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>
51 #include <linux/fsverity.h>
55 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
56 static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
57 struct writeback_control *wbc);
59 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
61 inline void touch_buffer(struct buffer_head *bh)
63 trace_block_touch_buffer(bh);
64 folio_mark_accessed(bh->b_folio);
66 EXPORT_SYMBOL(touch_buffer);
68 void __lock_buffer(struct buffer_head *bh)
70 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
72 EXPORT_SYMBOL(__lock_buffer);
74 void unlock_buffer(struct buffer_head *bh)
76 clear_bit_unlock(BH_Lock, &bh->b_state);
77 smp_mb__after_atomic();
78 wake_up_bit(&bh->b_state, BH_Lock);
80 EXPORT_SYMBOL(unlock_buffer);
83 * Returns if the folio has dirty or writeback buffers. If all the buffers
84 * are unlocked and clean then the folio_test_dirty information is stale. If
85 * any of the buffers are locked, it is assumed they are locked for IO.
87 void buffer_check_dirty_writeback(struct folio *folio,
88 bool *dirty, bool *writeback)
90 struct buffer_head *head, *bh;
94 BUG_ON(!folio_test_locked(folio));
96 head = folio_buffers(folio);
100 if (folio_test_writeback(folio))
105 if (buffer_locked(bh))
108 if (buffer_dirty(bh))
111 bh = bh->b_this_page;
112 } while (bh != head);
114 EXPORT_SYMBOL(buffer_check_dirty_writeback);
117 * Block until a buffer comes unlocked. This doesn't stop it
118 * from becoming locked again - you have to lock it yourself
119 * if you want to preserve its state.
121 void __wait_on_buffer(struct buffer_head * bh)
123 wait_on_bit_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
125 EXPORT_SYMBOL(__wait_on_buffer);
127 static void buffer_io_error(struct buffer_head *bh, char *msg)
129 if (!test_bit(BH_Quiet, &bh->b_state))
130 printk_ratelimited(KERN_ERR
131 "Buffer I/O error on dev %pg, logical block %llu%s\n",
132 bh->b_bdev, (unsigned long long)bh->b_blocknr, msg);
136 * End-of-IO handler helper function which does not touch the bh after
138 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
139 * a race there is benign: unlock_buffer() only use the bh's address for
140 * hashing after unlocking the buffer, so it doesn't actually touch the bh
143 static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
146 set_buffer_uptodate(bh);
148 /* This happens, due to failed read-ahead attempts. */
149 clear_buffer_uptodate(bh);
155 * Default synchronous end-of-IO handler.. Just mark it up-to-date and
158 void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
160 __end_buffer_read_notouch(bh, uptodate);
163 EXPORT_SYMBOL(end_buffer_read_sync);
165 void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
168 set_buffer_uptodate(bh);
170 buffer_io_error(bh, ", lost sync page write");
171 mark_buffer_write_io_error(bh);
172 clear_buffer_uptodate(bh);
177 EXPORT_SYMBOL(end_buffer_write_sync);
180 * Various filesystems appear to want __find_get_block to be non-blocking.
181 * But it's the page lock which protects the buffers. To get around this,
182 * we get exclusion from try_to_free_buffers with the blockdev mapping's
185 * Hack idea: for the blockdev mapping, private_lock contention
186 * may be quite high. This code could TryLock the page, and if that
187 * succeeds, there is no need to take private_lock.
