1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 1991, 1992, 2002 Linus Torvalds
9 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
11 * Removed a lot of unnecessary code and simplified things now that
12 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
14 * Speed up hash, lru, and free list operations. Use gfp() for allocating
15 * hash table, use SLAB cache for buffer heads. SMP threading. -DaveM
17 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
19 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
22 #include <linux/kernel.h>
23 #include <linux/sched/signal.h>
24 #include <linux/syscalls.h>
26 #include <linux/iomap.h>
28 #include <linux/percpu.h>
29 #include <linux/slab.h>
30 #include <linux/capability.h>
31 #include <linux/blkdev.h>
32 #include <linux/file.h>
33 #include <linux/quotaops.h>
34 #include <linux/highmem.h>
35 #include <linux/export.h>
36 #include <linux/backing-dev.h>
37 #include <linux/writeback.h>
38 #include <linux/hash.h>
39 #include <linux/suspend.h>
40 #include <linux/buffer_head.h>
41 #include <linux/task_io_accounting_ops.h>
42 #include <linux/bio.h>
43 #include <linux/cpu.h>
44 #include <linux/bitops.h>
45 #include <linux/mpage.h>
46 #include <linux/bit_spinlock.h>
47 #include <linux/pagevec.h>
48 #include <linux/sched/mm.h>
49 #include <trace/events/block.h>
50 #include <linux/fscrypt.h>
54 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
55 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
56 enum rw_hint hint, struct writeback_control *wbc);
58 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
60 inline void touch_buffer(struct buffer_head *bh)
62 trace_block_touch_buffer(bh);
63 mark_page_accessed(bh->b_page);
65 EXPORT_SYMBOL(touch_buffer);
67 void __lock_buffer(struct buffer_head *bh)
69 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
71 EXPORT_SYMBOL(__lock_buffer);
73 void unlock_buffer(struct buffer_head *bh)
75 clear_bit_unlock(BH_Lock, &bh->b_state);
76 smp_mb__after_atomic();
77 wake_up_bit(&bh->b_state, BH_Lock);
79 EXPORT_SYMBOL(unlock_buffer);
82 * Returns if the page has dirty or writeback buffers. If all the buffers
83 * are unlocked and clean then the PageDirty information is stale. If
84 * any of the pages are locked, it is assumed they are locked for IO.
86 void buffer_check_dirty_writeback(struct page *page,
87 bool *dirty, bool *writeback)
89 struct buffer_head *head, *bh;
93 BUG_ON(!PageLocked(page));
95 if (!page_has_buffers(page))
98 if (PageWriteback(page))
101 head = page_buffers(page);
104 if (buffer_locked(bh))
107 if (buffer_dirty(bh))
110 bh = bh->b_this_page;
111 } while (bh != head);
113 EXPORT_SYMBOL(buffer_check_dirty_writeback);
116 * Block until a buffer comes unlocked. This doesn't stop it
117 * from becoming locked again - you have to lock it yourself
118 * if you want to preserve its state.
120 void __wait_on_buffer(struct buffer_head * bh)
122 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
124 EXPORT_SYMBOL(__wait_on_buffer);
126 static void buffer_io_error(struct buffer_head *bh, char *msg)
128 if (!test_bit(BH_Quiet, &bh->b_state))
129 printk_ratelimited(KERN_ERR
130 "Buffer I/O error on dev %pg, logical block %llu%s\n",
131 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
135 * End-of-IO handler helper function which does not touch the bh after
137 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
138 * a race there is benign: unlock_buffer() only use the bh's address for
139 * hashing after unlocking the buffer, so it doesn't actually touch the bh
142 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
145 set_buffer_uptodate(bh);
147 /* This happens, due to failed read-ahead attempts. */
148 clear_buffer_uptodate(bh);
154 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
155 * unlock the buffer. This is what ll_rw_block uses too.
157 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
159 __end_buffer_read_notouch(bh, uptodate);
162 EXPORT_SYMBOL(end_buffer_read_sync);
164 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
167 set_buffer_uptodate(bh);
169 buffer_io_error(bh, ", lost sync page write");
170 mark_buffer_write_io_error(bh);
171 clear_buffer_uptodate(bh);
176 EXPORT_SYMBOL(end_buffer_write_sync);
179 * Various filesystems appear to want __find_get_block to be non-blocking.
180 * But it's the page lock which protects the buffers. To get around this,
181 * we get exclusion from try_to_free_buffers with the blockdev mapping's
184 * Hack idea: for the blockdev mapping, private_lock contention
185 * may be quite high. This code could TryLock the page, and if that
186 * succeeds, there is no need to take private_lock.
188 static struct buffer_head *
189 __find_get_block_slow(struct block_device *bdev, sector_t block)
191 struct inode *bd_inode = bdev->bd_inode;
192 struct address_space *bd_mapping = bd_inode->i_mapping;
193 struct buffer_head *ret = NULL;
195 struct buffer_head *bh;
196 struct buffer_head *head;
199 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
201 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
202 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
206 spin_lock(&bd_mapping->private_lock);
207 if (!page_has_buffers(page))
209 head = page_buffers(page);
212 if (!buffer_mapped(bh))
214 else if (bh->b_blocknr == block) {
219 bh = bh->b_this_page;
220 } while (bh != head);
222 /* we might be here because some of the buffers on this page are
223 * not mapped. This is due to various races between
224 * file io on the block device and getblk. It gets dealt with
225 * elsewhere, don't buffer_error if we had some unmapped buffers
227 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
228 if (all_mapped && __ratelimit(&last_warned)) {
229 printk("__find_get_block_slow() failed. block=%llu, "
230 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
231 "device %pg blocksize: %d\n",
232 (unsigned long long)block,
233 (unsigned long long)bh->b_blocknr,
234 bh->b_state, bh->b_size, bdev,
235 1 << bd_inode->i_blkbits);
238 spin_unlock(&bd_mapping->private_lock);
244 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
247 struct buffer_head *first;
248 struct buffer_head *tmp;
250 int page_uptodate = 1;
252 BUG_ON(!buffer_async_read(bh));
256 set_buffer_uptodate(bh);
258 clear_buffer_uptodate(bh);
259 buffer_io_error(bh, ", async page read");
264 * Be _very_ careful from here on. Bad things can happen if
265 * two buffer heads end IO at almost the same time and both
266 * decide that the page is now completely done.
268 first = page_buffers(page);
269 spin_lock_irqsave(&first->b_uptodate_lock, flags);
270 clear_buffer_async_read(bh);
274 if (!buffer_uptodate(tmp))
276 if (buffer_async_read(tmp)) {
277 BUG_ON(!buffer_locked(tmp));
280 tmp = tmp->b_this_page;
282 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
285 * If none of the buffers had errors and they are all
286 * uptodate then we can set the page uptodate.
288 if (page_uptodate && !PageError(page))
289 SetPageUptodate(page);
294 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
298 struct decrypt_bh_ctx {
299 struct work_struct work;
300 struct buffer_head *bh;
303 static void decrypt_bh(struct work_struct *work)
305 struct decrypt_bh_ctx *ctx =
306 container_of(work, struct decrypt_bh_ctx, work);
307 struct buffer_head *bh = ctx->bh;
310 err = fscrypt_decrypt_pagecache_blocks(bh->b_page, bh->b_size,
312 end_buffer_async_read(bh, err == 0);
317 * I/O completion handler for block_read_full_page() - pages
318 * which come unlocked at the end of I/O.
320 static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
322 /* Decrypt if needed */
324 fscrypt_inode_uses_fs_layer_crypto(bh->b_page->mapping->host)) {
325 struct decrypt_bh_ctx *ctx = kmalloc(sizeof(*ctx), GFP_ATOMIC);
328 INIT_WORK(&ctx->work, decrypt_bh);
330 fscrypt_enqueue_decrypt_work(&ctx->work);
335 end_buffer_async_read(bh, uptodate);
339 * Completion handler for block_write_full_page() - pages which are unlocked
340 * during I/O, and which have PageWriteback cleared upon I/O completion.
342 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
345 struct buffer_head *first;
346 struct buffer_head *tmp;
349 BUG_ON(!buffer_async_write(bh));
353 set_buffer_uptodate(bh);
355 buffer_io_error(bh, ", lost async page write");
356 mark_buffer_write_io_error(bh);
357 clear_buffer_uptodate(bh);
361 first = page_buffers(page);
362 spin_lock_irqsave(&first->b_uptodate_lock, flags);
364 clear_buffer_async_write(bh);
366 tmp = bh->b_this_page;
368 if (buffer_async_write(tmp)) {
369 BUG_ON(!buffer_locked(tmp));
372 tmp = tmp->b_this_page;
374 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
375 end_page_writeback(page);
379 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
382 EXPORT_SYMBOL(end_buffer_async_write);
385 * If a page's buffers are under async readin (end_buffer_async_read
386 * completion) then there is a possibility that another thread of
387 * control could lock one of the buffers after it has completed
388 * but while some of the other buffers have not completed. This
389 * locked buffer would confuse end_buffer_async_read() into not unlocking
390 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
391 * that this buffer is not under async I/O.
393 * The page comes unlocked when it has no locked buffer_async buffers
396 * PageLocked prevents anyone starting new async I/O reads any of
399 * PageWriteback is used to prevent simultaneous writeout of the same
402 * PageLocked prevents anyone from starting writeback of a page which is
403 * under read I/O (PageWriteback is only ever set against a locked page).
405 static void mark_buffer_async_read(struct buffer_head *bh)
407 bh->b_end_io = end_buffer_async_read_io;
408 set_buffer_async_read(bh);
411 static void mark_buffer_async_write_endio(struct buffer_head *bh,
412 bh_end_io_t *handler)
414 bh->b_end_io = handler;
415 set_buffer_async_write(bh);
418 void mark_buffer_async_write(struct buffer_head *bh)
420 mark_buffer_async_write_endio(bh, end_buffer_async_write);
422 EXPORT_SYMBOL(mark_buffer_async_write);
426 * fs/buffer.c contains helper functions for buffer-backed address space's
427 * fsync functions. A common requirement for buffer-based filesystems is
428 * that certain data from the backing blockdev needs to be written out for
429 * a successful fsync(). For example, ext2 indirect blocks need to be
430 * written back and waited upon before fsync() returns.
