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);
127 __clear_page_buffers(struct page *page)
129 ClearPagePrivate(page);
130 set_page_private(page, 0);
134 static void buffer_io_error(struct buffer_head *bh, char *msg)
136 if (!test_bit(BH_Quiet, &bh->b_state))
137 printk_ratelimited(KERN_ERR
138 "Buffer I/O error on dev %pg, logical block %llu%s\n",
139 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
143 * End-of-IO handler helper function which does not touch the bh after
145 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
146 * a race there is benign: unlock_buffer() only use the bh's address for
147 * hashing after unlocking the buffer, so it doesn't actually touch the bh
150 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
153 set_buffer_uptodate(bh);
155 /* This happens, due to failed read-ahead attempts. */
156 clear_buffer_uptodate(bh);
162 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
163 * unlock the buffer. This is what ll_rw_block uses too.
165 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
167 __end_buffer_read_notouch(bh, uptodate);
170 EXPORT_SYMBOL(end_buffer_read_sync);
172 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
175 set_buffer_uptodate(bh);
177 buffer_io_error(bh, ", lost sync page write");
178 mark_buffer_write_io_error(bh);
179 clear_buffer_uptodate(bh);
184 EXPORT_SYMBOL(end_buffer_write_sync);
187 * Various filesystems appear to want __find_get_block to be non-blocking.
188 * But it's the page lock which protects the buffers. To get around this,
189 * we get exclusion from try_to_free_buffers with the blockdev mapping's
192 * Hack idea: for the blockdev mapping, private_lock contention
193 * may be quite high. This code could TryLock the page, and if that
194 * succeeds, there is no need to take private_lock.
196 static struct buffer_head *
197 __find_get_block_slow(struct block_device *bdev, sector_t block)
199 struct inode *bd_inode = bdev->bd_inode;
200 struct address_space *bd_mapping = bd_inode->i_mapping;
201 struct buffer_head *ret = NULL;
203 struct buffer_head *bh;
204 struct buffer_head *head;
207 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
209 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
210 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
214 spin_lock(&bd_mapping->private_lock);
215 if (!page_has_buffers(page))
217 head = page_buffers(page);
220 if (!buffer_mapped(bh))
222 else if (bh->b_blocknr == block) {
227 bh = bh->b_this_page;
228 } while (bh != head);
230 /* we might be here because some of the buffers on this page are
231 * not mapped. This is due to various races between
232 * file io on the block device and getblk. It gets dealt with
233 * elsewhere, don't buffer_error if we had some unmapped buffers
235 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
236 if (all_mapped && __ratelimit(&last_warned)) {
237 printk("__find_get_block_slow() failed. block=%llu, "
238 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
239 "device %pg blocksize: %d\n",
240 (unsigned long long)block,
241 (unsigned long long)bh->b_blocknr,
242 bh->b_state, bh->b_size, bdev,
243 1 << bd_inode->i_blkbits);
246 spin_unlock(&bd_mapping->private_lock);
252 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
255 struct buffer_head *first;
256 struct buffer_head *tmp;
258 int page_uptodate = 1;
260 BUG_ON(!buffer_async_read(bh));
264 set_buffer_uptodate(bh);
266 clear_buffer_uptodate(bh);
267 buffer_io_error(bh, ", async page read");
272 * Be _very_ careful from here on. Bad things can happen if
273 * two buffer heads end IO at almost the same time and both
274 * decide that the page is now completely done.
276 first = page_buffers(page);
277 spin_lock_irqsave(&first->b_uptodate_lock, flags);
278 clear_buffer_async_read(bh);
282 if (!buffer_uptodate(tmp))
284 if (buffer_async_read(tmp)) {
285 BUG_ON(!buffer_locked(tmp));
288 tmp = tmp->b_this_page;
290 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
293 * If none of the buffers had errors and they are all
294 * uptodate then we can set the page uptodate.
296 if (page_uptodate && !PageError(page))
297 SetPageUptodate(page);
302 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
306 struct decrypt_bh_ctx {
307 struct work_struct work;
308 struct buffer_head *bh;
311 static void decrypt_bh(struct work_struct *work)
313 struct decrypt_bh_ctx *ctx =
314 container_of(work, struct decrypt_bh_ctx, work);
315 struct buffer_head *bh = ctx->bh;
318 err = fscrypt_decrypt_pagecache_blocks(bh->b_page, bh->b_size,
320 end_buffer_async_read(bh, err == 0);
325 * I/O completion handler for block_read_full_page() - pages
326 * which come unlocked at the end of I/O.
328 static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
330 /* Decrypt if needed */
331 if (uptodate && IS_ENABLED(CONFIG_FS_ENCRYPTION) &&
332 IS_ENCRYPTED(bh->b_page->mapping->host) &&
333 S_ISREG(bh->b_page->mapping->host->i_mode)) {
334 struct decrypt_bh_ctx *ctx = kmalloc(sizeof(*ctx), GFP_ATOMIC);
337 INIT_WORK(&ctx->work, decrypt_bh);
339 fscrypt_enqueue_decrypt_work(&ctx->work);
344 end_buffer_async_read(bh, uptodate);
348 * Completion handler for block_write_full_page() - pages which are unlocked
349 * during I/O, and which have PageWriteback cleared upon I/O completion.
351 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
354 struct buffer_head *first;
355 struct buffer_head *tmp;
358 BUG_ON(!buffer_async_write(bh));
362 set_buffer_uptodate(bh);
364 buffer_io_error(bh, ", lost async page write");
365 mark_buffer_write_io_error(bh);
366 clear_buffer_uptodate(bh);
370 first = page_buffers(page);
371 spin_lock_irqsave(&first->b_uptodate_lock, flags);
373 clear_buffer_async_write(bh);
375 tmp = bh->b_this_page;
377 if (buffer_async_write(tmp)) {
378 BUG_ON(!buffer_locked(tmp));
381 tmp = tmp->b_this_page;
383 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
384 end_page_writeback(page);
388 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
391 EXPORT_SYMBOL(end_buffer_async_write);
394 * If a page's buffers are under async readin (end_buffer_async_read
395 * completion) then there is a possibility that another thread of
396 * control could lock one of the buffers after it has completed
397 * but while some of the other buffers have not completed. This
398 * locked buffer would confuse end_buffer_async_read() into not unlocking
399 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
400 * that this buffer is not under async I/O.
402 * The page comes unlocked when it has no locked buffer_async buffers
405 * PageLocked prevents anyone starting new async I/O reads any of
408 * PageWriteback is used to prevent simultaneous writeout of the same
411 * PageLocked prevents anyone from starting writeback of a page which is
412 * under read I/O (PageWriteback is only ever set against a locked page).
414 static void mark_buffer_async_read(struct buffer_head *bh)
416 bh->b_end_io = end_buffer_async_read_io;
417 set_buffer_async_read(bh);
420 static void mark_buffer_async_write_endio(struct buffer_head *bh,
421 bh_end_io_t *handler)
423 bh->b_end_io = handler;
424 set_buffer_async_write(bh);
427 void mark_buffer_async_write(struct buffer_head *bh)
429 mark_buffer_async_write_endio(bh, end_buffer_async_write);
431 EXPORT_SYMBOL(mark_buffer_async_write);
435 * fs/buffer.c contains helper functions for buffer-backed address space's
436 * fsync functions. A common requirement for buffer-based filesystems is
437 * that certain data from the backing blockdev needs to be written out for
438 * a successful fsync(). For example, ext2 indirect blocks need to be
439 * written back and waited upon before fsync() returns.
441 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
442 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
443 * management of a list of dependent buffers at ->i_mapping->private_list.
445 * Locking is a little subtle: try_to_free_buffers() will remove buffers
446 * from their controlling inode's queue when they are being freed. But
447 * try_to_free_buffers() will be operating against the *blockdev* mapping
448 * at the time, not against the S_ISREG file which depends on those buffers.
449 * So the locking for private_list is via the private_lock in the address_space
450 * which backs the buffers. Which is different from the address_space
451 * against which the buffers are listed. So for a particular address_space,
452 * mapping->private_lock does *not* protect mapping->private_list! In fact,
453 * mapping->private_list will always be protected by the backing blockdev's
456 * Which introduces a requirement: all buffers on an address_space's
457 * ->private_list must be from the same address_space: the blockdev's.
459 * address_spaces which do not place buffers at ->private_list via these
460 * utility functions are free to use private_lock and private_list for
461 * whatever they want. The only requirement is that list_empty(private_list)
462 * be true at clear_inode() time.
464 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
465 * filesystems should do that. invalidate_inode_buffers() should just go
466 * BUG_ON(!list_empty).
468 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
469 * take an address_space, not an inode. And it should be called
470 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
473 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
474 * list if it is already on a list. Because if the buffer is on a list,
475 * it *must* already be on the right one. If not, the filesystem is being
476 * silly. This will save a ton of locking. But first we have to ensure
477 * that buffers are taken *off* the old inode's list when they are freed
478 * (presumably in truncate). That requires careful auditing of all
479 * filesystems (do it inside bforget()). It could also be done by bringing
484 * The buffer's backing address_space's private_lock must be held
486 static void __remove_assoc_queue(struct buffer_head *bh)
488 list_del_init(&bh->b_assoc_buffers);
489 WARN_ON(!bh->b_assoc_map);
490 bh->b_assoc_map = NULL;
493 int inode_has_buffers(struct inode *inode)
495 return !list_empty(&inode->i_data.private_list);
499 * osync is designed to support O_SYNC io. It waits synchronously for
500 * all already-submitted IO to complete, but does not queue any new
501 * writes to the disk.
503 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
504 * you dirty the buffers, and then use osync_inode_buffers to wait for
505 * completion. Any other dirty buffers which are not yet queued for
506 * write will not be flushed to disk by the osync.