189 static struct buffer_head *
190 __find_get_block_slow(struct block_device *bdev, sector_t block)
192 struct inode *bd_inode = bdev->bd_inode;
193 struct address_space *bd_mapping = bd_inode->i_mapping;
194 struct buffer_head *ret = NULL;
196 struct buffer_head *bh;
197 struct buffer_head *head;
200 static DEFINE_RATELIMIT_STATE(last_warned, HZ, 1);
202 index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
203 page = find_get_page_flags(bd_mapping, index, FGP_ACCESSED);
207 spin_lock(&bd_mapping->private_lock);
208 if (!page_has_buffers(page))
210 head = page_buffers(page);
213 if (!buffer_mapped(bh))
215 else if (bh->b_blocknr == block) {
220 bh = bh->b_this_page;
221 } while (bh != head);
223 /* we might be here because some of the buffers on this page are
224 * not mapped. This is due to various races between
225 * file io on the block device and getblk. It gets dealt with
226 * elsewhere, don't buffer_error if we had some unmapped buffers
228 ratelimit_set_flags(&last_warned, RATELIMIT_MSG_ON_RELEASE);
229 if (all_mapped && __ratelimit(&last_warned)) {
230 printk("__find_get_block_slow() failed. block=%llu, "
231 "b_blocknr=%llu, b_state=0x%08lx, b_size=%zu, "
232 "device %pg blocksize: %d\n",
233 (unsigned long long)block,
234 (unsigned long long)bh->b_blocknr,
235 bh->b_state, bh->b_size, bdev,
236 1 << bd_inode->i_blkbits);
239 spin_unlock(&bd_mapping->private_lock);
245 static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
248 struct buffer_head *first;
249 struct buffer_head *tmp;
251 int folio_uptodate = 1;
253 BUG_ON(!buffer_async_read(bh));
257 set_buffer_uptodate(bh);
259 clear_buffer_uptodate(bh);
260 buffer_io_error(bh, ", async page read");
261 folio_set_error(folio);
265 * Be _very_ careful from here on. Bad things can happen if
266 * two buffer heads end IO at almost the same time and both
267 * decide that the page is now completely done.
269 first = folio_buffers(folio);
270 spin_lock_irqsave(&first->b_uptodate_lock, flags);
271 clear_buffer_async_read(bh);
275 if (!buffer_uptodate(tmp))
277 if (buffer_async_read(tmp)) {
278 BUG_ON(!buffer_locked(tmp));
281 tmp = tmp->b_this_page;
283 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
286 * If all of the buffers are uptodate then we can set the page
290 folio_mark_uptodate(folio);
295 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
299 struct postprocess_bh_ctx {
300 struct work_struct work;
301 struct buffer_head *bh;
304 static void verify_bh(struct work_struct *work)
306 struct postprocess_bh_ctx *ctx =
307 container_of(work, struct postprocess_bh_ctx, work);
308 struct buffer_head *bh = ctx->bh;
311 valid = fsverity_verify_blocks(bh->b_folio, bh->b_size, bh_offset(bh));
312 end_buffer_async_read(bh, valid);
316 static bool need_fsverity(struct buffer_head *bh)
318 struct folio *folio = bh->b_folio;
319 struct inode *inode = folio->mapping->host;
321 return fsverity_active(inode) &&
323 folio->index < DIV_ROUND_UP(inode->i_size, PAGE_SIZE);
326 static void decrypt_bh(struct work_struct *work)
328 struct postprocess_bh_ctx *ctx =
329 container_of(work, struct postprocess_bh_ctx, work);
330 struct buffer_head *bh = ctx->bh;
333 err = fscrypt_decrypt_pagecache_blocks(bh->b_folio, bh->b_size,
335 if (err == 0 && need_fsverity(bh)) {
337 * We use different work queues for decryption and for verity
338 * because verity may require reading metadata pages that need
339 * decryption, and we shouldn't recurse to the same workqueue.
341 INIT_WORK(&ctx->work, verify_bh);
342 fsverity_enqueue_verify_work(&ctx->work);
345 end_buffer_async_read(bh, err == 0);
350 * I/O completion handler for block_read_full_folio() - pages
351 * which come unlocked at the end of I/O.