432 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
433 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
434 * management of a list of dependent buffers at ->i_mapping->private_list.
436 * Locking is a little subtle: try_to_free_buffers() will remove buffers
437 * from their controlling inode's queue when they are being freed. But
438 * try_to_free_buffers() will be operating against the *blockdev* mapping
439 * at the time, not against the S_ISREG file which depends on those buffers.
440 * So the locking for private_list is via the private_lock in the address_space
441 * which backs the buffers. Which is different from the address_space
442 * against which the buffers are listed. So for a particular address_space,
443 * mapping->private_lock does *not* protect mapping->private_list! In fact,
444 * mapping->private_list will always be protected by the backing blockdev's
447 * Which introduces a requirement: all buffers on an address_space's
448 * ->private_list must be from the same address_space: the blockdev's.
450 * address_spaces which do not place buffers at ->private_list via these
451 * utility functions are free to use private_lock and private_list for
452 * whatever they want. The only requirement is that list_empty(private_list)
453 * be true at clear_inode() time.
455 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
456 * filesystems should do that. invalidate_inode_buffers() should just go
457 * BUG_ON(!list_empty).
459 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
460 * take an address_space, not an inode. And it should be called
461 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
464 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
465 * list if it is already on a list. Because if the buffer is on a list,
466 * it *must* already be on the right one. If not, the filesystem is being
467 * silly. This will save a ton of locking. But first we have to ensure
468 * that buffers are taken *off* the old inode's list when they are freed
469 * (presumably in truncate). That requires careful auditing of all
470 * filesystems (do it inside bforget()). It could also be done by bringing
475 * The buffer's backing address_space's private_lock must be held
477 static void __remove_assoc_queue(struct buffer_head *bh)
479 list_del_init(&bh->b_assoc_buffers);
480 WARN_ON(!bh->b_assoc_map);
481 bh->b_assoc_map = NULL;
484 int inode_has_buffers(struct inode *inode)
486 return !list_empty(&inode->i_data.private_list);
490 * osync is designed to support O_SYNC io. It waits synchronously for
491 * all already-submitted IO to complete, but does not queue any new
492 * writes to the disk.
494 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
495 * you dirty the buffers, and then use osync_inode_buffers to wait for
496 * completion. Any other dirty buffers which are not yet queued for
497 * write will not be flushed to disk by the osync.
499 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
501 struct buffer_head *bh;
507 list_for_each_prev(p, list) {
509 if (buffer_locked(bh)) {
513 if (!buffer_uptodate(bh))
524 void emergency_thaw_bdev(struct super_block *sb)
526 while (sb->s_bdev && !thaw_bdev(sb->s_bdev))
527 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
531 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
532 * @mapping: the mapping which wants those buffers written
534 * Starts I/O against the buffers at mapping->private_list, and waits upon
537 * Basically, this is a convenience function for fsync().
538 * @mapping is a file or directory which needs those buffers to be written for
539 * a successful fsync().
541 int sync_mapping_buffers(struct address_space *mapping)
543 struct address_space *buffer_mapping = mapping->private_data;
545 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
548 return fsync_buffers_list(&buffer_mapping->private_lock,
549 &mapping->private_list);
551 EXPORT_SYMBOL(sync_mapping_buffers);
554 * Called when we've recently written block `bblock', and it is known that
555 * `bblock' was for a buffer_boundary() buffer. This means that the block at
556 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
557 * dirty, schedule it for IO. So that indirects merge nicely with their data.
559 void write_boundary_block(struct block_device *bdev,
560 sector_t bblock, unsigned blocksize)
562 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
564 if (buffer_dirty(bh))
565 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
570 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
572 struct address_space *mapping = inode->i_mapping;
573 struct address_space *buffer_mapping = bh->b_page->mapping;
575 mark_buffer_dirty(bh);
576 if (!mapping->private_data) {
577 mapping->private_data = buffer_mapping;
579 BUG_ON(mapping->private_data != buffer_mapping);
581 if (!bh->b_assoc_map) {
582 spin_lock(&buffer_mapping->private_lock);
583 list_move_tail(&bh->b_assoc_buffers,
584 &mapping->private_list);
585 bh->b_assoc_map = mapping;
586 spin_unlock(&buffer_mapping->private_lock);
589 EXPORT_SYMBOL(mark_buffer_dirty_inode);
592 * Mark the page dirty, and set it dirty in the page cache, and mark the inode
595 * If warn is true, then emit a warning if the page is not uptodate and has
596 * not been truncated.
598 * The caller must hold lock_page_memcg().
600 void __set_page_dirty(struct page *page, struct address_space *mapping,
605 xa_lock_irqsave(&mapping->i_pages, flags);
606 if (page->mapping) { /* Race with truncate? */
607 WARN_ON_ONCE(warn && !PageUptodate(page));
608 account_page_dirtied(page, mapping);
609 __xa_set_mark(&mapping->i_pages, page_index(page),
610 PAGECACHE_TAG_DIRTY);
612 xa_unlock_irqrestore(&mapping->i_pages, flags);
614 EXPORT_SYMBOL_GPL(__set_page_dirty);
617 * Add a page to the dirty page list.
619 * It is a sad fact of life that this function is called from several places
620 * deeply under spinlocking. It may not sleep.
622 * If the page has buffers, the uptodate buffers are set dirty, to preserve
623 * dirty-state coherency between the page and the buffers. It the page does
624 * not have buffers then when they are later attached they will all be set
627 * The buffers are dirtied before the page is dirtied. There's a small race
628 * window in which a writepage caller may see the page cleanness but not the
629 * buffer dirtiness. That's fine. If this code were to set the page dirty
630 * before the buffers, a concurrent writepage caller could clear the page dirty
631 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
632 * page on the dirty page list.
634 * We use private_lock to lock against try_to_free_buffers while using the
635 * page's buffer list. Also use this to protect against clean buffers being
636 * added to the page after it was set dirty.
638 * FIXME: may need to call ->reservepage here as well. That's rather up to the
639 * address_space though.
641 int __set_page_dirty_buffers(struct page *page)
644 struct address_space *mapping = page_mapping(page);
646 if (unlikely(!mapping))
647 return !TestSetPageDirty(page);
649 spin_lock(&mapping->private_lock);
650 if (page_has_buffers(page)) {
651 struct buffer_head *head = page_buffers(page);
652 struct buffer_head *bh = head;
655 set_buffer_dirty(bh);
656 bh = bh->b_this_page;
657 } while (bh != head);
660 * Lock out page's memcg migration to keep PageDirty
661 * synchronized with per-memcg dirty page counters.
663 lock_page_memcg(page);
664 newly_dirty = !TestSetPageDirty(page);
665 spin_unlock(&mapping->private_lock);
668 __set_page_dirty(page, mapping, 1);
670 unlock_page_memcg(page);
673 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
677 EXPORT_SYMBOL(__set_page_dirty_buffers);
680 * Write out and wait upon a list of buffers.
682 * We have conflicting pressures: we want to make sure that all
683 * initially dirty buffers get waited on, but that any subsequently
684 * dirtied buffers don't. After all, we don't want fsync to last
685 * forever if somebody is actively writing to the file.
687 * Do this in two main stages: first we copy dirty buffers to a
688 * temporary inode list, queueing the writes as we go. Then we clean
689 * up, waiting for those writes to complete.
691 * During this second stage, any subsequent updates to the file may end
692 * up refiling the buffer on the original inode's dirty list again, so
693 * there is a chance we will end up with a buffer queued for write but
694 * not yet completed on that list. So, as a final cleanup we go through
695 * the osync code to catch these locked, dirty buffers without requeuing
696 * any newly dirty buffers for write.
698 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
700 struct buffer_head *bh;
701 struct list_head tmp;
702 struct address_space *mapping;
704 struct blk_plug plug;
706 INIT_LIST_HEAD(&tmp);
707 blk_start_plug(&plug);
710 while (!list_empty(list)) {
711 bh = BH_ENTRY(list->next);
712 mapping = bh->b_assoc_map;
713 __remove_assoc_queue(bh);
714 /* Avoid race with mark_buffer_dirty_inode() which does
715 * a lockless check and we rely on seeing the dirty bit */
717 if (buffer_dirty(bh) || buffer_locked(bh)) {
718 list_add(&bh->b_assoc_buffers, &tmp);
719 bh->b_assoc_map = mapping;
720 if (buffer_dirty(bh)) {
724 * Ensure any pending I/O completes so that
725 * write_dirty_buffer() actually writes the
726 * current contents - it is a noop if I/O is
727 * still in flight on potentially older
730 write_dirty_buffer(bh, REQ_SYNC);
733 * Kick off IO for the previous mapping. Note
734 * that we will not run the very last mapping,
735 * wait_on_buffer() will do that for us
736 * through sync_buffer().
745 blk_finish_plug(&plug);
748 while (!list_empty(&tmp)) {
749 bh = BH_ENTRY(tmp.prev);
751 mapping = bh->b_assoc_map;
752 __remove_assoc_queue(bh);
753 /* Avoid race with mark_buffer_dirty_inode() which does
754 * a lockless check and we rely on seeing the dirty bit */
756 if (buffer_dirty(bh)) {
757 list_add(&bh->b_assoc_buffers,
758 &mapping->private_list);
759 bh->b_assoc_map = mapping;
763 if (!buffer_uptodate(bh))
770 err2 = osync_buffers_list(lock, list);
778 * Invalidate any and all dirty buffers on a given inode. We are
779 * probably unmounting the fs, but that doesn't mean we have already
780 * done a sync(). Just drop the buffers from the inode list.
782 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
783 * assumes that all the buffers are against the blockdev. Not true
786 void invalidate_inode_buffers(struct inode *inode)
788 if (inode_has_buffers(inode)) {
789 struct address_space *mapping = &inode->i_data;
790 struct list_head *list = &mapping->private_list;
791 struct address_space *buffer_mapping = mapping->private_data;
793 spin_lock(&buffer_mapping->private_lock);
794 while (!list_empty(list))
795 __remove_assoc_queue(BH_ENTRY(list->next));
796 spin_unlock(&buffer_mapping->private_lock);
799 EXPORT_SYMBOL(invalidate_inode_buffers);
802 * Remove any clean buffers from the inode's buffer list. This is called
803 * when we're trying to free the inode itself. Those buffers can pin it.
