer_mapped(bh)) {
2589 zero_user_segments(page, block_start, from,
2593 if (buffer_uptodate(bh))
2594 continue; /* reiserfs does this */
2595 if (block_start < from || block_end > to) {
2597 bh->b_end_io = end_buffer_read_nobh;
2598 submit_bh(READ, bh);
2605 * The page is locked, so these buffers are protected from
2606 * any VM or truncate activity. Hence we don't need to care
2607 * for the buffer_head refcounts.
2609 for (bh = head; bh; bh = bh->b_this_page) {
2611 if (!buffer_uptodate(bh))
2618 if (is_mapped_to_disk)
2619 SetPageMappedToDisk(page);
2621 *fsdata = head; /* to be released by nobh_write_end */
2628 * Error recovery is a bit difficult. We need to zero out blocks that
2629 * were newly allocated, and dirty them to ensure they get written out.
2630 * Buffers need to be attached to the page at this point, otherwise
2631 * the handling of potential IO errors during writeout would be hard
2632 * (could try doing synchronous writeout, but what if that fails too?)
2634 attach_nobh_buffers(page, head);
2635 page_zero_new_buffers(page, from, to);
2639 page_cache_release(page);
2644 EXPORT_SYMBOL(nobh_write_begin);
2646 int nobh_write_end(struct file *file, struct address_space *mapping,
2647 loff_t pos, unsigned len, unsigned copied,
2648 struct page *page, void *fsdata)
2650 struct inode *inode = page->mapping->host;
2651 struct buffer_head *head = fsdata;
2652 struct buffer_head *bh;
2653 BUG_ON(fsdata != NULL && page_has_buffers(page));
2655 if (unlikely(copied < len) && head)
2656 attach_nobh_buffers(page, head);
2657 if (page_has_buffers(page))
2658 return generic_write_end(file, mapping, pos, len,
2659 copied, page, fsdata);
2661 SetPageUptodate(page);
2662 set_page_dirty(page);
2663 if (pos+copied > inode->i_size) {
2664 i_size_write(inode, pos+copied);
2665 mark_inode_dirty(inode);
2669 page_cache_release(page);
2673 head = head->b_this_page;
2674 free_buffer_head(bh);
2679 EXPORT_SYMBOL(nobh_write_end);
2682 * nobh_writepage() - based on block_full_write_page() except
2683 * that it tries to operate without attaching bufferheads to
2686 int nobh_writepage(struct page *page, get_block_t *get_block,
2687 struct writeback_control *wbc)
2689 struct inode * const inode = page->mapping->host;
2690 loff_t i_size = i_size_read(inode);
2691 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2695 /* Is the page fully inside i_size? */
2696 if (page->index < end_index)
2699 /* Is the page fully outside i_size? (truncate in progress) */
2700 offset = i_size & (PAGE_CACHE_SIZE-1);
2701 if (page->index >= end_index+1 || !offset) {
2703 * The page may have dirty, unmapped buffers. For example,
2704 * they may have been added in ext3_writepage(). Make them
2705 * freeable here, so the page does not leak.
2708 /* Not really sure about this - do we need this ? */
2709 if (page->mapping->a_ops->invalidatepage)
2710 page->mapping->a_ops->invalidatepage(page, offset);
2713 return 0; /* don't care */
2717 * The page straddles i_size. It must be zeroed out on each and every
2718 * writepage invocation because it may be mmapped. "A file is mapped
2719 * in multiples of the page size. For a file that is not a multiple of
2720 * the page size, the remaining memory is zeroed when mapped, and
2721 * writes to that region are not written out to the file."
