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>
24 #include <linux/iomap.h>
26 #include <linux/percpu.h>
27 #include <linux/slab.h>
28 #include <linux/capability.h>
29 #include <linux/blkdev.h>
30 #include <linux/file.h>
31 #include <linux/quotaops.h>
32 #include <linux/highmem.h>
33 #include <linux/export.h>
34 #include <linux/backing-dev.h>
35 #include <linux/writeback.h>
36 #include <linux/hash.h>
37 #include <linux/suspend.h>
38 #include <linux/buffer_head.h>
39 #include <linux/task_io_accounting_ops.h>
40 #include <linux/bio.h>
41 #include <linux/notifier.h>
42 #include <linux/cpu.h>
43 #include <linux/bitops.h>
44 #include <linux/mpage.h>
45 #include <linux/bit_spinlock.h>
46 #include <trace/events/block.h>
48 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
49 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
50 unsigned long bio_flags,
51 struct writeback_control *wbc);
53 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
55 void init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
57 bh->b_end_io = handler;
58 bh->b_private = private;
60 EXPORT_SYMBOL(init_buffer);
62 inline void touch_buffer(struct buffer_head *bh)
64 trace_block_touch_buffer(bh);
65 mark_page_accessed(bh->b_page);
67 EXPORT_SYMBOL(touch_buffer);
69 void __lock_buffer(struct buffer_head *bh)
71 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
73 EXPORT_SYMBOL(__lock_buffer);
75 void unlock_buffer(struct buffer_head *bh)
77 clear_bit_unlock(BH_Lock, &bh->b_state);
78 smp_mb__after_atomic();
79 wake_up_bit(&bh->b_state, BH_Lock);
81 EXPORT_SYMBOL(unlock_buffer);
84 * Returns if the page has dirty or writeback buffers. If all the buffers
85 * are unlocked and clean then the PageDirty information is stale. If
86 * any of the pages are locked, it is assumed they are locked for IO.
88 void buffer_check_dirty_writeback(struct page *page,
89 bool *dirty, bool *writeback)
91 struct buffer_head *head, *bh;
95 BUG_ON(!PageLocked(page));
97 if (!page_has_buffers(page))
100 if (PageWriteback(page))
103 head = page_buffers(page);
106 if (buffer_locked(bh))
109 if (buffer_dirty(bh))
112 bh = bh->b_this_page;
113 } while (bh != head);
115 EXPORT_SYMBOL(buffer_check_dirty_writeback);
118 * Block until a buffer comes unlocked. This doesn't stop it
119 * from becoming locked again - you have to lock it yourself
120 * if you want to preserve its state.
122 void __wait_on_buffer(struct buffer_head * bh)
124 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
126 EXPORT_SYMBOL(__wait_on_buffer);
129 __clear_page_buffers(struct page *page)
131 ClearPagePrivate(page);
132 set_page_private(page, 0);
136 static void buffer_io_error(struct buffer_head *bh, char *msg)
138 if (!test_bit(BH_Quiet, &bh->b_state))
139 printk_ratelimited(KERN_ERR
140 "Buffer I/O error on dev %pg, logical block %llu%s\n",
141 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
145 * End-of-IO handler helper function which does not touch the bh after
147 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
148 * a race there is benign: unlock_buffer() only use the bh's address for
149 * hashing after unlocking the buffer, so it doesn't actually touch the bh
152 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
155 set_buffer_uptodate(bh);
157 /* This happens, due to failed read-ahead attempts. */
158 clear_buffer_uptodate(bh);
164 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
165 * unlock the buffer. This is what ll_rw_block uses too.
167 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
169 __end_buffer_read_notouch(bh, uptodate);
172 EXPORT_SYMBOL(end_buffer_read_sync);
174 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
177 set_buffer_uptodate(bh);
179 buffer_io_error(bh, ", lost sync page write");
180 set_buffer_write_io_error(bh);
181 clear_buffer_uptodate(bh);
186 EXPORT_SYMBOL(end_buffer_write_sync);
189 * Various filesystems appear to want __find_get_block to be non-blocking.
190 * But it's the page lock which protects the buffers. To get around this,
191 * we get exclusion from try_to_free_buffers with the blockdev mapping's
194 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
195 * may be quite high. This code could TryLock the page, and if that
196 * succeeds, there is no need to take private_lock. (But if
197 * private_lock is contended then so is mapping->tree_lock).
199 static struct buffer_head *
200 __find_get_block_slow(struct block_device *bdev, sector_t block)
202 struct inode *bd_inode = bdev->bd_inode;
203 struct address_space *bd_mapping = bd_inode->i_mapping;
204 struct buffer_head *ret = NULL;
206 struct buffer_head *bh;
207 struct buffer_head *head;
211 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
212 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
216 spin_lock(&bd_mapping->private_lock);
217 if (!page_has_buffers(page))
219 head = page_buffers(page);
222 if (!buffer_mapped(bh))
224 else if (bh->b_blocknr == block) {
229 bh = bh->b_this_page;
230 } while (bh != head);
232 /* we might be here because some of the buffers on this page are
233 * not mapped. This is due to various races between
234 * file io on the block device and getblk. It gets dealt with
235 * elsewhere, don't buffer_error if we had some unmapped buffers
238 printk("__find_get_block_slow() failed. "
239 "block=%llu, b_blocknr=%llu\n",
240 (unsigned long long)block,
241 (unsigned long long)bh->b_blocknr);
242 printk("b_state=0x%08lx, b_size=%zu\n",
243 bh->b_state, bh->b_size);
244 printk("device %pg blocksize: %d\n", bdev,
245 1 << bd_inode->i_blkbits);
248 spin_unlock(&bd_mapping->private_lock);
255 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
257 static void free_more_memory(void)
262 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
265 for_each_online_node(nid) {
267 z = first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
268 gfp_zone(GFP_NOFS), NULL);
270 try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
276 * I/O completion handler for block_read_full_page() - pages
277 * which come unlocked at the end of I/O.
279 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
282 struct buffer_head *first;
283 struct buffer_head *tmp;
285 int page_uptodate = 1;
287 BUG_ON(!buffer_async_read(bh));
291 set_buffer_uptodate(bh);
293 clear_buffer_uptodate(bh);
294 buffer_io_error(bh, ", async page read");
299 * Be _very_ careful from here on. Bad things can happen if
300 * two buffer heads end IO at almost the same time and both
301 * decide that the page is now completely done.
303 first = page_buffers(page);
304 flags = bh_uptodate_lock_irqsave(first);
305 clear_buffer_async_read(bh);
309 if (!buffer_uptodate(tmp))
311 if (buffer_async_read(tmp)) {
312 BUG_ON(!buffer_locked(tmp));
315 tmp = tmp->b_this_page;
317 bh_uptodate_unlock_irqrestore(first, flags);
320 * If none of the buffers had errors and they are all
321 * uptodate then we can set the page uptodate.
323 if (page_uptodate && !PageError(page))
324 SetPageUptodate(page);
329 bh_uptodate_unlock_irqrestore(first, flags);
333 * Completion handler for block_write_full_page() - pages which are unlocked
334 * during I/O, and which have PageWriteback cleared upon I/O completion.
336 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
339 struct buffer_head *first;
340 struct buffer_head *tmp;
343 BUG_ON(!buffer_async_write(bh));
347 set_buffer_uptodate(bh);
349 buffer_io_error(bh, ", lost async page write");
350 mapping_set_error(page->mapping, -EIO);
351 set_buffer_write_io_error(bh);
352 clear_buffer_uptodate(bh);
356 first = page_buffers(page);
357 flags = bh_uptodate_lock_irqsave(first);
359 clear_buffer_async_write(bh);
361 tmp = bh->b_this_page;
363 if (buffer_async_write(tmp)) {
364 BUG_ON(!buffer_locked(tmp));
367 tmp = tmp->b_this_page;
369 bh_uptodate_unlock_irqrestore(first, flags);
370 end_page_writeback(page);
374 bh_uptodate_unlock_irqrestore(first, flags);
376 EXPORT_SYMBOL(end_buffer_async_write);
379 * If a page's buffers are under async readin (end_buffer_async_read
380 * completion) then there is a possibility that another thread of
381 * control could lock one of the buffers after it has completed
382 * but while some of the other buffers have not completed. This
383 * locked buffer would confuse end_buffer_async_read() into not unlocking
384 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
385 * that this buffer is not under async I/O.
387 * The page comes unlocked when it has no locked buffer_async buffers
390 * PageLocked prevents anyone starting new async I/O reads any of
393 * PageWriteback is used to prevent simultaneous writeout of the same
396 * PageLocked prevents anyone from starting writeback of a page which is
397 * under read I/O (PageWriteback is only ever set against a locked page).
399 static void mark_buffer_async_read(struct buffer_head *bh)
401 bh->b_end_io = end_buffer_async_read;
402 set_buffer_async_read(bh);
405 static void mark_buffer_async_write_endio(struct buffer_head *bh,
406 bh_end_io_t *handler)
408 bh->b_end_io = handler;
409 set_buffer_async_write(bh);
412 void mark_buffer_async_write(struct buffer_head *bh)
414 mark_buffer_async_write_endio(bh, end_buffer_async_write);
416 EXPORT_SYMBOL(mark_buffer_async_write);
420 * fs/buffer.c contains helper functions for buffer-backed address space's
421 * fsync functions. A common requirement for buffer-based filesystems is
422 * that certain data from the backing blockdev needs to be written out for
423 * a successful fsync(). For example, ext2 indirect blocks need to be
424 * written back and waited upon before fsync() returns.
426 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
427 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
428 * management of a list of dependent buffers at ->i_mapping->private_list.
430 * Locking is a little subtle: try_to_free_buffers() will remove buffers
431 * from their controlling inode's queue when they are being freed. But
432 * try_to_free_buffers() will be operating against the *blockdev* mapping
433 * at the time, not against the S_ISREG file which depends on those buffers.
434 * So the locking for private_list is via the private_lock in the address_space
435 * which backs the buffers. Which is different from the address_space
436 * against which the buffers are listed. So for a particular address_space,
437 * mapping->private_lock does *not* protect mapping->private_list! In fact,
438 * mapping->private_list will always be protected by the backing blockdev's
441 * Which introduces a requirement: all buffers on an address_space's
442 * ->private_list must be from the same address_space: the blockdev's.
