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>
51 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
52 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
53 enum rw_hint hint, struct writeback_control *wbc);
55 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
57 inline void touch_buffer(struct buffer_head *bh)
59 trace_block_touch_buffer(bh);
60 mark_page_accessed(bh->b_page);
62 EXPORT_SYMBOL(touch_buffer);
64 void __lock_buffer(struct buffer_head *bh)
66 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
68 EXPORT_SYMBOL(__lock_buffer);
70 void unlock_buffer(struct buffer_head *bh)
72 clear_bit_unlock(BH_Lock, &bh->b_state);
73 smp_mb__after_atomic();
74 wake_up_bit(&bh->b_state, BH_Lock);
76 EXPORT_SYMBOL(unlock_buffer);
79 * Returns if the page has dirty or writeback buffers. If all the buffers
80 * are unlocked and clean then the PageDirty information is stale. If
81 * any of the pages are locked, it is assumed they are locked for IO.
83 void buffer_check_dirty_writeback(struct page *page,
84 bool *dirty, bool *writeback)
86 struct buffer_head *head, *bh;
90 BUG_ON(!PageLocked(page));
92 if (!page_has_buffers(page))
95 if (PageWriteback(page))
98 head = page_buffers(page);
101 if (buffer_locked(bh))
104 if (buffer_dirty(bh))
107 bh = bh->b_this_page;
108 } while (bh != head);
110 EXPORT_SYMBOL(buffer_check_dirty_writeback);
113 * Block until a buffer comes unlocked. This doesn't stop it
114 * from becoming locked again - you have to lock it yourself
115 * if you want to preserve its state.
117 void __wait_on_buffer(struct buffer_head * bh)
119 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
121 EXPORT_SYMBOL(__wait_on_buffer);
124 __clear_page_buffers(struct page *page)
126 ClearPagePrivate(page);
127 set_page_private(page, 0);
131 static void buffer_io_error(struct buffer_head *bh, char *msg)
133 if (!test_bit(BH_Quiet, &bh->b_state))
134 printk_ratelimited(KERN_ERR
135 "Buffer I/O error on dev %pg, logical block %llu%s\n",
136 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
140 * End-of-IO handler helper function which does not touch the bh after
142 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
143 * a race there is benign: unlock_buffer() only use the bh's address for
144 * hashing after unlocking the buffer, so it doesn't actually touch the bh
147 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
150 set_buffer_uptodate(bh);
152 /* This happens, due to failed read-ahead attempts. */
153 clear_buffer_uptodate(bh);
159 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
160 * unlock the buffer. This is what ll_rw_block uses too.
162 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
164 __end_buffer_read_notouch(bh, uptodate);
167 EXPORT_SYMBOL(end_buffer_read_sync);
169 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
172 set_buffer_uptodate(bh);
174 buffer_io_error(bh, ", lost sync page write");
175 mark_buffer_write_io_error(bh);
176 clear_buffer_uptodate(bh);
181 EXPORT_SYMBOL(end_buffer_write_sync);
184 * Various filesystems appear to want __find_get_block to be non-blocking.
185 * But it's the page lock which protects the buffers. To get around this,
186 * we get exclusion from try_to_free_buffers with the blockdev mapping's
189 * Hack idea: for the blockdev mapping, private_lock contention
190 * may be quite high. This code could TryLock the page, and if that
191 * succeeds, there is no need to take private_lock.
193 static struct buffer_head *
194 __find_get_block_slow(struct block_device *bdev, sector_t block)
196 struct inode *bd_inode = bdev->bd_inode;
197 struct address_space *bd_mapping = bd_inode->i_mapping;
198 struct buffer_head *ret = NULL;
200 struct buffer_head *bh;
201 struct buffer_head *head;
204 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
206 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
207 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
211 spin_lock(&bd_mapping->private_lock);
212 if (!page_has_buffers(page))
214 head = page_buffers(page);
217 if (!buffer_mapped(bh))
219 else if (bh->b_blocknr == block) {
224 bh = bh->b_this_page;
225 } while (bh != head);
227 /* we might be here because some of the buffers on this page are
228 * not mapped. This is due to various races between
229 * file io on the block device and getblk. It gets dealt with
230 * elsewhere, don't buffer_error if we had some unmapped buffers
232 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
233 if (all_mapped && __ratelimit(&last_warned)) {
234 printk("__find_get_block_slow() failed. block=%llu, "
235 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
236 "device %pg blocksize: %d\n",
237 (unsigned long long)block,
238 (unsigned long long)bh->b_blocknr,
239 bh->b_state, bh->b_size, bdev,
240 1 << bd_inode->i_blkbits);
243 spin_unlock(&bd_mapping->private_lock);
250 * I/O completion handler for block_read_full_page() - pages
251 * which come unlocked at the end of I/O.
253 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
256 struct buffer_head *first;
257 struct buffer_head *tmp;
259 int page_uptodate = 1;
261 BUG_ON(!buffer_async_read(bh));
265 set_buffer_uptodate(bh);
267 clear_buffer_uptodate(bh);
268 buffer_io_error(bh, ", async page read");
273 * Be _very_ careful from here on. Bad things can happen if
274 * two buffer heads end IO at almost the same time and both
275 * decide that the page is now completely done.
277 first = page_buffers(page);
278 local_irq_save(flags);
279 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
280 clear_buffer_async_read(bh);
284 if (!buffer_uptodate(tmp))
286 if (buffer_async_read(tmp)) {
287 BUG_ON(!buffer_locked(tmp));
290 tmp = tmp->b_this_page;
292 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
293 local_irq_restore(flags);
296 * If none of the buffers had errors and they are all
297 * uptodate then we can set the page uptodate.
299 if (page_uptodate && !PageError(page))
300 SetPageUptodate(page);
305 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
306 local_irq_restore(flags);
311 * Completion handler for block_write_full_page() - pages which are unlocked
312 * during I/O, and which have PageWriteback cleared upon I/O completion.
314 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
317 struct buffer_head *first;
318 struct buffer_head *tmp;
321 BUG_ON(!buffer_async_write(bh));
325 set_buffer_uptodate(bh);
327 buffer_io_error(bh, ", lost async page write");
328 mark_buffer_write_io_error(bh);
329 clear_buffer_uptodate(bh);
333 first = page_buffers(page);
334 local_irq_save(flags);
335 bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
337 clear_buffer_async_write(bh);
339 tmp = bh->b_this_page;
341 if (buffer_async_write(tmp)) {
342 BUG_ON(!buffer_locked(tmp));
345 tmp = tmp->b_this_page;
347 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
348 local_irq_restore(flags);
349 end_page_writeback(page);
353 bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
354 local_irq_restore(flags);
357 EXPORT_SYMBOL(end_buffer_async_write);
360 * If a page's buffers are under async readin (end_buffer_async_read
361 * completion) then there is a possibility that another thread of
362 * control could lock one of the buffers after it has completed
363 * but while some of the other buffers have not completed. This
364 * locked buffer would confuse end_buffer_async_read() into not unlocking
365 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
366 * that this buffer is not under async I/O.
368 * The page comes unlocked when it has no locked buffer_async buffers
371 * PageLocked prevents anyone starting new async I/O reads any of
374 * PageWriteback is used to prevent simultaneous writeout of the same
377 * PageLocked prevents anyone from starting writeback of a page which is
378 * under read I/O (PageWriteback is only ever set against a locked page).
380 static void mark_buffer_async_read(struct buffer_head *bh)
382 bh->b_end_io = end_buffer_async_read;
383 set_buffer_async_read(bh);
386 static void mark_buffer_async_write_endio(struct buffer_head *bh,
387 bh_end_io_t *handler)
389 bh->b_end_io = handler;
390 set_buffer_async_write(bh);
393 void mark_buffer_async_write(struct buffer_head *bh)
395 mark_buffer_async_write_endio(bh, end_buffer_async_write);
397 EXPORT_SYMBOL(mark_buffer_async_write);
401 * fs/buffer.c contains helper functions for buffer-backed address space's
402 * fsync functions. A common requirement for buffer-based filesystems is
403 * that certain data from the backing blockdev needs to be written out for
404 * a successful fsync(). For example, ext2 indirect blocks need to be
405 * written back and waited upon before fsync() returns.
407 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
408 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
409 * management of a list of dependent buffers at ->i_mapping->private_list.
411 * Locking is a little subtle: try_to_free_buffers() will remove buffers
412 * from their controlling inode's queue when they are being freed. But
413 * try_to_free_buffers() will be operating against the *blockdev* mapping
414 * at the time, not against the S_ISREG file which depends on those buffers.
415 * So the locking for private_list is via the private_lock in the address_space
416 * which backs the buffers. Which is different from the address_space
417 * against which the buffers are listed. So for a particular address_space,
418 * mapping->private_lock does *not* protect mapping->private_list! In fact,
419 * mapping->private_list will always be protected by the backing blockdev's
422 * Which introduces a requirement: all buffers on an address_space's
423 * ->private_list must be from the same address_space: the blockdev's.
425 * address_spaces which do not place buffers at ->private_list via these
426 * utility functions are free to use private_lock and private_list for
427 * whatever they want. The only requirement is that list_empty(private_list)
428 * be true at clear_inode() time.
430 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
431 * filesystems should do that. invalidate_inode_buffers() should just go
432 * BUG_ON(!list_empty).
434 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
435 * take an address_space, not an inode. And it should be called
436 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
439 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
440 * list if it is already on a list. Because if the buffer is on a list,
441 * it *must* already be on the right one. If not, the filesystem is being
442 * silly. This will save a ton of locking. But first we have to ensure
443 * that buffers are taken *off* the old inode's list when they are freed
444 * (presumably in truncate). That requires careful auditing of all
445 * filesystems (do it inside bforget()). It could also be done by bringing
450 * The buffer's backing address_space's private_lock must be held
452 static void __remove_assoc_queue(struct buffer_head *bh)
454 list_del_init(&bh->b_assoc_buffers);
455 WARN_ON(!bh->b_assoc_map);
456 bh->b_assoc_map = NULL;
459 int inode_has_buffers(struct inode *inode)
461 return !list_empty(&inode->i_data.private_list);
465 * osync is designed to support O_SYNC io. It waits synchronously for
466 * all already-submitted IO to complete, but does not queue any new
467 * writes to the disk.