353 static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
355 struct inode *inode = bh->b_folio->mapping->host;
356 bool decrypt = fscrypt_inode_uses_fs_layer_crypto(inode);
357 bool verify = need_fsverity(bh);
359 /* Decrypt (with fscrypt) and/or verify (with fsverity) if needed. */
360 if (uptodate && (decrypt || verify)) {
361 struct postprocess_bh_ctx *ctx =
362 kmalloc(sizeof(*ctx), GFP_ATOMIC);
367 INIT_WORK(&ctx->work, decrypt_bh);
368 fscrypt_enqueue_decrypt_work(&ctx->work);
370 INIT_WORK(&ctx->work, verify_bh);
371 fsverity_enqueue_verify_work(&ctx->work);
377 end_buffer_async_read(bh, uptodate);
381 * Completion handler for block_write_full_page() - pages which are unlocked
382 * during I/O, and which have PageWriteback cleared upon I/O completion.
384 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
387 struct buffer_head *first;
388 struct buffer_head *tmp;
391 BUG_ON(!buffer_async_write(bh));
395 set_buffer_uptodate(bh);
397 buffer_io_error(bh, ", lost async page write");
398 mark_buffer_write_io_error(bh);
399 clear_buffer_uptodate(bh);
400 folio_set_error(folio);
403 first = folio_buffers(folio);
404 spin_lock_irqsave(&first->b_uptodate_lock, flags);
406 clear_buffer_async_write(bh);
408 tmp = bh->b_this_page;
410 if (buffer_async_write(tmp)) {
411 BUG_ON(!buffer_locked(tmp));
414 tmp = tmp->b_this_page;
416 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
417 folio_end_writeback(folio);
421 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
424 EXPORT_SYMBOL(end_buffer_async_write);
427 * If a page's buffers are under async readin (end_buffer_async_read
428 * completion) then there is a possibility that another thread of
429 * control could lock one of the buffers after it has completed
430 * but while some of the other buffers have not completed. This
431 * locked buffer would confuse end_buffer_async_read() into not unlocking
432 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
433 * that this buffer is not under async I/O.
435 * The page comes unlocked when it has no locked buffer_async buffers
438 * PageLocked prevents anyone starting new async I/O reads any of
441 * PageWriteback is used to prevent simultaneous writeout of the same
444 * PageLocked prevents anyone from starting writeback of a page which is
445 * under read I/O (PageWriteback is only ever set against a locked page).
447 static void mark_buffer_async_read(struct buffer_head *bh)
449 bh->b_end_io = end_buffer_async_read_io;
450 set_buffer_async_read(bh);
453 static void mark_buffer_async_write_endio(struct buffer_head *bh,
454 bh_end_io_t *handler)
456 bh->b_end_io = handler;
457 set_buffer_async_write(bh);
460 void mark_buffer_async_write(struct buffer_head *bh)
462 mark_buffer_async_write_endio(bh, end_buffer_async_write);
464 EXPORT_SYMBOL(mark_buffer_async_write);
468 * fs/buffer.c contains helper functions for buffer-backed address space's
469 * fsync functions. A common requirement for buffer-based filesystems is
470 * that certain data from the backing blockdev needs to be written out for
471 * a successful fsync(). For example, ext2 indirect blocks need to be
472 * written back and waited upon before fsync() returns.
474 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
475 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
476 * management of a list of dependent buffers at ->i_mapping->private_list.
478 * Locking is a little subtle: try_to_free_buffers() will remove buffers
479 * from their controlling inode's queue when they are being freed. But
480 * try_to_free_buffers() will be operating against the *blockdev* mapping
481 * at the time, not against the S_ISREG file which depends on those buffers.
482 * So the locking for private_list is via the private_lock in the address_space
483 * which backs the buffers. Which is different from the address_space
484 * against which the buffers are listed. So for a particular address_space,
485 * mapping->private_lock does *not* protect mapping->private_list! In fact,
486 * mapping->private_list will always be protected by the backing blockdev's
489 * Which introduces a requirement: all buffers on an address_space's
490 * ->private_list must be from the same address_space: the blockdev's.
492 * address_spaces which do not place buffers at ->private_list via these
493 * utility functions are free to use private_lock and private_list for
494 * whatever they want. The only requirement is that list_empty(private_list)
495 * be true at clear_inode() time.