444 * address_spaces which do not place buffers at ->private_list via these
445 * utility functions are free to use private_lock and private_list for
446 * whatever they want. The only requirement is that list_empty(private_list)
447 * be true at clear_inode() time.
449 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
450 * filesystems should do that. invalidate_inode_buffers() should just go
451 * BUG_ON(!list_empty).
453 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
454 * take an address_space, not an inode. And it should be called
455 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
458 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
459 * list if it is already on a list. Because if the buffer is on a list,
460 * it *must* already be on the right one. If not, the filesystem is being
461 * silly. This will save a ton of locking. But first we have to ensure
462 * that buffers are taken *off* the old inode's list when they are freed
463 * (presumably in truncate). That requires careful auditing of all
464 * filesystems (do it inside bforget()). It could also be done by bringing
469 * The buffer's backing address_space's private_lock must be held
471 static void __remove_assoc_queue(struct buffer_head *bh)
473 list_del_init(&bh->b_assoc_buffers);
474 WARN_ON(!bh->b_assoc_map);
475 if (buffer_write_io_error(bh))
476 set_bit(AS_EIO, &bh->b_assoc_map->flags);
477 bh->b_assoc_map = NULL;
480 int inode_has_buffers(struct inode *inode)
482 return !list_empty(&inode->i_data.private_list);
486 * osync is designed to support O_SYNC io. It waits synchronously for
487 * all already-submitted IO to complete, but does not queue any new
488 * writes to the disk.
490 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
491 * you dirty the buffers, and then use osync_inode_buffers to wait for
492 * completion. Any other dirty buffers which are not yet queued for
493 * write will not be flushed to disk by the osync.
495 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
497 struct buffer_head *bh;
503 list_for_each_prev(p, list) {
505 if (buffer_locked(bh)) {
509 if (!buffer_uptodate(bh))
520 static void do_thaw_one(struct super_block *sb, void *unused)
522 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
523 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
526 static void do_thaw_all(struct work_struct *work)
528 iterate_supers(do_thaw_one, NULL);
530 printk(KERN_WARNING "Emergency Thaw complete\n");
534 * emergency_thaw_all -- forcibly thaw every frozen filesystem
536 * Used for emergency unfreeze of all filesystems via SysRq
538 void emergency_thaw_all(void)
540 struct work_struct *work;
542 work = kmalloc(sizeof(*work), GFP_ATOMIC);
544 INIT_WORK(work, do_thaw_all);
550 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
551 * @mapping: the mapping which wants those buffers written
553 * Starts I/O against the buffers at mapping->private_list, and waits upon
556 * Basically, this is a convenience function for fsync().
557 * @mapping is a file or directory which needs those buffers to be written for
558 * a successful fsync().
560 int sync_mapping_buffers(struct address_space *mapping)
562 struct address_space *buffer_mapping = mapping->private_data;
564 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
567 return fsync_buffers_list(&buffer_mapping->private_lock,
568 &mapping->private_list);
570 EXPORT_SYMBOL(sync_mapping_buffers);
573 * Called when we've recently written block `bblock', and it is known that
574 * `bblock' was for a buffer_boundary() buffer. This means that the block at
575 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
576 * dirty, schedule it for IO. So that indirects merge nicely with their data.
578 void write_boundary_block(struct block_device *bdev,
579 sector_t bblock, unsigned blocksize)
581 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
583 if (buffer_dirty(bh))
584 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
589 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
591 struct address_space *mapping = inode->i_mapping;
592 struct address_space *buffer_mapping = bh->b_page->mapping;
594 mark_buffer_dirty(bh);
595 if (!mapping->private_data) {
596 mapping->private_data = buffer_mapping;
598 BUG_ON(mapping->private_data != buffer_mapping);
600 if (!bh->b_assoc_map) {
601 spin_lock(&buffer_mapping->private_lock);
602 list_move_tail(&bh->b_assoc_buffers,
603 &mapping->private_list);
604 bh->b_assoc_map = mapping;
605 spin_unlock(&buffer_mapping->private_lock);
608 EXPORT_SYMBOL(mark_buffer_dirty_inode);
611 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
614 * If warn is true, then emit a warning if the page is not uptodate and has
615 * not been truncated.
617 * The caller must hold lock_page_memcg().
619 static void __set_page_dirty(struct page *page, struct address_space *mapping,
624 spin_lock_irqsave(&mapping->tree_lock, flags);
625 if (page->mapping) { /* Race with truncate? */
626 WARN_ON_ONCE(warn && !PageUptodate(page));
627 account_page_dirtied(page, mapping);
628 radix_tree_tag_set(&mapping->page_tree,
629 page_index(page), PAGECACHE_TAG_DIRTY);
631 spin_unlock_irqrestore(&mapping->tree_lock, flags);
635 * Add a page to the dirty page list.
637 * It is a sad fact of life that this function is called from several places
638 * deeply under spinlocking. It may not sleep.
640 * If the page has buffers, the uptodate buffers are set dirty, to preserve
641 * dirty-state coherency between the page and the buffers. It the page does
642 * not have buffers then when they are later attached they will all be set
645 * The buffers are dirtied before the page is dirtied. There's a small race
646 * window in which a writepage caller may see the page cleanness but not the
647 * buffer dirtiness. That's fine. If this code were to set the page dirty
648 * before the buffers, a concurrent writepage caller could clear the page dirty
649 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
650 * page on the dirty page list.
652 * We use private_lock to lock against try_to_free_buffers while using the
653 * page's buffer list. Also use this to protect against clean buffers being
654 * added to the page after it was set dirty.
656 * FIXME: may need to call ->reservepage here as well. That's rather up to the
657 * address_space though.
659 int __set_page_dirty_buffers(struct page *page)
662 struct address_space *mapping = page_mapping(page);
664 if (unlikely(!mapping))
665 return !TestSetPageDirty(page);
667 spin_lock(&mapping->private_lock);
668 if (page_has_buffers(page)) {
669 struct buffer_head *head = page_buffers(page);
670 struct buffer_head *bh = head;
673 set_buffer_dirty(bh);
674 bh = bh->b_this_page;
675 } while (bh != head);
678 * Lock out page->mem_cgroup migration to keep PageDirty
679 * synchronized with per-memcg dirty page counters.
681 lock_page_memcg(page);
682 newly_dirty = !TestSetPageDirty(page);
683 spin_unlock(&mapping->private_lock);
686 __set_page_dirty(page, mapping, 1);
688 unlock_page_memcg(page);
691 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
695 EXPORT_SYMBOL(__set_page_dirty_buffers);
698 * Write out and wait upon a list of buffers.
700 * We have conflicting pressures: we want to make sure that all
701 * initially dirty buffers get waited on, but that any subsequently
702 * dirtied buffers don't. After all, we don't want fsync to last
703 * forever if somebody is actively writing to the file.
705 * Do this in two main stages: first we copy dirty buffers to a
706 * temporary inode list, queueing the writes as we go. Then we clean
707 * up, waiting for those writes to complete.
709 * During this second stage, any subsequent updates to the file may end
710 * up refiling the buffer on the original inode's dirty list again, so
711 * there is a chance we will end up with a buffer queued for write but
712 * not yet completed on that list. So, as a final cleanup we go through
713 * the osync code to catch these locked, dirty buffers without requeuing
714 * any newly dirty buffers for write.
716 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
718 struct buffer_head *bh;
719 struct list_head tmp;
720 struct address_space *mapping;
722 struct blk_plug plug;
724 INIT_LIST_HEAD(&tmp);
725 blk_start_plug(&plug);
728 while (!list_empty(list)) {
729 bh = BH_ENTRY(list->next);
730 mapping = bh->b_assoc_map;
731 __remove_assoc_queue(bh);
732 /* Avoid race with mark_buffer_dirty_inode() which does
733 * a lockless check and we rely on seeing the dirty bit */
735 if (buffer_dirty(bh) || buffer_locked(bh)) {
736 list_add(&bh->b_assoc_buffers, &tmp);
737 bh->b_assoc_map = mapping;
738 if (buffer_dirty(bh)) {
742 * Ensure any pending I/O completes so that
743 * write_dirty_buffer() actually writes the
744 * current contents - it is a noop if I/O is
745 * still in flight on potentially older
748 write_dirty_buffer(bh, WRITE_SYNC);
751 * Kick off IO for the previous mapping. Note
752 * that we will not run the very last mapping,
753 * wait_on_buffer() will do that for us
754 * through sync_buffer().
763 blk_finish_plug(&plug);
766 while (!list_empty(&tmp)) {
767 bh = BH_ENTRY(tmp.prev);
769 mapping = bh->b_assoc_map;
770 __remove_assoc_queue(bh);
771 /* Avoid race with mark_buffer_dirty_inode() which does
772 * a lockless check and we rely on seeing the dirty bit */
774 if (buffer_dirty(bh)) {
775 list_add(&bh->b_assoc_buffers,
776 &mapping->private_list);
777 bh->b_assoc_map = mapping;
781 if (!buffer_uptodate(bh))
788 err2 = osync_buffers_list(lock, list);
796 * Invalidate any and all dirty buffers on a given inode. We are
797 * probably unmounting the fs, but that doesn't mean we have already
798 * done a sync(). Just drop the buffers from the inode list.
800 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
801 * assumes that all the buffers are against the blockdev. Not true
804 void invalidate_inode_buffers(struct inode *inode)
806 if (inode_has_buffers(inode)) {
807 struct address_space *mapping = &inode->i_data;
808 struct list_head *list = &mapping->private_list;
809 struct address_space *buffer_mapping = mapping->private_data;
811 spin_lock(&buffer_mapping->private_lock);
812 while (!list_empty(list))
813 __remove_assoc_queue(BH_ENTRY(list->next));
814 spin_unlock(&buffer_mapping->private_lock);
817 EXPORT_SYMBOL(invalidate_inode_buffers);
820 * Remove any clean buffers from the inode's buffer list. This is called
821 * when we're trying to free the inode itself. Those buffers can pin it.
823 * Returns true if all buffers were removed.