469 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
470 * you dirty the buffers, and then use osync_inode_buffers to wait for
471 * completion. Any other dirty buffers which are not yet queued for
472 * write will not be flushed to disk by the osync.
474 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
476 struct buffer_head *bh;
482 list_for_each_prev(p, list) {
484 if (buffer_locked(bh)) {
488 if (!buffer_uptodate(bh))
499 void emergency_thaw_bdev(struct super_block *sb)
501 while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
502 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
506 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
507 * @mapping: the mapping which wants those buffers written
509 * Starts I/O against the buffers at mapping->private_list, and waits upon
512 * Basically, this is a convenience function for fsync().
513 * @mapping is a file or directory which needs those buffers to be written for
514 * a successful fsync().
516 int sync_mapping_buffers(struct address_space *mapping)
518 struct address_space *buffer_mapping = mapping->private_data;
520 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
523 return fsync_buffers_list(&buffer_mapping->private_lock,
524 &mapping->private_list);
526 EXPORT_SYMBOL(sync_mapping_buffers);
529 * Called when we've recently written block `bblock', and it is known that
530 * `bblock' was for a buffer_boundary() buffer. This means that the block at
531 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
532 * dirty, schedule it for IO. So that indirects merge nicely with their data.
534 void write_boundary_block(struct block_device *bdev,
535 sector_t bblock, unsigned blocksize)
537 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
539 if (buffer_dirty(bh))
540 ll_rw_block(REQ_OP_WRITE, 0, 1, &bh);
545 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
547 struct address_space *mapping = inode->i_mapping;
548 struct address_space *buffer_mapping = bh->b_page->mapping;
550 mark_buffer_dirty(bh);
551 if (!mapping->private_data) {
552 mapping->private_data = buffer_mapping;
554 BUG_ON(mapping->private_data != buffer_mapping);
556 if (!bh->b_assoc_map) {
557 spin_lock(&buffer_mapping->private_lock);
558 list_move_tail(&bh->b_assoc_buffers,
559 &mapping->private_list);
560 bh->b_assoc_map = mapping;
561 spin_unlock(&buffer_mapping->private_lock);
564 EXPORT_SYMBOL(mark_buffer_dirty_inode);
567 * Mark the page dirty, and set it dirty in the page cache, and mark the inode
570 * If warn is true, then emit a warning if the page is not uptodate and has
571 * not been truncated.
573 * The caller must hold lock_page_memcg().
575 void __set_page_dirty(struct page *page, struct address_space *mapping,
580 xa_lock_irqsave(&mapping->i_pages, flags);
581 if (page->mapping) { /* Race with truncate? */
582 WARN_ON_ONCE(warn && !PageUptodate(page));
583 account_page_dirtied(page, mapping);
584 __xa_set_mark(&mapping->i_pages, page_index(page),
585 PAGECACHE_TAG_DIRTY);
587 xa_unlock_irqrestore(&mapping->i_pages, flags);
589 EXPORT_SYMBOL_GPL(__set_page_dirty);
592 * Add a page to the dirty page list.
594 * It is a sad fact of life that this function is called from several places
595 * deeply under spinlocking. It may not sleep.
597 * If the page has buffers, the uptodate buffers are set dirty, to preserve
598 * dirty-state coherency between the page and the buffers. It the page does
599 * not have buffers then when they are later attached they will all be set
602 * The buffers are dirtied before the page is dirtied. There's a small race
603 * window in which a writepage caller may see the page cleanness but not the
604 * buffer dirtiness. That's fine. If this code were to set the page dirty
605 * before the buffers, a concurrent writepage caller could clear the page dirty
606 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
607 * page on the dirty page list.
609 * We use private_lock to lock against try_to_free_buffers while using the
610 * page's buffer list. Also use this to protect against clean buffers being
611 * added to the page after it was set dirty.
613 * FIXME: may need to call ->reservepage here as well. That's rather up to the
614 * address_space though.
616 int __set_page_dirty_buffers(struct page *page)
619 struct address_space *mapping = page_mapping(page);
621 if (unlikely(!mapping))
622 return !TestSetPageDirty(page);
624 spin_lock(&mapping->private_lock);
625 if (page_has_buffers(page)) {
626 struct buffer_head *head = page_buffers(page);
627 struct buffer_head *bh = head;
630 set_buffer_dirty(bh);
631 bh = bh->b_this_page;
632 } while (bh != head);
635 * Lock out page->mem_cgroup migration to keep PageDirty
636 * synchronized with per-memcg dirty page counters.
638 lock_page_memcg(page);
639 newly_dirty = !TestSetPageDirty(page);
640 spin_unlock(&mapping->private_lock);
643 __set_page_dirty(page, mapping, 1);
645 unlock_page_memcg(page);
648 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
652 EXPORT_SYMBOL(__set_page_dirty_buffers);
655 * Write out and wait upon a list of buffers.
657 * We have conflicting pressures: we want to make sure that all
658 * initially dirty buffers get waited on, but that any subsequently
659 * dirtied buffers don't. After all, we don't want fsync to last
660 * forever if somebody is actively writing to the file.
662 * Do this in two main stages: first we copy dirty buffers to a
663 * temporary inode list, queueing the writes as we go. Then we clean
664 * up, waiting for those writes to complete.
666 * During this second stage, any subsequent updates to the file may end
667 * up refiling the buffer on the original inode's dirty list again, so
668 * there is a chance we will end up with a buffer queued for write but
669 * not yet completed on that list. So, as a final cleanup we go through
670 * the osync code to catch these locked, dirty buffers without requeuing
671 * any newly dirty buffers for write.
673 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
675 struct buffer_head *bh;
676 struct list_head tmp;
677 struct address_space *mapping;
679 struct blk_plug plug;
681 INIT_LIST_HEAD(&tmp);
682 blk_start_plug(&plug);
685 while (!list_empty(list)) {
686 bh = BH_ENTRY(list->next);
687 mapping = bh->b_assoc_map;
688 __remove_assoc_queue(bh);
689 /* Avoid race with mark_buffer_dirty_inode() which does
690 * a lockless check and we rely on seeing the dirty bit */
692 if (buffer_dirty(bh) || buffer_locked(bh)) {
693 list_add(&bh->b_assoc_buffers, &tmp);
694 bh->b_assoc_map = mapping;
695 if (buffer_dirty(bh)) {
699 * Ensure any pending I/O completes so that
700 * write_dirty_buffer() actually writes the
701 * current contents - it is a noop if I/O is
702 * still in flight on potentially older
705 write_dirty_buffer(bh, REQ_SYNC);
708 * Kick off IO for the previous mapping. Note
709 * that we will not run the very last mapping,
710 * wait_on_buffer() will do that for us
711 * through sync_buffer().
720 blk_finish_plug(&plug);
723 while (!list_empty(&tmp)) {
724 bh = BH_ENTRY(tmp.prev);
726 mapping = bh->b_assoc_map;
727 __remove_assoc_queue(bh);
728 /* Avoid race with mark_buffer_dirty_inode() which does
729 * a lockless check and we rely on seeing the dirty bit */
731 if (buffer_dirty(bh)) {
732 list_add(&bh->b_assoc_buffers,
733 &mapping->private_list);
734 bh->b_assoc_map = mapping;
738 if (!buffer_uptodate(bh))
745 err2 = osync_buffers_list(lock, list);
753 * Invalidate any and all dirty buffers on a given inode. We are
754 * probably unmounting the fs, but that doesn't mean we have already
755 * done a sync(). Just drop the buffers from the inode list.
757 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
758 * assumes that all the buffers are against the blockdev. Not true
761 void invalidate_inode_buffers(struct inode *inode)
763 if (inode_has_buffers(inode)) {
764 struct address_space *mapping = &inode->i_data;
765 struct list_head *list = &mapping->private_list;
766 struct address_space *buffer_mapping = mapping->private_data;
768 spin_lock(&buffer_mapping->private_lock);
769 while (!list_empty(list))
770 __remove_assoc_queue(BH_ENTRY(list->next));
771 spin_unlock(&buffer_mapping->private_lock);
774 EXPORT_SYMBOL(invalidate_inode_buffers);
777 * Remove any clean buffers from the inode's buffer list. This is called
778 * when we're trying to free the inode itself. Those buffers can pin it.
780 * Returns true if all buffers were removed.
782 int remove_inode_buffers(struct inode *inode)
786 if (inode_has_buffers(inode)) {
787 struct address_space *mapping = &inode->i_data;
788 struct list_head *list = &mapping->private_list;
789 struct address_space *buffer_mapping = mapping->private_data;
791 spin_lock(&buffer_mapping->private_lock);
792 while (!list_empty(list)) {
793 struct buffer_head *bh = BH_ENTRY(list->next);
794 if (buffer_dirty(bh)) {
798 __remove_assoc_queue(bh);
800 spin_unlock(&buffer_mapping->private_lock);
806 * Create the appropriate buffers when given a page for data area and
807 * the size of each buffer.. Use the bh->b_this_page linked list to
808 * follow the buffers created. Return NULL if unable to create more
811 * The retry flag is used to differentiate async IO (paging, swapping)
812 * which may not fail from ordinary buffer allocations.
814 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
817 struct buffer_head *bh, *head;
818 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
820 struct mem_cgroup *memcg;
825 memcg = get_mem_cgroup_from_page(page);
826 memalloc_use_memcg(memcg);
830 while ((offset -= size) >= 0) {
831 bh = alloc_buffer_head(gfp);
835 bh->b_this_page = head;
841 /* Link the buffer to its page */
842 set_bh_page(bh, page, offset);
845 memalloc_unuse_memcg();
846 mem_cgroup_put(memcg);
849 * In case anything failed, we just free everything we got.
855 head = head->b_this_page;
856 free_buffer_head(bh);
862 EXPORT_SYMBOL_GPL(alloc_page_buffers);
865 link_dev_buffers(struct page *page, struct buffer_head *head)
867 struct buffer_head *bh, *tail;
872 bh = bh->b_this_page;
874 tail->b_this_page = head;
875 attach_page_buffers(page, head);
878 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
880 sector_t retval = ~((sector_t)0);
881 loff_t sz = i_size_read(bdev->bd_inode);
884 unsigned int sizebits = blksize_bits(size);
885 retval = (sz >> sizebits);
891 * Initialise the state of a blockdev page's buffers.