805 * Returns true if all buffers were removed.
807 int remove_inode_buffers(struct inode *inode)
811 if (inode_has_buffers(inode)) {
812 struct address_space *mapping = &inode->i_data;
813 struct list_head *list = &mapping->private_list;
814 struct address_space *buffer_mapping = mapping->private_data;
816 spin_lock(&buffer_mapping->private_lock);
817 while (!list_empty(list)) {
818 struct buffer_head *bh = BH_ENTRY(list->next);
819 if (buffer_dirty(bh)) {
823 __remove_assoc_queue(bh);
825 spin_unlock(&buffer_mapping->private_lock);
831 * Create the appropriate buffers when given a page for data area and
832 * the size of each buffer.. Use the bh->b_this_page linked list to
833 * follow the buffers created. Return NULL if unable to create more
836 * The retry flag is used to differentiate async IO (paging, swapping)
837 * which may not fail from ordinary buffer allocations.
839 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
842 struct buffer_head *bh, *head;
843 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
845 struct mem_cgroup *memcg, *old_memcg;
850 /* The page lock pins the memcg */
851 memcg = page_memcg(page);
852 old_memcg = set_active_memcg(memcg);
856 while ((offset -= size) >= 0) {
857 bh = alloc_buffer_head(gfp);
861 bh->b_this_page = head;
867 /* Link the buffer to its page */
868 set_bh_page(bh, page, offset);
871 set_active_memcg(old_memcg);
874 * In case anything failed, we just free everything we got.
880 head = head->b_this_page;
881 free_buffer_head(bh);
887 EXPORT_SYMBOL_GPL(alloc_page_buffers);
890 link_dev_buffers(struct page *page, struct buffer_head *head)
892 struct buffer_head *bh, *tail;
897 bh = bh->b_this_page;
899 tail->b_this_page = head;
900 attach_page_private(page, head);
903 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
905 sector_t retval = ~((sector_t)0);
906 loff_t sz = i_size_read(bdev->bd_inode);
909 unsigned int sizebits = blksize_bits(size);
910 retval = (sz >> sizebits);
916 * Initialise the state of a blockdev page's buffers.
919 init_page_buffers(struct page *page, struct block_device *bdev,
920 sector_t block, int size)
922 struct buffer_head *head = page_buffers(page);
923 struct buffer_head *bh = head;
924 int uptodate = PageUptodate(page);
925 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
928 if (!buffer_mapped(bh)) {
930 bh->b_private = NULL;
932 bh->b_blocknr = block;
934 set_buffer_uptodate(bh);
935 if (block < end_block)
936 set_buffer_mapped(bh);
939 bh = bh->b_this_page;
940 } while (bh != head);
943 * Caller needs to validate requested block against end of device.
949 * Create the page-cache page that contains the requested block.
951 * This is used purely for blockdev mappings.
954 grow_dev_page(struct block_device *bdev, sector_t block,
955 pgoff_t index, int size, int sizebits, gfp_t gfp)
957 struct inode *inode = bdev->bd_inode;
959 struct buffer_head *bh;
964 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
967 * XXX: __getblk_slow() can not really deal with failure and
968 * will endlessly loop on improvised global reclaim. Prefer
969 * looping in the allocator rather than here, at least that
970 * code knows what it's doing.
972 gfp_mask |= __GFP_NOFAIL;
974 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
976 BUG_ON(!PageLocked(page));
978 if (page_has_buffers(page)) {
979 bh = page_buffers(page);
980 if (bh->b_size == size) {
981 end_block = init_page_buffers(page, bdev,
982 (sector_t)index << sizebits,
986 if (!try_to_free_buffers(page))
991 * Allocate some buffers for this page
993 bh = alloc_page_buffers(page, size, true);
996 * Link the page to the buffers and initialise them. Take the
997 * lock to be atomic wrt __find_get_block(), which does not
998 * run under the page lock.
1000 spin_lock(&inode->i_mapping->private_lock);
1001 link_dev_buffers(page, bh);
1002 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1004 spin_unlock(&inode->i_mapping->private_lock);
1006 ret = (block < end_block) ? 1 : -ENXIO;
1014 * Create buffers for the specified block device block's page. If
1015 * that page was dirty, the buffers are set dirty also.
1018 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1023 sizebits = PAGE_SHIFT - __ffs(size);
1024 index = block >> sizebits;
1027 * Check for a block which wants to lie outside our maximum possible
1028 * pagecache index. (this comparison is done using sector_t types).
1030 if (unlikely(index != block >> sizebits)) {
1031 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1033 __func__, (unsigned long long)block,
1038 /* Create a page with the proper size buffers.. */
1039 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1042 static struct buffer_head *
1043 __getblk_slow(struct block_device *bdev, sector_t block,
1044 unsigned size, gfp_t gfp)
1046 /* Size must be multiple of hard sectorsize */
1047 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1048 (size < 512 || size > PAGE_SIZE))) {
1049 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1051 printk(KERN_ERR "logical block size: %d\n",
1052 bdev_logical_block_size(bdev));
1059 struct buffer_head *bh;
1062 bh = __find_get_block(bdev, block, size);
1066 ret = grow_buffers(bdev, block, size, gfp);
1073 * The relationship between dirty buffers and dirty pages:
1075 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1076 * the page is tagged dirty in the page cache.
1078 * At all times, the dirtiness of the buffers represents the dirtiness of
1079 * subsections of the page. If the page has buffers, the page dirty bit is
1080 * merely a hint about the true dirty state.
1082 * When a page is set dirty in its entirety, all its buffers are marked dirty
1083 * (if the page has buffers).
1085 * When a buffer is marked dirty, its page is dirtied, but the page's other
1088 * Also. When blockdev buffers are explicitly read with bread(), they
1089 * individually become uptodate. But their backing page remains not
1090 * uptodate - even if all of its buffers are uptodate. A subsequent
1091 * block_read_full_page() against that page will discover all the uptodate
1092 * buffers, will set the page uptodate and will perform no I/O.
1096 * mark_buffer_dirty - mark a buffer_head as needing writeout
1097 * @bh: the buffer_head to mark dirty
1099 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1100 * its backing page dirty, then tag the page as dirty in the page cache
1101 * and then attach the address_space's inode to its superblock's dirty
1104 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1105 * i_pages lock and mapping->host->i_lock.
1107 void mark_buffer_dirty(struct buffer_head *bh)
1109 WARN_ON_ONCE(!buffer_uptodate(bh));
1111 trace_block_dirty_buffer(bh);
1114 * Very *carefully* optimize the it-is-already-dirty case.
1116 * Don't let the final "is it dirty" escape to before we
1117 * perhaps modified the buffer.
1119 if (buffer_dirty(bh)) {
1121 if (buffer_dirty(bh))
1125 if (!test_set_buffer_dirty(bh)) {
1126 struct page *page = bh->b_page;
1127 struct address_space *mapping = NULL;
1129 lock_page_memcg(page);
1130 if (!TestSetPageDirty(page)) {
1131 mapping = page_mapping(page);
1133 __set_page_dirty(page, mapping, 0);
1135 unlock_page_memcg(page);
1137 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1140 EXPORT_SYMBOL(mark_buffer_dirty);
1142 void mark_buffer_write_io_error(struct buffer_head *bh)
1144 struct super_block *sb;
1146 set_buffer_write_io_error(bh);
1147 /* FIXME: do we need to set this in both places? */
1148 if (bh->b_page && bh->b_page->mapping)
1149 mapping_set_error(bh->b_page->mapping, -EIO);
1150 if (bh->b_assoc_map)
1151 mapping_set_error(bh->b_assoc_map, -EIO);
1153 sb = READ_ONCE(bh->b_bdev->bd_super);
1155 errseq_set(&sb->s_wb_err, -EIO);
1158 EXPORT_SYMBOL(mark_buffer_write_io_error);
1161 * Decrement a buffer_head's reference count. If all buffers against a page
1162 * have zero reference count, are clean and unlocked, and if the page is clean
1163 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1164 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1165 * a page but it ends up not being freed, and buffers may later be reattached).
1167 void __brelse(struct buffer_head * buf)
1169 if (atomic_read(&buf->b_count)) {
1173 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1175 EXPORT_SYMBOL(__brelse);
1178 * bforget() is like brelse(), except it discards any
1179 * potentially dirty data.
1181 void __bforget(struct buffer_head *bh)
1183 clear_buffer_dirty(bh);
1184 if (bh->b_assoc_map) {
1185 struct address_space *buffer_mapping = bh->b_page->mapping;
1187 spin_lock(&buffer_mapping->private_lock);
1188 list_del_init(&bh->b_assoc_buffers);
1189 bh->b_assoc_map = NULL;
1190 spin_unlock(&buffer_mapping->private_lock);
1194 EXPORT_SYMBOL(__bforget);
1196 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1199 if (buffer_uptodate(bh)) {
1204 bh->b_end_io = end_buffer_read_sync;
1205 submit_bh(REQ_OP_READ, 0, bh);
1207 if (buffer_uptodate(bh))
1215 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1216 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1217 * refcount elevated by one when they're in an LRU. A buffer can only appear
1218 * once in a particular CPU's LRU. A single buffer can be present in multiple
1219 * CPU's LRUs at the same time.
1221 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1222 * sb_find_get_block().
1224 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1225 * a local interrupt disable for that.
1228 #define BH_LRU_SIZE 16
1231 struct buffer_head *bhs[BH_LRU_SIZE];
1234 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1237 #define bh_lru_lock() local_irq_disable()
1238 #define bh_lru_unlock() local_irq_enable()
1240 #define bh_lru_lock() preempt_disable()
1241 #define bh_lru_unlock() preempt_enable()
1244 static inline void check_irqs_on(void)
1246 #ifdef irqs_disabled
1247 BUG_ON(irqs_disabled());
1252 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1253 * inserted at the front, and the buffer_head at the back if any is evicted.