508 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
510 struct buffer_head *bh;
516 list_for_each_prev(p, list) {
518 if (buffer_locked(bh)) {
522 if (!buffer_uptodate(bh))
533 void emergency_thaw_bdev(struct super_block *sb)
535 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
536 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
540 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
541 * @mapping: the mapping which wants those buffers written
543 * Starts I/O against the buffers at mapping->private_list, and waits upon
546 * Basically, this is a convenience function for fsync().
547 * @mapping is a file or directory which needs those buffers to be written for
548 * a successful fsync().
550 int sync_mapping_buffers(struct address_space *mapping)
552 struct address_space *buffer_mapping = mapping->private_data;
554 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
557 return fsync_buffers_list(&buffer_mapping->private_lock,
558 &mapping->private_list);
560 EXPORT_SYMBOL(sync_mapping_buffers);
563 * Called when we've recently written block `bblock', and it is known that
564 * `bblock' was for a buffer_boundary() buffer. This means that the block at
565 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
566 * dirty, schedule it for IO. So that indirects merge nicely with their data.
568 void write_boundary_block(struct block_device *bdev,
569 sector_t bblock, unsigned blocksize)
571 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
573 if (buffer_dirty(bh))
574 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
579 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
581 struct address_space *mapping = inode->i_mapping;
582 struct address_space *buffer_mapping = bh->b_page->mapping;
584 mark_buffer_dirty(bh);
585 if (!mapping->private_data) {
586 mapping->private_data = buffer_mapping;
588 BUG_ON(mapping->private_data != buffer_mapping);
590 if (!bh->b_assoc_map) {
591 spin_lock(&buffer_mapping->private_lock);
592 list_move_tail(&bh->b_assoc_buffers,
593 &mapping->private_list);
594 bh->b_assoc_map = mapping;
595 spin_unlock(&buffer_mapping->private_lock);
598 EXPORT_SYMBOL(mark_buffer_dirty_inode);
601 * Mark the page dirty, and set it dirty in the page cache, and mark the inode
604 * If warn is true, then emit a warning if the page is not uptodate and has
605 * not been truncated.
607 * The caller must hold lock_page_memcg().
609 void __set_page_dirty(struct page *page, struct address_space *mapping,
614 xa_lock_irqsave(&mapping->i_pages, flags);
615 if (page->mapping) { /* Race with truncate? */
616 WARN_ON_ONCE(warn && !PageUptodate(page));
617 account_page_dirtied(page, mapping);
618 __xa_set_mark(&mapping->i_pages, page_index(page),
619 PAGECACHE_TAG_DIRTY);
621 xa_unlock_irqrestore(&mapping->i_pages, flags);
623 EXPORT_SYMBOL_GPL(__set_page_dirty);
626 * Add a page to the dirty page list.
628 * It is a sad fact of life that this function is called from several places
629 * deeply under spinlocking. It may not sleep.
631 * If the page has buffers, the uptodate buffers are set dirty, to preserve
632 * dirty-state coherency between the page and the buffers. It the page does
633 * not have buffers then when they are later attached they will all be set
636 * The buffers are dirtied before the page is dirtied. There's a small race
637 * window in which a writepage caller may see the page cleanness but not the
638 * buffer dirtiness. That's fine. If this code were to set the page dirty
639 * before the buffers, a concurrent writepage caller could clear the page dirty
640 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
641 * page on the dirty page list.
643 * We use private_lock to lock against try_to_free_buffers while using the
644 * page's buffer list. Also use this to protect against clean buffers being
645 * added to the page after it was set dirty.
647 * FIXME: may need to call ->reservepage here as well. That's rather up to the
648 * address_space though.
650 int __set_page_dirty_buffers(struct page *page)
653 struct address_space *mapping = page_mapping(page);
655 if (unlikely(!mapping))
656 return !TestSetPageDirty(page);
658 spin_lock(&mapping->private_lock);
659 if (page_has_buffers(page)) {
660 struct buffer_head *head = page_buffers(page);
661 struct buffer_head *bh = head;
664 set_buffer_dirty(bh);
665 bh = bh->b_this_page;
666 } while (bh != head);
669 * Lock out page->mem_cgroup migration to keep PageDirty
670 * synchronized with per-memcg dirty page counters.
672 lock_page_memcg(page);
673 newly_dirty = !TestSetPageDirty(page);
674 spin_unlock(&mapping->private_lock);
677 __set_page_dirty(page, mapping, 1);
679 unlock_page_memcg(page);
682 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
686 EXPORT_SYMBOL(__set_page_dirty_buffers);
689 * Write out and wait upon a list of buffers.
691 * We have conflicting pressures: we want to make sure that all
692 * initially dirty buffers get waited on, but that any subsequently
693 * dirtied buffers don't. After all, we don't want fsync to last
694 * forever if somebody is actively writing to the file.
696 * Do this in two main stages: first we copy dirty buffers to a
697 * temporary inode list, queueing the writes as we go. Then we clean
698 * up, waiting for those writes to complete.
700 * During this second stage, any subsequent updates to the file may end
701 * up refiling the buffer on the original inode's dirty list again, so
702 * there is a chance we will end up with a buffer queued for write but
703 * not yet completed on that list. So, as a final cleanup we go through
704 * the osync code to catch these locked, dirty buffers without requeuing
705 * any newly dirty buffers for write.
707 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
709 struct buffer_head *bh;
710 struct list_head tmp;
711 struct address_space *mapping;
713 struct blk_plug plug;
715 INIT_LIST_HEAD(&tmp);
716 blk_start_plug(&plug);
719 while (!list_empty(list)) {
720 bh = BH_ENTRY(list->next);
721 mapping = bh->b_assoc_map;
722 __remove_assoc_queue(bh);
723 /* Avoid race with mark_buffer_dirty_inode() which does
724 * a lockless check and we rely on seeing the dirty bit */
726 if (buffer_dirty(bh) || buffer_locked(bh)) {
727 list_add(&bh->b_assoc_buffers, &tmp);
728 bh->b_assoc_map = mapping;
729 if (buffer_dirty(bh)) {
733 * Ensure any pending I/O completes so that
734 * write_dirty_buffer() actually writes the
735 * current contents - it is a noop if I/O is
736 * still in flight on potentially older
739 write_dirty_buffer(bh, REQ_SYNC);
742 * Kick off IO for the previous mapping. Note
743 * that we will not run the very last mapping,
744 * wait_on_buffer() will do that for us
745 * through sync_buffer().
754 blk_finish_plug(&plug);
757 while (!list_empty(&tmp)) {
758 bh = BH_ENTRY(tmp.prev);
760 mapping = bh->b_assoc_map;
761 __remove_assoc_queue(bh);
762 /* Avoid race with mark_buffer_dirty_inode() which does
763 * a lockless check and we rely on seeing the dirty bit */
765 if (buffer_dirty(bh)) {
766 list_add(&bh->b_assoc_buffers,
767 &mapping->private_list);
768 bh->b_assoc_map = mapping;
772 if (!buffer_uptodate(bh))
779 err2 = osync_buffers_list(lock, list);
787 * Invalidate any and all dirty buffers on a given inode. We are
788 * probably unmounting the fs, but that doesn't mean we have already
789 * done a sync(). Just drop the buffers from the inode list.
791 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
792 * assumes that all the buffers are against the blockdev. Not true
795 void invalidate_inode_buffers(struct inode *inode)
797 if (inode_has_buffers(inode)) {
798 struct address_space *mapping = &inode->i_data;
799 struct list_head *list = &mapping->private_list;
800 struct address_space *buffer_mapping = mapping->private_data;
802 spin_lock(&buffer_mapping->private_lock);
803 while (!list_empty(list))
804 __remove_assoc_queue(BH_ENTRY(list->next));
805 spin_unlock(&buffer_mapping->private_lock);
808 EXPORT_SYMBOL(invalidate_inode_buffers);
811 * Remove any clean buffers from the inode's buffer list. This is called
812 * when we're trying to free the inode itself. Those buffers can pin it.
814 * Returns true if all buffers were removed.
816 int remove_inode_buffers(struct inode *inode)
820 if (inode_has_buffers(inode)) {
821 struct address_space *mapping = &inode->i_data;
822 struct list_head *list = &mapping->private_list;
823 struct address_space *buffer_mapping = mapping->private_data;
825 spin_lock(&buffer_mapping->private_lock);
826 while (!list_empty(list)) {
827 struct buffer_head *bh = BH_ENTRY(list->next);
828 if (buffer_dirty(bh)) {
832 __remove_assoc_queue(bh);
834 spin_unlock(&buffer_mapping->private_lock);
840 * Create the appropriate buffers when given a page for data area and
841 * the size of each buffer.. Use the bh->b_this_page linked list to
842 * follow the buffers created. Return NULL if unable to create more
845 * The retry flag is used to differentiate async IO (paging, swapping)
846 * which may not fail from ordinary buffer allocations.
848 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
851 struct buffer_head *bh, *head;
852 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
854 struct mem_cgroup *memcg;
859 memcg = get_mem_cgroup_from_page(page);
860 memalloc_use_memcg(memcg);
864 while ((offset -= size) >= 0) {
865 bh = alloc_buffer_head(gfp);
869 bh->b_this_page = head;
875 /* Link the buffer to its page */
876 set_bh_page(bh, page, offset);
879 memalloc_unuse_memcg();
880 mem_cgroup_put(memcg);
883 * In case anything failed, we just free everything we got.
889 head = head->b_this_page;
890 free_buffer_head(bh);
896 EXPORT_SYMBOL_GPL(alloc_page_buffers);
899 link_dev_buffers(struct page *page, struct buffer_head *head)
901 struct buffer_head *bh, *tail;
906 bh = bh->b_this_page;
908 tail->b_this_page = head;
909 attach_page_buffers(page, head);
912 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
914 sector_t retval = ~((sector_t)0);
915 loff_t sz = i_size_read(bdev->bd_inode);
918 unsigned int sizebits = blksize_bits(size);
919 retval = (sz >> sizebits);
925 * Initialise the state of a blockdev page's buffers.