894 init_page_buffers(struct page *page, struct block_device *bdev,
895 sector_t block, int size)
897 struct buffer_head *head = page_buffers(page);
898 struct buffer_head *bh = head;
899 int uptodate = PageUptodate(page);
900 sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
903 if (!buffer_mapped(bh)) {
905 bh->b_private = NULL;
907 bh->b_blocknr = block;
909 set_buffer_uptodate(bh);
910 if (block < end_block)
911 set_buffer_mapped(bh);
914 bh = bh->b_this_page;
915 } while (bh != head);
918 * Caller needs to validate requested block against end of device.
924 * Create the page-cache page that contains the requested block.
926 * This is used purely for blockdev mappings.
929 grow_dev_page(struct block_device *bdev, sector_t block,
930 pgoff_t index, int size, int sizebits, gfp_t gfp)
932 struct inode *inode = bdev->bd_inode;
934 struct buffer_head *bh;
936 int ret = 0; /* Will call free_more_memory() */
939 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
942 * XXX: __getblk_slow() can not really deal with failure and
943 * will endlessly loop on improvised global reclaim. Prefer
944 * looping in the allocator rather than here, at least that
945 * code knows what it's doing.
947 gfp_mask |= __GFP_NOFAIL;
949 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
951 BUG_ON(!PageLocked(page));
953 if (page_has_buffers(page)) {
954 bh = page_buffers(page);
955 if (bh->b_size == size) {
956 end_block = init_page_buffers(page, bdev,
957 (sector_t)index << sizebits,
961 if (!try_to_free_buffers(page))
966 * Allocate some buffers for this page
968 bh = alloc_page_buffers(page, size, true);
971 * Link the page to the buffers and initialise them. Take the
972 * lock to be atomic wrt __find_get_block(), which does not
973 * run under the page lock.
975 spin_lock(&inode->i_mapping->private_lock);
976 link_dev_buffers(page, bh);
977 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
979 spin_unlock(&inode->i_mapping->private_lock);
981 ret = (block < end_block) ? 1 : -ENXIO;
989 * Create buffers for the specified block device block's page. If
990 * that page was dirty, the buffers are set dirty also.
993 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1001 } while ((size << sizebits) < PAGE_SIZE);
1003 index = block >> sizebits;
1006 * Check for a block which wants to lie outside our maximum possible
1007 * pagecache index. (this comparison is done using sector_t types).
1009 if (unlikely(index != block >> sizebits)) {
1010 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1012 __func__, (unsigned long long)block,
1017 /* Create a page with the proper size buffers.. */
1018 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1021 static struct buffer_head *
1022 __getblk_slow(struct block_device *bdev, sector_t block,
1023 unsigned size, gfp_t gfp)
1025 /* Size must be multiple of hard sectorsize */
1026 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1027 (size < 512 || size > PAGE_SIZE))) {
1028 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1030 printk(KERN_ERR "logical block size: %d\n",
1031 bdev_logical_block_size(bdev));
1038 struct buffer_head *bh;
1041 bh = __find_get_block(bdev, block, size);
1045 ret = grow_buffers(bdev, block, size, gfp);
1052 * The relationship between dirty buffers and dirty pages:
1054 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1055 * the page is tagged dirty in the page cache.
1057 * At all times, the dirtiness of the buffers represents the dirtiness of
1058 * subsections of the page. If the page has buffers, the page dirty bit is
1059 * merely a hint about the true dirty state.
1061 * When a page is set dirty in its entirety, all its buffers are marked dirty
1062 * (if the page has buffers).
1064 * When a buffer is marked dirty, its page is dirtied, but the page's other
1067 * Also. When blockdev buffers are explicitly read with bread(), they
1068 * individually become uptodate. But their backing page remains not
1069 * uptodate - even if all of its buffers are uptodate. A subsequent
1070 * block_read_full_page() against that page will discover all the uptodate
1071 * buffers, will set the page uptodate and will perform no I/O.
1075 * mark_buffer_dirty - mark a buffer_head as needing writeout
1076 * @bh: the buffer_head to mark dirty
1078 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1079 * its backing page dirty, then tag the page as dirty in the page cache
1080 * and then attach the address_space's inode to its superblock's dirty
1083 * mark_buffer_dirty() is atomic. It takes bh->b_page->mapping->private_lock,
1084 * i_pages lock and mapping->host->i_lock.
1086 void mark_buffer_dirty(struct buffer_head *bh)
1088 WARN_ON_ONCE(!buffer_uptodate(bh));
1090 trace_block_dirty_buffer(bh);
1093 * Very *carefully* optimize the it-is-already-dirty case.
1095 * Don't let the final "is it dirty" escape to before we
1096 * perhaps modified the buffer.
1098 if (buffer_dirty(bh)) {
1100 if (buffer_dirty(bh))
1104 if (!test_set_buffer_dirty(bh)) {
1105 struct page *page = bh->b_page;
1106 struct address_space *mapping = NULL;
1108 lock_page_memcg(page);
1109 if (!TestSetPageDirty(page)) {
1110 mapping = page_mapping(page);
1112 __set_page_dirty(page, mapping, 0);
1114 unlock_page_memcg(page);
1116 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1119 EXPORT_SYMBOL(mark_buffer_dirty);
1121 void mark_buffer_write_io_error(struct buffer_head *bh)
1123 set_buffer_write_io_error(bh);
1124 /* FIXME: do we need to set this in both places? */
1125 if (bh->b_page && bh->b_page->mapping)
1126 mapping_set_error(bh->b_page->mapping, -EIO);
1127 if (bh->b_assoc_map)
1128 mapping_set_error(bh->b_assoc_map, -EIO);
1130 EXPORT_SYMBOL(mark_buffer_write_io_error);
1133 * Decrement a buffer_head's reference count. If all buffers against a page
1134 * have zero reference count, are clean and unlocked, and if the page is clean
1135 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1136 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1137 * a page but it ends up not being freed, and buffers may later be reattached).
1139 void __brelse(struct buffer_head * buf)
1141 if (atomic_read(&buf->b_count)) {
1145 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1147 EXPORT_SYMBOL(__brelse);
1150 * bforget() is like brelse(), except it discards any
1151 * potentially dirty data.
1153 void __bforget(struct buffer_head *bh)
1155 clear_buffer_dirty(bh);
1156 if (bh->b_assoc_map) {
1157 struct address_space *buffer_mapping = bh->b_page->mapping;
1159 spin_lock(&buffer_mapping->private_lock);
1160 list_del_init(&bh->b_assoc_buffers);
1161 bh->b_assoc_map = NULL;
1162 spin_unlock(&buffer_mapping->private_lock);
1166 EXPORT_SYMBOL(__bforget);
1168 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1171 if (buffer_uptodate(bh)) {
1176 bh->b_end_io = end_buffer_read_sync;
1177 submit_bh(REQ_OP_READ, 0, bh);
1179 if (buffer_uptodate(bh))
1187 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1188 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1189 * refcount elevated by one when they're in an LRU. A buffer can only appear
1190 * once in a particular CPU's LRU. A single buffer can be present in multiple
1191 * CPU's LRUs at the same time.
1193 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1194 * sb_find_get_block().
1196 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1197 * a local interrupt disable for that.
1200 #define BH_LRU_SIZE 16
1203 struct buffer_head *bhs[BH_LRU_SIZE];
1206 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1209 #define bh_lru_lock() local_irq_disable()
1210 #define bh_lru_unlock() local_irq_enable()
1212 #define bh_lru_lock() preempt_disable()
1213 #define bh_lru_unlock() preempt_enable()
1216 static inline void check_irqs_on(void)
1218 #ifdef irqs_disabled
1219 BUG_ON(irqs_disabled());
1224 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1225 * inserted at the front, and the buffer_head at the back if any is evicted.
1226 * Or, if already in the LRU it is moved to the front.
1228 static void bh_lru_install(struct buffer_head *bh)
1230 struct buffer_head *evictee = bh;
1237 b = this_cpu_ptr(&bh_lrus);
1238 for (i = 0; i < BH_LRU_SIZE; i++) {
1239 swap(evictee, b->bhs[i]);
1240 if (evictee == bh) {
1252 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1254 static struct buffer_head *
1255 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1257 struct buffer_head *ret = NULL;
1262 for (i = 0; i < BH_LRU_SIZE; i++) {
1263 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1265 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1266 bh->b_size == size) {
1269 __this_cpu_write(bh_lrus.bhs[i],
1270 __this_cpu_read(bh_lrus.bhs[i - 1]));
1273 __this_cpu_write(bh_lrus.bhs[0], bh);
1285 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1286 * it in the LRU and mark it as accessed. If it is not present then return
1289 struct buffer_head *
1290 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1292 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1295 /* __find_get_block_slow will mark the page accessed */
1296 bh = __find_get_block_slow(bdev, block);
1304 EXPORT_SYMBOL(__find_get_block);
1307 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1308 * which corresponds to the passed block_device, block and size. The
1309 * returned buffer has its reference count incremented.
1311 * __getblk_gfp() will lock up the machine if grow_dev_page's
1312 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1314 struct buffer_head *
1315 __getblk_gfp(struct block_device *bdev, sector_t block,
1316 unsigned size, gfp_t gfp)
1318 struct buffer_head *bh = __find_get_block(bdev, block, size);
1322 bh = __getblk_slow(bdev, block, size, gfp);
1325 EXPORT_SYMBOL(__getblk_gfp);
1328 * Do async read-ahead on a buffer..
1330 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1332 struct buffer_head *bh = __getblk(bdev, block, size);
1334 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1338 EXPORT_SYMBOL(__breadahead);
1340 void __breadahead_gfp(struct block_device *bdev, sector_t block, unsigned size,
1343 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1345 ll_rw_block(REQ_OP_READ, REQ_RAHEAD, 1, &bh);
1349 EXPORT_SYMBOL(__breadahead_gfp);
1352 * __bread_gfp() - reads a specified block and returns the bh
1353 * @bdev: the block_device to read from
1354 * @block: number of block
1355 * @size: size (in bytes) to read
1356 * @gfp: page allocation flag
1358 * Reads a specified block, and returns buffer head that contains it.