1254 * Or, if already in the LRU it is moved to the front.
1256 static void bh_lru_install(struct buffer_head *bh)
1258 struct buffer_head *evictee = bh;
1265 b = this_cpu_ptr(&bh_lrus);
1266 for (i = 0; i < BH_LRU_SIZE; i++) {
1267 swap(evictee, b->bhs[i]);
1268 if (evictee == bh) {
1280 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1282 static struct buffer_head *
1283 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1285 struct buffer_head *ret = NULL;
1290 for (i = 0; i < BH_LRU_SIZE; i++) {
1291 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1293 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1294 bh->b_size == size) {
1297 __this_cpu_write(bh_lrus.bhs[i],
1298 __this_cpu_read(bh_lrus.bhs[i - 1]));
1301 __this_cpu_write(bh_lrus.bhs[0], bh);
1313 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1314 * it in the LRU and mark it as accessed. If it is not present then return
1317 struct buffer_head *
1318 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1320 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1323 /* __find_get_block_slow will mark the page accessed */
1324 bh = __find_get_block_slow(bdev, block);
1332 EXPORT_SYMBOL(__find_get_block);
1335 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1336 * which corresponds to the passed block_device, block and size. The
1337 * returned buffer has its reference count incremented.
1339 * __getblk_gfp() will lock up the machine if grow_dev_page's
1340 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1342 struct buffer_head *
1343 __getblk_gfp(struct block_device *bdev, sector_t block,
1344 unsigned size, gfp_t gfp)
1346 struct buffer_head *bh = __find_get_block(bdev, block, size);
1350 bh = __getblk_slow(bdev, block, size, gfp);
1353 EXPORT_SYMBOL(__getblk_gfp);
1356 * Do async read-ahead on a buffer..
1358 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1360 struct buffer_head *bh = __getblk(bdev, block, size);
1362 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1366 EXPORT_SYMBOL(__breadahead);
1368 void __breadahead_gfp(struct block_device *bdev, sector_t block, unsigned size,
1371 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1373 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1377 EXPORT_SYMBOL(__breadahead_gfp);
1380 * __bread_gfp() - reads a specified block and returns the bh
1381 * @bdev: the block_device to read from
1382 * @block: number of block
1383 * @size: size (in bytes) to read
1384 * @gfp: page allocation flag
1386 * Reads a specified block, and returns buffer head that contains it.
1387 * The page cache can be allocated from non-movable area
1388 * not to prevent page migration if you set gfp to zero.
1389 * It returns NULL if the block was unreadable.
1391 struct buffer_head *
1392 __bread_gfp(struct block_device *bdev, sector_t block,
1393 unsigned size, gfp_t gfp)
1395 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1397 if (likely(bh) && !buffer_uptodate(bh))
1398 bh = __bread_slow(bh);
1401 EXPORT_SYMBOL(__bread_gfp);
1404 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1405 * This doesn't race because it runs in each cpu either in irq
1406 * or with preempt disabled.
1408 static void invalidate_bh_lru(void *arg)
1410 struct bh_lru *b = &get_cpu_var(bh_lrus);
1413 for (i = 0; i < BH_LRU_SIZE; i++) {
1417 put_cpu_var(bh_lrus);
1420 static bool has_bh_in_lru(int cpu, void *dummy)
1422 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1425 for (i = 0; i < BH_LRU_SIZE; i++) {
1433 void invalidate_bh_lrus(void)
1435 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1437 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1439 void set_bh_page(struct buffer_head *bh,
1440 struct page *page, unsigned long offset)
1443 BUG_ON(offset >= PAGE_SIZE);
1444 if (PageHighMem(page))
1446 * This catches illegal uses and preserves the offset:
1448 bh->b_data = (char *)(0 + offset);
1450 bh->b_data = page_address(page) + offset;
1452 EXPORT_SYMBOL(set_bh_page);
1455 * Called when truncating a buffer on a page completely.
1458 /* Bits that are cleared during an invalidate */
1459 #define BUFFER_FLAGS_DISCARD \
1460 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1461 1 << BH_Delay | 1 << BH_Unwritten)
1463 static void discard_buffer(struct buffer_head * bh)
1465 unsigned long b_state, b_state_old;
1468 clear_buffer_dirty(bh);
1470 b_state = bh->b_state;
1472 b_state_old = cmpxchg(&bh->b_state, b_state,
1473 (b_state & ~BUFFER_FLAGS_DISCARD));
1474 if (b_state_old == b_state)
1476 b_state = b_state_old;
1482 * block_invalidatepage - invalidate part or all of a buffer-backed page
1484 * @page: the page which is affected
1485 * @offset: start of the range to invalidate
1486 * @length: length of the range to invalidate
1488 * block_invalidatepage() is called when all or part of the page has become
1489 * invalidated by a truncate operation.
1491 * block_invalidatepage() does not have to release all buffers, but it must
1492 * ensure that no dirty buffer is left outside @offset and that no I/O
1493 * is underway against any of the blocks which are outside the truncation
1494 * point. Because the caller is about to free (and possibly reuse) those
1497 void block_invalidatepage(struct page *page, unsigned int offset,
1498 unsigned int length)
1500 struct buffer_head *head, *bh, *next;
1501 unsigned int curr_off = 0;
1502 unsigned int stop = length + offset;
1504 BUG_ON(!PageLocked(page));
1505 if (!page_has_buffers(page))
1509 * Check for overflow
1511 BUG_ON(stop > PAGE_SIZE || stop < length);
1513 head = page_buffers(page);
1516 unsigned int next_off = curr_off + bh->b_size;
1517 next = bh->b_this_page;
1520 * Are we still fully in range ?
1522 if (next_off > stop)
1526 * is this block fully invalidated?
1528 if (offset <= curr_off)
1530 curr_off = next_off;
1532 } while (bh != head);
1535 * We release buffers only if the entire page is being invalidated.
1536 * The get_block cached value has been unconditionally invalidated,
1537 * so real IO is not possible anymore.
1539 if (length == PAGE_SIZE)
1540 try_to_release_page(page, 0);
1544 EXPORT_SYMBOL(block_invalidatepage);
1548 * We attach and possibly dirty the buffers atomically wrt
1549 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1550 * is already excluded via the page lock.
1552 void create_empty_buffers(struct page *page,
1553 unsigned long blocksize, unsigned long b_state)
1555 struct buffer_head *bh, *head, *tail;
1557 head = alloc_page_buffers(page, blocksize, true);
1560 bh->b_state |= b_state;
1562 bh = bh->b_this_page;
1564 tail->b_this_page = head;
1566 spin_lock(&page->mapping->private_lock);
1567 if (PageUptodate(page) || PageDirty(page)) {
1570 if (PageDirty(page))
1571 set_buffer_dirty(bh);
1572 if (PageUptodate(page))
1573 set_buffer_uptodate(bh);
1574 bh = bh->b_this_page;
1575 } while (bh != head);
1577 attach_page_private(page, head);
1578 spin_unlock(&page->mapping->private_lock);
1580 EXPORT_SYMBOL(create_empty_buffers);
1583 * clean_bdev_aliases: clean a range of buffers in block device
1584 * @bdev: Block device to clean buffers in
1585 * @block: Start of a range of blocks to clean
1586 * @len: Number of blocks to clean
1588 * We are taking a range of blocks for data and we don't want writeback of any
1589 * buffer-cache aliases starting from return from this function and until the
1590 * moment when something will explicitly mark the buffer dirty (hopefully that
1591 * will not happen until we will free that block ;-) We don't even need to mark
1592 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1593 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1594 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1595 * would confuse anyone who might pick it with bread() afterwards...
1597 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1598 * writeout I/O going on against recently-freed buffers. We don't wait on that
1599 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1600 * need to. That happens here.
1602 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1604 struct inode *bd_inode = bdev->bd_inode;
1605 struct address_space *bd_mapping = bd_inode->i_mapping;
1606 struct pagevec pvec;
1607 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1610 struct buffer_head *bh;
1611 struct buffer_head *head;
1613 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1614 pagevec_init(&pvec);
1615 while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1616 count = pagevec_count(&pvec);
1617 for (i = 0; i < count; i++) {
1618 struct page *page = pvec.pages[i];
1620 if (!page_has_buffers(page))
1623 * We use page lock instead of bd_mapping->private_lock
1624 * to pin buffers here since we can afford to sleep and
1625 * it scales better than a global spinlock lock.
1628 /* Recheck when the page is locked which pins bhs */
1629 if (!page_has_buffers(page))
1631 head = page_buffers(page);
1634 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1636 if (bh->b_blocknr >= block + len)
1638 clear_buffer_dirty(bh);
1640 clear_buffer_req(bh);
1642 bh = bh->b_this_page;
1643 } while (bh != head);
1647 pagevec_release(&pvec);
1649 /* End of range already reached? */
1650 if (index > end || !index)
1654 EXPORT_SYMBOL(clean_bdev_aliases);
1657 * Size is a power-of-two in the range 512..PAGE_SIZE,
1658 * and the case we care about most is PAGE_SIZE.
1660 * So this *could* possibly be written with those
1661 * constraints in mind (relevant mostly if some
1662 * architecture has a slow bit-scan instruction)
1664 static inline int block_size_bits(unsigned int blocksize)
1666 return ilog2(blocksize);
1669 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1671 BUG_ON(!PageLocked(page));
1673 if (!page_has_buffers(page))
1674 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1676 return page_buffers(page);
1680 * NOTE! All mapped/uptodate combinations are valid:
1682 * Mapped Uptodate Meaning
1684 * No No "unknown" - must do get_block()
1685 * No Yes "hole" - zero-filled
1686 * Yes No "allocated" - allocated on disk, not read in
1687 * Yes Yes "valid" - allocated and up-to-date in memory.
1689 * "Dirty" is valid only with the last case (mapped+uptodate).
1693 * While block_write_full_page is writing back the dirty buffers under
1694 * the page lock, whoever dirtied the buffers may decide to clean them
1695 * again at any time. We handle that by only looking at the buffer
1696 * state inside lock_buffer().