825 int remove_inode_buffers(struct inode *inode)
829 if (inode_has_buffers(inode)) {
830 struct address_space *mapping = &inode->i_data;
831 struct list_head *list = &mapping->private_list;
832 struct address_space *buffer_mapping = mapping->private_data;
834 spin_lock(&buffer_mapping->private_lock);
835 while (!list_empty(list)) {
836 struct buffer_head *bh = BH_ENTRY(list->next);
837 if (buffer_dirty(bh)) {
841 __remove_assoc_queue(bh);
843 spin_unlock(&buffer_mapping->private_lock);
849 * Create the appropriate buffers when given a page for data area and
850 * the size of each buffer.. Use the bh->b_this_page linked list to
851 * follow the buffers created. Return NULL if unable to create more
854 * The retry flag is used to differentiate async IO (paging, swapping)
855 * which may not fail from ordinary buffer allocations.
857 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
860 struct buffer_head *bh, *head;
866 while ((offset -= size) >= 0) {
867 bh = alloc_buffer_head(GFP_NOFS);
871 bh->b_this_page = head;
877 /* Link the buffer to its page */
878 set_bh_page(bh, page, offset);
882 * In case anything failed, we just free everything we got.
888 head = head->b_this_page;
889 free_buffer_head(bh);
894 * Return failure for non-async IO requests. Async IO requests
895 * are not allowed to fail, so we have to wait until buffer heads
896 * become available. But we don't want tasks sleeping with
897 * partially complete buffers, so all were released above.
902 /* We're _really_ low on memory. Now we just
903 * wait for old buffer heads to become free due to
904 * finishing IO. Since this is an async request and
905 * the reserve list is empty, we're sure there are
906 * async buffer heads in use.
911 EXPORT_SYMBOL_GPL(alloc_page_buffers);
914 link_dev_buffers(struct page *page, struct buffer_head *head)
916 struct buffer_head *bh, *tail;
921 bh = bh->b_this_page;
923 tail->b_this_page = head;
924 attach_page_buffers(page, head);
927 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
929 sector_t retval = ~((sector_t)0);
930 loff_t sz = i_size_read(bdev->bd_inode);
933 unsigned int sizebits = blksize_bits(size);
934 retval = (sz >> sizebits);
940 * Initialise the state of a blockdev page's buffers.
943 init_page_buffers(struct page *page, struct block_device *bdev,
944 sector_t block, int size)
946 struct buffer_head *head = page_buffers(page);
947 struct buffer_head *bh = head;
948 int uptodate = PageUptodate(page);
949 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
952 if (!buffer_mapped(bh)) {
953 init_buffer(bh, NULL, NULL);
955 bh->b_blocknr = block;
957 set_buffer_uptodate(bh);
958 if (block < end_block)
959 set_buffer_mapped(bh);
962 bh = bh->b_this_page;
963 } while (bh != head);
966 * Caller needs to validate requested block against end of device.
972 * Create the page-cache page that contains the requested block.
974 * This is used purely for blockdev mappings.
977 grow_dev_page(struct block_device *bdev, sector_t block,
978 pgoff_t index, int size, int sizebits, gfp_t gfp)
980 struct inode *inode = bdev->bd_inode;
982 struct buffer_head *bh;
984 int ret = 0; /* Will call free_more_memory() */
987 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
990 * XXX: __getblk_slow() can not really deal with failure and
991 * will endlessly loop on improvised global reclaim. Prefer
992 * looping in the allocator rather than here, at least that
993 * code knows what it's doing.
995 gfp_mask |= __GFP_NOFAIL;
997 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
1001 BUG_ON(!PageLocked(page));
1003 if (page_has_buffers(page)) {
1004 bh = page_buffers(page);
1005 if (bh->b_size == size) {
1006 end_block = init_page_buffers(page, bdev,
1007 (sector_t)index << sizebits,
1011 if (!try_to_free_buffers(page))
1016 * Allocate some buffers for this page
1018 bh = alloc_page_buffers(page, size, 0);
1023 * Link the page to the buffers and initialise them. Take the
1024 * lock to be atomic wrt __find_get_block(), which does not
1025 * run under the page lock.
1027 spin_lock(&inode->i_mapping->private_lock);
1028 link_dev_buffers(page, bh);
1029 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1031 spin_unlock(&inode->i_mapping->private_lock);
1033 ret = (block < end_block) ? 1 : -ENXIO;
1041 * Create buffers for the specified block device block's page. If
1042 * that page was dirty, the buffers are set dirty also.
1045 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1053 } while ((size << sizebits) < PAGE_SIZE);
1055 index = block >> sizebits;
1058 * Check for a block which wants to lie outside our maximum possible
1059 * pagecache index. (this comparison is done using sector_t types).
1061 if (unlikely(index != block >> sizebits)) {
1062 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1064 __func__, (unsigned long long)block,
1069 /* Create a page with the proper size buffers.. */
1070 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1073 static struct buffer_head *
1074 __getblk_slow(struct block_device *bdev, sector_t block,
1075 unsigned size, gfp_t gfp)
1077 /* Size must be multiple of hard sectorsize */
1078 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1079 (size < 512 || size > PAGE_SIZE))) {
1080 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1082 printk(KERN_ERR "logical block size: %d\n",
1083 bdev_logical_block_size(bdev));
1090 struct buffer_head *bh;
1093 bh = __find_get_block(bdev, block, size);
1097 ret = grow_buffers(bdev, block, size, gfp);
1106 * The relationship between dirty buffers and dirty pages:
1108 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1109 * the page is tagged dirty in its radix tree.
1111 * At all times, the dirtiness of the buffers represents the dirtiness of
1112 * subsections of the page. If the page has buffers, the page dirty bit is
1113 * merely a hint about the true dirty state.
1115 * When a page is set dirty in its entirety, all its buffers are marked dirty
1116 * (if the page has buffers).
1118 * When a buffer is marked dirty, its page is dirtied, but the page's other
1121 * Also. When blockdev buffers are explicitly read with bread(), they
1122 * individually become uptodate. But their backing page remains not
1123 * uptodate - even if all of its buffers are uptodate. A subsequent
1124 * block_read_full_page() against that page will discover all the uptodate
1125 * buffers, will set the page uptodate and will perform no I/O.
1129 * mark_buffer_dirty - mark a buffer_head as needing writeout
1130 * @bh: the buffer_head to mark dirty
1132 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1133 * backing page dirty, then tag the page as dirty in its address_space's radix
1134 * tree and then attach the address_space's inode to its superblock's dirty
1137 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1138 * mapping->tree_lock and mapping->host->i_lock.
1140 void mark_buffer_dirty(struct buffer_head *bh)
1142 WARN_ON_ONCE(!buffer_uptodate(bh));
1144 trace_block_dirty_buffer(bh);
1147 * Very *carefully* optimize the it-is-already-dirty case.
1149 * Don't let the final "is it dirty" escape to before we
1150 * perhaps modified the buffer.
1152 if (buffer_dirty(bh)) {
1154 if (buffer_dirty(bh))
1158 if (!test_set_buffer_dirty(bh)) {
1159 struct page *page = bh->b_page;
1160 struct address_space *mapping = NULL;
1162 lock_page_memcg(page);
1163 if (!TestSetPageDirty(page)) {
1164 mapping = page_mapping(page);
1166 __set_page_dirty(page, mapping, 0);
1168 unlock_page_memcg(page);
1170 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1173 EXPORT_SYMBOL(mark_buffer_dirty);
1176 * Decrement a buffer_head's reference count. If all buffers against a page
1177 * have zero reference count, are clean and unlocked, and if the page is clean
1178 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1179 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1180 * a page but it ends up not being freed, and buffers may later be reattached).
1182 void __brelse(struct buffer_head * buf)
1184 if (atomic_read(&buf->b_count)) {
1188 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1190 EXPORT_SYMBOL(__brelse);
1193 * bforget() is like brelse(), except it discards any
1194 * potentially dirty data.
1196 void __bforget(struct buffer_head *bh)
1198 clear_buffer_dirty(bh);
1199 if (bh->b_assoc_map) {
1200 struct address_space *buffer_mapping = bh->b_page->mapping;
1202 spin_lock(&buffer_mapping->private_lock);
1203 list_del_init(&bh->b_assoc_buffers);
1204 bh->b_assoc_map = NULL;
1205 spin_unlock(&buffer_mapping->private_lock);
1209 EXPORT_SYMBOL(__bforget);
1211 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1214 if (buffer_uptodate(bh)) {
1219 bh->b_end_io = end_buffer_read_sync;
1220 submit_bh(REQ_OP_READ, 0, bh);
1222 if (buffer_uptodate(bh))
1230 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1231 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1232 * refcount elevated by one when they're in an LRU. A buffer can only appear
1233 * once in a particular CPU's LRU. A single buffer can be present in multiple
1234 * CPU's LRUs at the same time.
1236 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1237 * sb_find_get_block().
1239 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1240 * a local interrupt disable for that.
1243 #define BH_LRU_SIZE 16
1246 struct buffer_head *bhs[BH_LRU_SIZE];
1249 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1252 #define bh_lru_lock() local_irq_disable()
1253 #define bh_lru_unlock() local_irq_enable()
1255 #define bh_lru_lock() preempt_disable()
1256 #define bh_lru_unlock() preempt_enable()
1259 static inline void check_irqs_on(void)
1261 #ifdef irqs_disabled
1262 BUG_ON(irqs_disabled());
1267 * The LRU management algorithm is dopey-but-simple. Sorry.
1269 static void bh_lru_install(struct buffer_head *bh)
1271 struct buffer_head *evictee = NULL;
1275 if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1276 struct buffer_head *bhs[BH_LRU_SIZE];
1282 for (in = 0; in < BH_LRU_SIZE; in++) {
1283 struct buffer_head *bh2 =
1284 __this_cpu_read(bh_lrus.bhs[in]);
1289 if (out >= BH_LRU_SIZE) {
1290 BUG_ON(evictee != NULL);
1297 while (out < BH_LRU_SIZE)
1299 memcpy(this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1308 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1310 static struct buffer_head *
1311 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1313 struct buffer_head *ret = NULL;
1318 for (i = 0; i < BH_LRU_SIZE; i++) {
1319 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1321 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1322 bh->b_size == size) {
1325 __this_cpu_write(bh_lrus.bhs[i],
1326 __this_cpu_read(bh_lrus.bhs[i - 1]));
1329 __this_cpu_write(bh_lrus.bhs[0], bh);
1341 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1342 * it in the LRU and mark it as accessed. If it is not present then return
1345 struct buffer_head *
1346 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1348 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1351 /* __find_get_block_slow will mark the page accessed */
1352 bh = __find_get_block_slow(bdev, block);
1360 EXPORT_SYMBOL(__find_get_block);
1363 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1364 * which corresponds to the passed block_device, block and size. The
1365 * returned buffer has its reference count incremented.