928 init_page_buffers(struct page *page, struct block_device *bdev,
929 sector_t block, int size)
931 struct buffer_head *head = page_buffers(page);
932 struct buffer_head *bh = head;
933 int uptodate = PageUptodate(page);
934 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
937 if (!buffer_mapped(bh)) {
939 bh->b_private = NULL;
941 bh->b_blocknr = block;
943 set_buffer_uptodate(bh);
944 if (block < end_block)
945 set_buffer_mapped(bh);
948 bh = bh->b_this_page;
949 } while (bh != head);
952 * Caller needs to validate requested block against end of device.
958 * Create the page-cache page that contains the requested block.
960 * This is used purely for blockdev mappings.
963 grow_dev_page(struct block_device *bdev, sector_t block,
964 pgoff_t index, int size, int sizebits, gfp_t gfp)
966 struct inode *inode = bdev->bd_inode;
968 struct buffer_head *bh;
973 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
976 * XXX: __getblk_slow() can not really deal with failure and
977 * will endlessly loop on improvised global reclaim. Prefer
978 * looping in the allocator rather than here, at least that
979 * code knows what it's doing.
981 gfp_mask |= __GFP_NOFAIL;
983 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
985 BUG_ON(!PageLocked(page));
987 if (page_has_buffers(page)) {
988 bh = page_buffers(page);
989 if (bh->b_size == size) {
990 end_block = init_page_buffers(page, bdev,
991 (sector_t)index << sizebits,
995 if (!try_to_free_buffers(page))
1000 * Allocate some buffers for this page
1002 bh = alloc_page_buffers(page, size, true);
1005 * Link the page to the buffers and initialise them. Take the
1006 * lock to be atomic wrt __find_get_block(), which does not
1007 * run under the page lock.
1009 spin_lock(&inode->i_mapping->private_lock);
1010 link_dev_buffers(page, bh);
1011 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1013 spin_unlock(&inode->i_mapping->private_lock);
1015 ret = (block < end_block) ? 1 : -ENXIO;
1023 * Create buffers for the specified block device block's page. If
1024 * that page was dirty, the buffers are set dirty also.
1027 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1035 } while ((size << sizebits) < PAGE_SIZE);
1037 index = block >> sizebits;
1040 * Check for a block which wants to lie outside our maximum possible
1041 * pagecache index. (this comparison is done using sector_t types).
1043 if (unlikely(index != block >> sizebits)) {
1044 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1046 __func__, (unsigned long long)block,
1051 /* Create a page with the proper size buffers.. */
1052 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1055 static struct buffer_head *
1056 __getblk_slow(struct block_device *bdev, sector_t block,
1057 unsigned size, gfp_t gfp)
1059 /* Size must be multiple of hard sectorsize */
1060 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1061 (size < 512 || size > PAGE_SIZE))) {
1062 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1064 printk(KERN_ERR "logical block size: %d\n",
1065 bdev_logical_block_size(bdev));
1072 struct buffer_head *bh;
1075 bh = __find_get_block(bdev, block, size);
1079 ret = grow_buffers(bdev, block, size, gfp);
1086 * The relationship between dirty buffers and dirty pages:
1088 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1089 * the page is tagged dirty in the page cache.
1091 * At all times, the dirtiness of the buffers represents the dirtiness of
1092 * subsections of the page. If the page has buffers, the page dirty bit is
1093 * merely a hint about the true dirty state.
1095 * When a page is set dirty in its entirety, all its buffers are marked dirty
1096 * (if the page has buffers).
1098 * When a buffer is marked dirty, its page is dirtied, but the page's other
1101 * Also. When blockdev buffers are explicitly read with bread(), they
1102 * individually become uptodate. But their backing page remains not
1103 * uptodate - even if all of its buffers are uptodate. A subsequent
1104 * block_read_full_page() against that page will discover all the uptodate
1105 * buffers, will set the page uptodate and will perform no I/O.
1109 * mark_buffer_dirty - mark a buffer_head as needing writeout
1110 * @bh: the buffer_head to mark dirty
1112 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1113 * its backing page dirty, then tag the page as dirty in the page cache
1114 * and then attach the address_space's inode to its superblock's dirty
1117 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1118 * i_pages lock and mapping->host->i_lock.
1120 void mark_buffer_dirty(struct buffer_head *bh)
1122 WARN_ON_ONCE(!buffer_uptodate(bh));
1124 trace_block_dirty_buffer(bh);
1127 * Very *carefully* optimize the it-is-already-dirty case.
1129 * Don't let the final "is it dirty" escape to before we
1130 * perhaps modified the buffer.
1132 if (buffer_dirty(bh)) {
1134 if (buffer_dirty(bh))
1138 if (!test_set_buffer_dirty(bh)) {
1139 struct page *page = bh->b_page;
1140 struct address_space *mapping = NULL;
1142 lock_page_memcg(page);
1143 if (!TestSetPageDirty(page)) {
1144 mapping = page_mapping(page);
1146 __set_page_dirty(page, mapping, 0);
1148 unlock_page_memcg(page);
1150 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1153 EXPORT_SYMBOL(mark_buffer_dirty);
1155 void mark_buffer_write_io_error(struct buffer_head *bh)
1157 struct super_block *sb;
1159 set_buffer_write_io_error(bh);
1160 /* FIXME: do we need to set this in both places? */
1161 if (bh->b_page && bh->b_page->mapping)
1162 mapping_set_error(bh->b_page->mapping, -EIO);
1163 if (bh->b_assoc_map)
1164 mapping_set_error(bh->b_assoc_map, -EIO);
1166 sb = READ_ONCE(bh->b_bdev->bd_super);
1168 errseq_set(&sb->s_wb_err, -EIO);
1171 EXPORT_SYMBOL(mark_buffer_write_io_error);
1174 * Decrement a buffer_head's reference count. If all buffers against a page
1175 * have zero reference count, are clean and unlocked, and if the page is clean
1176 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1177 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1178 * a page but it ends up not being freed, and buffers may later be reattached).
1180 void __brelse(struct buffer_head * buf)
1182 if (atomic_read(&buf->b_count)) {
1186 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1188 EXPORT_SYMBOL(__brelse);
1191 * bforget() is like brelse(), except it discards any
1192 * potentially dirty data.
1194 void __bforget(struct buffer_head *bh)
1196 clear_buffer_dirty(bh);
1197 if (bh->b_assoc_map) {
1198 struct address_space *buffer_mapping = bh->b_page->mapping;
1200 spin_lock(&buffer_mapping->private_lock);
1201 list_del_init(&bh->b_assoc_buffers);
1202 bh->b_assoc_map = NULL;
1203 spin_unlock(&buffer_mapping->private_lock);
1207 EXPORT_SYMBOL(__bforget);
1209 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1212 if (buffer_uptodate(bh)) {
1217 bh->b_end_io = end_buffer_read_sync;
1218 submit_bh(REQ_OP_READ, 0, bh);
1220 if (buffer_uptodate(bh))
1228 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1229 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1230 * refcount elevated by one when they're in an LRU. A buffer can only appear
1231 * once in a particular CPU's LRU. A single buffer can be present in multiple
1232 * CPU's LRUs at the same time.
1234 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1235 * sb_find_get_block().
1237 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1238 * a local interrupt disable for that.
1241 #define BH_LRU_SIZE 16
1244 struct buffer_head *bhs[BH_LRU_SIZE];
1247 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1250 #define bh_lru_lock() local_irq_disable()
1251 #define bh_lru_unlock() local_irq_enable()
1253 #define bh_lru_lock() preempt_disable()
1254 #define bh_lru_unlock() preempt_enable()
1257 static inline void check_irqs_on(void)
1259 #ifdef irqs_disabled
1260 BUG_ON(irqs_disabled());
1265 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1266 * inserted at the front, and the buffer_head at the back if any is evicted.
1267 * Or, if already in the LRU it is moved to the front.
1269 static void bh_lru_install(struct buffer_head *bh)
1271 struct buffer_head *evictee = bh;
1278 b = this_cpu_ptr(&bh_lrus);
1279 for (i = 0; i < BH_LRU_SIZE; i++) {
1280 swap(evictee, b->bhs[i]);
1281 if (evictee == bh) {
1293 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1295 static struct buffer_head *
1296 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1298 struct buffer_head *ret = NULL;
1303 for (i = 0; i < BH_LRU_SIZE; i++) {
1304 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1306 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1307 bh->b_size == size) {
1310 __this_cpu_write(bh_lrus.bhs[i],
1311 __this_cpu_read(bh_lrus.bhs[i - 1]));
1314 __this_cpu_write(bh_lrus.bhs[0], bh);
1326 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1327 * it in the LRU and mark it as accessed. If it is not present then return
1330 struct buffer_head *
1331 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1333 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1336 /* __find_get_block_slow will mark the page accessed */
1337 bh = __find_get_block_slow(bdev, block);
1345 EXPORT_SYMBOL(__find_get_block);
1348 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1349 * which corresponds to the passed block_device, block and size. The
1350 * returned buffer has its reference count incremented.
1352 * __getblk_gfp() will lock up the machine if grow_dev_page's
1353 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1355 struct buffer_head *
1356 __getblk_gfp(struct block_device *bdev, sector_t block,
1357 unsigned size, gfp_t gfp)
1359 struct buffer_head *bh = __find_get_block(bdev, block, size);
1363 bh = __getblk_slow(bdev, block, size, gfp);
1366 EXPORT_SYMBOL(__getblk_gfp);
1369 * Do async read-ahead on a buffer..