1359 * The page cache can be allocated from non-movable area
1360 * not to prevent page migration if you set gfp to zero.
1361 * It returns NULL if the block was unreadable.
1363 struct buffer_head *
1364 __bread_gfp(struct block_device *bdev, sector_t block,
1365 unsigned size, gfp_t gfp)
1367 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1369 if (likely(bh) && !buffer_uptodate(bh))
1370 bh = __bread_slow(bh);
1373 EXPORT_SYMBOL(__bread_gfp);
1376 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1377 * This doesn't race because it runs in each cpu either in irq
1378 * or with preempt disabled.
1380 static void invalidate_bh_lru(void *arg)
1382 struct bh_lru *b = &get_cpu_var(bh_lrus);
1385 for (i = 0; i < BH_LRU_SIZE; i++) {
1389 put_cpu_var(bh_lrus);
1392 static bool has_bh_in_lru(int cpu, void *dummy)
1394 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1397 for (i = 0; i < BH_LRU_SIZE; i++) {
1405 void invalidate_bh_lrus(void)
1407 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1409 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1411 void set_bh_page(struct buffer_head *bh,
1412 struct page *page, unsigned long offset)
1415 BUG_ON(offset >= PAGE_SIZE);
1416 if (PageHighMem(page))
1418 * This catches illegal uses and preserves the offset:
1420 bh->b_data = (char *)(0 + offset);
1422 bh->b_data = page_address(page) + offset;
1424 EXPORT_SYMBOL(set_bh_page);
1427 * Called when truncating a buffer on a page completely.
1430 /* Bits that are cleared during an invalidate */
1431 #define BUFFER_FLAGS_DISCARD \
1432 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1433 1 << BH_Delay | 1 << BH_Unwritten)
1435 static void discard_buffer(struct buffer_head * bh)
1437 unsigned long b_state, b_state_old;
1440 clear_buffer_dirty(bh);
1442 b_state = bh->b_state;
1444 b_state_old = cmpxchg(&bh->b_state, b_state,
1445 (b_state & ~BUFFER_FLAGS_DISCARD));
1446 if (b_state_old == b_state)
1448 b_state = b_state_old;
1454 * block_invalidatepage - invalidate part or all of a buffer-backed page
1456 * @page: the page which is affected
1457 * @offset: start of the range to invalidate
1458 * @length: length of the range to invalidate
1460 * block_invalidatepage() is called when all or part of the page has become
1461 * invalidated by a truncate operation.
1463 * block_invalidatepage() does not have to release all buffers, but it must
1464 * ensure that no dirty buffer is left outside @offset and that no I/O
1465 * is underway against any of the blocks which are outside the truncation
1466 * point. Because the caller is about to free (and possibly reuse) those
1469 void block_invalidatepage(struct page *page, unsigned int offset,
1470 unsigned int length)
1472 struct buffer_head *head, *bh, *next;
1473 unsigned int curr_off = 0;
1474 unsigned int stop = length + offset;
1476 BUG_ON(!PageLocked(page));
1477 if (!page_has_buffers(page))
1481 * Check for overflow
1483 BUG_ON(stop > PAGE_SIZE || stop < length);
1485 head = page_buffers(page);
1488 unsigned int next_off = curr_off + bh->b_size;
1489 next = bh->b_this_page;
1492 * Are we still fully in range ?
1494 if (next_off > stop)
1498 * is this block fully invalidated?
1500 if (offset <= curr_off)
1502 curr_off = next_off;
1504 } while (bh != head);
1507 * We release buffers only if the entire page is being invalidated.
1508 * The get_block cached value has been unconditionally invalidated,
1509 * so real IO is not possible anymore.
1511 if (length == PAGE_SIZE)
1512 try_to_release_page(page, 0);
1516 EXPORT_SYMBOL(block_invalidatepage);
1520 * We attach and possibly dirty the buffers atomically wrt
1521 * __set_page_dirty_buffers() via private_lock. try_to_free_buffers
1522 * is already excluded via the page lock.
1524 void create_empty_buffers(struct page *page,
1525 unsigned long blocksize, unsigned long b_state)
1527 struct buffer_head *bh, *head, *tail;
1529 head = alloc_page_buffers(page, blocksize, true);
1532 bh->b_state |= b_state;
1534 bh = bh->b_this_page;
1536 tail->b_this_page = head;
1538 spin_lock(&page->mapping->private_lock);
1539 if (PageUptodate(page) || PageDirty(page)) {
1542 if (PageDirty(page))
1543 set_buffer_dirty(bh);
1544 if (PageUptodate(page))
1545 set_buffer_uptodate(bh);
1546 bh = bh->b_this_page;
1547 } while (bh != head);
1549 attach_page_buffers(page, head);
1550 spin_unlock(&page->mapping->private_lock);
1552 EXPORT_SYMBOL(create_empty_buffers);
1555 * clean_bdev_aliases: clean a range of buffers in block device
1556 * @bdev: Block device to clean buffers in
1557 * @block: Start of a range of blocks to clean
1558 * @len: Number of blocks to clean
1560 * We are taking a range of blocks for data and we don't want writeback of any
1561 * buffer-cache aliases starting from return from this function and until the
1562 * moment when something will explicitly mark the buffer dirty (hopefully that
1563 * will not happen until we will free that block ;-) We don't even need to mark
1564 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1565 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1566 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1567 * would confuse anyone who might pick it with bread() afterwards...
1569 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1570 * writeout I/O going on against recently-freed buffers. We don't wait on that
1571 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1572 * need to. That happens here.
1574 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1576 struct inode *bd_inode = bdev->bd_inode;
1577 struct address_space *bd_mapping = bd_inode->i_mapping;
1578 struct pagevec pvec;
1579 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1582 struct buffer_head *bh;
1583 struct buffer_head *head;
1585 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1586 pagevec_init(&pvec);
1587 while (pagevec_lookup_range(&pvec, bd_mapping, &index, end)) {
1588 count = pagevec_count(&pvec);
1589 for (i = 0; i < count; i++) {
1590 struct page *page = pvec.pages[i];
1592 if (!page_has_buffers(page))
1595 * We use page lock instead of bd_mapping->private_lock
1596 * to pin buffers here since we can afford to sleep and
1597 * it scales better than a global spinlock lock.
1600 /* Recheck when the page is locked which pins bhs */
1601 if (!page_has_buffers(page))
1603 head = page_buffers(page);
1606 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1608 if (bh->b_blocknr >= block + len)
1610 clear_buffer_dirty(bh);
1612 clear_buffer_req(bh);
1614 bh = bh->b_this_page;
1615 } while (bh != head);
1619 pagevec_release(&pvec);
1621 /* End of range already reached? */
1622 if (index > end || !index)
1626 EXPORT_SYMBOL(clean_bdev_aliases);
1629 * Size is a power-of-two in the range 512..PAGE_SIZE,
1630 * and the case we care about most is PAGE_SIZE.
1632 * So this *could* possibly be written with those
1633 * constraints in mind (relevant mostly if some
1634 * architecture has a slow bit-scan instruction)
1636 static inline int block_size_bits(unsigned int blocksize)
1638 return ilog2(blocksize);
1641 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1643 BUG_ON(!PageLocked(page));
1645 if (!page_has_buffers(page))
1646 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1648 return page_buffers(page);
1652 * NOTE! All mapped/uptodate combinations are valid:
1654 * Mapped Uptodate Meaning
1656 * No No "unknown" - must do get_block()
1657 * No Yes "hole" - zero-filled
1658 * Yes No "allocated" - allocated on disk, not read in
1659 * Yes Yes "valid" - allocated and up-to-date in memory.
1661 * "Dirty" is valid only with the last case (mapped+uptodate).
1665 * While block_write_full_page is writing back the dirty buffers under
1666 * the page lock, whoever dirtied the buffers may decide to clean them
1667 * again at any time. We handle that by only looking at the buffer
1668 * state inside lock_buffer().
1670 * If block_write_full_page() is called for regular writeback
1671 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1672 * locked buffer. This only can happen if someone has written the buffer
1673 * directly, with submit_bh(). At the address_space level PageWriteback
1674 * prevents this contention from occurring.
1676 * If block_write_full_page() is called with wbc->sync_mode ==
1677 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1678 * causes the writes to be flagged as synchronous writes.
1680 int __block_write_full_page(struct inode *inode, struct page *page,
1681 get_block_t *get_block, struct writeback_control *wbc,
1682 bh_end_io_t *handler)
1686 sector_t last_block;
1687 struct buffer_head *bh, *head;
1688 unsigned int blocksize, bbits;
1689 int nr_underway = 0;
1690 int write_flags = wbc_to_write_flags(wbc);
1692 head = create_page_buffers(page, inode,
1693 (1 << BH_Dirty)|(1 << BH_Uptodate));
1696 * Be very careful. We have no exclusion from __set_page_dirty_buffers
1697 * here, and the (potentially unmapped) buffers may become dirty at
1698 * any time. If a buffer becomes dirty here after we've inspected it
1699 * then we just miss that fact, and the page stays dirty.
1701 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1702 * handle that here by just cleaning them.
1706 blocksize = bh->b_size;
1707 bbits = block_size_bits(blocksize);
1709 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1710 last_block = (i_size_read(inode) - 1) >> bbits;
1713 * Get all the dirty buffers mapped to disk addresses and
1714 * handle any aliases from the underlying blockdev's mapping.
1717 if (block > last_block) {
1719 * mapped buffers outside i_size will occur, because
1720 * this page can be outside i_size when there is a
1721 * truncate in progress.
1724 * The buffer was zeroed by block_write_full_page()
1726 clear_buffer_dirty(bh);
1727 set_buffer_uptodate(bh);
1728 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1730 WARN_ON(bh->b_size != blocksize);
1731 err = get_block(inode, block, bh, 1);
1734 clear_buffer_delay(bh);
1735 if (buffer_new(bh)) {
1736 /* blockdev mappings never come here */
1737 clear_buffer_new(bh);
1738 clean_bdev_bh_alias(bh);
1741 bh = bh->b_this_page;
1743 } while (bh != head);
1746 if (!buffer_mapped(bh))
1749 * If it's a fully non-blocking write attempt and we cannot
1750 * lock the buffer then redirty the page. Note that this can
1751 * potentially cause a busy-wait loop from writeback threads
1752 * and kswapd activity, but those code paths have their own
1753 * higher-level throttling.