1698 * If block_write_full_page() is called for regular writeback
1699 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1700 * locked buffer. This only can happen if someone has written the buffer
1701 * directly, with submit_bh(). At the address_space level PageWriteback
1702 * prevents this contention from occurring.
1704 * If block_write_full_page() is called with wbc->sync_mode ==
1705 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1706 * causes the writes to be flagged as synchronous writes.
1708 int __block_write_full_page(struct inode *inode, struct page *page,
1709 get_block_t *get_block, struct writeback_control *wbc,
1710 bh_end_io_t *handler)
1714 sector_t last_block;
1715 struct buffer_head *bh, *head;
1716 unsigned int blocksize, bbits;
1717 int nr_underway = 0;
1718 int write_flags = wbc_to_write_flags(wbc);
1720 head = create_page_buffers(page, inode,
1721 (1 << BH_Dirty)|(1 << BH_Uptodate));
1724 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1725 * here, and the (potentially unmapped) buffers may become dirty at
1726 * any time. If a buffer becomes dirty here after we've inspected it
1727 * then we just miss that fact, and the page stays dirty.
1729 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1730 * handle that here by just cleaning them.
1734 blocksize = bh->b_size;
1735 bbits = block_size_bits(blocksize);
1737 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1738 last_block = (i_size_read(inode) - 1) >> bbits;
1741 * Get all the dirty buffers mapped to disk addresses and
1742 * handle any aliases from the underlying blockdev's mapping.
1745 if (block > last_block) {
1747 * mapped buffers outside i_size will occur, because
1748 * this page can be outside i_size when there is a
1749 * truncate in progress.
1752 * The buffer was zeroed by block_write_full_page()
1754 clear_buffer_dirty(bh);
1755 set_buffer_uptodate(bh);
1756 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1758 WARN_ON(bh->b_size != blocksize);
1759 err = get_block(inode, block, bh, 1);
1762 clear_buffer_delay(bh);
1763 if (buffer_new(bh)) {
1764 /* blockdev mappings never come here */
1765 clear_buffer_new(bh);
1766 clean_bdev_bh_alias(bh);
1769 bh = bh->b_this_page;
1771 } while (bh != head);
1774 if (!buffer_mapped(bh))
1777 * If it's a fully non-blocking write attempt and we cannot
1778 * lock the buffer then redirty the page. Note that this can
1779 * potentially cause a busy-wait loop from writeback threads
1780 * and kswapd activity, but those code paths have their own
1781 * higher-level throttling.
1783 if (wbc->sync_mode != WB_SYNC_NONE) {
1785 } else if (!trylock_buffer(bh)) {
1786 redirty_page_for_writepage(wbc, page);
1789 if (test_clear_buffer_dirty(bh)) {
1790 mark_buffer_async_write_endio(bh, handler);
1794 } while ((bh = bh->b_this_page) != head);
1797 * The page and its buffers are protected by PageWriteback(), so we can
1798 * drop the bh refcounts early.
1800 BUG_ON(PageWriteback(page));
1801 set_page_writeback(page);
1804 struct buffer_head *next = bh->b_this_page;
1805 if (buffer_async_write(bh)) {
1806 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1807 inode->i_write_hint, wbc);
1811 } while (bh != head);
1816 if (nr_underway == 0) {
1818 * The page was marked dirty, but the buffers were
1819 * clean. Someone wrote them back by hand with
1820 * ll_rw_block/submit_bh. A rare case.
1822 end_page_writeback(page);
1825 * The page and buffer_heads can be released at any time from
1833 * ENOSPC, or some other error. We may already have added some
1834 * blocks to the file, so we need to write these out to avoid
1835 * exposing stale data.
1836 * The page is currently locked and not marked for writeback
1839 /* Recovery: lock and submit the mapped buffers */
1841 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1842 !buffer_delay(bh)) {
1844 mark_buffer_async_write_endio(bh, handler);
1847 * The buffer may have been set dirty during
1848 * attachment to a dirty page.
1850 clear_buffer_dirty(bh);
1852 } while ((bh = bh->b_this_page) != head);
1854 BUG_ON(PageWriteback(page));
1855 mapping_set_error(page->mapping, err);
1856 set_page_writeback(page);
1858 struct buffer_head *next = bh->b_this_page;
1859 if (buffer_async_write(bh)) {
1860 clear_buffer_dirty(bh);
1861 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1862 inode->i_write_hint, wbc);
1866 } while (bh != head);
1870 EXPORT_SYMBOL(__block_write_full_page);
1873 * If a page has any new buffers, zero them out here, and mark them uptodate
1874 * and dirty so they'll be written out (in order to prevent uninitialised
1875 * block data from leaking). And clear the new bit.
1877 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1879 unsigned int block_start, block_end;
1880 struct buffer_head *head, *bh;
1882 BUG_ON(!PageLocked(page));
1883 if (!page_has_buffers(page))
1886 bh = head = page_buffers(page);
1889 block_end = block_start + bh->b_size;
1891 if (buffer_new(bh)) {
1892 if (block_end > from && block_start < to) {
1893 if (!PageUptodate(page)) {
1894 unsigned start, size;
1896 start = max(from, block_start);
1897 size = min(to, block_end) - start;
1899 zero_user(page, start, size);
1900 set_buffer_uptodate(bh);
1903 clear_buffer_new(bh);
1904 mark_buffer_dirty(bh);
1908 block_start = block_end;
1909 bh = bh->b_this_page;
1910 } while (bh != head);
1912 EXPORT_SYMBOL(page_zero_new_buffers);
1915 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1916 struct iomap *iomap)
1918 loff_t offset = block << inode->i_blkbits;
1920 bh->b_bdev = iomap->bdev;
1923 * Block points to offset in file we need to map, iomap contains
1924 * the offset at which the map starts. If the map ends before the
1925 * current block, then do not map the buffer and let the caller
1928 BUG_ON(offset >= iomap->offset + iomap->length);
1930 switch (iomap->type) {
1933 * If the buffer is not up to date or beyond the current EOF,
1934 * we need to mark it as new to ensure sub-block zeroing is
1935 * executed if necessary.
1937 if (!buffer_uptodate(bh) ||
1938 (offset >= i_size_read(inode)))
1941 case IOMAP_DELALLOC:
1942 if (!buffer_uptodate(bh) ||
1943 (offset >= i_size_read(inode)))
1945 set_buffer_uptodate(bh);
1946 set_buffer_mapped(bh);
1947 set_buffer_delay(bh);
1949 case IOMAP_UNWRITTEN:
1951 * For unwritten regions, we always need to ensure that regions
1952 * in the block we are not writing to are zeroed. Mark the
1953 * buffer as new to ensure this.
1956 set_buffer_unwritten(bh);
1959 if ((iomap->flags & IOMAP_F_NEW) ||
1960 offset >= i_size_read(inode))
1962 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1964 set_buffer_mapped(bh);
1969 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1970 get_block_t *get_block, struct iomap *iomap)
1972 unsigned from = pos & (PAGE_SIZE - 1);
1973 unsigned to = from + len;
1974 struct inode *inode = page->mapping->host;
1975 unsigned block_start, block_end;
1978 unsigned blocksize, bbits;
1979 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1981 BUG_ON(!PageLocked(page));
1982 BUG_ON(from > PAGE_SIZE);
1983 BUG_ON(to > PAGE_SIZE);
1986 head = create_page_buffers(page, inode, 0);
1987 blocksize = head->b_size;
1988 bbits = block_size_bits(blocksize);
1990 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1992 for(bh = head, block_start = 0; bh != head || !block_start;
1993 block++, block_start=block_end, bh = bh->b_this_page) {
1994 block_end = block_start + blocksize;
1995 if (block_end <= from || block_start >= to) {
1996 if (PageUptodate(page)) {
1997 if (!buffer_uptodate(bh))
1998 set_buffer_uptodate(bh);
2003 clear_buffer_new(bh);
2004 if (!buffer_mapped(bh)) {
2005 WARN_ON(bh->b_size != blocksize);
2007 err = get_block(inode, block, bh, 1);
2011 iomap_to_bh(inode, block, bh, iomap);
2014 if (buffer_new(bh)) {
2015 clean_bdev_bh_alias(bh);
2016 if (PageUptodate(page)) {
2017 clear_buffer_new(bh);
2018 set_buffer_uptodate(bh);
2019 mark_buffer_dirty(bh);
2022 if (block_end > to || block_start < from)
2023 zero_user_segments(page,
2029 if (PageUptodate(page)) {
2030 if (!buffer_uptodate(bh))
2031 set_buffer_uptodate(bh);
2034 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2035 !buffer_unwritten(bh) &&
2036 (block_start < from || block_end > to)) {
2037 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2042 * If we issued read requests - let them complete.
2044 while(wait_bh > wait) {
2045 wait_on_buffer(*--wait_bh);
2046 if (!buffer_uptodate(*wait_bh))
2050 page_zero_new_buffers(page, from, to);
2054 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2055 get_block_t *get_block)
2057 return __block_write_begin_int(page, pos, len, get_block, NULL);
2059 EXPORT_SYMBOL(__block_write_begin);
2061 static int __block_commit_write(struct inode *inode, struct page *page,
2062 unsigned from, unsigned to)
2064 unsigned block_start, block_end;
2067 struct buffer_head *bh, *head;
2069 bh = head = page_buffers(page);
2070 blocksize = bh->b_size;
2074 block_end = block_start + blocksize;
2075 if (block_end <= from || block_start >= to) {
2076 if (!buffer_uptodate(bh))
2079 set_buffer_uptodate(bh);
2080 mark_buffer_dirty(bh);
2083 clear_buffer_new(bh);
2085 block_start = block_end;
2086 bh = bh->b_this_page;
2087 } while (bh != head);
2090 * If this is a partial write which happened to make all buffers
2091 * uptodate then we can optimize away a bogus readpage() for
2092 * the next read(). Here we 'discover' whether the page went
2093 * uptodate as a result of this (potentially partial) write.
2096 SetPageUptodate(page);
2101 * block_write_begin takes care of the basic task of block allocation and
2102 * bringing partial write blocks uptodate first.
2104 * The filesystem needs to handle block truncation upon failure.