1367 * __getblk_gfp() will lock up the machine if grow_dev_page's
1368 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1370 struct buffer_head *
1371 __getblk_gfp(struct block_device *bdev, sector_t block,
1372 unsigned size, gfp_t gfp)
1374 struct buffer_head *bh = __find_get_block(bdev, block, size);
1378 bh = __getblk_slow(bdev, block, size, gfp);
1381 EXPORT_SYMBOL(__getblk_gfp);
1384 * Do async read-ahead on a buffer..
1386 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1388 struct buffer_head *bh = __getblk(bdev, block, size);
1390 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1394 EXPORT_SYMBOL(__breadahead);
1397 * __bread_gfp() - reads a specified block and returns the bh
1398 * @bdev: the block_device to read from
1399 * @block: number of block
1400 * @size: size (in bytes) to read
1401 * @gfp: page allocation flag
1403 * Reads a specified block, and returns buffer head that contains it.
1404 * The page cache can be allocated from non-movable area
1405 * not to prevent page migration if you set gfp to zero.
1406 * It returns NULL if the block was unreadable.
1408 struct buffer_head *
1409 __bread_gfp(struct block_device *bdev, sector_t block,
1410 unsigned size, gfp_t gfp)
1412 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1414 if (likely(bh) && !buffer_uptodate(bh))
1415 bh = __bread_slow(bh);
1418 EXPORT_SYMBOL(__bread_gfp);
1421 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1422 * This doesn't race because it runs in each cpu either in irq
1423 * or with preempt disabled.
1425 static void invalidate_bh_lru(void *arg)
1427 struct bh_lru *b = &get_cpu_var(bh_lrus);
1430 for (i = 0; i < BH_LRU_SIZE; i++) {
1434 put_cpu_var(bh_lrus);
1437 static bool has_bh_in_lru(int cpu, void *dummy)
1439 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1442 for (i = 0; i < BH_LRU_SIZE; i++) {
1450 void invalidate_bh_lrus(void)
1452 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1454 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1456 void set_bh_page(struct buffer_head *bh,
1457 struct page *page, unsigned long offset)
1460 BUG_ON(offset >= PAGE_SIZE);
1461 if (PageHighMem(page))
1463 * This catches illegal uses and preserves the offset:
1465 bh->b_data = (char *)(0 + offset);
1467 bh->b_data = page_address(page) + offset;
1469 EXPORT_SYMBOL(set_bh_page);
1472 * Called when truncating a buffer on a page completely.
1475 /* Bits that are cleared during an invalidate */
1476 #define BUFFER_FLAGS_DISCARD \
1477 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1478 1 << BH_Delay | 1 << BH_Unwritten)
1480 static void discard_buffer(struct buffer_head * bh)
1482 unsigned long b_state, b_state_old;
1485 clear_buffer_dirty(bh);
1487 b_state = bh->b_state;
1489 b_state_old = cmpxchg(&bh->b_state, b_state,
1490 (b_state & ~BUFFER_FLAGS_DISCARD));
1491 if (b_state_old == b_state)
1493 b_state = b_state_old;
1499 * block_invalidatepage - invalidate part or all of a buffer-backed page
1501 * @page: the page which is affected
1502 * @offset: start of the range to invalidate
1503 * @length: length of the range to invalidate
1505 * block_invalidatepage() is called when all or part of the page has become
1506 * invalidated by a truncate operation.
1508 * block_invalidatepage() does not have to release all buffers, but it must
1509 * ensure that no dirty buffer is left outside @offset and that no I/O
1510 * is underway against any of the blocks which are outside the truncation
1511 * point. Because the caller is about to free (and possibly reuse) those
1514 void block_invalidatepage(struct page *page, unsigned int offset,
1515 unsigned int length)
1517 struct buffer_head *head, *bh, *next;
1518 unsigned int curr_off = 0;
1519 unsigned int stop = length + offset;
1521 BUG_ON(!PageLocked(page));
1522 if (!page_has_buffers(page))
1526 * Check for overflow
1528 BUG_ON(stop > PAGE_SIZE || stop < length);
1530 head = page_buffers(page);
1533 unsigned int next_off = curr_off + bh->b_size;
1534 next = bh->b_this_page;
1537 * Are we still fully in range ?
1539 if (next_off > stop)
1543 * is this block fully invalidated?
1545 if (offset <= curr_off)
1547 curr_off = next_off;
1549 } while (bh != head);
1552 * We release buffers only if the entire page is being invalidated.
1553 * The get_block cached value has been unconditionally invalidated,
1554 * so real IO is not possible anymore.
1557 try_to_release_page(page, 0);
1561 EXPORT_SYMBOL(block_invalidatepage);
1565 * We attach and possibly dirty the buffers atomically wrt
1566 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1567 * is already excluded via the page lock.
1569 void create_empty_buffers(struct page *page,
1570 unsigned long blocksize, unsigned long b_state)
1572 struct buffer_head *bh, *head, *tail;
1574 head = alloc_page_buffers(page, blocksize, 1);
1577 bh->b_state |= b_state;
1579 bh = bh->b_this_page;
1581 tail->b_this_page = head;
1583 spin_lock(&page->mapping->private_lock);
1584 if (PageUptodate(page) || PageDirty(page)) {
1587 if (PageDirty(page))
1588 set_buffer_dirty(bh);
1589 if (PageUptodate(page))
1590 set_buffer_uptodate(bh);
1591 bh = bh->b_this_page;
1592 } while (bh != head);
1594 attach_page_buffers(page, head);
1595 spin_unlock(&page->mapping->private_lock);
1597 EXPORT_SYMBOL(create_empty_buffers);
1600 * We are taking a block for data and we don't want any output from any
1601 * buffer-cache aliases starting from return from that function and
1602 * until the moment when something will explicitly mark the buffer
1603 * dirty (hopefully that will not happen until we will free that block ;-)
1604 * We don't even need to mark it not-uptodate - nobody can expect
1605 * anything from a newly allocated buffer anyway. We used to used
1606 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1607 * don't want to mark the alias unmapped, for example - it would confuse
1608 * anyone who might pick it with bread() afterwards...
1610 * Also.. Note that bforget() doesn't lock the buffer. So there can
1611 * be writeout I/O going on against recently-freed buffers. We don't
1612 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1613 * only if we really need to. That happens here.
1615 void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1617 struct buffer_head *old_bh;
1621 old_bh = __find_get_block_slow(bdev, block);
1623 clear_buffer_dirty(old_bh);
1624 wait_on_buffer(old_bh);
1625 clear_buffer_req(old_bh);
1629 EXPORT_SYMBOL(unmap_underlying_metadata);
1632 * Size is a power-of-two in the range 512..PAGE_SIZE,
1633 * and the case we care about most is PAGE_SIZE.
1635 * So this *could* possibly be written with those
1636 * constraints in mind (relevant mostly if some
1637 * architecture has a slow bit-scan instruction)
1639 static inline int block_size_bits(unsigned int blocksize)
1641 return ilog2(blocksize);
1644 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1646 BUG_ON(!PageLocked(page));
1648 if (!page_has_buffers(page))
1649 create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1650 return page_buffers(page);
1654 * NOTE! All mapped/uptodate combinations are valid:
1656 * Mapped Uptodate Meaning
1658 * No No "unknown" - must do get_block()
1659 * No Yes "hole" - zero-filled
1660 * Yes No "allocated" - allocated on disk, not read in
1661 * Yes Yes "valid" - allocated and up-to-date in memory.
1663 * "Dirty" is valid only with the last case (mapped+uptodate).
1667 * While block_write_full_page is writing back the dirty buffers under
1668 * the page lock, whoever dirtied the buffers may decide to clean them
1669 * again at any time. We handle that by only looking at the buffer
1670 * state inside lock_buffer().
1672 * If block_write_full_page() is called for regular writeback
1673 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1674 * locked buffer. This only can happen if someone has written the buffer
1675 * directly, with submit_bh(). At the address_space level PageWriteback
1676 * prevents this contention from occurring.
1678 * If block_write_full_page() is called with wbc->sync_mode ==
1679 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1680 * causes the writes to be flagged as synchronous writes.
1682 int __block_write_full_page(struct inode *inode, struct page *page,
1683 get_block_t *get_block, struct writeback_control *wbc,
1684 bh_end_io_t *handler)
1688 sector_t last_block;
1689 struct buffer_head *bh, *head;
1690 unsigned int blocksize, bbits;
1691 int nr_underway = 0;
1692 int write_flags = (wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : 0);
1694 head = create_page_buffers(page, inode,
1695 (1 << BH_Dirty)|(1 << BH_Uptodate));
1698 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1699 * here, and the (potentially unmapped) buffers may become dirty at
1700 * any time. If a buffer becomes dirty here after we've inspected it
1701 * then we just miss that fact, and the page stays dirty.
1703 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1704 * handle that here by just cleaning them.
1708 blocksize = bh->b_size;
1709 bbits = block_size_bits(blocksize);
1711 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1712 last_block = (i_size_read(inode) - 1) >> bbits;
1715 * Get all the dirty buffers mapped to disk addresses and
1716 * handle any aliases from the underlying blockdev's mapping.
1719 if (block > last_block) {
1721 * mapped buffers outside i_size will occur, because
1722 * this page can be outside i_size when there is a
1723 * truncate in progress.
1726 * The buffer was zeroed by block_write_full_page()
1728 clear_buffer_dirty(bh);
1729 set_buffer_uptodate(bh);
1730 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1732 WARN_ON(bh->b_size != blocksize);
1733 err = get_block(inode, block, bh, 1);
1736 clear_buffer_delay(bh);
1737 if (buffer_new(bh)) {
1738 /* blockdev mappings never come here */
1739 clear_buffer_new(bh);
1740 unmap_underlying_metadata(bh->b_bdev,
1744 bh = bh->b_this_page;
1746 } while (bh != head);
1749 if (!buffer_mapped(bh))
1752 * If it's a fully non-blocking write attempt and we cannot
1753 * lock the buffer then redirty the page. Note that this can
1754 * potentially cause a busy-wait loop from writeback threads
1755 * and kswapd activity, but those code paths have their own
1756 * higher-level throttling.