1755 if (wbc->sync_mode != WB_SYNC_NONE) {
1757 } else if (!trylock_buffer(bh)) {
1758 redirty_page_for_writepage(wbc, page);
1761 if (test_clear_buffer_dirty(bh)) {
1762 mark_buffer_async_write_endio(bh, handler);
1766 } while ((bh = bh->b_this_page) != head);
1769 * The page and its buffers are protected by PageWriteback(), so we can
1770 * drop the bh refcounts early.
1772 BUG_ON(PageWriteback(page));
1773 set_page_writeback(page);
1776 struct buffer_head *next = bh->b_this_page;
1777 if (buffer_async_write(bh)) {
1778 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1779 inode->i_write_hint, wbc);
1783 } while (bh != head);
1788 if (nr_underway == 0) {
1790 * The page was marked dirty, but the buffers were
1791 * clean. Someone wrote them back by hand with
1792 * ll_rw_block/submit_bh. A rare case.
1794 end_page_writeback(page);
1797 * The page and buffer_heads can be released at any time from
1805 * ENOSPC, or some other error. We may already have added some
1806 * blocks to the file, so we need to write these out to avoid
1807 * exposing stale data.
1808 * The page is currently locked and not marked for writeback
1811 /* Recovery: lock and submit the mapped buffers */
1813 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1814 !buffer_delay(bh)) {
1816 mark_buffer_async_write_endio(bh, handler);
1819 * The buffer may have been set dirty during
1820 * attachment to a dirty page.
1822 clear_buffer_dirty(bh);
1824 } while ((bh = bh->b_this_page) != head);
1826 BUG_ON(PageWriteback(page));
1827 mapping_set_error(page->mapping, err);
1828 set_page_writeback(page);
1830 struct buffer_head *next = bh->b_this_page;
1831 if (buffer_async_write(bh)) {
1832 clear_buffer_dirty(bh);
1833 submit_bh_wbc(REQ_OP_WRITE, write_flags, bh,
1834 inode->i_write_hint, wbc);
1838 } while (bh != head);
1842 EXPORT_SYMBOL(__block_write_full_page);
1845 * If a page has any new buffers, zero them out here, and mark them uptodate
1846 * and dirty so they'll be written out (in order to prevent uninitialised
1847 * block data from leaking). And clear the new bit.
1849 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1851 unsigned int block_start, block_end;
1852 struct buffer_head *head, *bh;
1854 BUG_ON(!PageLocked(page));
1855 if (!page_has_buffers(page))
1858 bh = head = page_buffers(page);
1861 block_end = block_start + bh->b_size;
1863 if (buffer_new(bh)) {
1864 if (block_end > from && block_start < to) {
1865 if (!PageUptodate(page)) {
1866 unsigned start, size;
1868 start = max(from, block_start);
1869 size = min(to, block_end) - start;
1871 zero_user(page, start, size);
1872 set_buffer_uptodate(bh);
1875 clear_buffer_new(bh);
1876 mark_buffer_dirty(bh);
1880 block_start = block_end;
1881 bh = bh->b_this_page;
1882 } while (bh != head);
1884 EXPORT_SYMBOL(page_zero_new_buffers);
1887 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1888 struct iomap *iomap)
1890 loff_t offset = block << inode->i_blkbits;
1892 bh->b_bdev = iomap->bdev;
1895 * Block points to offset in file we need to map, iomap contains
1896 * the offset at which the map starts. If the map ends before the
1897 * current block, then do not map the buffer and let the caller
1900 BUG_ON(offset >= iomap->offset + iomap->length);
1902 switch (iomap->type) {
1905 * If the buffer is not up to date or beyond the current EOF,
1906 * we need to mark it as new to ensure sub-block zeroing is
1907 * executed if necessary.
1909 if (!buffer_uptodate(bh) ||
1910 (offset >= i_size_read(inode)))
1913 case IOMAP_DELALLOC:
1914 if (!buffer_uptodate(bh) ||
1915 (offset >= i_size_read(inode)))
1917 set_buffer_uptodate(bh);
1918 set_buffer_mapped(bh);
1919 set_buffer_delay(bh);
1921 case IOMAP_UNWRITTEN:
1923 * For unwritten regions, we always need to ensure that regions
1924 * in the block we are not writing to are zeroed. Mark the
1925 * buffer as new to ensure this.
1928 set_buffer_unwritten(bh);
1931 if ((iomap->flags & IOMAP_F_NEW) ||
1932 offset >= i_size_read(inode))
1934 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1936 set_buffer_mapped(bh);
1941 int __block_write_begin_int(struct page *page, loff_t pos, unsigned len,
1942 get_block_t *get_block, struct iomap *iomap)
1944 unsigned from = pos & (PAGE_SIZE - 1);
1945 unsigned to = from + len;
1946 struct inode *inode = page->mapping->host;
1947 unsigned block_start, block_end;
1950 unsigned blocksize, bbits;
1951 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1953 BUG_ON(!PageLocked(page));
1954 BUG_ON(from > PAGE_SIZE);
1955 BUG_ON(to > PAGE_SIZE);
1958 head = create_page_buffers(page, inode, 0);
1959 blocksize = head->b_size;
1960 bbits = block_size_bits(blocksize);
1962 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1964 for(bh = head, block_start = 0; bh != head || !block_start;
1965 block++, block_start=block_end, bh = bh->b_this_page) {
1966 block_end = block_start + blocksize;
1967 if (block_end <= from || block_start >= to) {
1968 if (PageUptodate(page)) {
1969 if (!buffer_uptodate(bh))
1970 set_buffer_uptodate(bh);
1975 clear_buffer_new(bh);
1976 if (!buffer_mapped(bh)) {
1977 WARN_ON(bh->b_size != blocksize);
1979 err = get_block(inode, block, bh, 1);
1983 iomap_to_bh(inode, block, bh, iomap);
1986 if (buffer_new(bh)) {
1987 clean_bdev_bh_alias(bh);
1988 if (PageUptodate(page)) {
1989 clear_buffer_new(bh);
1990 set_buffer_uptodate(bh);
1991 mark_buffer_dirty(bh);
1994 if (block_end > to || block_start < from)
1995 zero_user_segments(page,
2001 if (PageUptodate(page)) {
2002 if (!buffer_uptodate(bh))
2003 set_buffer_uptodate(bh);
2006 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2007 !buffer_unwritten(bh) &&
2008 (block_start < from || block_end > to)) {
2009 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2014 * If we issued read requests - let them complete.
2016 while(wait_bh > wait) {
2017 wait_on_buffer(*--wait_bh);
2018 if (!buffer_uptodate(*wait_bh))
2022 page_zero_new_buffers(page, from, to);
2026 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2027 get_block_t *get_block)
2029 return __block_write_begin_int(page, pos, len, get_block, NULL);
2031 EXPORT_SYMBOL(__block_write_begin);
2033 static int __block_commit_write(struct inode *inode, struct page *page,
2034 unsigned from, unsigned to)
2036 unsigned block_start, block_end;
2039 struct buffer_head *bh, *head;
2041 bh = head = page_buffers(page);
2042 blocksize = bh->b_size;
2046 block_end = block_start + blocksize;
2047 if (block_end <= from || block_start >= to) {
2048 if (!buffer_uptodate(bh))
2051 set_buffer_uptodate(bh);
2052 mark_buffer_dirty(bh);
2054 clear_buffer_new(bh);
2056 block_start = block_end;
2057 bh = bh->b_this_page;
2058 } while (bh != head);
2061 * If this is a partial write which happened to make all buffers
2062 * uptodate then we can optimize away a bogus readpage() for
2063 * the next read(). Here we 'discover' whether the page went
2064 * uptodate as a result of this (potentially partial) write.
2067 SetPageUptodate(page);
2072 * block_write_begin takes care of the basic task of block allocation and
2073 * bringing partial write blocks uptodate first.
2075 * The filesystem needs to handle block truncation upon failure.
2077 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2078 unsigned flags, struct page **pagep, get_block_t *get_block)
2080 pgoff_t index = pos >> PAGE_SHIFT;
2084 page = grab_cache_page_write_begin(mapping, index, flags);
2088 status = __block_write_begin(page, pos, len, get_block);
2089 if (unlikely(status)) {
2098 EXPORT_SYMBOL(block_write_begin);
2100 int block_write_end(struct file *file, struct address_space *mapping,
2101 loff_t pos, unsigned len, unsigned copied,
2102 struct page *page, void *fsdata)
2104 struct inode *inode = mapping->host;
2107 start = pos & (PAGE_SIZE - 1);
2109 if (unlikely(copied < len)) {
2111 * The buffers that were written will now be uptodate, so we
2112 * don't have to worry about a readpage reading them and
2113 * overwriting a partial write. However if we have encountered
2114 * a short write and only partially written into a buffer, it
2115 * will not be marked uptodate, so a readpage might come in and
2116 * destroy our partial write.
2118 * Do the simplest thing, and just treat any short write to a
2119 * non uptodate page as a zero-length write, and force the
2120 * caller to redo the whole thing.
2122 if (!PageUptodate(page))
2125 page_zero_new_buffers(page, start+copied, start+len);
2127 flush_dcache_page(page);
2129 /* This could be a short (even 0-length) commit */
2130 __block_commit_write(inode, page, start, start+copied);
2134 EXPORT_SYMBOL(block_write_end);
2136 int generic_write_end(struct file *file, struct address_space *mapping,
2137 loff_t pos, unsigned len, unsigned copied,
2138 struct page *page, void *fsdata)
2140 struct inode *inode = mapping->host;
2141 loff_t old_size = inode->i_size;
2142 bool i_size_changed = false;
2144 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2147 * No need to use i_size_read() here, the i_size cannot change under us
2148 * because we hold i_rwsem.
2150 * But it's important to update i_size while still holding page lock:
2151 * page writeout could otherwise come in and zero beyond i_size.