1371 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1373 struct buffer_head *bh = __getblk(bdev, block, size);
1375 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1379 EXPORT_SYMBOL(__breadahead);
1381 void __breadahead_gfp(struct block_device *bdev, sector_t block, unsigned size,
1384 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1386 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1390 EXPORT_SYMBOL(__breadahead_gfp);
1393 * __bread_gfp() - reads a specified block and returns the bh
1394 * @bdev: the block_device to read from
1395 * @block: number of block
1396 * @size: size (in bytes) to read
1397 * @gfp: page allocation flag
1399 * Reads a specified block, and returns buffer head that contains it.
1400 * The page cache can be allocated from non-movable area
1401 * not to prevent page migration if you set gfp to zero.
1402 * It returns NULL if the block was unreadable.
1404 struct buffer_head *
1405 __bread_gfp(struct block_device *bdev, sector_t block,
1406 unsigned size, gfp_t gfp)
1408 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1410 if (likely(bh) && !buffer_uptodate(bh))
1411 bh = __bread_slow(bh);
1414 EXPORT_SYMBOL(__bread_gfp);
1417 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1418 * This doesn't race because it runs in each cpu either in irq
1419 * or with preempt disabled.
1421 static void invalidate_bh_lru(void *arg)
1423 struct bh_lru *b = &get_cpu_var(bh_lrus);
1426 for (i = 0; i < BH_LRU_SIZE; i++) {
1430 put_cpu_var(bh_lrus);
1433 static bool has_bh_in_lru(int cpu, void *dummy)
1435 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1438 for (i = 0; i < BH_LRU_SIZE; i++) {
1446 void invalidate_bh_lrus(void)
1448 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1450 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1452 void set_bh_page(struct buffer_head *bh,
1453 struct page *page, unsigned long offset)
1456 BUG_ON(offset >= PAGE_SIZE);
1457 if (PageHighMem(page))
1459 * This catches illegal uses and preserves the offset:
1461 bh->b_data = (char *)(0 + offset);
1463 bh->b_data = page_address(page) + offset;
1465 EXPORT_SYMBOL(set_bh_page);
1468 * Called when truncating a buffer on a page completely.
1471 /* Bits that are cleared during an invalidate */
1472 #define BUFFER_FLAGS_DISCARD \
1473 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1474 1 << BH_Delay | 1 << BH_Unwritten)
1476 static void discard_buffer(struct buffer_head * bh)
1478 unsigned long b_state, b_state_old;
1481 clear_buffer_dirty(bh);
1483 b_state = bh->b_state;
1485 b_state_old = cmpxchg(&bh->b_state, b_state,
1486 (b_state & ~BUFFER_FLAGS_DISCARD));
1487 if (b_state_old == b_state)
1489 b_state = b_state_old;
1495 * block_invalidatepage - invalidate part or all of a buffer-backed page
1497 * @page: the page which is affected
1498 * @offset: start of the range to invalidate
1499 * @length: length of the range to invalidate
1501 * block_invalidatepage() is called when all or part of the page has become
1502 * invalidated by a truncate operation.
1504 * block_invalidatepage() does not have to release all buffers, but it must
1505 * ensure that no dirty buffer is left outside @offset and that no I/O
1506 * is underway against any of the blocks which are outside the truncation
1507 * point. Because the caller is about to free (and possibly reuse) those
1510 void block_invalidatepage(struct page *page, unsigned int offset,
1511 unsigned int length)
1513 struct buffer_head *head, *bh, *next;
1514 unsigned int curr_off = 0;
1515 unsigned int stop = length + offset;
1517 BUG_ON(!PageLocked(page));
1518 if (!page_has_buffers(page))
1522 * Check for overflow
1524 BUG_ON(stop > PAGE_SIZE || stop < length);
1526 head = page_buffers(page);
1529 unsigned int next_off = curr_off + bh->b_size;
1530 next = bh->b_this_page;
1533 * Are we still fully in range ?
1535 if (next_off > stop)
1539 * is this block fully invalidated?
1541 if (offset <= curr_off)
1543 curr_off = next_off;
1545 } while (bh != head);
1548 * We release buffers only if the entire page is being invalidated.
1549 * The get_block cached value has been unconditionally invalidated,
1550 * so real IO is not possible anymore.
1552 if (length == PAGE_SIZE)
1553 try_to_release_page(page, 0);
1557 EXPORT_SYMBOL(block_invalidatepage);
1561 * We attach and possibly dirty the buffers atomically wrt
1562 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1563 * is already excluded via the page lock.
1565 void create_empty_buffers(struct page *page,
1566 unsigned long blocksize, unsigned long b_state)
1568 struct buffer_head *bh, *head, *tail;
1570 head = alloc_page_buffers(page, blocksize, true);
1573 bh->b_state |= b_state;
1575 bh = bh->b_this_page;
1577 tail->b_this_page = head;
1579 spin_lock(&page->mapping->private_lock);
1580 if (PageUptodate(page) || PageDirty(page)) {
1583 if (PageDirty(page))
1584 set_buffer_dirty(bh);
1585 if (PageUptodate(page))
1586 set_buffer_uptodate(bh);
1587 bh = bh->b_this_page;
1588 } while (bh != head);
1590 attach_page_buffers(page, head);
1591 spin_unlock(&page->mapping->private_lock);
1593 EXPORT_SYMBOL(create_empty_buffers);
1596 * clean_bdev_aliases: clean a range of buffers in block device
1597 * @bdev: Block device to clean buffers in
1598 * @block: Start of a range of blocks to clean
1599 * @len: Number of blocks to clean
1601 * We are taking a range of blocks for data and we don't want writeback of any
1602 * buffer-cache aliases starting from return from this function and until the
1603 * moment when something will explicitly mark the buffer dirty (hopefully that
1604 * will not happen until we will free that block ;-) We don't even need to mark
1605 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1606 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1607 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1608 * would confuse anyone who might pick it with bread() afterwards...
1610 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1611 * writeout I/O going on against recently-freed buffers. We don't wait on that
1612 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1613 * need to. That happens here.
1615 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1617 struct inode *bd_inode = bdev->bd_inode;
1618 struct address_space *bd_mapping = bd_inode->i_mapping;
1619 struct pagevec pvec;
1620 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1623 struct buffer_head *bh;
1624 struct buffer_head *head;
1626 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1627 pagevec_init(&pvec);
1628 while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1629 count = pagevec_count(&pvec);
1630 for (i = 0; i < count; i++) {
1631 struct page *page = pvec.pages[i];
1633 if (!page_has_buffers(page))
1636 * We use page lock instead of bd_mapping->private_lock
1637 * to pin buffers here since we can afford to sleep and
1638 * it scales better than a global spinlock lock.
1641 /* Recheck when the page is locked which pins bhs */
1642 if (!page_has_buffers(page))
1644 head = page_buffers(page);
1647 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1649 if (bh->b_blocknr >= block + len)
1651 clear_buffer_dirty(bh);
1653 clear_buffer_req(bh);
1655 bh = bh->b_this_page;
1656 } while (bh != head);
1660 pagevec_release(&pvec);
1662 /* End of range already reached? */
1663 if (index > end || !index)
1667 EXPORT_SYMBOL(clean_bdev_aliases);
1670 * Size is a power-of-two in the range 512..PAGE_SIZE,
1671 * and the case we care about most is PAGE_SIZE.
1673 * So this *could* possibly be written with those
1674 * constraints in mind (relevant mostly if some
1675 * architecture has a slow bit-scan instruction)
1677 static inline int block_size_bits(unsigned int blocksize)
1679 return ilog2(blocksize);
1682 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1684 BUG_ON(!PageLocked(page));
1686 if (!page_has_buffers(page))
1687 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1689 return page_buffers(page);
1693 * NOTE! All mapped/uptodate combinations are valid:
1695 * Mapped Uptodate Meaning
1697 * No No "unknown" - must do get_block()
1698 * No Yes "hole" - zero-filled
1699 * Yes No "allocated" - allocated on disk, not read in
1700 * Yes Yes "valid" - allocated and up-to-date in memory.
1702 * "Dirty" is valid only with the last case (mapped+uptodate).
1706 * While block_write_full_page is writing back the dirty buffers under
1707 * the page lock, whoever dirtied the buffers may decide to clean them
1708 * again at any time. We handle that by only looking at the buffer
1709 * state inside lock_buffer().
1711 * If block_write_full_page() is called for regular writeback
1712 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1713 * locked buffer. This only can happen if someone has written the buffer
1714 * directly, with submit_bh(). At the address_space level PageWriteback
1715 * prevents this contention from occurring.
1717 * If block_write_full_page() is called with wbc->sync_mode ==
1718 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1719 * causes the writes to be flagged as synchronous writes.
1721 int __block_write_full_page(struct inode *inode, struct page *page,
1722 get_block_t *get_block, struct writeback_control *wbc,
1723 bh_end_io_t *handler)
1727 sector_t last_block;
1728 struct buffer_head *bh, *head;
1729 unsigned int blocksize, bbits;
1730 int nr_underway = 0;
1731 int write_flags = wbc_to_write_flags(wbc);
1733 head = create_page_buffers(page, inode,
1734 (1 << BH_Dirty)|(1 << BH_Uptodate));
1737 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1738 * here, and the (potentially unmapped) buffers may become dirty at
1739 * any time. If a buffer becomes dirty here after we've inspected it
1740 * then we just miss that fact, and the page stays dirty.
1742 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1743 * handle that here by just cleaning them.
1747 blocksize = bh->b_size;
1748 bbits = block_size_bits(blocksize);
1750 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1751 last_block = (i_size_read(inode) - 1) >> bbits;
1754 * Get all the dirty buffers mapped to disk addresses and
1755 * handle any aliases from the underlying blockdev's mapping.
1758 if (block > last_block) {
1760 * mapped buffers outside i_size will occur, because
1761 * this page can be outside i_size when there is a
1762 * truncate in progress.