2106 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2107 unsigned flags, struct page **pagep, get_block_t *get_block)
2109 pgoff_t index = pos >> PAGE_SHIFT;
2113 page = grab_cache_page_write_begin(mapping, index, flags);
2117 status = __block_write_begin(page, pos, len, get_block);
2118 if (unlikely(status)) {
2127 EXPORT_SYMBOL(block_write_begin);
2129 int block_write_end(struct file *file, struct address_space *mapping,
2130 loff_t pos, unsigned len, unsigned copied,
2131 struct page *page, void *fsdata)
2133 struct inode *inode = mapping->host;
2136 start = pos & (PAGE_SIZE - 1);
2138 if (unlikely(copied < len)) {
2140 * The buffers that were written will now be uptodate, so we
2141 * don't have to worry about a readpage reading them and
2142 * overwriting a partial write. However if we have encountered
2143 * a short write and only partially written into a buffer, it
2144 * will not be marked uptodate, so a readpage might come in and
2145 * destroy our partial write.
2147 * Do the simplest thing, and just treat any short write to a
2148 * non uptodate page as a zero-length write, and force the
2149 * caller to redo the whole thing.
2151 if (!PageUptodate(page))
2154 page_zero_new_buffers(page, start+copied, start+len);
2156 flush_dcache_page(page);
2158 /* This could be a short (even 0-length) commit */
2159 __block_commit_write(inode, page, start, start+copied);
2163 EXPORT_SYMBOL(block_write_end);
2165 int generic_write_end(struct file *file, struct address_space *mapping,
2166 loff_t pos, unsigned len, unsigned copied,
2167 struct page *page, void *fsdata)
2169 struct inode *inode = mapping->host;
2170 loff_t old_size = inode->i_size;
2171 bool i_size_changed = false;
2173 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2176 * No need to use i_size_read() here, the i_size cannot change under us
2177 * because we hold i_rwsem.
2179 * But it's important to update i_size while still holding page lock:
2180 * page writeout could otherwise come in and zero beyond i_size.
2182 if (pos + copied > inode->i_size) {
2183 i_size_write(inode, pos + copied);
2184 i_size_changed = true;
2191 pagecache_isize_extended(inode, old_size, pos);
2193 * Don't mark the inode dirty under page lock. First, it unnecessarily
2194 * makes the holding time of page lock longer. Second, it forces lock
2195 * ordering of page lock and transaction start for journaling
2199 mark_inode_dirty(inode);
2202 EXPORT_SYMBOL(generic_write_end);
2205 * block_is_partially_uptodate checks whether buffers within a page are
2208 * Returns true if all buffers which correspond to a file portion
2209 * we want to read are uptodate.
2211 int block_is_partially_uptodate(struct page *page, unsigned long from,
2212 unsigned long count)
2214 unsigned block_start, block_end, blocksize;
2216 struct buffer_head *bh, *head;
2219 if (!page_has_buffers(page))
2222 head = page_buffers(page);
2223 blocksize = head->b_size;
2224 to = min_t(unsigned, PAGE_SIZE - from, count);
2226 if (from < blocksize && to > PAGE_SIZE - blocksize)
2232 block_end = block_start + blocksize;
2233 if (block_end > from && block_start < to) {
2234 if (!buffer_uptodate(bh)) {
2238 if (block_end >= to)
2241 block_start = block_end;
2242 bh = bh->b_this_page;
2243 } while (bh != head);
2247 EXPORT_SYMBOL(block_is_partially_uptodate);
2250 * Generic "read page" function for block devices that have the normal
2251 * get_block functionality. This is most of the block device filesystems.
2252 * Reads the page asynchronously --- the unlock_buffer() and
2253 * set/clear_buffer_uptodate() functions propagate buffer state into the
2254 * page struct once IO has completed.
2256 int block_read_full_page(struct page *page, get_block_t *get_block)
2258 struct inode *inode = page->mapping->host;
2259 sector_t iblock, lblock;
2260 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2261 unsigned int blocksize, bbits;
2263 int fully_mapped = 1;
2265 head = create_page_buffers(page, inode, 0);
2266 blocksize = head->b_size;
2267 bbits = block_size_bits(blocksize);
2269 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2270 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2276 if (buffer_uptodate(bh))
2279 if (!buffer_mapped(bh)) {
2283 if (iblock < lblock) {
2284 WARN_ON(bh->b_size != blocksize);
2285 err = get_block(inode, iblock, bh, 0);
2289 if (!buffer_mapped(bh)) {
2290 zero_user(page, i * blocksize, blocksize);
2292 set_buffer_uptodate(bh);
2296 * get_block() might have updated the buffer
2299 if (buffer_uptodate(bh))
2303 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2306 SetPageMappedToDisk(page);
2310 * All buffers are uptodate - we can set the page uptodate
2311 * as well. But not if get_block() returned an error.
2313 if (!PageError(page))
2314 SetPageUptodate(page);
2319 /* Stage two: lock the buffers */
2320 for (i = 0; i < nr; i++) {
2323 mark_buffer_async_read(bh);
2327 * Stage 3: start the IO. Check for uptodateness
2328 * inside the buffer lock in case another process reading
2329 * the underlying blockdev brought it uptodate (the sct fix).
2331 for (i = 0; i < nr; i++) {
2333 if (buffer_uptodate(bh))
2334 end_buffer_async_read(bh, 1);
2336 submit_bh(REQ_OP_READ, 0, bh);
2340 EXPORT_SYMBOL(block_read_full_page);
2342 /* utility function for filesystems that need to do work on expanding
2343 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2344 * deal with the hole.
2346 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2348 struct address_space *mapping = inode->i_mapping;
2353 err = inode_newsize_ok(inode, size);
2357 err = pagecache_write_begin(NULL, mapping, size, 0,
2358 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2362 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2368 EXPORT_SYMBOL(generic_cont_expand_simple);
2370 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2371 loff_t pos, loff_t *bytes)
2373 struct inode *inode = mapping->host;
2374 unsigned int blocksize = i_blocksize(inode);
2377 pgoff_t index, curidx;
2379 unsigned zerofrom, offset, len;
2382 index = pos >> PAGE_SHIFT;
2383 offset = pos & ~PAGE_MASK;
2385 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2386 zerofrom = curpos & ~PAGE_MASK;
2387 if (zerofrom & (blocksize-1)) {
2388 *bytes |= (blocksize-1);
2391 len = PAGE_SIZE - zerofrom;
2393 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2397 zero_user(page, zerofrom, len);
2398 err = pagecache_write_end(file, mapping, curpos, len, len,
2405 balance_dirty_pages_ratelimited(mapping);
2407 if (fatal_signal_pending(current)) {
2413 /* page covers the boundary, find the boundary offset */
2414 if (index == curidx) {
2415 zerofrom = curpos & ~PAGE_MASK;
2416 /* if we will expand the thing last block will be filled */
2417 if (offset <= zerofrom) {
2420 if (zerofrom & (blocksize-1)) {
2421 *bytes |= (blocksize-1);
2424 len = offset - zerofrom;
2426 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2430 zero_user(page, zerofrom, len);
2431 err = pagecache_write_end(file, mapping, curpos, len, len,
2443 * For moronic filesystems that do not allow holes in file.
2444 * We may have to extend the file.
2446 int cont_write_begin(struct file *file, struct address_space *mapping,
2447 loff_t pos, unsigned len, unsigned flags,
2448 struct page **pagep, void **fsdata,
2449 get_block_t *get_block, loff_t *bytes)
2451 struct inode *inode = mapping->host;
2452 unsigned int blocksize = i_blocksize(inode);
2453 unsigned int zerofrom;
2456 err = cont_expand_zero(file, mapping, pos, bytes);
2460 zerofrom = *bytes & ~PAGE_MASK;
2461 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2462 *bytes |= (blocksize-1);
2466 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2468 EXPORT_SYMBOL(cont_write_begin);
2470 int block_commit_write(struct page *page, unsigned from, unsigned to)
2472 struct inode *inode = page->mapping->host;
2473 __block_commit_write(inode,page,from,to);
2476 EXPORT_SYMBOL(block_commit_write);
2479 * block_page_mkwrite() is not allowed to change the file size as it gets
2480 * called from a page fault handler when a page is first dirtied. Hence we must
2481 * be careful to check for EOF conditions here. We set the page up correctly
2482 * for a written page which means we get ENOSPC checking when writing into
2483 * holes and correct delalloc and unwritten extent mapping on filesystems that
2484 * support these features.
2486 * We are not allowed to take the i_mutex here so we have to play games to
2487 * protect against truncate races as the page could now be beyond EOF. Because
2488 * truncate writes the inode size before removing pages, once we have the
2489 * page lock we can determine safely if the page is beyond EOF. If it is not
2490 * beyond EOF, then the page is guaranteed safe against truncation until we
2493 * Direct callers of this function should protect against filesystem freezing
2494 * using sb_start_pagefault() - sb_end_pagefault() functions.
2496 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2497 get_block_t get_block)
2499 struct page *page = vmf->page;
2500 struct inode *inode = file_inode(vma->vm_file);
2506 size = i_size_read(inode);
2507 if ((page->mapping != inode->i_mapping) ||
2508 (page_offset(page) > size)) {
2509 /* We overload EFAULT to mean page got truncated */
2514 /* page is wholly or partially inside EOF */
2515 if (((page->index + 1) << PAGE_SHIFT) > size)
2516 end = size & ~PAGE_MASK;
2520 ret = __block_write_begin(page, 0, end, get_block);
2522 ret = block_commit_write(page, 0, end);
2524 if (unlikely(ret < 0))
2526 set_page_dirty(page);
2527 wait_for_stable_page(page);
2533 EXPORT_SYMBOL(block_page_mkwrite);
2536 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2537 * immediately, while under the page lock. So it needs a special end_io
2538 * handler which does not touch the bh after unlocking it.