2153 if (pos + copied > inode->i_size) {
2154 i_size_write(inode, pos + copied);
2155 i_size_changed = true;
2162 pagecache_isize_extended(inode, old_size, pos);
2164 * Don't mark the inode dirty under page lock. First, it unnecessarily
2165 * makes the holding time of page lock longer. Second, it forces lock
2166 * ordering of page lock and transaction start for journaling
2170 mark_inode_dirty(inode);
2173 EXPORT_SYMBOL(generic_write_end);
2176 * block_is_partially_uptodate checks whether buffers within a page are
2179 * Returns true if all buffers which correspond to a file portion
2180 * we want to read are uptodate.
2182 int block_is_partially_uptodate(struct page *page, unsigned long from,
2183 unsigned long count)
2185 unsigned block_start, block_end, blocksize;
2187 struct buffer_head *bh, *head;
2190 if (!page_has_buffers(page))
2193 head = page_buffers(page);
2194 blocksize = head->b_size;
2195 to = min_t(unsigned, PAGE_SIZE - from, count);
2197 if (from < blocksize && to > PAGE_SIZE - blocksize)
2203 block_end = block_start + blocksize;
2204 if (block_end > from && block_start < to) {
2205 if (!buffer_uptodate(bh)) {
2209 if (block_end >= to)
2212 block_start = block_end;
2213 bh = bh->b_this_page;
2214 } while (bh != head);
2218 EXPORT_SYMBOL(block_is_partially_uptodate);
2221 * Generic "read page" function for block devices that have the normal
2222 * get_block functionality. This is most of the block device filesystems.
2223 * Reads the page asynchronously --- the unlock_buffer() and
2224 * set/clear_buffer_uptodate() functions propagate buffer state into the
2225 * page struct once IO has completed.
2227 int block_read_full_page(struct page *page, get_block_t *get_block)
2229 struct inode *inode = page->mapping->host;
2230 sector_t iblock, lblock;
2231 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2232 unsigned int blocksize, bbits;
2234 int fully_mapped = 1;
2236 head = create_page_buffers(page, inode, 0);
2237 blocksize = head->b_size;
2238 bbits = block_size_bits(blocksize);
2240 iblock = (sector_t)page->index << (PAGE_SHIFT - bbits);
2241 lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2247 if (buffer_uptodate(bh))
2250 if (!buffer_mapped(bh)) {
2254 if (iblock < lblock) {
2255 WARN_ON(bh->b_size != blocksize);
2256 err = get_block(inode, iblock, bh, 0);
2260 if (!buffer_mapped(bh)) {
2261 zero_user(page, i * blocksize, blocksize);
2263 set_buffer_uptodate(bh);
2267 * get_block() might have updated the buffer
2270 if (buffer_uptodate(bh))
2274 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2277 SetPageMappedToDisk(page);
2281 * All buffers are uptodate - we can set the page uptodate
2282 * as well. But not if get_block() returned an error.
2284 if (!PageError(page))
2285 SetPageUptodate(page);
2290 /* Stage two: lock the buffers */
2291 for (i = 0; i < nr; i++) {
2294 mark_buffer_async_read(bh);
2298 * Stage 3: start the IO. Check for uptodateness
2299 * inside the buffer lock in case another process reading
2300 * the underlying blockdev brought it uptodate (the sct fix).
2302 for (i = 0; i < nr; i++) {
2304 if (buffer_uptodate(bh))
2305 end_buffer_async_read(bh, 1);
2307 submit_bh(REQ_OP_READ, 0, bh);
2311 EXPORT_SYMBOL(block_read_full_page);
2313 /* utility function for filesystems that need to do work on expanding
2314 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2315 * deal with the hole.
2317 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2319 struct address_space *mapping = inode->i_mapping;
2321 void *fsdata = NULL;
2324 err = inode_newsize_ok(inode, size);
2328 err = pagecache_write_begin(NULL, mapping, size, 0,
2329 AOP_FLAG_CONT_EXPAND, &page, &fsdata);
2333 err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2339 EXPORT_SYMBOL(generic_cont_expand_simple);
2341 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2342 loff_t pos, loff_t *bytes)
2344 struct inode *inode = mapping->host;
2345 unsigned int blocksize = i_blocksize(inode);
2347 void *fsdata = NULL;
2348 pgoff_t index, curidx;
2350 unsigned zerofrom, offset, len;
2353 index = pos >> PAGE_SHIFT;
2354 offset = pos & ~PAGE_MASK;
2356 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2357 zerofrom = curpos & ~PAGE_MASK;
2358 if (zerofrom & (blocksize-1)) {
2359 *bytes |= (blocksize-1);
2362 len = PAGE_SIZE - zerofrom;
2364 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2368 zero_user(page, zerofrom, len);
2369 err = pagecache_write_end(file, mapping, curpos, len, len,
2376 balance_dirty_pages_ratelimited(mapping);
2378 if (fatal_signal_pending(current)) {
2384 /* page covers the boundary, find the boundary offset */
2385 if (index == curidx) {
2386 zerofrom = curpos & ~PAGE_MASK;
2387 /* if we will expand the thing last block will be filled */
2388 if (offset <= zerofrom) {
2391 if (zerofrom & (blocksize-1)) {
2392 *bytes |= (blocksize-1);
2395 len = offset - zerofrom;
2397 err = pagecache_write_begin(file, mapping, curpos, len, 0,
2401 zero_user(page, zerofrom, len);
2402 err = pagecache_write_end(file, mapping, curpos, len, len,
2414 * For moronic filesystems that do not allow holes in file.
2415 * We may have to extend the file.
2417 int cont_write_begin(struct file *file, struct address_space *mapping,
2418 loff_t pos, unsigned len, unsigned flags,
2419 struct page **pagep, void **fsdata,
2420 get_block_t *get_block, loff_t *bytes)
2422 struct inode *inode = mapping->host;
2423 unsigned int blocksize = i_blocksize(inode);
2424 unsigned int zerofrom;
2427 err = cont_expand_zero(file, mapping, pos, bytes);
2431 zerofrom = *bytes & ~PAGE_MASK;
2432 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2433 *bytes |= (blocksize-1);
2437 return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2439 EXPORT_SYMBOL(cont_write_begin);
2441 int block_commit_write(struct page *page, unsigned from, unsigned to)
2443 struct inode *inode = page->mapping->host;
2444 __block_commit_write(inode,page,from,to);
2447 EXPORT_SYMBOL(block_commit_write);
2450 * block_page_mkwrite() is not allowed to change the file size as it gets
2451 * called from a page fault handler when a page is first dirtied. Hence we must
2452 * be careful to check for EOF conditions here. We set the page up correctly
2453 * for a written page which means we get ENOSPC checking when writing into
2454 * holes and correct delalloc and unwritten extent mapping on filesystems that
2455 * support these features.
2457 * We are not allowed to take the i_mutex here so we have to play games to
2458 * protect against truncate races as the page could now be beyond EOF. Because
2459 * truncate writes the inode size before removing pages, once we have the
2460 * page lock we can determine safely if the page is beyond EOF. If it is not
2461 * beyond EOF, then the page is guaranteed safe against truncation until we
2464 * Direct callers of this function should protect against filesystem freezing
2465 * using sb_start_pagefault() - sb_end_pagefault() functions.
2467 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2468 get_block_t get_block)
2470 struct page *page = vmf->page;
2471 struct inode *inode = file_inode(vma->vm_file);
2477 size = i_size_read(inode);
2478 if ((page->mapping != inode->i_mapping) ||
2479 (page_offset(page) > size)) {
2480 /* We overload EFAULT to mean page got truncated */
2485 /* page is wholly or partially inside EOF */
2486 if (((page->index + 1) << PAGE_SHIFT) > size)
2487 end = size & ~PAGE_MASK;
2491 ret = __block_write_begin(page, 0, end, get_block);
2493 ret = block_commit_write(page, 0, end);
2495 if (unlikely(ret < 0))
2497 set_page_dirty(page);
2498 wait_for_stable_page(page);
2504 EXPORT_SYMBOL(block_page_mkwrite);
2507 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2508 * immediately, while under the page lock. So it needs a special end_io
2509 * handler which does not touch the bh after unlocking it.
2511 static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2513 __end_buffer_read_notouch(bh, uptodate);
2517 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2518 * the page (converting it to circular linked list and taking care of page
2521 static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2523 struct buffer_head *bh;
2525 BUG_ON(!PageLocked(page));
2527 spin_lock(&page->mapping->private_lock);
2530 if (PageDirty(page))
2531 set_buffer_dirty(bh);
2532 if (!bh->b_this_page)
2533 bh->b_this_page = head;
2534 bh = bh->b_this_page;
2535 } while (bh != head);
2536 attach_page_buffers(page, head);
2537 spin_unlock(&page->mapping->private_lock);
2541 * On entry, the page is fully not uptodate.
2542 * On exit the page is fully uptodate in the areas outside (from,to)
2543 * The filesystem needs to handle block truncation upon failure.
2545 int nobh_write_begin(struct address_space *mapping,
2546 loff_t pos, unsigned len, unsigned flags,
2547 struct page **pagep, void **fsdata,
2548 get_block_t *get_block)
2550 struct inode *inode = mapping->host;
2551 const unsigned blkbits = inode->i_blkbits;
2552 const unsigned blocksize = 1 << blkbits;
2553 struct buffer_head *head, *bh;
2557 unsigned block_in_page;
2558 unsigned block_start, block_end;
2559 sector_t block_in_file;
2562 int is_mapped_to_disk = 1;
2564 index = pos >> PAGE_SHIFT;
2565 from = pos & (PAGE_SIZE - 1);
2568 page = grab_cache_page_write_begin(mapping, index, flags);
2574 if (page_has_buffers(page)) {
2575 ret = __block_write_begin(page, pos, len, get_block);
2581 if (PageMappedToDisk(page))
2585 * Allocate buffers so that we can keep track of state, and potentially
2586 * attach them to the page if an error occurs. In the common case of
2587 * no error, they will just be freed again without ever being attached
2588 * to the page (which is all OK, because we're under the page lock).
2590 * Be careful: the buffer linked list is a NULL terminated one, rather
2591 * than the circular one we're used to.