1765 * The buffer was zeroed by block_write_full_page()
1767 clear_buffer_dirty(bh);
1768 set_buffer_uptodate(bh);
1769 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1771 WARN_ON(bh->b_size != blocksize);
1772 err = get_block(inode, block, bh, 1);
1775 clear_buffer_delay(bh);
1776 if (buffer_new(bh)) {
1777 /* blockdev mappings never come here */
1778 clear_buffer_new(bh);
1779 clean_bdev_bh_alias(bh);
1782 bh = bh->b_this_page;
1784 } while (bh != head);
1787 if (!buffer_mapped(bh))
1790 * If it's a fully non-blocking write attempt and we cannot
1791 * lock the buffer then redirty the page. Note that this can
1792 * potentially cause a busy-wait loop from writeback threads
1793 * and kswapd activity, but those code paths have their own
1794 * higher-level throttling.
1796 if (wbc->sync_mode != WB_SYNC_NONE) {
1798 } else if (!trylock_buffer(bh)) {
1799 redirty_page_for_writepage(wbc, page);
1802 if (test_clear_buffer_dirty(bh)) {
1803 mark_buffer_async_write_endio(bh, handler);
1807 } while ((bh = bh->b_this_page) != head);
1810 * The page and its buffers are protected by PageWriteback(), so we can
1811 * drop the bh refcounts early.
1813 BUG_ON(PageWriteback(page));
1814 set_page_writeback(page);
1817 struct buffer_head *next = bh->b_this_page;
1818 if (buffer_async_write(bh)) {
1819 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1820 inode->i_write_hint, wbc);
1824 } while (bh != head);
1829 if (nr_underway == 0) {
1831 * The page was marked dirty, but the buffers were
1832 * clean. Someone wrote them back by hand with
1833 * ll_rw_block/submit_bh. A rare case.
1835 end_page_writeback(page);
1838 * The page and buffer_heads can be released at any time from
1846 * ENOSPC, or some other error. We may already have added some
1847 * blocks to the file, so we need to write these out to avoid
1848 * exposing stale data.
1849 * The page is currently locked and not marked for writeback
1852 /* Recovery: lock and submit the mapped buffers */
1854 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1855 !buffer_delay(bh)) {
1857 mark_buffer_async_write_endio(bh, handler);
1860 * The buffer may have been set dirty during
1861 * attachment to a dirty page.
1863 clear_buffer_dirty(bh);
1865 } while ((bh = bh->b_this_page) != head);
1867 BUG_ON(PageWriteback(page));
1868 mapping_set_error(page->mapping, err);
1869 set_page_writeback(page);
1871 struct buffer_head *next = bh->b_this_page;
1872 if (buffer_async_write(bh)) {
1873 clear_buffer_dirty(bh);
1874 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1875 inode->i_write_hint, wbc);
1879 } while (bh != head);
1883 EXPORT_SYMBOL(__block_write_full_page);
1886 * If a page has any new buffers, zero them out here, and mark them uptodate
1887 * and dirty so they'll be written out (in order to prevent uninitialised
1888 * block data from leaking). And clear the new bit.
1890 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1892 unsigned int block_start, block_end;
1893 struct buffer_head *head, *bh;
1895 BUG_ON(!PageLocked(page));
1896 if (!page_has_buffers(page))
1899 bh = head = page_buffers(page);
1902 block_end = block_start + bh->b_size;
1904 if (buffer_new(bh)) {
1905 if (block_end > from && block_start < to) {
1906 if (!PageUptodate(page)) {
1907 unsigned start, size;
1909 start = max(from, block_start);
1910 size = min(to, block_end) - start;
1912 zero_user(page, start, size);
1913 set_buffer_uptodate(bh);
1916 clear_buffer_new(bh);
1917 mark_buffer_dirty(bh);
1921 block_start = block_end;
1922 bh = bh->b_this_page;
1923 } while (bh != head);
1925 EXPORT_SYMBOL(page_zero_new_buffers);
1928 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1929 struct iomap *iomap)
1931 loff_t offset = block << inode->i_blkbits;
1933 bh->b_bdev = iomap->bdev;
1936 * Block points to offset in file we need to map, iomap contains
1937 * the offset at which the map starts. If the map ends before the
1938 * current block, then do not map the buffer and let the caller
1941 BUG_ON(offset >= iomap->offset + iomap->length);
1943 switch (iomap->type) {
1946 * If the buffer is not up to date or beyond the current EOF,
1947 * we need to mark it as new to ensure sub-block zeroing is
1948 * executed if necessary.
1950 if (!buffer_uptodate(bh) ||
1951 (offset >= i_size_read(inode)))
1954 case IOMAP_DELALLOC:
1955 if (!buffer_uptodate(bh) ||
1956 (offset >= i_size_read(inode)))
1958 set_buffer_uptodate(bh);
1959 set_buffer_mapped(bh);
1960 set_buffer_delay(bh);
1962 case IOMAP_UNWRITTEN:
1964 * For unwritten regions, we always need to ensure that regions
1965 * in the block we are not writing to are zeroed. Mark the
1966 * buffer as new to ensure this.
1969 set_buffer_unwritten(bh);
1972 if ((iomap->flags & IOMAP_F_NEW) ||
1973 offset >= i_size_read(inode))
1975 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1977 set_buffer_mapped(bh);
1982 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1983 get_block_t *get_block, struct iomap *iomap)
1985 unsigned from = pos & (PAGE_SIZE - 1);
1986 unsigned to = from + len;
1987 struct inode *inode = page->mapping->host;
1988 unsigned block_start, block_end;
1991 unsigned blocksize, bbits;
1992 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1994 BUG_ON(!PageLocked(page));
1995 BUG_ON(from > PAGE_SIZE);
1996 BUG_ON(to > PAGE_SIZE);
1999 head = create_page_buffers(page, inode, 0);
2000 blocksize = head->b_size;
2001 bbits = block_size_bits(blocksize);
2003 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
2005 for(bh = head, block_start = 0; bh != head || !block_start;
2006 block++, block_start=block_end, bh = bh->b_this_page) {
2007 block_end = block_start + blocksize;
2008 if (block_end <= from || block_start >= to) {
2009 if (PageUptodate(page)) {
2010 if (!buffer_uptodate(bh))
2011 set_buffer_uptodate(bh);
2016 clear_buffer_new(bh);
2017 if (!buffer_mapped(bh)) {
2018 WARN_ON(bh->b_size != blocksize);
2020 err = get_block(inode, block, bh, 1);
2024 iomap_to_bh(inode, block, bh, iomap);
2027 if (buffer_new(bh)) {
2028 clean_bdev_bh_alias(bh);
2029 if (PageUptodate(page)) {
2030 clear_buffer_new(bh);
2031 set_buffer_uptodate(bh);
2032 mark_buffer_dirty(bh);
2035 if (block_end > to || block_start < from)
2036 zero_user_segments(page,
2042 if (PageUptodate(page)) {
2043 if (!buffer_uptodate(bh))
2044 set_buffer_uptodate(bh);
2047 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2048 !buffer_unwritten(bh) &&
2049 (block_start < from || block_end > to)) {
2050 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2055 * If we issued read requests - let them complete.
2057 while(wait_bh > wait) {
2058 wait_on_buffer(*--wait_bh);
2059 if (!buffer_uptodate(*wait_bh))
2063 page_zero_new_buffers(page, from, to);
2067 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2068 get_block_t *get_block)
2070 return __block_write_begin_int(page, pos, len, get_block, NULL);
2072 EXPORT_SYMBOL(__block_write_begin);
2074 static int __block_commit_write(struct inode *inode, struct page *page,
2075 unsigned from, unsigned to)
2077 unsigned block_start, block_end;
2080 struct buffer_head *bh, *head;
2082 bh = head = page_buffers(page);
2083 blocksize = bh->b_size;
2087 block_end = block_start + blocksize;
2088 if (block_end <= from || block_start >= to) {
2089 if (!buffer_uptodate(bh))
2092 set_buffer_uptodate(bh);
2093 mark_buffer_dirty(bh);
2095 clear_buffer_new(bh);
2097 block_start = block_end;
2098 bh = bh->b_this_page;
2099 } while (bh != head);
2102 * If this is a partial write which happened to make all buffers
2103 * uptodate then we can optimize away a bogus readpage() for
2104 * the next read(). Here we 'discover' whether the page went
2105 * uptodate as a result of this (potentially partial) write.
2108 SetPageUptodate(page);
2113 * block_write_begin takes care of the basic task of block allocation and
2114 * bringing partial write blocks uptodate first.
2116 * The filesystem needs to handle block truncation upon failure.
2118 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2119 unsigned flags, struct page **pagep, get_block_t *get_block)
2121 pgoff_t index = pos >> PAGE_SHIFT;
2125 page = grab_cache_page_write_begin(mapping, index, flags);
2129 status = __block_write_begin(page, pos, len, get_block);
2130 if (unlikely(status)) {
2139 EXPORT_SYMBOL(block_write_begin);
2141 int block_write_end(struct file *file, struct address_space *mapping,
2142 loff_t pos, unsigned len, unsigned copied,
2143 struct page *page, void *fsdata)
2145 struct inode *inode = mapping->host;
2148 start = pos & (PAGE_SIZE - 1);
2150 if (unlikely(copied < len)) {
2152 * The buffers that were written will now be uptodate, so we
2153 * don't have to worry about a readpage reading them and
2154 * overwriting a partial write. However if we have encountered
2155 * a short write and only partially written into a buffer, it
2156 * will not be marked uptodate, so a readpage might come in and
2157 * destroy our partial write.
2159 * Do the simplest thing, and just treat any short write to a
2160 * non uptodate page as a zero-length write, and force the
2161 * caller to redo the whole thing.