1758 if (wbc->sync_mode != WB_SYNC_NONE) {
1760 } else if (!trylock_buffer(bh)) {
1761 redirty_page_for_writepage(wbc, page);
1764 if (test_clear_buffer_dirty(bh)) {
1765 mark_buffer_async_write_endio(bh, handler);
1769 } while ((bh = bh->b_this_page) != head);
1772 * The page and its buffers are protected by PageWriteback(), so we can
1773 * drop the bh refcounts early.
1775 BUG_ON(PageWriteback(page));
1776 set_page_writeback(page);
1779 struct buffer_head *next = bh->b_this_page;
1780 if (buffer_async_write(bh)) {
1781 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1785 } while (bh != head);
1790 if (nr_underway == 0) {
1792 * The page was marked dirty, but the buffers were
1793 * clean. Someone wrote them back by hand with
1794 * ll_rw_block/submit_bh. A rare case.
1796 end_page_writeback(page);
1799 * The page and buffer_heads can be released at any time from
1807 * ENOSPC, or some other error. We may already have added some
1808 * blocks to the file, so we need to write these out to avoid
1809 * exposing stale data.
1810 * The page is currently locked and not marked for writeback
1813 /* Recovery: lock and submit the mapped buffers */
1815 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1816 !buffer_delay(bh)) {
1818 mark_buffer_async_write_endio(bh, handler);
1821 * The buffer may have been set dirty during
1822 * attachment to a dirty page.
1824 clear_buffer_dirty(bh);
1826 } while ((bh = bh->b_this_page) != head);
1828 BUG_ON(PageWriteback(page));
1829 mapping_set_error(page->mapping, err);
1830 set_page_writeback(page);
1832 struct buffer_head *next = bh->b_this_page;
1833 if (buffer_async_write(bh)) {
1834 clear_buffer_dirty(bh);
1835 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, 0, wbc);
1839 } while (bh != head);
1843 EXPORT_SYMBOL(__block_write_full_page);
1846 * If a page has any new buffers, zero them out here, and mark them uptodate
1847 * and dirty so they'll be written out (in order to prevent uninitialised
1848 * block data from leaking). And clear the new bit.
1850 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1852 unsigned int block_start, block_end;
1853 struct buffer_head *head, *bh;
1855 BUG_ON(!PageLocked(page));
1856 if (!page_has_buffers(page))
1859 bh = head = page_buffers(page);
1862 block_end = block_start + bh->b_size;
1864 if (buffer_new(bh)) {
1865 if (block_end > from && block_start < to) {
1866 if (!PageUptodate(page)) {
1867 unsigned start, size;
1869 start = max(from, block_start);
1870 size = min(to, block_end) - start;
1872 zero_user(page, start, size);
1873 set_buffer_uptodate(bh);
1876 clear_buffer_new(bh);
1877 mark_buffer_dirty(bh);
1881 block_start = block_end;
1882 bh = bh->b_this_page;
1883 } while (bh != head);
1885 EXPORT_SYMBOL(page_zero_new_buffers);
1888 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1889 struct iomap *iomap)
1891 loff_t offset = block << inode->i_blkbits;
1893 bh->b_bdev = iomap->bdev;
1896 * Block points to offset in file we need to map, iomap contains
1897 * the offset at which the map starts. If the map ends before the
1898 * current block, then do not map the buffer and let the caller
1901 BUG_ON(offset >= iomap->offset + iomap->length);
1903 switch (iomap->type) {
1906 * If the buffer is not up to date or beyond the current EOF,
1907 * we need to mark it as new to ensure sub-block zeroing is
1908 * executed if necessary.
1910 if (!buffer_uptodate(bh) ||
1911 (offset >= i_size_read(inode)))
1914 case IOMAP_DELALLOC:
1915 if (!buffer_uptodate(bh) ||
1916 (offset >= i_size_read(inode)))
1918 set_buffer_uptodate(bh);
1919 set_buffer_mapped(bh);
1920 set_buffer_delay(bh);
1922 case IOMAP_UNWRITTEN:
1924 * For unwritten regions, we always need to ensure that
1925 * sub-block writes cause the regions in the block we are not
1926 * writing to are zeroed. Set the buffer as new to ensure this.
1929 set_buffer_unwritten(bh);
1932 if (offset >= i_size_read(inode))
1934 bh->b_blocknr = (iomap->blkno >> (inode->i_blkbits - 9)) +
1935 ((offset - iomap->offset) >> inode->i_blkbits);
1936 set_buffer_mapped(bh);
1941 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1942 get_block_t *get_block, struct iomap *iomap)
1944 unsigned from = pos & (PAGE_SIZE - 1);
1945 unsigned to = from + len;
1946 struct inode *inode = page->mapping->host;
1947 unsigned block_start, block_end;
1950 unsigned blocksize, bbits;
1951 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1953 BUG_ON(!PageLocked(page));
1954 BUG_ON(from > PAGE_SIZE);
1955 BUG_ON(to > PAGE_SIZE);
1958 head = create_page_buffers(page, inode, 0);
1959 blocksize = head->b_size;
1960 bbits = block_size_bits(blocksize);
1962 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1964 for(bh = head, block_start = 0; bh != head || !block_start;
1965 block++, block_start=block_end, bh = bh->b_this_page) {
1966 block_end = block_start + blocksize;
1967 if (block_end <= from || block_start >= to) {
1968 if (PageUptodate(page)) {
1969 if (!buffer_uptodate(bh))
1970 set_buffer_uptodate(bh);
1975 clear_buffer_new(bh);
1976 if (!buffer_mapped(bh)) {
1977 WARN_ON(bh->b_size != blocksize);
1979 err = get_block(inode, block, bh, 1);
1983 iomap_to_bh(inode, block, bh, iomap);
1986 if (buffer_new(bh)) {
1987 unmap_underlying_metadata(bh->b_bdev,
1989 if (PageUptodate(page)) {
1990 clear_buffer_new(bh);
1991 set_buffer_uptodate(bh);
1992 mark_buffer_dirty(bh);
1995 if (block_end > to || block_start < from)
1996 zero_user_segments(page,
2002 if (PageUptodate(page)) {
2003 if (!buffer_uptodate(bh))
2004 set_buffer_uptodate(bh);
2007 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2008 !buffer_unwritten(bh) &&
2009 (block_start < from || block_end > to)) {
2010 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2015 * If we issued read requests - let them complete.
2017 while(wait_bh > wait) {
2018 wait_on_buffer(*--wait_bh);
2019 if (!buffer_uptodate(*wait_bh))
2023 page_zero_new_buffers(page, from, to);
2027 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2028 get_block_t *get_block)
2030 return __block_write_begin_int(page, pos, len, get_block, NULL);
2032 EXPORT_SYMBOL(__block_write_begin);
2034 static int __block_commit_write(struct inode *inode, struct page *page,
2035 unsigned from, unsigned to)
2037 unsigned block_start, block_end;
2040 struct buffer_head *bh, *head;
2042 bh = head = page_buffers(page);
2043 blocksize = bh->b_size;
2047 block_end = block_start + blocksize;
2048 if (block_end <= from || block_start >= to) {
2049 if (!buffer_uptodate(bh))
2052 set_buffer_uptodate(bh);
2053 mark_buffer_dirty(bh);
2055 clear_buffer_new(bh);
2057 block_start = block_end;
2058 bh = bh->b_this_page;
2059 } while (bh != head);
2062 * If this is a partial write which happened to make all buffers
2063 * uptodate then we can optimize away a bogus readpage() for
2064 * the next read(). Here we 'discover' whether the page went
2065 * uptodate as a result of this (potentially partial) write.
2068 SetPageUptodate(page);
2073 * block_write_begin takes care of the basic task of block allocation and
2074 * bringing partial write blocks uptodate first.
2076 * The filesystem needs to handle block truncation upon failure.
2078 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2079 unsigned flags, struct page **pagep, get_block_t *get_block)
2081 pgoff_t index = pos >> PAGE_SHIFT;
2085 page = grab_cache_page_write_begin(mapping, index, flags);
2089 status = __block_write_begin(page, pos, len, get_block);
2090 if (unlikely(status)) {
2099 EXPORT_SYMBOL(block_write_begin);
2101 int block_write_end(struct file *file, struct address_space *mapping,
2102 loff_t pos, unsigned len, unsigned copied,
2103 struct page *page, void *fsdata)
2105 struct inode *inode = mapping->host;
2108 start = pos & (PAGE_SIZE - 1);
2110 if (unlikely(copied < len)) {
2112 * The buffers that were written will now be uptodate, so we
2113 * don't have to worry about a readpage reading them and
2114 * overwriting a partial write. However if we have encountered
2115 * a short write and only partially written into a buffer, it
2116 * will not be marked uptodate, so a readpage might come in and
2117 * destroy our partial write.
2119 * Do the simplest thing, and just treat any short write to a
2120 * non uptodate page as a zero-length write, and force the
2121 * caller to redo the whole thing.
2123 if (!PageUptodate(page))
2126 page_zero_new_buffers(page, start+copied, start+len);
2128 flush_dcache_page(page);
2130 /* This could be a short (even 0-length) commit */
2131 __block_commit_write(inode, page, start, start+copied);
2135 EXPORT_SYMBOL(block_write_end);
2137 int generic_write_end(struct file *file, struct address_space *mapping,
2138 loff_t pos, unsigned len, unsigned copied,
2139 struct page *page, void *fsdata)
2141 struct inode *inode = mapping->host;
2142 loff_t old_size = inode->i_size;
2143 int i_size_changed = 0;
2145 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2148 * No need to use i_size_read() here, the i_size
2149 * cannot change under us because we hold i_mutex.
2151 * But it's important to update i_size while still holding page lock:
2152 * page writeout could otherwise come in and zero beyond i_size.