497 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
498 * filesystems should do that. invalidate_inode_buffers() should just go
499 * BUG_ON(!list_empty).
501 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
502 * take an address_space, not an inode. And it should be called
503 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
506 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
507 * list if it is already on a list. Because if the buffer is on a list,
508 * it *must* already be on the right one. If not, the filesystem is being
509 * silly. This will save a ton of locking. But first we have to ensure
510 * that buffers are taken *off* the old inode's list when they are freed
511 * (presumably in truncate). That requires careful auditing of all
512 * filesystems (do it inside bforget()). It could also be done by bringing
517 * The buffer's backing address_space's private_lock must be held
519 static void __remove_assoc_queue(struct buffer_head *bh)
521 list_del_init(&bh->b_assoc_buffers);
522 WARN_ON(!bh->b_assoc_map);
523 bh->b_assoc_map = NULL;
526 int inode_has_buffers(struct inode *inode)
528 return !list_empty(&inode->i_data.private_list);
532 * osync is designed to support O_SYNC io. It waits synchronously for
533 * all already-submitted IO to complete, but does not queue any new
534 * writes to the disk.
536 * To do O_SYNC writes, just queue the buffer writes with write_dirty_buffer
537 * as you dirty the buffers, and then use osync_inode_buffers to wait for
538 * completion. Any other dirty buffers which are not yet queued for
539 * write will not be flushed to disk by the osync.
541 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
543 struct buffer_head *bh;
549 list_for_each_prev(p, list) {
551 if (buffer_locked(bh)) {
555 if (!buffer_uptodate(bh))
566 void emergency_thaw_bdev(struct super_block *sb)
568 while (sb->s_bdev && !thaw_bdev(sb->s_bdev))
569 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
573 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
574 * @mapping: the mapping which wants those buffers written
576 * Starts I/O against the buffers at mapping->private_list, and waits upon
579 * Basically, this is a convenience function for fsync().
580 * @mapping is a file or directory which needs those buffers to be written for
581 * a successful fsync().
583 int sync_mapping_buffers(struct address_space *mapping)
585 struct address_space *buffer_mapping = mapping->private_data;
587 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
590 return fsync_buffers_list(&buffer_mapping->private_lock,
591 &mapping->private_list);
593 EXPORT_SYMBOL(sync_mapping_buffers);
596 * Called when we've recently written block `bblock', and it is known that
597 * `bblock' was for a buffer_boundary() buffer. This means that the block at
598 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
599 * dirty, schedule it for IO. So that indirects merge nicely with their data.
601 void write_boundary_block(struct block_device *bdev,
602 sector_t bblock, unsigned blocksize)
604 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
606 if (buffer_dirty(bh))
607 write_dirty_buffer(bh, 0);
612 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
614 struct address_space *mapping = inode->i_mapping;
615 struct address_space *buffer_mapping = bh->b_folio->mapping;
617 mark_buffer_dirty(bh);
618 if (!mapping->private_data) {
619 mapping->private_data = buffer_mapping;
621 BUG_ON(mapping->private_data != buffer_mapping);
623 if (!bh->b_assoc_map) {
624 spin_lock(&buffer_mapping->private_lock);
625 list_move_tail(&bh->b_assoc_buffers,
626 &mapping->private_list);
627 bh->b_assoc_map = mapping;
628 spin_unlock(&buffer_mapping->private_lock);
631 EXPORT_SYMBOL(mark_buffer_dirty_inode);
634 * Add a page to the dirty page list.
636 * It is a sad fact of life that this function is called from several places
637 * deeply under spinlocking. It may not sleep.
639 * If the page has buffers, the uptodate buffers are set dirty, to preserve
640 * dirty-state coherency between the page and the buffers. It the page does
641 * not have buffers then when they are later attached they will all be set
644 * The buffers are dirtied before the page is dirtied. There's a small race
645 * window in which a writepage caller may see the page cleanness but not the
646 * buffer dirtiness. That's fine. If this code were to set the page dirty
647 * before the buffers, a concurrent writepage caller could clear the page dirty
648 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
649 * page on the dirty page list.