2723 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2725 ret = mpage_writepage(page, get_block, wbc);
2727 ret = __block_write_full_page(inode, page, get_block, wbc,
2728 end_buffer_async_write);
2731 EXPORT_SYMBOL(nobh_writepage);
2733 int nobh_truncate_page(struct address_space *mapping,
2734 loff_t from, get_block_t *get_block)
2736 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2737 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2740 unsigned length, pos;
2741 struct inode *inode = mapping->host;
2743 struct buffer_head map_bh;
2746 blocksize = 1 << inode->i_blkbits;
2747 length = offset & (blocksize - 1);
2749 /* Block boundary? Nothing to do */
2753 length = blocksize - length;
2754 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2756 page = grab_cache_page(mapping, index);
2761 if (page_has_buffers(page)) {
2764 page_cache_release(page);
2765 return block_truncate_page(mapping, from, get_block);
2768 /* Find the buffer that contains "offset" */
2770 while (offset >= pos) {
2775 map_bh.b_size = blocksize;
2777 err = get_block(inode, iblock, &map_bh, 0);
2780 /* unmapped? It's a hole - nothing to do */
2781 if (!buffer_mapped(&map_bh))
2784 /* Ok, it's mapped. Make sure it's up-to-date */
2785 if (!PageUptodate(page)) {
2786 err = mapping->a_ops->readpage(NULL, page);
2788 page_cache_release(page);
2792 if (!PageUptodate(page)) {
2796 if (page_has_buffers(page))
2799 zero_user(page, offset, length);
2800 set_page_dirty(page);
2805 page_cache_release(page);
2809 EXPORT_SYMBOL(nobh_truncate_page);
2811 int block_truncate_page(struct address_space *mapping,
2812 loff_t from, get_block_t *get_block)
2814 pgoff_t index = from >> PAGE_CACHE_SHIFT;
2815 unsigned offset = from & (PAGE_CACHE_SIZE-1);
2818 unsigned length, pos;
2819 struct inode *inode = mapping->host;
2821 struct buffer_head *bh;
2824 blocksize = 1 << inode->i_blkbits;
2825 length = offset & (blocksize - 1);
2827 /* Block boundary? Nothing to do */
2831 length = blocksize - length;
2832 iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2834 page = grab_cache_page(mapping, index);
2839 if (!page_has_buffers(page))
2840 create_empty_buffers(page, blocksize, 0);
2842 /* Find the buffer that contains "offset" */
2843 bh = page_buffers(page);
2845 while (offset >= pos) {
2846 bh = bh->b_this_page;
2852 if (!buffer_mapped(bh)) {
2853 WARN_ON(bh->b_size != blocksize);
2854 err = get_block(inode, iblock, bh, 0);
2857 /* unmapped? It's a hole - nothing to do */
2858 if (!buffer_mapped(bh))
2862 /* Ok, it's mapped. Make sure it's up-to-date */
2863 if (PageUptodate(page))
2864 set_buffer_uptodate(bh);
2866 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2868 ll_rw_block(READ, 1, &bh);
2870 /* Uhhuh. Read error. Complain and punt. */
2871 if (!buffer_uptodate(bh))
2875 zero_user(page, offset, length);
2876 mark_buffer_dirty(bh);
2881 page_cache_release(page);
2885 EXPORT_SYMBOL(block_truncate_page);
2888 * The generic ->writepage function for buffer-backed address_spaces
2889 * this form passes in the end_io handler used to finish the IO.
2891 int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2892 struct writeback_control *wbc, bh_end_io_t *handler)
2894 struct inode * const inode = page->mapping->host;
2895 loff_t i_size = i_size_read(inode);
2896 const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2899 /* Is the page fully inside i_size? */
2900 if (page->index < end_index)
2901 return __block_write_full_page(inode, page, get_block, wbc,
2904 /* Is the page fully outside i_size? (truncate in progress) */
2905 offset = i_size & (PAGE_CACHE_SIZE-1);
2906 if (page->index >= end_index+1 || !offset) {
2908 * The page may have dirty, unmapped buffers. For example,
2909 * they may have been added in ext3_writepage(). Make them
2910 * freeable here, so the page does not leak.
2912 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
2914 return 0; /* don't care */
2918 * The page straddles i_size. It must be zeroed out on each and every
2919 * writepage invocation because it may be mmapped. "A file is mapped
2920 * in multiples of the page size. For a file that is not a multiple of
2921 * the page size, the remaining memory is zeroed when mapped, and
2922 * writes to that region are not written out to the file."