2154 if (pos+copied > inode->i_size) {
2155 i_size_write(inode, pos+copied);
2163 pagecache_isize_extended(inode, old_size, pos);
2165 * Don't mark the inode dirty under page lock. First, it unnecessarily
2166 * makes the holding time of page lock longer. Second, it forces lock
2167 * ordering of page lock and transaction start for journaling
2171 mark_inode_dirty(inode);
2175 EXPORT_SYMBOL(generic_write_end);
2178 * block_is_partially_uptodate checks whether buffers within a page are
2181 * Returns true if all buffers which correspond to a file portion
2182 * we want to read are uptodate.
2184 int block_is_partially_uptodate(struct page *page, unsigned long from,
2185 unsigned long count)
2187 unsigned block_start, block_end, blocksize;
2189 struct buffer_head *bh, *head;
2192 if (!page_has_buffers(page))
2195 head = page_buffers(page);
2196 blocksize = head->b_size;
2197 to = min_t(unsigned, PAGE_SIZE - from, count);
2199 if (from < blocksize && to > PAGE_SIZE - blocksize)
2205 block_end = block_start + blocksize;
2206 if (block_end > from && block_start < to) {
2207 if (!buffer_uptodate(bh)) {
2211 if (block_end >= to)
2214 block_start = block_end;
2215 bh = bh->b_this_page;
2216 } while (bh != head);
2220 EXPORT_SYMBOL(block_is_partially_uptodate);
2223 * Generic "read page" function for block devices that have the normal
2224 * get_block functionality. This is most of the block device filesystems.
2225 * Reads the page asynchronously --- the unlock_buffer() and
2226 * set/clear_buffer_uptodate() functions propagate buffer state into the
2227 * page struct once IO has completed.
2229 int block_read_full_page(struct page *page, get_block_t *get_block)
2231 struct inode *inode = page->mapping->host;
2232 sector_t iblock, lblock;
2233 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2234 unsigned int blocksize, bbits;
2236 int fully_mapped = 1;
2238 head = create_page_buffers(page, inode, 0);
2239 blocksize = head->b_size;
2240 bbits = block_size_bits(blocksize);
2242 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2243 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2249 if (buffer_uptodate(bh))
2252 if (!buffer_mapped(bh)) {
2256 if (iblock < lblock) {
2257 WARN_ON(bh->b_size != blocksize);
2258 err = get_block(inode, iblock, bh, 0);
2262 if (!buffer_mapped(bh)) {
2263 zero_user(page, i * blocksize, blocksize);
2265 set_buffer_uptodate(bh);
2269 * get_block() might have updated the buffer
2272 if (buffer_uptodate(bh))
2276 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2279 SetPageMappedToDisk(page);
2283 * All buffers are uptodate - we can set the page uptodate
2284 * as well. But not if get_block() returned an error.
2286 if (!PageError(page))
2287 SetPageUptodate(page);
2292 /* Stage two: lock the buffers */
2293 for (i = 0; i < nr; i++) {
2296 mark_buffer_async_read(bh);
2300 * Stage 3: start the IO. Check for uptodateness
2301 * inside the buffer lock in case another process reading
2302 * the underlying blockdev brought it uptodate (the sct fix).
2304 for (i = 0; i < nr; i++) {
2306 if (buffer_uptodate(bh))
2307 end_buffer_async_read(bh, 1);
2309 submit_bh(REQ_OP_READ, 0, bh);
2313 EXPORT_SYMBOL(block_read_full_page);
2315 /* utility function for filesystems that need to do work on expanding
2316 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2317 * deal with the hole.
2319 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2321 struct address_space *mapping = inode->i_mapping;
2326 err = inode_newsize_ok(inode, size);
2330 err = pagecache_write_begin(NULL, mapping, size, 0,
2331 AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2336 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2342 EXPORT_SYMBOL(generic_cont_expand_simple);
2344 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2345 loff_t pos, loff_t *bytes)
2347 struct inode *inode = mapping->host;
2348 unsigned blocksize = 1 << inode->i_blkbits;
2351 pgoff_t index, curidx;
2353 unsigned zerofrom, offset, len;
2356 index = pos >> PAGE_SHIFT;
2357 offset = pos & ~PAGE_MASK;
2359 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2360 zerofrom = curpos & ~PAGE_MASK;
2361 if (zerofrom & (blocksize-1)) {
2362 *bytes |= (blocksize-1);
2365 len = PAGE_SIZE - zerofrom;
2367 err = pagecache_write_begin(file, mapping, curpos, len,
2368 AOP_FLAG_UNINTERRUPTIBLE,
2372 zero_user(page, zerofrom, len);
2373 err = pagecache_write_end(file, mapping, curpos, len, len,
2380 balance_dirty_pages_ratelimited(mapping);
2382 if (unlikely(fatal_signal_pending(current))) {
2388 /* page covers the boundary, find the boundary offset */
2389 if (index == curidx) {
2390 zerofrom = curpos & ~PAGE_MASK;
2391 /* if we will expand the thing last block will be filled */
2392 if (offset <= zerofrom) {
2395 if (zerofrom & (blocksize-1)) {
2396 *bytes |= (blocksize-1);
2399 len = offset - zerofrom;
2401 err = pagecache_write_begin(file, mapping, curpos, len,
2402 AOP_FLAG_UNINTERRUPTIBLE,
2406 zero_user(page, zerofrom, len);
2407 err = pagecache_write_end(file, mapping, curpos, len, len,
2419 * For moronic filesystems that do not allow holes in file.
2420 * We may have to extend the file.
2422 int cont_write_begin(struct file *file, struct address_space *mapping,
2423 loff_t pos, unsigned len, unsigned flags,
2424 struct page **pagep, void **fsdata,
2425 get_block_t *get_block, loff_t *bytes)
2427 struct inode *inode = mapping->host;
2428 unsigned blocksize = 1 << inode->i_blkbits;
2432 err = cont_expand_zero(file, mapping, pos, bytes);
2436 zerofrom = *bytes & ~PAGE_MASK;
2437 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2438 *bytes |= (blocksize-1);
2442 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2444 EXPORT_SYMBOL(cont_write_begin);
2446 int block_commit_write(struct page *page, unsigned from, unsigned to)
2448 struct inode *inode = page->mapping->host;
2449 __block_commit_write(inode,page,from,to);
2452 EXPORT_SYMBOL(block_commit_write);
2455 * block_page_mkwrite() is not allowed to change the file size as it gets
2456 * called from a page fault handler when a page is first dirtied. Hence we must
2457 * be careful to check for EOF conditions here. We set the page up correctly
2458 * for a written page which means we get ENOSPC checking when writing into
2459 * holes and correct delalloc and unwritten extent mapping on filesystems that
2460 * support these features.
2462 * We are not allowed to take the i_mutex here so we have to play games to
2463 * protect against truncate races as the page could now be beyond EOF. Because
2464 * truncate writes the inode size before removing pages, once we have the
2465 * page lock we can determine safely if the page is beyond EOF. If it is not
2466 * beyond EOF, then the page is guaranteed safe against truncation until we
2469 * Direct callers of this function should protect against filesystem freezing
2470 * using sb_start_pagefault() - sb_end_pagefault() functions.
2472 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2473 get_block_t get_block)
2475 struct page *page = vmf->page;
2476 struct inode *inode = file_inode(vma->vm_file);
2482 size = i_size_read(inode);
2483 if ((page->mapping != inode->i_mapping) ||
2484 (page_offset(page) > size)) {
2485 /* We overload EFAULT to mean page got truncated */
2490 /* page is wholly or partially inside EOF */
2491 if (((page->index + 1) << PAGE_SHIFT) > size)
2492 end = size & ~PAGE_MASK;
2496 ret = __block_write_begin(page, 0, end, get_block);
2498 ret = block_commit_write(page, 0, end);
2500 if (unlikely(ret < 0))
2502 set_page_dirty(page);
2503 wait_for_stable_page(page);
2509 EXPORT_SYMBOL(block_page_mkwrite);
2512 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2513 * immediately, while under the page lock. So it needs a special end_io
2514 * handler which does not touch the bh after unlocking it.
2516 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2518 __end_buffer_read_notouch(bh, uptodate);
2522 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2523 * the page (converting it to circular linked list and taking care of page
2526 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2528 struct buffer_head *bh;
2530 BUG_ON(!PageLocked(page));
2532 spin_lock(&page->mapping->private_lock);
2535 if (PageDirty(page))
2536 set_buffer_dirty(bh);
2537 if (!bh->b_this_page)
2538 bh->b_this_page = head;
2539 bh = bh->b_this_page;
2540 } while (bh != head);
2541 attach_page_buffers(page, head);
2542 spin_unlock(&page->mapping->private_lock);
2546 * On entry, the page is fully not uptodate.
2547 * On exit the page is fully uptodate in the areas outside (from,to)
2548 * The filesystem needs to handle block truncation upon failure.
2550 int nobh_write_begin(struct address_space *mapping,
2551 loff_t pos, unsigned len, unsigned flags,
2552 struct page **pagep, void **fsdata,
2553 get_block_t *get_block)
2555 struct inode *inode = mapping->host;
2556 const unsigned blkbits = inode->i_blkbits;
2557 const unsigned blocksize = 1 << blkbits;
2558 struct buffer_head *head, *bh;
2562 unsigned block_in_page;
2563 unsigned block_start, block_end;
2564 sector_t block_in_file;
2567 int is_mapped_to_disk = 1;
2569 index = pos >> PAGE_SHIFT;
2570 from = pos & (PAGE_SIZE - 1);
2573 page = grab_cache_page_write_begin(mapping, index, flags);
2579 if (page_has_buffers(page)) {
2580 ret = __block_write_begin(page, pos, len, get_block);
2586 if (PageMappedToDisk(page))
2590 * Allocate buffers so that we can keep track of state, and potentially
2591 * attach them to the page if an error occurs. In the common case of
2592 * no error, they will just be freed again without ever being attached
2593 * to the page (which is all OK, because we're under the page lock).
2595 * Be careful: the buffer linked list is a NULL terminated one, rather
2596 * than the circular one we're used to.
2598 head = alloc_page_buffers(page, blocksize, 0);
2604 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2607 * We loop across all blocks in the page, whether or not they are
2608 * part of the affected region. This is so we can discover if the
2609 * page is fully mapped-to-disk.