2593 head = alloc_page_buffers(page, blocksize, false);
2599 block_in_file = (sector_t)page->index << (PAGE_SHIFT - blkbits);
2602 * We loop across all blocks in the page, whether or not they are
2603 * part of the affected region. This is so we can discover if the
2604 * page is fully mapped-to-disk.
2606 for (block_start = 0, block_in_page = 0, bh = head;
2607 block_start < PAGE_SIZE;
2608 block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2611 block_end = block_start + blocksize;
2614 if (block_start >= to)
2616 ret = get_block(inode, block_in_file + block_in_page,
2620 if (!buffer_mapped(bh))
2621 is_mapped_to_disk = 0;
2623 clean_bdev_bh_alias(bh);
2624 if (PageUptodate(page)) {
2625 set_buffer_uptodate(bh);
2628 if (buffer_new(bh) || !buffer_mapped(bh)) {
2629 zero_user_segments(page, block_start, from,
2633 if (buffer_uptodate(bh))
2634 continue; /* reiserfs does this */
2635 if (block_start < from || block_end > to) {
2637 bh->b_end_io = end_buffer_read_nobh;
2638 submit_bh(REQ_OP_READ, 0, bh);
2645 * The page is locked, so these buffers are protected from
2646 * any VM or truncate activity. Hence we don't need to care
2647 * for the buffer_head refcounts.
2649 for (bh = head; bh; bh = bh->b_this_page) {
2651 if (!buffer_uptodate(bh))
2658 if (is_mapped_to_disk)
2659 SetPageMappedToDisk(page);
2661 *fsdata = head; /* to be released by nobh_write_end */
2668 * Error recovery is a bit difficult. We need to zero out blocks that
2669 * were newly allocated, and dirty them to ensure they get written out.
2670 * Buffers need to be attached to the page at this point, otherwise
2671 * the handling of potential IO errors during writeout would be hard
2672 * (could try doing synchronous writeout, but what if that fails too?)
2674 attach_nobh_buffers(page, head);
2675 page_zero_new_buffers(page, from, to);
2684 EXPORT_SYMBOL(nobh_write_begin);
2686 int nobh_write_end(struct file *file, struct address_space *mapping,
2687 loff_t pos, unsigned len, unsigned copied,
2688 struct page *page, void *fsdata)
2690 struct inode *inode = page->mapping->host;
2691 struct buffer_head *head = fsdata;
2692 struct buffer_head *bh;
2693 BUG_ON(fsdata != NULL && page_has_buffers(page));
2695 if (unlikely(copied < len) && head)
2696 attach_nobh_buffers(page, head);
2697 if (page_has_buffers(page))
2698 return generic_write_end(file, mapping, pos, len,
2699 copied, page, fsdata);
2701 SetPageUptodate(page);
2702 set_page_dirty(page);
2703 if (pos+copied > inode->i_size) {
2704 i_size_write(inode, pos+copied);
2705 mark_inode_dirty(inode);
2713 head = head->b_this_page;
2714 free_buffer_head(bh);
2719 EXPORT_SYMBOL(nobh_write_end);
2722 * nobh_writepage() - based on block_full_write_page() except
2723 * that it tries to operate without attaching bufferheads to
2726 int nobh_writepage(struct page *page, get_block_t *get_block,
2727 struct writeback_control *wbc)
2729 struct inode * const inode = page->mapping->host;
2730 loff_t i_size = i_size_read(inode);
2731 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2735 /* Is the page fully inside i_size? */
2736 if (page->index < end_index)
2739 /* Is the page fully outside i_size? (truncate in progress) */
2740 offset = i_size & (PAGE_SIZE-1);
2741 if (page->index >= end_index+1 || !offset) {
2743 return 0; /* don't care */
2747 * The page straddles i_size. It must be zeroed out on each and every
2748 * writepage invocation because it may be mmapped. "A file is mapped
2749 * in multiples of the page size. For a file that is not a multiple of
2750 * the page size, the remaining memory is zeroed when mapped, and
2751 * writes to that region are not written out to the file."
2753 zero_user_segment(page, offset, PAGE_SIZE);
2755 ret = mpage_writepage(page, get_block, wbc);
2757 ret = __block_write_full_page(inode, page, get_block, wbc,
2758 end_buffer_async_write);
2761 EXPORT_SYMBOL(nobh_writepage);
2763 int nobh_truncate_page(struct address_space *mapping,
2764 loff_t from, get_block_t *get_block)
2766 pgoff_t index = from >> PAGE_SHIFT;
2767 unsigned offset = from & (PAGE_SIZE-1);
2770 unsigned length, pos;
2771 struct inode *inode = mapping->host;
2773 struct buffer_head map_bh;
2776 blocksize = i_blocksize(inode);
2777 length = offset & (blocksize - 1);
2779 /* Block boundary? Nothing to do */
2783 length = blocksize - length;
2784 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2786 page = grab_cache_page(mapping, index);
2791 if (page_has_buffers(page)) {
2795 return block_truncate_page(mapping, from, get_block);
2798 /* Find the buffer that contains "offset" */
2800 while (offset >= pos) {
2805 map_bh.b_size = blocksize;
2807 err = get_block(inode, iblock, &map_bh, 0);
2810 /* unmapped? It's a hole - nothing to do */
2811 if (!buffer_mapped(&map_bh))
2814 /* Ok, it's mapped. Make sure it's up-to-date */
2815 if (!PageUptodate(page)) {
2816 err = mapping->a_ops->readpage(NULL, page);
2822 if (!PageUptodate(page)) {
2826 if (page_has_buffers(page))
2829 zero_user(page, offset, length);
2830 set_page_dirty(page);
2839 EXPORT_SYMBOL(nobh_truncate_page);
2841 int block_truncate_page(struct address_space *mapping,
2842 loff_t from, get_block_t *get_block)
2844 pgoff_t index = from >> PAGE_SHIFT;
2845 unsigned offset = from & (PAGE_SIZE-1);
2848 unsigned length, pos;
2849 struct inode *inode = mapping->host;
2851 struct buffer_head *bh;
2854 blocksize = i_blocksize(inode);
2855 length = offset & (blocksize - 1);
2857 /* Block boundary? Nothing to do */
2861 length = blocksize - length;
2862 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2864 page = grab_cache_page(mapping, index);
2869 if (!page_has_buffers(page))
2870 create_empty_buffers(page, blocksize, 0);
2872 /* Find the buffer that contains "offset" */
2873 bh = page_buffers(page);
2875 while (offset >= pos) {
2876 bh = bh->b_this_page;
2882 if (!buffer_mapped(bh)) {
2883 WARN_ON(bh->b_size != blocksize);
2884 err = get_block(inode, iblock, bh, 0);
2887 /* unmapped? It's a hole - nothing to do */
2888 if (!buffer_mapped(bh))
2892 /* Ok, it's mapped. Make sure it's up-to-date */
2893 if (PageUptodate(page))
2894 set_buffer_uptodate(bh);
2896 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2898 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
2900 /* Uhhuh. Read error. Complain and punt. */
2901 if (!buffer_uptodate(bh))
2905 zero_user(page, offset, length);
2906 mark_buffer_dirty(bh);
2915 EXPORT_SYMBOL(block_truncate_page);
2918 * The generic ->writepage function for buffer-backed address_spaces
2920 int block_write_full_page(struct page *page, get_block_t *get_block,
2921 struct writeback_control *wbc)
2923 struct inode * const inode = page->mapping->host;
2924 loff_t i_size = i_size_read(inode);
2925 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2928 /* Is the page fully inside i_size? */
2929 if (page->index < end_index)
2930 return __block_write_full_page(inode, page, get_block, wbc,
2931 end_buffer_async_write);
2933 /* Is the page fully outside i_size? (truncate in progress) */
2934 offset = i_size & (PAGE_SIZE-1);
2935 if (page->index >= end_index+1 || !offset) {
2937 return 0; /* don't care */
2941 * The page straddles i_size. It must be zeroed out on each and every
2942 * writepage invocation because it may be mmapped. "A file is mapped
2943 * in multiples of the page size. For a file that is not a multiple of
2944 * the page size, the remaining memory is zeroed when mapped, and
2945 * writes to that region are not written out to the file."
2947 zero_user_segment(page, offset, PAGE_SIZE);
2948 return __block_write_full_page(inode, page, get_block, wbc,
2949 end_buffer_async_write);
2951 EXPORT_SYMBOL(block_write_full_page);
2953 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2954 get_block_t *get_block)
2956 struct inode *inode = mapping->host;
2957 struct buffer_head tmp = {
2958 .b_size = i_blocksize(inode),
2961 get_block(inode, block, &tmp, 0);
2962 return tmp.b_blocknr;
2964 EXPORT_SYMBOL(generic_block_bmap);
2966 static void end_bio_bh_io_sync(struct bio *bio)
2968 struct buffer_head *bh = bio->bi_private;
2970 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2971 set_bit(BH_Quiet, &bh->b_state);
2973 bh->b_end_io(bh, !bio->bi_status);
2978 * This allows us to do IO even on the odd last sectors
2979 * of a device, even if the block size is some multiple
2980 * of the physical sector size.
2982 * We'll just truncate the bio to the size of the device,
2983 * and clear the end of the buffer head manually.
2985 * Truly out-of-range accesses will turn into actual IO
2986 * errors, this only handles the "we need to be able to
2987 * do IO at the final sector" case.