2924 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2925 return __block_write_full_page(inode, page, get_block, wbc, handler);
2927 EXPORT_SYMBOL(block_write_full_page_endio);
2930 * The generic ->writepage function for buffer-backed address_spaces
2932 int block_write_full_page(struct page *page, get_block_t *get_block,
2933 struct writeback_control *wbc)
2935 return block_write_full_page_endio(page, get_block, wbc,
2936 end_buffer_async_write);
2938 EXPORT_SYMBOL(block_write_full_page);
2940 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2941 get_block_t *get_block)
2943 struct buffer_head tmp;
2944 struct inode *inode = mapping->host;
2947 tmp.b_size = 1 << inode->i_blkbits;
2948 get_block(inode, block, &tmp, 0);
2949 return tmp.b_blocknr;
2951 EXPORT_SYMBOL(generic_block_bmap);
2953 static void end_bio_bh_io_sync(struct bio *bio, int err)
2955 struct buffer_head *bh = bio->bi_private;
2957 if (err == -EOPNOTSUPP) {
2958 set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2961 if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2962 set_bit(BH_Quiet, &bh->b_state);
2964 bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2969 * This allows us to do IO even on the odd last sectors
2970 * of a device, even if the bh block size is some multiple
2971 * of the physical sector size.
2973 * We'll just truncate the bio to the size of the device,
2974 * and clear the end of the buffer head manually.
2976 * Truly out-of-range accesses will turn into actual IO
2977 * errors, this only handles the "we need to be able to
2978 * do IO at the final sector" case.
2980 static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2985 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2990 * If the *whole* IO is past the end of the device,
2991 * let it through, and the IO layer will turn it into
2994 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
2997 maxsector -= bio->bi_iter.bi_sector;
2998 bytes = bio->bi_iter.bi_size;
2999 if (likely((bytes >> 9) <= maxsector))
3002 /* Uhhuh. We've got a bh that straddles the device size! */
3003 bytes = maxsector << 9;
3005 /* Truncate the bio.. */
3006 bio->bi_iter.bi_size = bytes;
3007 bio->bi_io_vec[0].bv_len = bytes;
3009 /* ..and clear the end of the buffer for reads */
3010 if ((rw & RW_MASK) == READ) {
3011 void *kaddr = kmap_atomic(bh->b_page);
3012 memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
3013 kunmap_atomic(kaddr);
3014 flush_dcache_page(bh->b_page);
3018 int _submit_bh(int rw, struct buffer_head *bh, unsigned long bio_flags)
3023 BUG_ON(!buffer_locked(bh));
3024 BUG_ON(!buffer_mapped(bh));
3025 BUG_ON(!bh->b_end_io);
3026 BUG_ON(buffer_delay(bh));
3027 BUG_ON(buffer_unwritten(bh));
3030 * Only clear out a write error when rewriting
3032 if (test_set_buffer_req(bh) && (rw & WRITE))
3033 clear_buffer_write_io_error(bh);
3036 * from here on down, it's all bio -- do the initial mapping,
3037 * submit_bio -> generic_make_request may further map this bio around
3039 bio = bio_alloc(GFP_NOIO, 1);
3041 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3042 bio->bi_bdev = bh->b_bdev;
3043 bio->bi_io_vec[0].bv_page = bh->b_page;
3044 bio->bi_io_vec[0].bv_len = bh->b_size;
3045 bio->bi_io_vec[0].bv_offset = bh_offset(bh);
3048 bio->bi_iter.bi_size = bh->b_size;
3050 bio->bi_end_io = end_bio_bh_io_sync;
3051 bio->bi_private = bh;
3052 bio->bi_flags |= bio_flags;
3054 /* Take care of bh's that straddle the end of the device */
3055 guard_bh_eod(rw, bio, bh);
3057 if (buffer_meta(bh))
3059 if (buffer_prio(bh))
3063 submit_bio(rw, bio);
3065 if (bio_flagged(bio, BIO_EOPNOTSUPP))
3071 EXPORT_SYMBOL_GPL(_submit_bh);
3073 int submit_bh(int rw, struct buffer_head *bh)
3075 return _submit_bh(rw, bh, 0);
3077 EXPORT_SYMBOL(submit_bh);
3080 * ll_rw_block: low-level access to block devices (DEPRECATED)
3081 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
3082 * @nr: number of &struct buffer_heads in the array
3083 * @bhs: array of pointers to &struct buffer_head
3085 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3086 * requests an I/O operation on them, either a %READ or a %WRITE. The third
3087 * %READA option is described in the documentation for generic_make_request()
3088 * which ll_rw_block() calls.