2611 for (block_start = 0, block_in_page = 0, bh = head;
2612 block_start < PAGE_SIZE;
2613 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2616 block_end = block_start + blocksize;
2619 if (block_start >= to)
2621 ret = get_block(inode, block_in_file + block_in_page,
2625 if (!buffer_mapped(bh))
2626 is_mapped_to_disk = 0;
2628 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2629 if (PageUptodate(page)) {
2630 set_buffer_uptodate(bh);
2633 if (buffer_new(bh) || !buffer_mapped(bh)) {
2634 zero_user_segments(page, block_start, from,
2638 if (buffer_uptodate(bh))
2639 continue; /* reiserfs does this */
2640 if (block_start < from || block_end > to) {
2642 bh->b_end_io = end_buffer_read_nobh;
2643 submit_bh(REQ_OP_READ, 0, bh);
2650 * The page is locked, so these buffers are protected from
2651 * any VM or truncate activity. Hence we don't need to care
2652 * for the buffer_head refcounts.
2654 for (bh = head; bh; bh = bh->b_this_page) {
2656 if (!buffer_uptodate(bh))
2663 if (is_mapped_to_disk)
2664 SetPageMappedToDisk(page);
2666 *fsdata = head; /* to be released by nobh_write_end */
2673 * Error recovery is a bit difficult. We need to zero out blocks that
2674 * were newly allocated, and dirty them to ensure they get written out.
2675 * Buffers need to be attached to the page at this point, otherwise
2676 * the handling of potential IO errors during writeout would be hard
2677 * (could try doing synchronous writeout, but what if that fails too?)
2679 attach_nobh_buffers(page, head);
2680 page_zero_new_buffers(page, from, to);
2689 EXPORT_SYMBOL(nobh_write_begin);
2691 int nobh_write_end(struct file *file, struct address_space *mapping,
2692 loff_t pos, unsigned len, unsigned copied,
2693 struct page *page, void *fsdata)
2695 struct inode *inode = page->mapping->host;
2696 struct buffer_head *head = fsdata;
2697 struct buffer_head *bh;
2698 BUG_ON(fsdata != NULL && page_has_buffers(page));
2700 if (unlikely(copied < len) && head)
2701 attach_nobh_buffers(page, head);
2702 if (page_has_buffers(page))
2703 return generic_write_end(file, mapping, pos, len,
2704 copied, page, fsdata);
2706 SetPageUptodate(page);
2707 set_page_dirty(page);
2708 if (pos+copied > inode->i_size) {
2709 i_size_write(inode, pos+copied);
2710 mark_inode_dirty(inode);
2718 head = head->b_this_page;
2719 free_buffer_head(bh);
2724 EXPORT_SYMBOL(nobh_write_end);
2727 * nobh_writepage() - based on block_full_write_page() except
2728 * that it tries to operate without attaching bufferheads to
2731 int nobh_writepage(struct page *page, get_block_t *get_block,
2732 struct writeback_control *wbc)
2734 struct inode * const inode = page->mapping->host;
2735 loff_t i_size = i_size_read(inode);
2736 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2740 /* Is the page fully inside i_size? */
2741 if (page->index < end_index)
2744 /* Is the page fully outside i_size? (truncate in progress) */
2745 offset = i_size & (PAGE_SIZE-1);
2746 if (page->index >= end_index+1 || !offset) {
2748 * The page may have dirty, unmapped buffers. For example,
2749 * they may have been added in ext3_writepage(). Make them
2750 * freeable here, so the page does not leak.
2753 /* Not really sure about this - do we need this ? */
2754 if (page->mapping->a_ops->invalidatepage)
2755 page->mapping->a_ops->invalidatepage(page, offset);
2758 return 0; /* don't care */
2762 * The page straddles i_size. It must be zeroed out on each and every
2763 * writepage invocation because it may be mmapped. "A file is mapped
2764 * in multiples of the page size. For a file that is not a multiple of
2765 * the page size, the remaining memory is zeroed when mapped, and
2766 * writes to that region are not written out to the file."
2768 zero_user_segment(page, offset, PAGE_SIZE);
2770 ret = mpage_writepage(page, get_block, wbc);
2772 ret = __block_write_full_page(inode, page, get_block, wbc,
2773 end_buffer_async_write);
2776 EXPORT_SYMBOL(nobh_writepage);
2778 int nobh_truncate_page(struct address_space *mapping,
2779 loff_t from, get_block_t *get_block)
2781 pgoff_t index = from >> PAGE_SHIFT;
2782 unsigned offset = from & (PAGE_SIZE-1);
2785 unsigned length, pos;
2786 struct inode *inode = mapping->host;
2788 struct buffer_head map_bh;
2791 blocksize = 1 << inode->i_blkbits;
2792 length = offset & (blocksize - 1);
2794 /* Block boundary? Nothing to do */
2798 length = blocksize - length;
2799 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2801 page = grab_cache_page(mapping, index);
2806 if (page_has_buffers(page)) {
2810 return block_truncate_page(mapping, from, get_block);
2813 /* Find the buffer that contains "offset" */
2815 while (offset >= pos) {
2820 map_bh.b_size = blocksize;
2822 err = get_block(inode, iblock, &map_bh, 0);
2825 /* unmapped? It's a hole - nothing to do */
2826 if (!buffer_mapped(&map_bh))
2829 /* Ok, it's mapped. Make sure it's up-to-date */
2830 if (!PageUptodate(page)) {
2831 err = mapping->a_ops->readpage(NULL, page);
2837 if (!PageUptodate(page)) {
2841 if (page_has_buffers(page))
2844 zero_user(page, offset, length);
2845 set_page_dirty(page);
2854 EXPORT_SYMBOL(nobh_truncate_page);
2856 int block_truncate_page(struct address_space *mapping,
2857 loff_t from, get_block_t *get_block)
2859 pgoff_t index = from >> PAGE_SHIFT;
2860 unsigned offset = from & (PAGE_SIZE-1);
2863 unsigned length, pos;
2864 struct inode *inode = mapping->host;
2866 struct buffer_head *bh;
2869 blocksize = 1 << inode->i_blkbits;
2870 length = offset & (blocksize - 1);
2872 /* Block boundary? Nothing to do */
2876 length = blocksize - length;
2877 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2879 page = grab_cache_page(mapping, index);
2884 if (!page_has_buffers(page))
2885 create_empty_buffers(page, blocksize, 0);
2887 /* Find the buffer that contains "offset" */
2888 bh = page_buffers(page);
2890 while (offset >= pos) {
2891 bh = bh->b_this_page;
2897 if (!buffer_mapped(bh)) {
2898 WARN_ON(bh->b_size != blocksize);
2899 err = get_block(inode, iblock, bh, 0);
2902 /* unmapped? It's a hole - nothing to do */
2903 if (!buffer_mapped(bh))
2907 /* Ok, it's mapped. Make sure it's up-to-date */
2908 if (PageUptodate(page))
2909 set_buffer_uptodate(bh);
2911 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2913 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2915 /* Uhhuh. Read error. Complain and punt. */
2916 if (!buffer_uptodate(bh))
2920 zero_user(page, offset, length);
2921 mark_buffer_dirty(bh);
2930 EXPORT_SYMBOL(block_truncate_page);
2933 * The generic ->writepage function for buffer-backed address_spaces
2935 int block_write_full_page(struct page *page, get_block_t *get_block,
2936 struct writeback_control *wbc)
2938 struct inode * const inode = page->mapping->host;
2939 loff_t i_size = i_size_read(inode);
2940 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2943 /* Is the page fully inside i_size? */
2944 if (page->index < end_index)
2945 return __block_write_full_page(inode, page, get_block, wbc,
2946 end_buffer_async_write);
2948 /* Is the page fully outside i_size? (truncate in progress) */
2949 offset = i_size & (PAGE_SIZE-1);
2950 if (page->index >= end_index+1 || !offset) {
2952 * The page may have dirty, unmapped buffers. For example,
2953 * they may have been added in ext3_writepage(). Make them
2954 * freeable here, so the page does not leak.
2956 do_invalidatepage(page, 0, PAGE_SIZE);
2958 return 0; /* don't care */
2962 * The page straddles i_size. It must be zeroed out on each and every
2963 * writepage invocation because it may be mmapped. "A file is mapped
2964 * in multiples of the page size. For a file that is not a multiple of
2965 * the page size, the remaining memory is zeroed when mapped, and
2966 * writes to that region are not written out to the file."
2968 zero_user_segment(page, offset, PAGE_SIZE);
2969 return __block_write_full_page(inode, page, get_block, wbc,
2970 end_buffer_async_write);
2972 EXPORT_SYMBOL(block_write_full_page);
2974 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2975 get_block_t *get_block)
2977 struct buffer_head tmp;
2978 struct inode *inode = mapping->host;
2981 tmp.b_size = 1 << inode->i_blkbits;
2982 get_block(inode, block, &tmp, 0);
2983 return tmp.b_blocknr;
2985 EXPORT_SYMBOL(generic_block_bmap);
2987 static void end_bio_bh_io_sync(struct bio *bio)
2989 struct buffer_head *bh = bio->bi_private;
2991 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2992 set_bit(BH_Quiet, &bh->b_state);
2994 bh->b_end_io(bh, !bio->bi_error);
2999 * This allows us to do IO even on the odd last sectors
3000 * of a device, even if the block size is some multiple
3001 * of the physical sector size.
3003 * We'll just truncate the bio to the size of the device,
3004 * and clear the end of the buffer head manually.
3006 * Truly out-of-range accesses will turn into actual IO
3007 * errors, this only handles the "we need to be able to
3008 * do IO at the final sector" case.