2540 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2542 __end_buffer_read_notouch(bh, uptodate);
2546 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2547 * the page (converting it to circular linked list and taking care of page
2550 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2552 struct buffer_head *bh;
2554 BUG_ON(!PageLocked(page));
2556 spin_lock(&page->mapping->private_lock);
2559 if (PageDirty(page))
2560 set_buffer_dirty(bh);
2561 if (!bh->b_this_page)
2562 bh->b_this_page = head;
2563 bh = bh->b_this_page;
2564 } while (bh != head);
2565 attach_page_private(page, head);
2566 spin_unlock(&page->mapping->private_lock);
2570 * On entry, the page is fully not uptodate.
2571 * On exit the page is fully uptodate in the areas outside (from,to)
2572 * The filesystem needs to handle block truncation upon failure.
2574 int nobh_write_begin(struct address_space *mapping,
2575 loff_t pos, unsigned len, unsigned flags,
2576 struct page **pagep, void **fsdata,
2577 get_block_t *get_block)
2579 struct inode *inode = mapping->host;
2580 const unsigned blkbits = inode->i_blkbits;
2581 const unsigned blocksize = 1 << blkbits;
2582 struct buffer_head *head, *bh;
2586 unsigned block_in_page;
2587 unsigned block_start, block_end;
2588 sector_t block_in_file;
2591 int is_mapped_to_disk = 1;
2593 index = pos >> PAGE_SHIFT;
2594 from = pos & (PAGE_SIZE - 1);
2597 page = grab_cache_page_write_begin(mapping, index, flags);
2603 if (page_has_buffers(page)) {
2604 ret = __block_write_begin(page, pos, len, get_block);
2610 if (PageMappedToDisk(page))
2614 * Allocate buffers so that we can keep track of state, and potentially
2615 * attach them to the page if an error occurs. In the common case of
2616 * no error, they will just be freed again without ever being attached
2617 * to the page (which is all OK, because we're under the page lock).
2619 * Be careful: the buffer linked list is a NULL terminated one, rather
2620 * than the circular one we're used to.
2622 head = alloc_page_buffers(page, blocksize, false);
2628 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2631 * We loop across all blocks in the page, whether or not they are
2632 * part of the affected region. This is so we can discover if the
2633 * page is fully mapped-to-disk.
2635 for (block_start = 0, block_in_page = 0, bh = head;
2636 block_start < PAGE_SIZE;
2637 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2640 block_end = block_start + blocksize;
2643 if (block_start >= to)
2645 ret = get_block(inode, block_in_file + block_in_page,
2649 if (!buffer_mapped(bh))
2650 is_mapped_to_disk = 0;
2652 clean_bdev_bh_alias(bh);
2653 if (PageUptodate(page)) {
2654 set_buffer_uptodate(bh);
2657 if (buffer_new(bh) || !buffer_mapped(bh)) {
2658 zero_user_segments(page, block_start, from,
2662 if (buffer_uptodate(bh))
2663 continue; /* reiserfs does this */
2664 if (block_start < from || block_end > to) {
2666 bh->b_end_io = end_buffer_read_nobh;
2667 submit_bh(REQ_OP_READ, 0, bh);
2674 * The page is locked, so these buffers are protected from
2675 * any VM or truncate activity. Hence we don't need to care
2676 * for the buffer_head refcounts.
2678 for (bh = head; bh; bh = bh->b_this_page) {
2680 if (!buffer_uptodate(bh))
2687 if (is_mapped_to_disk)
2688 SetPageMappedToDisk(page);
2690 *fsdata = head; /* to be released by nobh_write_end */
2697 * Error recovery is a bit difficult. We need to zero out blocks that
2698 * were newly allocated, and dirty them to ensure they get written out.
2699 * Buffers need to be attached to the page at this point, otherwise
2700 * the handling of potential IO errors during writeout would be hard
2701 * (could try doing synchronous writeout, but what if that fails too?)
2703 attach_nobh_buffers(page, head);
2704 page_zero_new_buffers(page, from, to);
2713 EXPORT_SYMBOL(nobh_write_begin);
2715 int nobh_write_end(struct file *file, struct address_space *mapping,
2716 loff_t pos, unsigned len, unsigned copied,
2717 struct page *page, void *fsdata)
2719 struct inode *inode = page->mapping->host;
2720 struct buffer_head *head = fsdata;
2721 struct buffer_head *bh;
2722 BUG_ON(fsdata != NULL && page_has_buffers(page));
2724 if (unlikely(copied < len) && head)
2725 attach_nobh_buffers(page, head);
2726 if (page_has_buffers(page))
2727 return generic_write_end(file, mapping, pos, len,
2728 copied, page, fsdata);
2730 SetPageUptodate(page);
2731 set_page_dirty(page);
2732 if (pos+copied > inode->i_size) {
2733 i_size_write(inode, pos+copied);
2734 mark_inode_dirty(inode);
2742 head = head->b_this_page;
2743 free_buffer_head(bh);
2748 EXPORT_SYMBOL(nobh_write_end);
2751 * nobh_writepage() - based on block_full_write_page() except
2752 * that it tries to operate without attaching bufferheads to
2755 int nobh_writepage(struct page *page, get_block_t *get_block,
2756 struct writeback_control *wbc)
2758 struct inode * const inode = page->mapping->host;
2759 loff_t i_size = i_size_read(inode);
2760 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2764 /* Is the page fully inside i_size? */
2765 if (page->index < end_index)
2768 /* Is the page fully outside i_size? (truncate in progress) */
2769 offset = i_size & (PAGE_SIZE-1);
2770 if (page->index >= end_index+1 || !offset) {
2772 return 0; /* don't care */
2776 * The page straddles i_size. It must be zeroed out on each and every
2777 * writepage invocation because it may be mmapped. "A file is mapped
2778 * in multiples of the page size. For a file that is not a multiple of
2779 * the page size, the remaining memory is zeroed when mapped, and
2780 * writes to that region are not written out to the file."
2782 zero_user_segment(page, offset, PAGE_SIZE);
2784 ret = mpage_writepage(page, get_block, wbc);
2786 ret = __block_write_full_page(inode, page, get_block, wbc,
2787 end_buffer_async_write);
2790 EXPORT_SYMBOL(nobh_writepage);
2792 int nobh_truncate_page(struct address_space *mapping,
2793 loff_t from, get_block_t *get_block)
2795 pgoff_t index = from >> PAGE_SHIFT;
2796 unsigned offset = from & (PAGE_SIZE-1);
2799 unsigned length, pos;
2800 struct inode *inode = mapping->host;
2802 struct buffer_head map_bh;
2805 blocksize = i_blocksize(inode);
2806 length = offset & (blocksize - 1);
2808 /* Block boundary? Nothing to do */
2812 length = blocksize - length;
2813 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2815 page = grab_cache_page(mapping, index);
2820 if (page_has_buffers(page)) {
2824 return block_truncate_page(mapping, from, get_block);
2827 /* Find the buffer that contains "offset" */
2829 while (offset >= pos) {
2834 map_bh.b_size = blocksize;
2836 err = get_block(inode, iblock, &map_bh, 0);
2839 /* unmapped? It's a hole - nothing to do */
2840 if (!buffer_mapped(&map_bh))
2843 /* Ok, it's mapped. Make sure it's up-to-date */
2844 if (!PageUptodate(page)) {
2845 err = mapping->a_ops->readpage(NULL, page);
2851 if (!PageUptodate(page)) {
2855 if (page_has_buffers(page))
2858 zero_user(page, offset, length);
2859 set_page_dirty(page);
2868 EXPORT_SYMBOL(nobh_truncate_page);
2870 int block_truncate_page(struct address_space *mapping,
2871 loff_t from, get_block_t *get_block)
2873 pgoff_t index = from >> PAGE_SHIFT;
2874 unsigned offset = from & (PAGE_SIZE-1);
2877 unsigned length, pos;
2878 struct inode *inode = mapping->host;
2880 struct buffer_head *bh;
2883 blocksize = i_blocksize(inode);
2884 length = offset & (blocksize - 1);
2886 /* Block boundary? Nothing to do */
2890 length = blocksize - length;
2891 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2893 page = grab_cache_page(mapping, index);
2898 if (!page_has_buffers(page))
2899 create_empty_buffers(page, blocksize, 0);
2901 /* Find the buffer that contains "offset" */
2902 bh = page_buffers(page);
2904 while (offset >= pos) {
2905 bh = bh->b_this_page;
2911 if (!buffer_mapped(bh)) {
2912 WARN_ON(bh->b_size != blocksize);
2913 err = get_block(inode, iblock, bh, 0);
2916 /* unmapped? It's a hole - nothing to do */
2917 if (!buffer_mapped(bh))
2921 /* Ok, it's mapped. Make sure it's up-to-date */
2922 if (PageUptodate(page))
2923 set_buffer_uptodate(bh);
2925 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2927 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2929 /* Uhhuh. Read error. Complain and punt. */
2930 if (!buffer_uptodate(bh))
2934 zero_user(page, offset, length);
2935 mark_buffer_dirty(bh);
2944 EXPORT_SYMBOL(block_truncate_page);
2947 * The generic ->writepage function for buffer-backed address_spaces
2949 int block_write_full_page(struct page *page, get_block_t *get_block,
2950 struct writeback_control *wbc)
2952 struct inode * const inode = page->mapping->host;
2953 loff_t i_size = i_size_read(inode);
2954 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2957 /* Is the page fully inside i_size? */
2958 if (page->index < end_index)
2959 return __block_write_full_page(inode, page, get_block, wbc,
2960 end_buffer_async_write);
2962 /* Is the page fully outside i_size? (truncate in progress) */
2963 offset = i_size & (PAGE_SIZE-1);
2964 if (page->index >= end_index+1 || !offset) {
2966 return 0; /* don't care */
2970 * The page straddles i_size. It must be zeroed out on each and every
2971 * writepage invocation because it may be mmapped. "A file is mapped
2972 * in multiples of the page size. For a file that is not a multiple of
2973 * the page size, the remaining memory is zeroed when mapped, and
2974 * writes to that region are not written out to the file."