651 * We use private_lock to lock against try_to_free_buffers while using the
652 * page's buffer list. Also use this to protect against clean buffers being
653 * added to the page after it was set dirty.
655 * FIXME: may need to call ->reservepage here as well. That's rather up to the
656 * address_space though.
658 bool block_dirty_folio(struct address_space *mapping, struct folio *folio)
660 struct buffer_head *head;
663 spin_lock(&mapping->private_lock);
664 head = folio_buffers(folio);
666 struct buffer_head *bh = head;
669 set_buffer_dirty(bh);
670 bh = bh->b_this_page;
671 } while (bh != head);
674 * Lock out page's memcg migration to keep PageDirty
675 * synchronized with per-memcg dirty page counters.
677 folio_memcg_lock(folio);
678 newly_dirty = !folio_test_set_dirty(folio);
679 spin_unlock(&mapping->private_lock);
682 __folio_mark_dirty(folio, mapping, 1);
684 folio_memcg_unlock(folio);
687 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
691 EXPORT_SYMBOL(block_dirty_folio);
694 * Write out and wait upon a list of buffers.
696 * We have conflicting pressures: we want to make sure that all
697 * initially dirty buffers get waited on, but that any subsequently
698 * dirtied buffers don't. After all, we don't want fsync to last
699 * forever if somebody is actively writing to the file.
701 * Do this in two main stages: first we copy dirty buffers to a
702 * temporary inode list, queueing the writes as we go. Then we clean
703 * up, waiting for those writes to complete.
705 * During this second stage, any subsequent updates to the file may end
706 * up refiling the buffer on the original inode's dirty list again, so
707 * there is a chance we will end up with a buffer queued for write but
708 * not yet completed on that list. So, as a final cleanup we go through
709 * the osync code to catch these locked, dirty buffers without requeuing
710 * any newly dirty buffers for write.
712 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
714 struct buffer_head *bh;
715 struct list_head tmp;
716 struct address_space *mapping;
718 struct blk_plug plug;
720 INIT_LIST_HEAD(&tmp);
721 blk_start_plug(&plug);
724 while (!list_empty(list)) {
725 bh = BH_ENTRY(list->next);
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) || buffer_locked(bh)) {
732 list_add(&bh->b_assoc_buffers, &tmp);
733 bh->b_assoc_map = mapping;
734 if (buffer_dirty(bh)) {
738 * Ensure any pending I/O completes so that
739 * write_dirty_buffer() actually writes the
740 * current contents - it is a noop if I/O is
741 * still in flight on potentially older
744 write_dirty_buffer(bh, REQ_SYNC);
747 * Kick off IO for the previous mapping. Note
748 * that we will not run the very last mapping,
749 * wait_on_buffer() will do that for us
750 * through sync_buffer().
759 blk_finish_plug(&plug);
762 while (!list_empty(&tmp)) {
763 bh = BH_ENTRY(tmp.prev);
765 mapping = bh->b_assoc_map;
766 __remove_assoc_queue(bh);
767 /* Avoid race with mark_buffer_dirty_inode() which does
768 * a lockless check and we rely on seeing the dirty bit */
770 if (buffer_dirty(bh)) {
771 list_add(&bh->b_assoc_buffers,
772 &mapping->private_list);
773 bh->b_assoc_map = mapping;
777 if (!buffer_uptodate(bh))
784 err2 = osync_buffers_list(lock, list);
792 * Invalidate any and all dirty buffers on a given inode. We are
793 * probably unmounting the fs, but that doesn't mean we have already
794 * done a sync(). Just drop the buffers from the inode list.