2163 if (!PageUptodate(page))
2166 page_zero_new_buffers(page, start+copied, start+len);
2168 flush_dcache_page(page);
2170 /* This could be a short (even 0-length) commit */
2171 __block_commit_write(inode, page, start, start+copied);
2175 EXPORT_SYMBOL(block_write_end);
2177 int generic_write_end(struct file *file, struct address_space *mapping,
2178 loff_t pos, unsigned len, unsigned copied,
2179 struct page *page, void *fsdata)
2181 struct inode *inode = mapping->host;
2182 loff_t old_size = inode->i_size;
2183 bool i_size_changed = false;
2185 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2188 * No need to use i_size_read() here, the i_size cannot change under us
2189 * because we hold i_rwsem.
2191 * But it's important to update i_size while still holding page lock:
2192 * page writeout could otherwise come in and zero beyond i_size.
2194 if (pos + copied > inode->i_size) {
2195 i_size_write(inode, pos + copied);
2196 i_size_changed = true;
2203 pagecache_isize_extended(inode, old_size, pos);
2205 * Don't mark the inode dirty under page lock. First, it unnecessarily
2206 * makes the holding time of page lock longer. Second, it forces lock
2207 * ordering of page lock and transaction start for journaling
2211 mark_inode_dirty(inode);
2214 EXPORT_SYMBOL(generic_write_end);
2217 * block_is_partially_uptodate checks whether buffers within a page are
2220 * Returns true if all buffers which correspond to a file portion
2221 * we want to read are uptodate.
2223 int block_is_partially_uptodate(struct page *page, unsigned long from,
2224 unsigned long count)
2226 unsigned block_start, block_end, blocksize;
2228 struct buffer_head *bh, *head;
2231 if (!page_has_buffers(page))
2234 head = page_buffers(page);
2235 blocksize = head->b_size;
2236 to = min_t(unsigned, PAGE_SIZE - from, count);
2238 if (from < blocksize && to > PAGE_SIZE - blocksize)
2244 block_end = block_start + blocksize;
2245 if (block_end > from && block_start < to) {
2246 if (!buffer_uptodate(bh)) {
2250 if (block_end >= to)
2253 block_start = block_end;
2254 bh = bh->b_this_page;
2255 } while (bh != head);
2259 EXPORT_SYMBOL(block_is_partially_uptodate);
2262 * Generic "read page" function for block devices that have the normal
2263 * get_block functionality. This is most of the block device filesystems.
2264 * Reads the page asynchronously --- the unlock_buffer() and
2265 * set/clear_buffer_uptodate() functions propagate buffer state into the
2266 * page struct once IO has completed.
2268 int block_read_full_page(struct page *page, get_block_t *get_block)
2270 struct inode *inode = page->mapping->host;
2271 sector_t iblock, lblock;
2272 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2273 unsigned int blocksize, bbits;
2275 int fully_mapped = 1;
2277 head = create_page_buffers(page, inode, 0);
2278 blocksize = head->b_size;
2279 bbits = block_size_bits(blocksize);
2281 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2282 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2288 if (buffer_uptodate(bh))
2291 if (!buffer_mapped(bh)) {
2295 if (iblock < lblock) {
2296 WARN_ON(bh->b_size != blocksize);
2297 err = get_block(inode, iblock, bh, 0);
2301 if (!buffer_mapped(bh)) {
2302 zero_user(page, i * blocksize, blocksize);
2304 set_buffer_uptodate(bh);
2308 * get_block() might have updated the buffer
2311 if (buffer_uptodate(bh))
2315 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2318 SetPageMappedToDisk(page);
2322 * All buffers are uptodate - we can set the page uptodate
2323 * as well. But not if get_block() returned an error.
2325 if (!PageError(page))
2326 SetPageUptodate(page);
2331 /* Stage two: lock the buffers */
2332 for (i = 0; i < nr; i++) {
2335 mark_buffer_async_read(bh);
2339 * Stage 3: start the IO. Check for uptodateness
2340 * inside the buffer lock in case another process reading
2341 * the underlying blockdev brought it uptodate (the sct fix).
2343 for (i = 0; i < nr; i++) {
2345 if (buffer_uptodate(bh))
2346 end_buffer_async_read(bh, 1);
2348 submit_bh(REQ_OP_READ, 0, bh);
2352 EXPORT_SYMBOL(block_read_full_page);
2354 /* utility function for filesystems that need to do work on expanding
2355 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2356 * deal with the hole.
2358 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2360 struct address_space *mapping = inode->i_mapping;
2365 err = inode_newsize_ok(inode, size);
2369 err = pagecache_write_begin(NULL, mapping, size, 0,
2370 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2374 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2380 EXPORT_SYMBOL(generic_cont_expand_simple);
2382 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2383 loff_t pos, loff_t *bytes)
2385 struct inode *inode = mapping->host;
2386 unsigned int blocksize = i_blocksize(inode);
2389 pgoff_t index, curidx;
2391 unsigned zerofrom, offset, len;
2394 index = pos >> PAGE_SHIFT;
2395 offset = pos & ~PAGE_MASK;
2397 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2398 zerofrom = curpos & ~PAGE_MASK;
2399 if (zerofrom & (blocksize-1)) {
2400 *bytes |= (blocksize-1);
2403 len = PAGE_SIZE - zerofrom;
2405 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2409 zero_user(page, zerofrom, len);
2410 err = pagecache_write_end(file, mapping, curpos, len, len,
2417 balance_dirty_pages_ratelimited(mapping);
2419 if (fatal_signal_pending(current)) {
2425 /* page covers the boundary, find the boundary offset */
2426 if (index == curidx) {
2427 zerofrom = curpos & ~PAGE_MASK;
2428 /* if we will expand the thing last block will be filled */
2429 if (offset <= zerofrom) {
2432 if (zerofrom & (blocksize-1)) {
2433 *bytes |= (blocksize-1);
2436 len = offset - zerofrom;
2438 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2442 zero_user(page, zerofrom, len);
2443 err = pagecache_write_end(file, mapping, curpos, len, len,
2455 * For moronic filesystems that do not allow holes in file.
2456 * We may have to extend the file.
2458 int cont_write_begin(struct file *file, struct address_space *mapping,
2459 loff_t pos, unsigned len, unsigned flags,
2460 struct page **pagep, void **fsdata,
2461 get_block_t *get_block, loff_t *bytes)
2463 struct inode *inode = mapping->host;
2464 unsigned int blocksize = i_blocksize(inode);
2465 unsigned int zerofrom;
2468 err = cont_expand_zero(file, mapping, pos, bytes);
2472 zerofrom = *bytes & ~PAGE_MASK;
2473 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2474 *bytes |= (blocksize-1);
2478 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2480 EXPORT_SYMBOL(cont_write_begin);
2482 int block_commit_write(struct page *page, unsigned from, unsigned to)
2484 struct inode *inode = page->mapping->host;
2485 __block_commit_write(inode,page,from,to);
2488 EXPORT_SYMBOL(block_commit_write);
2491 * block_page_mkwrite() is not allowed to change the file size as it gets
2492 * called from a page fault handler when a page is first dirtied. Hence we must
2493 * be careful to check for EOF conditions here. We set the page up correctly
2494 * for a written page which means we get ENOSPC checking when writing into
2495 * holes and correct delalloc and unwritten extent mapping on filesystems that
2496 * support these features.
2498 * We are not allowed to take the i_mutex here so we have to play games to
2499 * protect against truncate races as the page could now be beyond EOF. Because
2500 * truncate writes the inode size before removing pages, once we have the
2501 * page lock we can determine safely if the page is beyond EOF. If it is not
2502 * beyond EOF, then the page is guaranteed safe against truncation until we
2505 * Direct callers of this function should protect against filesystem freezing
2506 * using sb_start_pagefault() - sb_end_pagefault() functions.
2508 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2509 get_block_t get_block)
2511 struct page *page = vmf->page;
2512 struct inode *inode = file_inode(vma->vm_file);
2518 size = i_size_read(inode);
2519 if ((page->mapping != inode->i_mapping) ||
2520 (page_offset(page) > size)) {
2521 /* We overload EFAULT to mean page got truncated */
2526 /* page is wholly or partially inside EOF */
2527 if (((page->index + 1) << PAGE_SHIFT) > size)
2528 end = size & ~PAGE_MASK;
2532 ret = __block_write_begin(page, 0, end, get_block);
2534 ret = block_commit_write(page, 0, end);
2536 if (unlikely(ret < 0))
2538 set_page_dirty(page);
2539 wait_for_stable_page(page);
2545 EXPORT_SYMBOL(block_page_mkwrite);
2548 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2549 * immediately, while under the page lock. So it needs a special end_io
2550 * handler which does not touch the bh after unlocking it.
2552 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2554 __end_buffer_read_notouch(bh, uptodate);
2558 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2559 * the page (converting it to circular linked list and taking care of page
2562 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2564 struct buffer_head *bh;
2566 BUG_ON(!PageLocked(page));
2568 spin_lock(&page->mapping->private_lock);
2571 if (PageDirty(page))
2572 set_buffer_dirty(bh);
2573 if (!bh->b_this_page)
2574 bh->b_this_page = head;
2575 bh = bh->b_this_page;
2576 } while (bh != head);
2577 attach_page_buffers(page, head);
2578 spin_unlock(&page->mapping->private_lock);
2582 * On entry, the page is fully not uptodate.
2583 * On exit the page is fully uptodate in the areas outside (from,to)
2584 * The filesystem needs to handle block truncation upon failure.
2586 int nobh_write_begin(struct address_space *mapping,
2587 loff_t pos, unsigned len, unsigned flags,
2588 struct page **pagep, void **fsdata,
2589 get_block_t *get_block)
2591 struct inode *inode = mapping->host;
2592 const unsigned blkbits = inode->i_blkbits;
2593 const unsigned blocksize = 1 << blkbits;
2594 struct buffer_head *head, *bh;
2598 unsigned block_in_page;
2599 unsigned block_start, block_end;
2600 sector_t block_in_file;
2603 int is_mapped_to_disk = 1;
2605 index = pos >> PAGE_SHIFT;
2606 from = pos & (PAGE_SIZE - 1);
2609 page = grab_cache_page_write_begin(mapping, index, flags);
2615 if (page_has_buffers(page)) {
2616 ret = __block_write_begin(page, pos, len, get_block);
2622 if (PageMappedToDisk(page))
2626 * Allocate buffers so that we can keep track of state, and potentially
2627 * attach them to the page if an error occurs. In the common case of
2628 * no error, they will just be freed again without ever being attached
2629 * to the page (which is all OK, because we're under the page lock).