3010 void guard_bio_eod(int op, struct bio *bio)
3013 struct bio_vec *bvec = &bio->bi_io_vec[bio->bi_vcnt - 1];
3014 unsigned truncated_bytes;
3016 maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
3021 * If the *whole* IO is past the end of the device,
3022 * let it through, and the IO layer will turn it into
3025 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3028 maxsector -= bio->bi_iter.bi_sector;
3029 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3032 /* Uhhuh. We've got a bio that straddles the device size! */
3033 truncated_bytes = bio->bi_iter.bi_size - (maxsector << 9);
3035 /* Truncate the bio.. */
3036 bio->bi_iter.bi_size -= truncated_bytes;
3037 bvec->bv_len -= truncated_bytes;
3039 /* ..and clear the end of the buffer for reads */
3040 if (op == REQ_OP_READ) {
3041 zero_user(bvec->bv_page, bvec->bv_offset + bvec->bv_len,
3046 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3047 unsigned long bio_flags, struct writeback_control *wbc)
3051 BUG_ON(!buffer_locked(bh));
3052 BUG_ON(!buffer_mapped(bh));
3053 BUG_ON(!bh->b_end_io);
3054 BUG_ON(buffer_delay(bh));
3055 BUG_ON(buffer_unwritten(bh));
3058 * Only clear out a write error when rewriting
3060 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3061 clear_buffer_write_io_error(bh);
3064 * from here on down, it's all bio -- do the initial mapping,
3065 * submit_bio -> generic_make_request may further map this bio around
3067 bio = bio_alloc(GFP_NOIO, 1);
3070 wbc_init_bio(wbc, bio);
3071 wbc_account_io(wbc, bh->b_page, bh->b_size);
3074 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3075 bio->bi_bdev = bh->b_bdev;
3077 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3078 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3080 bio->bi_end_io = end_bio_bh_io_sync;
3081 bio->bi_private = bh;
3082 bio->bi_flags |= bio_flags;
3084 /* Take care of bh's that straddle the end of the device */
3085 guard_bio_eod(op, bio);
3087 if (buffer_meta(bh))
3088 op_flags |= REQ_META;
3089 if (buffer_prio(bh))
3090 op_flags |= REQ_PRIO;
3091 bio_set_op_attrs(bio, op, op_flags);
3097 int _submit_bh(int op, int op_flags, struct buffer_head *bh,
3098 unsigned long bio_flags)
3100 return submit_bh_wbc(op, op_flags, bh, bio_flags, NULL);
3102 EXPORT_SYMBOL_GPL(_submit_bh);
3104 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3106 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3108 EXPORT_SYMBOL(submit_bh);
3111 * ll_rw_block: low-level access to block devices (DEPRECATED)
3112 * @op: whether to %READ or %WRITE
3113 * @op_flags: rq_flag_bits
3114 * @nr: number of &struct buffer_heads in the array
3115 * @bhs: array of pointers to &struct buffer_head
3117 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3118 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3119 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3122 * This function drops any buffer that it cannot get a lock on (with the
3123 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3124 * request, and any buffer that appears to be up-to-date when doing read
3125 * request. Further it marks as clean buffers that are processed for
3126 * writing (the buffer cache won't assume that they are actually clean
3127 * until the buffer gets unlocked).
3129 * ll_rw_block sets b_end_io to simple completion handler that marks
3130 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3133 * All of the buffers must be for the same device, and must also be a
3134 * multiple of the current approved size for the device.
3136 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3140 for (i = 0; i < nr; i++) {
3141 struct buffer_head *bh = bhs[i];
3143 if (!trylock_buffer(bh))
3146 if (test_clear_buffer_dirty(bh)) {
3147 bh->b_end_io = end_buffer_write_sync;
3149 submit_bh(op, op_flags, bh);
3153 if (!buffer_uptodate(bh)) {
3154 bh->b_end_io = end_buffer_read_sync;
3156 submit_bh(op, op_flags, bh);
3163 EXPORT_SYMBOL(ll_rw_block);
3165 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3168 if (!test_clear_buffer_dirty(bh)) {
3172 bh->b_end_io = end_buffer_write_sync;
3174 submit_bh(REQ_OP_WRITE, op_flags, bh);
3176 EXPORT_SYMBOL(write_dirty_buffer);
3179 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3180 * and then start new I/O and then wait upon it. The caller must have a ref on
3183 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3187 WARN_ON(atomic_read(&bh->b_count) < 1);
3189 if (test_clear_buffer_dirty(bh)) {
3191 bh->b_end_io = end_buffer_write_sync;
3192 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3194 if (!ret && !buffer_uptodate(bh))
3201 EXPORT_SYMBOL(__sync_dirty_buffer);
3203 int sync_dirty_buffer(struct buffer_head *bh)
3205 return __sync_dirty_buffer(bh, WRITE_SYNC);
3207 EXPORT_SYMBOL(sync_dirty_buffer);
3210 * try_to_free_buffers() checks if all the buffers on this particular page
3211 * are unused, and releases them if so.
3213 * Exclusion against try_to_free_buffers may be obtained by either
3214 * locking the page or by holding its mapping's private_lock.
3216 * If the page is dirty but all the buffers are clean then we need to
3217 * be sure to mark the page clean as well. This is because the page
3218 * may be against a block device, and a later reattachment of buffers
3219 * to a dirty page will set *all* buffers dirty. Which would corrupt
3220 * filesystem data on the same device.
3222 * The same applies to regular filesystem pages: if all the buffers are
3223 * clean then we set the page clean and proceed. To do that, we require
3224 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3227 * try_to_free_buffers() is non-blocking.
3229 static inline int buffer_busy(struct buffer_head *bh)
3231 return atomic_read(&bh->b_count) |
3232 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3236 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3238 struct buffer_head *head = page_buffers(page);
3239 struct buffer_head *bh;
3243 if (buffer_write_io_error(bh) && page->mapping)
3244 mapping_set_error(page->mapping, -EIO);
3245 if (buffer_busy(bh))
3247 bh = bh->b_this_page;
3248 } while (bh != head);
3251 struct buffer_head *next = bh->b_this_page;
3253 if (bh->b_assoc_map)
3254 __remove_assoc_queue(bh);
3256 } while (bh != head);
3257 *buffers_to_free = head;
3258 __clear_page_buffers(page);
3264 int try_to_free_buffers(struct page *page)
3266 struct address_space * const mapping = page->mapping;
3267 struct buffer_head *buffers_to_free = NULL;
3270 BUG_ON(!PageLocked(page));
3271 if (PageWriteback(page))
3274 if (mapping == NULL) { /* can this still happen? */
3275 ret = drop_buffers(page, &buffers_to_free);
3279 spin_lock(&mapping->private_lock);
3280 ret = drop_buffers(page, &buffers_to_free);
3283 * If the filesystem writes its buffers by hand (eg ext3)
3284 * then we can have clean buffers against a dirty page. We
3285 * clean the page here; otherwise the VM will never notice
3286 * that the filesystem did any IO at all.
3288 * Also, during truncate, discard_buffer will have marked all
3289 * the page's buffers clean. We discover that here and clean
3292 * private_lock must be held over this entire operation in order
3293 * to synchronise against __set_page_dirty_buffers and prevent the
3294 * dirty bit from being lost.
3297 cancel_dirty_page(page);
3298 spin_unlock(&mapping->private_lock);
3300 if (buffers_to_free) {
3301 struct buffer_head *bh = buffers_to_free;
3304 struct buffer_head *next = bh->b_this_page;
3305 free_buffer_head(bh);
3307 } while (bh != buffers_to_free);
3311 EXPORT_SYMBOL(try_to_free_buffers);
3314 * There are no bdflush tunables left. But distributions are
3315 * still running obsolete flush daemons, so we terminate them here.
3317 * Use of bdflush() is deprecated and will be removed in a future kernel.
3318 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3320 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3322 static int msg_count;
3324 if (!capable(CAP_SYS_ADMIN))
3327 if (msg_count < 5) {
3330 "warning: process `%s' used the obsolete bdflush"
3331 " system call\n", current->comm);
3332 printk(KERN_INFO "Fix your initscripts?\n");
3341 * Buffer-head allocation
3343 static struct kmem_cache *bh_cachep __read_mostly;
3346 * Once the number of bh's in the machine exceeds this level, we start
3347 * stripping them in writeback.
3349 static unsigned long max_buffer_heads;
3351 int buffer_heads_over_limit;
3353 struct bh_accounting {
3354 int nr; /* Number of live bh's */
3355 int ratelimit; /* Limit cacheline bouncing */
3358 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3360 static void recalc_bh_state(void)
3365 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3367 __this_cpu_write(bh_accounting.ratelimit, 0);
3368 for_each_online_cpu(i)
3369 tot += per_cpu(bh_accounting, i).nr;
3370 buffer_heads_over_limit = (tot > max_buffer_heads);
3373 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3375 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3377 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3378 buffer_head_init_locks(ret);
3380 __this_cpu_inc(bh_accounting.nr);
3386 EXPORT_SYMBOL(alloc_buffer_head);
3388 void free_buffer_head(struct buffer_head *bh)
3390 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3391 kmem_cache_free(bh_cachep, bh);
3393 __this_cpu_dec(bh_accounting.nr);
3397 EXPORT_SYMBOL(free_buffer_head);
3399 static void buffer_exit_cpu(int cpu)
3402 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3404 for (i = 0; i < BH_LRU_SIZE; i++) {
3408 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3409 per_cpu(bh_accounting, cpu).nr = 0;
3412 static int buffer_cpu_notify(struct notifier_block *self,
3413 unsigned long action, void *hcpu)
3415 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3416 buffer_exit_cpu((unsigned long)hcpu);
3421 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3422 * @bh: struct buffer_head
3424 * Return true if the buffer is up-to-date and false,
3425 * with the buffer locked, if not.
3427 int bh_uptodate_or_lock(struct buffer_head *bh)
3429 if (!buffer_uptodate(bh)) {
3431 if (!buffer_uptodate(bh))
3437 EXPORT_SYMBOL(bh_uptodate_or_lock);
3440 * bh_submit_read - Submit a locked buffer for reading
3441 * @bh: struct buffer_head
3443 * Returns zero on success and -EIO on error.
3445 int bh_submit_read(struct buffer_head *bh)
3447 BUG_ON(!buffer_locked(bh));
3449 if (buffer_uptodate(bh)) {
3455 bh->b_end_io = end_buffer_read_sync;
3456 submit_bh(REQ_OP_READ, 0, bh);
3458 if (buffer_uptodate(bh))
3462 EXPORT_SYMBOL(bh_submit_read);
3464 void __init buffer_init(void)
3466 unsigned long nrpages;
3468 bh_cachep = kmem_cache_create("buffer_head",
3469 sizeof(struct buffer_head), 0,
3470 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3475 * Limit the bh occupancy to 10% of ZONE_NORMAL
3477 nrpages = (nr_free_buffer_pages() * 10) / 100;
3478 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3479 hotcpu_notifier(buffer_cpu_notify, 0);