796 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
797 * assumes that all the buffers are against the blockdev. Not true
800 void invalidate_inode_buffers(struct inode *inode)
802 if (inode_has_buffers(inode)) {
803 struct address_space *mapping = &inode->i_data;
804 struct list_head *list = &mapping->private_list;
805 struct address_space *buffer_mapping = mapping->private_data;
807 spin_lock(&buffer_mapping->private_lock);
808 while (!list_empty(list))
809 __remove_assoc_queue(BH_ENTRY(list->next));
810 spin_unlock(&buffer_mapping->private_lock);
813 EXPORT_SYMBOL(invalidate_inode_buffers);
816 * Remove any clean buffers from the inode's buffer list. This is called
817 * when we're trying to free the inode itself. Those buffers can pin it.
819 * Returns true if all buffers were removed.
821 int remove_inode_buffers(struct inode *inode)
825 if (inode_has_buffers(inode)) {
826 struct address_space *mapping = &inode->i_data;
827 struct list_head *list = &mapping->private_list;
828 struct address_space *buffer_mapping = mapping->private_data;
830 spin_lock(&buffer_mapping->private_lock);
831 while (!list_empty(list)) {
832 struct buffer_head *bh = BH_ENTRY(list->next);
833 if (buffer_dirty(bh)) {
837 __remove_assoc_queue(bh);
839 spin_unlock(&buffer_mapping->private_lock);
845 * Create the appropriate buffers when given a page for data area and
846 * the size of each buffer.. Use the bh->b_this_page linked list to
847 * follow the buffers created. Return NULL if unable to create more
850 * The retry flag is used to differentiate async IO (paging, swapping)
851 * which may not fail from ordinary buffer allocations.
853 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
856 struct buffer_head *bh, *head;
857 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
859 struct mem_cgroup *memcg, *old_memcg;
864 /* The page lock pins the memcg */
865 memcg = page_memcg(page);
866 old_memcg = set_active_memcg(memcg);
870 while ((offset -= size) >= 0) {
871 bh = alloc_buffer_head(gfp);
875 bh->b_this_page = head;
881 /* Link the buffer to its page */
882 set_bh_page(bh, page, offset);
885 set_active_memcg(old_memcg);
888 * In case anything failed, we just free everything we got.
894 head = head->b_this_page;
895 free_buffer_head(bh);
901 EXPORT_SYMBOL_GPL(alloc_page_buffers);
904 link_dev_buffers(struct page *page, struct buffer_head *head)
906 struct buffer_head *bh, *tail;
911 bh = bh->b_this_page;
913 tail->b_this_page = head;
914 attach_page_private(page, head);
917 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
919 sector_t retval = ~((sector_t)0);
920 loff_t sz = bdev_nr_bytes(bdev);
923 unsigned int sizebits = blksize_bits(size);
924 retval = (sz >> sizebits);
930 * Initialise the state of a blockdev page's buffers.
933 init_page_buffers(struct page *page, struct block_device *bdev,
934 sector_t block, int size)
936 struct buffer_head *head = page_buffers(page);
937 struct buffer_head *bh = head;
938 int uptodate = PageUptodate(page);
939 sector_t end_block = blkdev_max_block(bdev, size);
942 if (!buffer_mapped(bh)) {
944 bh->b_private = NULL;
946 bh->b_blocknr = block;
948 set_buffer_uptodate(bh);
949 if (block < end_block)
950 set_buffer_mapped(bh);
953 bh = bh->b_this_page;
954 } while (bh != head);
957 * Caller needs to validate requested block against end of device.
963 * Create the page-cache page that contains the requested block.
965 * This is used purely for blockdev mappings.
968 grow_dev_page(struct block_device *bdev, sector_t block,
969 pgoff_t index, int size, int sizebits, gfp_t gfp)
971 struct inode *inode = bdev->bd_inode;
973 struct buffer_head *bh;
978 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
981 * XXX: __getblk_slow() can not really deal with failure and
982 * will endlessly loop on improvised global reclaim. Prefer
983 * looping in the allocator rather than here, at least that
984 * code knows what it's doing.