2631 * Be careful: the buffer linked list is a NULL terminated one, rather
2632 * than the circular one we're used to.
2634 head = alloc_page_buffers(page, blocksize, false);
2640 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2643 * We loop across all blocks in the page, whether or not they are
2644 * part of the affected region. This is so we can discover if the
2645 * page is fully mapped-to-disk.
2647 for (block_start = 0, block_in_page = 0, bh = head;
2648 block_start < PAGE_SIZE;
2649 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2652 block_end = block_start + blocksize;
2655 if (block_start >= to)
2657 ret = get_block(inode, block_in_file + block_in_page,
2661 if (!buffer_mapped(bh))
2662 is_mapped_to_disk = 0;
2664 clean_bdev_bh_alias(bh);
2665 if (PageUptodate(page)) {
2666 set_buffer_uptodate(bh);
2669 if (buffer_new(bh) || !buffer_mapped(bh)) {
2670 zero_user_segments(page, block_start, from,
2674 if (buffer_uptodate(bh))
2675 continue; /* reiserfs does this */
2676 if (block_start < from || block_end > to) {
2678 bh->b_end_io = end_buffer_read_nobh;
2679 submit_bh(REQ_OP_READ, 0, bh);
2686 * The page is locked, so these buffers are protected from
2687 * any VM or truncate activity. Hence we don't need to care
2688 * for the buffer_head refcounts.
2690 for (bh = head; bh; bh = bh->b_this_page) {
2692 if (!buffer_uptodate(bh))
2699 if (is_mapped_to_disk)
2700 SetPageMappedToDisk(page);
2702 *fsdata = head; /* to be released by nobh_write_end */
2709 * Error recovery is a bit difficult. We need to zero out blocks that
2710 * were newly allocated, and dirty them to ensure they get written out.
2711 * Buffers need to be attached to the page at this point, otherwise
2712 * the handling of potential IO errors during writeout would be hard
2713 * (could try doing synchronous writeout, but what if that fails too?)
2715 attach_nobh_buffers(page, head);
2716 page_zero_new_buffers(page, from, to);
2725 EXPORT_SYMBOL(nobh_write_begin);
2727 int nobh_write_end(struct file *file, struct address_space *mapping,
2728 loff_t pos, unsigned len, unsigned copied,
2729 struct page *page, void *fsdata)
2731 struct inode *inode = page->mapping->host;
2732 struct buffer_head *head = fsdata;
2733 struct buffer_head *bh;
2734 BUG_ON(fsdata != NULL && page_has_buffers(page));
2736 if (unlikely(copied < len) && head)
2737 attach_nobh_buffers(page, head);
2738 if (page_has_buffers(page))
2739 return generic_write_end(file, mapping, pos, len,
2740 copied, page, fsdata);
2742 SetPageUptodate(page);
2743 set_page_dirty(page);
2744 if (pos+copied > inode->i_size) {
2745 i_size_write(inode, pos+copied);
2746 mark_inode_dirty(inode);
2754 head = head->b_this_page;
2755 free_buffer_head(bh);
2760 EXPORT_SYMBOL(nobh_write_end);
2763 * nobh_writepage() - based on block_full_write_page() except
2764 * that it tries to operate without attaching bufferheads to
2767 int nobh_writepage(struct page *page, get_block_t *get_block,
2768 struct writeback_control *wbc)
2770 struct inode * const inode = page->mapping->host;
2771 loff_t i_size = i_size_read(inode);
2772 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2776 /* Is the page fully inside i_size? */
2777 if (page->index < end_index)
2780 /* Is the page fully outside i_size? (truncate in progress) */
2781 offset = i_size & (PAGE_SIZE-1);
2782 if (page->index >= end_index+1 || !offset) {
2784 * The page may have dirty, unmapped buffers. For example,
2785 * they may have been added in ext3_writepage(). Make them
2786 * freeable here, so the page does not leak.
2789 /* Not really sure about this - do we need this ? */
2790 if (page->mapping->a_ops->invalidatepage)
2791 page->mapping->a_ops->invalidatepage(page, offset);
2794 return 0; /* don't care */
2798 * The page straddles i_size. It must be zeroed out on each and every
2799 * writepage invocation because it may be mmapped. "A file is mapped
2800 * in multiples of the page size. For a file that is not a multiple of
2801 * the page size, the remaining memory is zeroed when mapped, and
2802 * writes to that region are not written out to the file."
2804 zero_user_segment(page, offset, PAGE_SIZE);
2806 ret = mpage_writepage(page, get_block, wbc);
2808 ret = __block_write_full_page(inode, page, get_block, wbc,
2809 end_buffer_async_write);
2812 EXPORT_SYMBOL(nobh_writepage);
2814 int nobh_truncate_page(struct address_space *mapping,
2815 loff_t from, get_block_t *get_block)
2817 pgoff_t index = from >> PAGE_SHIFT;
2818 unsigned offset = from & (PAGE_SIZE-1);
2821 unsigned length, pos;
2822 struct inode *inode = mapping->host;
2824 struct buffer_head map_bh;
2827 blocksize = i_blocksize(inode);
2828 length = offset & (blocksize - 1);
2830 /* Block boundary? Nothing to do */
2834 length = blocksize - length;
2835 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2837 page = grab_cache_page(mapping, index);
2842 if (page_has_buffers(page)) {
2846 return block_truncate_page(mapping, from, get_block);
2849 /* Find the buffer that contains "offset" */
2851 while (offset >= pos) {
2856 map_bh.b_size = blocksize;
2858 err = get_block(inode, iblock, &map_bh, 0);
2861 /* unmapped? It's a hole - nothing to do */
2862 if (!buffer_mapped(&map_bh))
2865 /* Ok, it's mapped. Make sure it's up-to-date */
2866 if (!PageUptodate(page)) {
2867 err = mapping->a_ops->readpage(NULL, page);
2873 if (!PageUptodate(page)) {
2877 if (page_has_buffers(page))
2880 zero_user(page, offset, length);
2881 set_page_dirty(page);
2890 EXPORT_SYMBOL(nobh_truncate_page);
2892 int block_truncate_page(struct address_space *mapping,
2893 loff_t from, get_block_t *get_block)
2895 pgoff_t index = from >> PAGE_SHIFT;
2896 unsigned offset = from & (PAGE_SIZE-1);
2899 unsigned length, pos;
2900 struct inode *inode = mapping->host;
2902 struct buffer_head *bh;
2905 blocksize = i_blocksize(inode);
2906 length = offset & (blocksize - 1);
2908 /* Block boundary? Nothing to do */
2912 length = blocksize - length;
2913 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2915 page = grab_cache_page(mapping, index);
2920 if (!page_has_buffers(page))
2921 create_empty_buffers(page, blocksize, 0);
2923 /* Find the buffer that contains "offset" */
2924 bh = page_buffers(page);
2926 while (offset >= pos) {
2927 bh = bh->b_this_page;
2933 if (!buffer_mapped(bh)) {
2934 WARN_ON(bh->b_size != blocksize);
2935 err = get_block(inode, iblock, bh, 0);
2938 /* unmapped? It's a hole - nothing to do */
2939 if (!buffer_mapped(bh))
2943 /* Ok, it's mapped. Make sure it's up-to-date */
2944 if (PageUptodate(page))
2945 set_buffer_uptodate(bh);
2947 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2949 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2951 /* Uhhuh. Read error. Complain and punt. */
2952 if (!buffer_uptodate(bh))
2956 zero_user(page, offset, length);
2957 mark_buffer_dirty(bh);
2966 EXPORT_SYMBOL(block_truncate_page);
2969 * The generic ->writepage function for buffer-backed address_spaces
2971 int block_write_full_page(struct page *page, get_block_t *get_block,
2972 struct writeback_control *wbc)
2974 struct inode * const inode = page->mapping->host;
2975 loff_t i_size = i_size_read(inode);
2976 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2979 /* Is the page fully inside i_size? */
2980 if (page->index < end_index)
2981 return __block_write_full_page(inode, page, get_block, wbc,
2982 end_buffer_async_write);
2984 /* Is the page fully outside i_size? (truncate in progress) */
2985 offset = i_size & (PAGE_SIZE-1);
2986 if (page->index >= end_index+1 || !offset) {
2988 * The page may have dirty, unmapped buffers. For example,
2989 * they may have been added in ext3_writepage(). Make them
2990 * freeable here, so the page does not leak.
2992 do_invalidatepage(page, 0, PAGE_SIZE);
2994 return 0; /* don't care */
2998 * The page straddles i_size. It must be zeroed out on each and every
2999 * writepage invocation because it may be mmapped. "A file is mapped
3000 * in multiples of the page size. For a file that is not a multiple of
3001 * the page size, the remaining memory is zeroed when mapped, and
3002 * writes to that region are not written out to the file."