3090 * This function drops any buffer that it cannot get a lock on (with the
3091 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3092 * request, and any buffer that appears to be up-to-date when doing read
3093 * request. Further it marks as clean buffers that are processed for
3094 * writing (the buffer cache won't assume that they are actually clean
3095 * until the buffer gets unlocked).
3097 * ll_rw_block sets b_end_io to simple completion handler that marks
3098 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3101 * All of the buffers must be for the same device, and must also be a
3102 * multiple of the current approved size for the device.
3104 void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3108 for (i = 0; i < nr; i++) {
3109 struct buffer_head *bh = bhs[i];
3111 if (!trylock_buffer(bh))
3114 if (test_clear_buffer_dirty(bh)) {
3115 bh->b_end_io = end_buffer_write_sync;
3117 submit_bh(WRITE, bh);
3121 if (!buffer_uptodate(bh)) {
3122 bh->b_end_io = end_buffer_read_sync;
3131 EXPORT_SYMBOL(ll_rw_block);
3133 void write_dirty_buffer(struct buffer_head *bh, int rw)
3136 if (!test_clear_buffer_dirty(bh)) {
3140 bh->b_end_io = end_buffer_write_sync;
3144 EXPORT_SYMBOL(write_dirty_buffer);
3147 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3148 * and then start new I/O and then wait upon it. The caller must have a ref on
3151 int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3155 WARN_ON(atomic_read(&bh->b_count) < 1);
3157 if (test_clear_buffer_dirty(bh)) {
3159 bh->b_end_io = end_buffer_write_sync;
3160 ret = submit_bh(rw, bh);
3162 if (!ret && !buffer_uptodate(bh))
3169 EXPORT_SYMBOL(__sync_dirty_buffer);
3171 int sync_dirty_buffer(struct buffer_head *bh)
3173 return __sync_dirty_buffer(bh, WRITE_SYNC);
3175 EXPORT_SYMBOL(sync_dirty_buffer);
3178 * try_to_free_buffers() checks if all the buffers on this particular page
3179 * are unused, and releases them if so.
3181 * Exclusion against try_to_free_buffers may be obtained by either
3182 * locking the page or by holding its mapping's private_lock.
3184 * If the page is dirty but all the buffers are clean then we need to
3185 * be sure to mark the page clean as well. This is because the page
3186 * may be against a block device, and a later reattachment of buffers
3187 * to a dirty page will set *all* buffers dirty. Which would corrupt
3188 * filesystem data on the same device.
3190 * The same applies to regular filesystem pages: if all the buffers are
3191 * clean then we set the page clean and proceed. To do that, we require
3192 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3195 * try_to_free_buffers() is non-blocking.
3197 static inline int buffer_busy(struct buffer_head *bh)
3199 return atomic_read(&bh->b_count) |
3200 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3204 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3206 struct buffer_head *head = page_buffers(page);
3207 struct buffer_head *bh;
3211 if (buffer_write_io_error(bh) && page->mapping)
3212 set_bit(AS_EIO, &page->mapping->flags);
3213 if (buffer_busy(bh))
3215 bh = bh->b_this_page;
3216 } while (bh != head);
3219 struct buffer_head *next = bh->b_this_page;
3221 if (bh->b_assoc_map)
3222 __remove_assoc_queue(bh);
3224 } while (bh != head);
3225 *buffers_to_free = head;
3226 __clear_page_buffers(page);
3232 int try_to_free_buffers(struct page *page)
3234 struct address_space * const mapping = page->mapping;
3235 struct buffer_head *buffers_to_free = NULL;
3238 BUG_ON(!PageLocked(page));
3239 if (PageWriteback(page))
3242 if (mapping == NULL) { /* can this still happen? */
3243 ret = drop_buffers(page, &buffers_to_free);
3247 spin_lock(&mapping->private_lock);
3248 ret = drop_buffers(page, &buffers_to_free);
3251 * If the filesystem writes its buffers by hand (eg ext3)
3252 * then we can have clean buffers against a dirty page. We
3253 * clean the page here; otherwise the VM will never notice
3254 * that the filesystem did any IO at all.