2976 zero_user_segment(page, offset, PAGE_SIZE);
2977 return __block_write_full_page(inode, page, get_block, wbc,
2978 end_buffer_async_write);
2980 EXPORT_SYMBOL(block_write_full_page);
2982 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2983 get_block_t *get_block)
2985 struct inode *inode = mapping->host;
2986 struct buffer_head tmp = {
2987 .b_size = i_blocksize(inode),
2990 get_block(inode, block, &tmp, 0);
2991 return tmp.b_blocknr;
2993 EXPORT_SYMBOL(generic_block_bmap);
2995 static void end_bio_bh_io_sync(struct bio *bio)
2997 struct buffer_head *bh = bio->bi_private;
2999 if (unlikely(bio_flagged(bio, BIO_QUIET)))
3000 set_bit(BH_Quiet, &bh->b_state);
3002 bh->b_end_io(bh, !bio->bi_status);
3006 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3007 enum rw_hint write_hint, struct writeback_control *wbc)
3011 BUG_ON(!buffer_locked(bh));
3012 BUG_ON(!buffer_mapped(bh));
3013 BUG_ON(!bh->b_end_io);
3014 BUG_ON(buffer_delay(bh));
3015 BUG_ON(buffer_unwritten(bh));
3018 * Only clear out a write error when rewriting
3020 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3021 clear_buffer_write_io_error(bh);
3023 bio = bio_alloc(GFP_NOIO, 1);
3025 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
3027 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3028 bio_set_dev(bio, bh->b_bdev);
3029 bio->bi_write_hint = write_hint;
3031 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3032 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3034 bio->bi_end_io = end_bio_bh_io_sync;
3035 bio->bi_private = bh;
3037 if (buffer_meta(bh))
3038 op_flags |= REQ_META;
3039 if (buffer_prio(bh))
3040 op_flags |= REQ_PRIO;
3041 bio_set_op_attrs(bio, op, op_flags);
3043 /* Take care of bh's that straddle the end of the device */
3047 wbc_init_bio(wbc, bio);
3048 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
3055 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3057 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3059 EXPORT_SYMBOL(submit_bh);
3062 * ll_rw_block: low-level access to block devices (DEPRECATED)
3063 * @op: whether to %READ or %WRITE
3064 * @op_flags: req_flag_bits
3065 * @nr: number of &struct buffer_heads in the array
3066 * @bhs: array of pointers to &struct buffer_head
3068 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3069 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3070 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3073 * This function drops any buffer that it cannot get a lock on (with the
3074 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3075 * request, and any buffer that appears to be up-to-date when doing read
3076 * request. Further it marks as clean buffers that are processed for
3077 * writing (the buffer cache won't assume that they are actually clean
3078 * until the buffer gets unlocked).
3080 * ll_rw_block sets b_end_io to simple completion handler that marks
3081 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3084 * All of the buffers must be for the same device, and must also be a
3085 * multiple of the current approved size for the device.
3087 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3091 for (i = 0; i < nr; i++) {
3092 struct buffer_head *bh = bhs[i];
3094 if (!trylock_buffer(bh))
3097 if (test_clear_buffer_dirty(bh)) {
3098 bh->b_end_io = end_buffer_write_sync;
3100 submit_bh(op, op_flags, bh);
3104 if (!buffer_uptodate(bh)) {
3105 bh->b_end_io = end_buffer_read_sync;
3107 submit_bh(op, op_flags, bh);
3114 EXPORT_SYMBOL(ll_rw_block);
3116 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3119 if (!test_clear_buffer_dirty(bh)) {
3123 bh->b_end_io = end_buffer_write_sync;
3125 submit_bh(REQ_OP_WRITE, op_flags, bh);
3127 EXPORT_SYMBOL(write_dirty_buffer);
3130 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3131 * and then start new I/O and then wait upon it. The caller must have a ref on
3134 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3138 WARN_ON(atomic_read(&bh->b_count) < 1);
3140 if (test_clear_buffer_dirty(bh)) {
3142 * The bh should be mapped, but it might not be if the
3143 * device was hot-removed. Not much we can do but fail the I/O.
3145 if (!buffer_mapped(bh)) {
3151 bh->b_end_io = end_buffer_write_sync;
3152 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3154 if (!ret && !buffer_uptodate(bh))
3161 EXPORT_SYMBOL(__sync_dirty_buffer);
3163 int sync_dirty_buffer(struct buffer_head *bh)
3165 return __sync_dirty_buffer(bh, REQ_SYNC);
3167 EXPORT_SYMBOL(sync_dirty_buffer);
3170 * try_to_free_buffers() checks if all the buffers on this particular page
3171 * are unused, and releases them if so.
3173 * Exclusion against try_to_free_buffers may be obtained by either
3174 * locking the page or by holding its mapping's private_lock.
3176 * If the page is dirty but all the buffers are clean then we need to
3177 * be sure to mark the page clean as well. This is because the page
3178 * may be against a block device, and a later reattachment of buffers
3179 * to a dirty page will set *all* buffers dirty. Which would corrupt
3180 * filesystem data on the same device.
3182 * The same applies to regular filesystem pages: if all the buffers are
3183 * clean then we set the page clean and proceed. To do that, we require
3184 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3187 * try_to_free_buffers() is non-blocking.
3189 static inline int buffer_busy(struct buffer_head *bh)
3191 return atomic_read(&bh->b_count) |
3192 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3196 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3198 struct buffer_head *head = page_buffers(page);
3199 struct buffer_head *bh;
3203 if (buffer_busy(bh))
3205 bh = bh->b_this_page;
3206 } while (bh != head);
3209 struct buffer_head *next = bh->b_this_page;
3211 if (bh->b_assoc_map)
3212 __remove_assoc_queue(bh);
3214 } while (bh != head);
3215 *buffers_to_free = head;
3216 detach_page_private(page);
3222 int try_to_free_buffers(struct page *page)
3224 struct address_space * const mapping = page->mapping;
3225 struct buffer_head *buffers_to_free = NULL;
3228 BUG_ON(!PageLocked(page));
3229 if (PageWriteback(page))
3232 if (mapping == NULL) { /* can this still happen? */
3233 ret = drop_buffers(page, &buffers_to_free);
3237 spin_lock(&mapping->private_lock);
3238 ret = drop_buffers(page, &buffers_to_free);
3241 * If the filesystem writes its buffers by hand (eg ext3)
3242 * then we can have clean buffers against a dirty page. We
3243 * clean the page here; otherwise the VM will never notice
3244 * that the filesystem did any IO at all.
3246 * Also, during truncate, discard_buffer will have marked all
3247 * the page's buffers clean. We discover that here and clean
3250 * private_lock must be held over this entire operation in order
3251 * to synchronise against __set_page_dirty_buffers and prevent the
3252 * dirty bit from being lost.
3255 cancel_dirty_page(page);
3256 spin_unlock(&mapping->private_lock);
3258 if (buffers_to_free) {
3259 struct buffer_head *bh = buffers_to_free;
3262 struct buffer_head *next = bh->b_this_page;
3263 free_buffer_head(bh);
3265 } while (bh != buffers_to_free);
3269 EXPORT_SYMBOL(try_to_free_buffers);
3272 * There are no bdflush tunables left. But distributions are
3273 * still running obsolete flush daemons, so we terminate them here.
3275 * Use of bdflush() is deprecated and will be removed in a future kernel.
3276 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3278 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3280 static int msg_count;
3282 if (!capable(CAP_SYS_ADMIN))
3285 if (msg_count < 5) {
3288 "warning: process `%s' used the obsolete bdflush"
3289 " system call\n", current->comm);
3290 printk(KERN_INFO "Fix your initscripts?\n");
3299 * Buffer-head allocation
3301 static struct kmem_cache *bh_cachep __read_mostly;
3304 * Once the number of bh's in the machine exceeds this level, we start
3305 * stripping them in writeback.
3307 static unsigned long max_buffer_heads;
3309 int buffer_heads_over_limit;
3311 struct bh_accounting {
3312 int nr; /* Number of live bh's */
3313 int ratelimit; /* Limit cacheline bouncing */
3316 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3318 static void recalc_bh_state(void)
3323 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3325 __this_cpu_write(bh_accounting.ratelimit, 0);
3326 for_each_online_cpu(i)
3327 tot += per_cpu(bh_accounting, i).nr;
3328 buffer_heads_over_limit = (tot > max_buffer_heads);
3331 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3333 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3335 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3336 spin_lock_init(&ret->b_uptodate_lock);
3338 __this_cpu_inc(bh_accounting.nr);
3344 EXPORT_SYMBOL(alloc_buffer_head);
3346 void free_buffer_head(struct buffer_head *bh)
3348 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3349 kmem_cache_free(bh_cachep, bh);
3351 __this_cpu_dec(bh_accounting.nr);
3355 EXPORT_SYMBOL(free_buffer_head);
3357 static int buffer_exit_cpu_dead(unsigned int cpu)
3360 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3362 for (i = 0; i < BH_LRU_SIZE; i++) {
3366 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3367 per_cpu(bh_accounting, cpu).nr = 0;
3372 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3373 * @bh: struct buffer_head
3375 * Return true if the buffer is up-to-date and false,
3376 * with the buffer locked, if not.
3378 int bh_uptodate_or_lock(struct buffer_head *bh)
3380 if (!buffer_uptodate(bh)) {
3382 if (!buffer_uptodate(bh))
3388 EXPORT_SYMBOL(bh_uptodate_or_lock);
3391 * bh_submit_read - Submit a locked buffer for reading
3392 * @bh: struct buffer_head
3394 * Returns zero on success and -EIO on error.
3396 int bh_submit_read(struct buffer_head *bh)
3398 BUG_ON(!buffer_locked(bh));
3400 if (buffer_uptodate(bh)) {
3406 bh->b_end_io = end_buffer_read_sync;
3407 submit_bh(REQ_OP_READ, 0, bh);
3409 if (buffer_uptodate(bh))
3413 EXPORT_SYMBOL(bh_submit_read);
3415 void __init buffer_init(void)
3417 unsigned long nrpages;
3420 bh_cachep = kmem_cache_create("buffer_head",
3421 sizeof(struct buffer_head), 0,
3422 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3427 * Limit the bh occupancy to 10% of ZONE_NORMAL
3429 nrpages = (nr_free_buffer_pages() * 10) / 100;
3430 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3431 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3432 NULL, buffer_exit_cpu_dead);