3004 zero_user_segment(page, offset, PAGE_SIZE);
3005 return __block_write_full_page(inode, page, get_block, wbc,
3006 end_buffer_async_write);
3008 EXPORT_SYMBOL(block_write_full_page);
3010 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
3011 get_block_t *get_block)
3013 struct inode *inode = mapping->host;
3014 struct buffer_head tmp = {
3015 .b_size = i_blocksize(inode),
3018 get_block(inode, block, &tmp, 0);
3019 return tmp.b_blocknr;
3021 EXPORT_SYMBOL(generic_block_bmap);
3023 static void end_bio_bh_io_sync(struct bio *bio)
3025 struct buffer_head *bh = bio->bi_private;
3027 if (unlikely(bio_flagged(bio, BIO_QUIET)))
3028 set_bit(BH_Quiet, &bh->b_state);
3030 bh->b_end_io(bh, !bio->bi_status);
3034 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3035 enum rw_hint write_hint, struct writeback_control *wbc)
3039 BUG_ON(!buffer_locked(bh));
3040 BUG_ON(!buffer_mapped(bh));
3041 BUG_ON(!bh->b_end_io);
3042 BUG_ON(buffer_delay(bh));
3043 BUG_ON(buffer_unwritten(bh));
3046 * Only clear out a write error when rewriting
3048 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3049 clear_buffer_write_io_error(bh);
3052 * from here on down, it's all bio -- do the initial mapping,
3053 * submit_bio -> generic_make_request may further map this bio around
3055 bio = bio_alloc(GFP_NOIO, 1);
3057 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3058 bio_set_dev(bio, bh->b_bdev);
3059 bio->bi_write_hint = write_hint;
3061 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3062 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3064 bio->bi_end_io = end_bio_bh_io_sync;
3065 bio->bi_private = bh;
3067 if (buffer_meta(bh))
3068 op_flags |= REQ_META;
3069 if (buffer_prio(bh))
3070 op_flags |= REQ_PRIO;
3071 bio_set_op_attrs(bio, op, op_flags);
3073 /* Take care of bh's that straddle the end of the device */
3077 wbc_init_bio(wbc, bio);
3078 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
3085 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3087 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3089 EXPORT_SYMBOL(submit_bh);
3092 * ll_rw_block: low-level access to block devices (DEPRECATED)
3093 * @op: whether to %READ or %WRITE
3094 * @op_flags: req_flag_bits
3095 * @nr: number of &struct buffer_heads in the array
3096 * @bhs: array of pointers to &struct buffer_head
3098 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3099 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3100 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3103 * This function drops any buffer that it cannot get a lock on (with the
3104 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3105 * request, and any buffer that appears to be up-to-date when doing read
3106 * request. Further it marks as clean buffers that are processed for
3107 * writing (the buffer cache won't assume that they are actually clean
3108 * until the buffer gets unlocked).
3110 * ll_rw_block sets b_end_io to simple completion handler that marks
3111 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3114 * All of the buffers must be for the same device, and must also be a
3115 * multiple of the current approved size for the device.
3117 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3121 for (i = 0; i < nr; i++) {
3122 struct buffer_head *bh = bhs[i];
3124 if (!trylock_buffer(bh))
3127 if (test_clear_buffer_dirty(bh)) {
3128 bh->b_end_io = end_buffer_write_sync;
3130 submit_bh(op, op_flags, bh);
3134 if (!buffer_uptodate(bh)) {
3135 bh->b_end_io = end_buffer_read_sync;
3137 submit_bh(op, op_flags, bh);
3144 EXPORT_SYMBOL(ll_rw_block);
3146 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3149 if (!test_clear_buffer_dirty(bh)) {
3153 bh->b_end_io = end_buffer_write_sync;
3155 submit_bh(REQ_OP_WRITE, op_flags, bh);
3157 EXPORT_SYMBOL(write_dirty_buffer);
3160 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3161 * and then start new I/O and then wait upon it. The caller must have a ref on
3164 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3168 WARN_ON(atomic_read(&bh->b_count) < 1);
3170 if (test_clear_buffer_dirty(bh)) {
3172 bh->b_end_io = end_buffer_write_sync;
3173 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3175 if (!ret && !buffer_uptodate(bh))
3182 EXPORT_SYMBOL(__sync_dirty_buffer);
3184 int sync_dirty_buffer(struct buffer_head *bh)
3186 return __sync_dirty_buffer(bh, REQ_SYNC);
3188 EXPORT_SYMBOL(sync_dirty_buffer);
3191 * try_to_free_buffers() checks if all the buffers on this particular page
3192 * are unused, and releases them if so.
3194 * Exclusion against try_to_free_buffers may be obtained by either
3195 * locking the page or by holding its mapping's private_lock.
3197 * If the page is dirty but all the buffers are clean then we need to
3198 * be sure to mark the page clean as well. This is because the page
3199 * may be against a block device, and a later reattachment of buffers
3200 * to a dirty page will set *all* buffers dirty. Which would corrupt
3201 * filesystem data on the same device.
3203 * The same applies to regular filesystem pages: if all the buffers are
3204 * clean then we set the page clean and proceed. To do that, we require
3205 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3208 * try_to_free_buffers() is non-blocking.
3210 static inline int buffer_busy(struct buffer_head *bh)
3212 return atomic_read(&bh->b_count) |
3213 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3217 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3219 struct buffer_head *head = page_buffers(page);
3220 struct buffer_head *bh;
3224 if (buffer_busy(bh))
3226 bh = bh->b_this_page;
3227 } while (bh != head);
3230 struct buffer_head *next = bh->b_this_page;
3232 if (bh->b_assoc_map)
3233 __remove_assoc_queue(bh);
3235 } while (bh != head);
3236 *buffers_to_free = head;
3237 __clear_page_buffers(page);
3243 int try_to_free_buffers(struct page *page)
3245 struct address_space * const mapping = page->mapping;
3246 struct buffer_head *buffers_to_free = NULL;
3249 BUG_ON(!PageLocked(page));
3250 if (PageWriteback(page))
3253 if (mapping == NULL) { /* can this still happen? */
3254 ret = drop_buffers(page, &buffers_to_free);
3258 spin_lock(&mapping->private_lock);
3259 ret = drop_buffers(page, &buffers_to_free);
3262 * If the filesystem writes its buffers by hand (eg ext3)
3263 * then we can have clean buffers against a dirty page. We
3264 * clean the page here; otherwise the VM will never notice
3265 * that the filesystem did any IO at all.
3267 * Also, during truncate, discard_buffer will have marked all
3268 * the page's buffers clean. We discover that here and clean
3271 * private_lock must be held over this entire operation in order
3272 * to synchronise against __set_page_dirty_buffers and prevent the
3273 * dirty bit from being lost.
3276 cancel_dirty_page(page);
3277 spin_unlock(&mapping->private_lock);
3279 if (buffers_to_free) {
3280 struct buffer_head *bh = buffers_to_free;
3283 struct buffer_head *next = bh->b_this_page;
3284 free_buffer_head(bh);
3286 } while (bh != buffers_to_free);
3290 EXPORT_SYMBOL(try_to_free_buffers);
3293 * There are no bdflush tunables left. But distributions are
3294 * still running obsolete flush daemons, so we terminate them here.
3296 * Use of bdflush() is deprecated and will be removed in a future kernel.
3297 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3299 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3301 static int msg_count;
3303 if (!capable(CAP_SYS_ADMIN))
3306 if (msg_count < 5) {
3309 "warning: process `%s' used the obsolete bdflush"
3310 " system call\n", current->comm);
3311 printk(KERN_INFO "Fix your initscripts?\n");
3320 * Buffer-head allocation
3322 static struct kmem_cache *bh_cachep __read_mostly;
3325 * Once the number of bh's in the machine exceeds this level, we start
3326 * stripping them in writeback.
3328 static unsigned long max_buffer_heads;
3330 int buffer_heads_over_limit;
3332 struct bh_accounting {
3333 int nr; /* Number of live bh's */
3334 int ratelimit; /* Limit cacheline bouncing */
3337 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3339 static void recalc_bh_state(void)
3344 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3346 __this_cpu_write(bh_accounting.ratelimit, 0);
3347 for_each_online_cpu(i)
3348 tot += per_cpu(bh_accounting, i).nr;
3349 buffer_heads_over_limit = (tot > max_buffer_heads);
3352 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3354 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3356 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3357 spin_lock_init(&ret->b_uptodate_lock);
3359 __this_cpu_inc(bh_accounting.nr);
3365 EXPORT_SYMBOL(alloc_buffer_head);
3367 void free_buffer_head(struct buffer_head *bh)
3369 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3370 kmem_cache_free(bh_cachep, bh);
3372 __this_cpu_dec(bh_accounting.nr);
3376 EXPORT_SYMBOL(free_buffer_head);
3378 static int buffer_exit_cpu_dead(unsigned int cpu)
3381 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3383 for (i = 0; i < BH_LRU_SIZE; i++) {
3387 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3388 per_cpu(bh_accounting, cpu).nr = 0;
3393 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3394 * @bh: struct buffer_head
3396 * Return true if the buffer is up-to-date and false,
3397 * with the buffer locked, if not.
3399 int bh_uptodate_or_lock(struct buffer_head *bh)
3401 if (!buffer_uptodate(bh)) {
3403 if (!buffer_uptodate(bh))
3409 EXPORT_SYMBOL(bh_uptodate_or_lock);
3412 * bh_submit_read - Submit a locked buffer for reading
3413 * @bh: struct buffer_head
3415 * Returns zero on success and -EIO on error.
3417 int bh_submit_read(struct buffer_head *bh)
3419 BUG_ON(!buffer_locked(bh));
3421 if (buffer_uptodate(bh)) {
3427 bh->b_end_io = end_buffer_read_sync;
3428 submit_bh(REQ_OP_READ, 0, bh);
3430 if (buffer_uptodate(bh))
3434 EXPORT_SYMBOL(bh_submit_read);
3436 void __init buffer_init(void)
3438 unsigned long nrpages;
3441 bh_cachep = kmem_cache_create("buffer_head",
3442 sizeof(struct buffer_head), 0,
3443 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3448 * Limit the bh occupancy to 10% of ZONE_NORMAL
3450 nrpages = (nr_free_buffer_pages() * 10) / 100;
3451 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3452 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3453 NULL, buffer_exit_cpu_dead);