3256 * Also, during truncate, discard_buffer will have marked all
3257 * the page's buffers clean. We discover that here and clean
3260 * private_lock must be held over this entire operation in order
3261 * to synchronise against __set_page_dirty_buffers and prevent the
3262 * dirty bit from being lost.
3265 cancel_dirty_page(page, PAGE_CACHE_SIZE);
3266 spin_unlock(&mapping->private_lock);
3268 if (buffers_to_free) {
3269 struct buffer_head *bh = buffers_to_free;
3272 struct buffer_head *next = bh->b_this_page;
3273 free_buffer_head(bh);
3275 } while (bh != buffers_to_free);
3279 EXPORT_SYMBOL(try_to_free_buffers);
3282 * There are no bdflush tunables left. But distributions are
3283 * still running obsolete flush daemons, so we terminate them here.
3285 * Use of bdflush() is deprecated and will be removed in a future kernel.
3286 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3288 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3290 static int msg_count;
3292 if (!capable(CAP_SYS_ADMIN))
3295 if (msg_count < 5) {
3298 "warning: process `%s' used the obsolete bdflush"
3299 " system call\n", current->comm);
3300 printk(KERN_INFO "Fix your initscripts?\n");
3309 * Buffer-head allocation
3311 static struct kmem_cache *bh_cachep __read_mostly;
3314 * Once the number of bh's in the machine exceeds this level, we start
3315 * stripping them in writeback.
3317 static unsigned long max_buffer_heads;
3319 int buffer_heads_over_limit;
3321 struct bh_accounting {
3322 int nr; /* Number of live bh's */
3323 int ratelimit; /* Limit cacheline bouncing */
3326 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3328 static void recalc_bh_state(void)
3333 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3335 __this_cpu_write(bh_accounting.ratelimit, 0);
3336 for_each_online_cpu(i)
3337 tot += per_cpu(bh_accounting, i).nr;
3338 buffer_heads_over_limit = (tot > max_buffer_heads);
3341 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3343 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3345 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3347 __this_cpu_inc(bh_accounting.nr);
3353 EXPORT_SYMBOL(alloc_buffer_head);
3355 void free_buffer_head(struct buffer_head *bh)
3357 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3358 kmem_cache_free(bh_cachep, bh);
3360 __this_cpu_dec(bh_accounting.nr);
3364 EXPORT_SYMBOL(free_buffer_head);
3366 static void buffer_exit_cpu(int cpu)
3369 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3371 for (i = 0; i < BH_LRU_SIZE; i++) {
3375 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3376 per_cpu(bh_accounting, cpu).nr = 0;
3379 static int buffer_cpu_notify(struct notifier_block *self,
3380 unsigned long action, void *hcpu)
3382 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3383 buffer_exit_cpu((unsigned long)hcpu);
3388 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3389 * @bh: struct buffer_head
3391 * Return true if the buffer is up-to-date and false,
3392 * with the buffer locked, if not.
3394 int bh_uptodate_or_lock(struct buffer_head *bh)
3396 if (!buffer_uptodate(bh)) {
3398 if (!buffer_uptodate(bh))
3404 EXPORT_SYMBOL(bh_uptodate_or_lock);
3407 * bh_submit_read - Submit a locked buffer for reading
3408 * @bh: struct buffer_head
3410 * Returns zero on success and -EIO on error.
3412 int bh_submit_read(struct buffer_head *bh)
3414 BUG_ON(!buffer_locked(bh));
3416 if (buffer_uptodate(bh)) {
3422 bh->b_end_io = end_buffer_read_sync;
3423 submit_bh(READ, bh);
3425 if (buffer_uptodate(bh))
3429 EXPORT_SYMBOL(bh_submit_read);
3431 void __init buffer_init(void)
3433 unsigned long nrpages;
3435 bh_cachep = kmem_cache_create("buffer_head",
3436 sizeof(struct buffer_head), 0,
3437 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3442 * Limit the bh occupancy to 10% of ZONE_NORMAL
3444 nrpages = (nr_free_buffer_pages() * 10) / 100;
3445 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3446 hotcpu_notifier(buffer_cpu_notify, 0);