2989 void guard_bio_eod(struct bio *bio)
2992 struct hd_struct *part;
2995 part = __disk_get_part(bio->bi_disk, bio->bi_partno);
2997 maxsector = part_nr_sects_read(part);
2999 maxsector = get_capacity(bio->bi_disk);
3006 * If the *whole* IO is past the end of the device,
3007 * let it through, and the IO layer will turn it into
3010 if (unlikely(bio->bi_iter.bi_sector >= maxsector))
3013 maxsector -= bio->bi_iter.bi_sector;
3014 if (likely((bio->bi_iter.bi_size >> 9) <= maxsector))
3017 bio_truncate(bio, maxsector << 9);
3020 static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,
3021 enum rw_hint write_hint, struct writeback_control *wbc)
3025 BUG_ON(!buffer_locked(bh));
3026 BUG_ON(!buffer_mapped(bh));
3027 BUG_ON(!bh->b_end_io);
3028 BUG_ON(buffer_delay(bh));
3029 BUG_ON(buffer_unwritten(bh));
3032 * Only clear out a write error when rewriting
3034 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
3035 clear_buffer_write_io_error(bh);
3038 * from here on down, it's all bio -- do the initial mapping,
3039 * submit_bio -> generic_make_request may further map this bio around
3041 bio = bio_alloc(GFP_NOIO, 1);
3043 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
3044 bio_set_dev(bio, bh->b_bdev);
3045 bio->bi_write_hint = write_hint;
3047 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
3048 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
3050 bio->bi_end_io = end_bio_bh_io_sync;
3051 bio->bi_private = bh;
3053 if (buffer_meta(bh))
3054 op_flags |= REQ_META;
3055 if (buffer_prio(bh))
3056 op_flags |= REQ_PRIO;
3057 bio_set_op_attrs(bio, op, op_flags);
3059 /* Take care of bh's that straddle the end of the device */
3063 wbc_init_bio(wbc, bio);
3064 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
3071 int submit_bh(int op, int op_flags, struct buffer_head *bh)
3073 return submit_bh_wbc(op, op_flags, bh, 0, NULL);
3075 EXPORT_SYMBOL(submit_bh);
3078 * ll_rw_block: low-level access to block devices (DEPRECATED)
3079 * @op: whether to %READ or %WRITE
3080 * @op_flags: req_flag_bits
3081 * @nr: number of &struct buffer_heads in the array
3082 * @bhs: array of pointers to &struct buffer_head
3084 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3085 * requests an I/O operation on them, either a %REQ_OP_READ or a %REQ_OP_WRITE.
3086 * @op_flags contains flags modifying the detailed I/O behavior, most notably
3089 * This function drops any buffer that it cannot get a lock on (with the
3090 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3091 * request, and any buffer that appears to be up-to-date when doing read
3092 * request. Further it marks as clean buffers that are processed for
3093 * writing (the buffer cache won't assume that they are actually clean
3094 * until the buffer gets unlocked).
3096 * ll_rw_block sets b_end_io to simple completion handler that marks
3097 * the buffer up-to-date (if appropriate), unlocks the buffer and wakes
3100 * All of the buffers must be for the same device, and must also be a
3101 * multiple of the current approved size for the device.
3103 void ll_rw_block(int op, int op_flags, int nr, struct buffer_head *bhs[])
3107 for (i = 0; i < nr; i++) {
3108 struct buffer_head *bh = bhs[i];
3110 if (!trylock_buffer(bh))
3113 if (test_clear_buffer_dirty(bh)) {
3114 bh->b_end_io = end_buffer_write_sync;
3116 submit_bh(op, op_flags, bh);
3120 if (!buffer_uptodate(bh)) {
3121 bh->b_end_io = end_buffer_read_sync;
3123 submit_bh(op, op_flags, bh);
3130 EXPORT_SYMBOL(ll_rw_block);
3132 void write_dirty_buffer(struct buffer_head *bh, int op_flags)
3135 if (!test_clear_buffer_dirty(bh)) {
3139 bh->b_end_io = end_buffer_write_sync;
3141 submit_bh(REQ_OP_WRITE, op_flags, bh);
3143 EXPORT_SYMBOL(write_dirty_buffer);
3146 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3147 * and then start new I/O and then wait upon it. The caller must have a ref on
3150 int __sync_dirty_buffer(struct buffer_head *bh, int op_flags)
3154 WARN_ON(atomic_read(&bh->b_count) < 1);
3156 if (test_clear_buffer_dirty(bh)) {
3158 * The bh should be mapped, but it might not be if the
3159 * device was hot-removed. Not much we can do but fail the I/O.
3161 if (!buffer_mapped(bh)) {
3167 bh->b_end_io = end_buffer_write_sync;
3168 ret = submit_bh(REQ_OP_WRITE, op_flags, bh);
3170 if (!ret && !buffer_uptodate(bh))
3177 EXPORT_SYMBOL(__sync_dirty_buffer);
3179 int sync_dirty_buffer(struct buffer_head *bh)
3181 return __sync_dirty_buffer(bh, REQ_SYNC);
3183 EXPORT_SYMBOL(sync_dirty_buffer);
3186 * try_to_free_buffers() checks if all the buffers on this particular page
3187 * are unused, and releases them if so.
3189 * Exclusion against try_to_free_buffers may be obtained by either
3190 * locking the page or by holding its mapping's private_lock.
3192 * If the page is dirty but all the buffers are clean then we need to
3193 * be sure to mark the page clean as well. This is because the page
3194 * may be against a block device, and a later reattachment of buffers
3195 * to a dirty page will set *all* buffers dirty. Which would corrupt
3196 * filesystem data on the same device.
3198 * The same applies to regular filesystem pages: if all the buffers are
3199 * clean then we set the page clean and proceed. To do that, we require
3200 * total exclusion from __set_page_dirty_buffers(). That is obtained with
3203 * try_to_free_buffers() is non-blocking.
3205 static inline int buffer_busy(struct buffer_head *bh)
3207 return atomic_read(&bh->b_count) |
3208 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3212 drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3214 struct buffer_head *head = page_buffers(page);
3215 struct buffer_head *bh;
3219 if (buffer_busy(bh))
3221 bh = bh->b_this_page;
3222 } while (bh != head);
3225 struct buffer_head *next = bh->b_this_page;
3227 if (bh->b_assoc_map)
3228 __remove_assoc_queue(bh);
3230 } while (bh != head);
3231 *buffers_to_free = head;
3232 __clear_page_buffers(page);
3238 int try_to_free_buffers(struct page *page)
3240 struct address_space * const mapping = page->mapping;
3241 struct buffer_head *buffers_to_free = NULL;
3244 BUG_ON(!PageLocked(page));
3245 if (PageWriteback(page))
3248 if (mapping == NULL) { /* can this still happen? */
3249 ret = drop_buffers(page, &buffers_to_free);
3253 spin_lock(&mapping->private_lock);
3254 ret = drop_buffers(page, &buffers_to_free);
3257 * If the filesystem writes its buffers by hand (eg ext3)
3258 * then we can have clean buffers against a dirty page. We
3259 * clean the page here; otherwise the VM will never notice
3260 * that the filesystem did any IO at all.
3262 * Also, during truncate, discard_buffer will have marked all
3263 * the page's buffers clean. We discover that here and clean
3266 * private_lock must be held over this entire operation in order
3267 * to synchronise against __set_page_dirty_buffers and prevent the
3268 * dirty bit from being lost.
3271 cancel_dirty_page(page);
3272 spin_unlock(&mapping->private_lock);
3274 if (buffers_to_free) {
3275 struct buffer_head *bh = buffers_to_free;
3278 struct buffer_head *next = bh->b_this_page;
3279 free_buffer_head(bh);
3281 } while (bh != buffers_to_free);
3285 EXPORT_SYMBOL(try_to_free_buffers);
3288 * There are no bdflush tunables left. But distributions are
3289 * still running obsolete flush daemons, so we terminate them here.
3291 * Use of bdflush() is deprecated and will be removed in a future kernel.
3292 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3294 SYSCALL_DEFINE2(bdflush, int, func, long, data)
3296 static int msg_count;
3298 if (!capable(CAP_SYS_ADMIN))
3301 if (msg_count < 5) {
3304 "warning: process `%s' used the obsolete bdflush"
3305 " system call\n", current->comm);
3306 printk(KERN_INFO "Fix your initscripts?\n");
3315 * Buffer-head allocation
3317 static struct kmem_cache *bh_cachep __read_mostly;
3320 * Once the number of bh's in the machine exceeds this level, we start
3321 * stripping them in writeback.
3323 static unsigned long max_buffer_heads;
3325 int buffer_heads_over_limit;
3327 struct bh_accounting {
3328 int nr; /* Number of live bh's */
3329 int ratelimit; /* Limit cacheline bouncing */
3332 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3334 static void recalc_bh_state(void)
3339 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3341 __this_cpu_write(bh_accounting.ratelimit, 0);
3342 for_each_online_cpu(i)
3343 tot += per_cpu(bh_accounting, i).nr;
3344 buffer_heads_over_limit = (tot > max_buffer_heads);
3347 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3349 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3351 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3353 __this_cpu_inc(bh_accounting.nr);
3359 EXPORT_SYMBOL(alloc_buffer_head);
3361 void free_buffer_head(struct buffer_head *bh)
3363 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3364 kmem_cache_free(bh_cachep, bh);
3366 __this_cpu_dec(bh_accounting.nr);
3370 EXPORT_SYMBOL(free_buffer_head);
3372 static int buffer_exit_cpu_dead(unsigned int cpu)
3375 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3377 for (i = 0; i < BH_LRU_SIZE; i++) {
3381 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3382 per_cpu(bh_accounting, cpu).nr = 0;
3387 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3388 * @bh: struct buffer_head
3390 * Return true if the buffer is up-to-date and false,
3391 * with the buffer locked, if not.
3393 int bh_uptodate_or_lock(struct buffer_head *bh)
3395 if (!buffer_uptodate(bh)) {
3397 if (!buffer_uptodate(bh))
3403 EXPORT_SYMBOL(bh_uptodate_or_lock);
3406 * bh_submit_read - Submit a locked buffer for reading
3407 * @bh: struct buffer_head
3409 * Returns zero on success and -EIO on error.
3411 int bh_submit_read(struct buffer_head *bh)
3413 BUG_ON(!buffer_locked(bh));
3415 if (buffer_uptodate(bh)) {
3421 bh->b_end_io = end_buffer_read_sync;
3422 submit_bh(REQ_OP_READ, 0, bh);
3424 if (buffer_uptodate(bh))
3428 EXPORT_SYMBOL(bh_submit_read);
3430 void __init buffer_init(void)
3432 unsigned long nrpages;
3435 bh_cachep = kmem_cache_create("buffer_head",
3436 sizeof(struct buffer_head), 0,
3437 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3442 * Limit the bh occupancy to 10% of ZONE_NORMAL
3444 nrpages = (nr_free_buffer_pages() * 10) / 100;
3445 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3446 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3447 NULL, buffer_exit_cpu_dead);