986 gfp_mask |= __GFP_NOFAIL;
988 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
990 BUG_ON(!PageLocked(page));
992 if (page_has_buffers(page)) {
993 bh = page_buffers(page);
994 if (bh->b_size == size) {
995 end_block = init_page_buffers(page, bdev,
996 (sector_t)index << sizebits,
1000 if (!try_to_free_buffers(page_folio(page)))
1005 * Allocate some buffers for this page
1007 bh = alloc_page_buffers(page, size, true);
1010 * Link the page to the buffers and initialise them. Take the
1011 * lock to be atomic wrt __find_get_block(), which does not
1012 * run under the page lock.
1014 spin_lock(&inode->i_mapping->private_lock);
1015 link_dev_buffers(page, bh);
1016 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1018 spin_unlock(&inode->i_mapping->private_lock);
1020 ret = (block < end_block) ? 1 : -ENXIO;
1028 * Create buffers for the specified block device block's page. If
1029 * that page was dirty, the buffers are set dirty also.
1032 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1037 sizebits = PAGE_SHIFT - __ffs(size);
1038 index = block >> sizebits;
1041 * Check for a block which wants to lie outside our maximum possible
1042 * pagecache index. (this comparison is done using sector_t types).
1044 if (unlikely(index != block >> sizebits)) {
1045 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1047 __func__, (unsigned long long)block,
1052 /* Create a page with the proper size buffers.. */
1053 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1056 static struct buffer_head *
1057 __getblk_slow(struct block_device *bdev, sector_t block,
1058 unsigned size, gfp_t gfp)
1060 /* Size must be multiple of hard sectorsize */
1061 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1062 (size < 512 || size > PAGE_SIZE))) {
1063 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1065 printk(KERN_ERR "logical block size: %d\n",
1066 bdev_logical_block_size(bdev));
1073 struct buffer_head *bh;
1076 bh = __find_get_block(bdev, block, size);
1080 ret = grow_buffers(bdev, block, size, gfp);
1087 * The relationship between dirty buffers and dirty pages:
1089 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1090 * the page is tagged dirty in the page cache.
1092 * At all times, the dirtiness of the buffers represents the dirtiness of
1093 * subsections of the page. If the page has buffers, the page dirty bit is
1094 * merely a hint about the true dirty state.
1096 * When a page is set dirty in its entirety, all its buffers are marked dirty
1097 * (if the page has buffers).
1099 * When a buffer is marked dirty, its page is dirtied, but the page's other
1102 * Also. When blockdev buffers are explicitly read with bread(), they
1103 * individually become uptodate. But their backing page remains not
1104 * uptodate - even if all of its buffers are uptodate. A subsequent
1105 * block_read_full_folio() against that folio will discover all the uptodate
1106 * buffers, will set the folio uptodate and will perform no I/O.
1110 * mark_buffer_dirty - mark a buffer_head as needing writeout
1111 * @bh: the buffer_head to mark dirty
1113 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1114 * its backing page dirty, then tag the page as dirty in the page cache
1115 * and then attach the address_space's inode to its superblock's dirty
1118 * mark_buffer_dirty() is atomic. It takes bh->b_folio->mapping->private_lock,
1119 * i_pages lock and mapping->host->i_lock.
1121 void mark_buffer_dirty(struct buffer_head *bh)
1123 WARN_ON_ONCE(!buffer_uptodate(bh));
1125 trace_block_dirty_buffer(bh);
1128 * Very *carefully* optimize the it-is-already-dirty case.
1130 * Don't let the final "is it dirty" escape to before we
1131 * perhaps modified the buffer.
1133 if (buffer_dirty(bh)) {
1135 if (buffer_dirty(bh))
1139 if (!test_set_buffer_dirty(bh)) {
1140 struct folio *folio = bh->b_folio;
1141 struct address_space *mapping = NULL;
1143 folio_memcg_lock(folio);
1144 if (!folio_test_set_dirty(folio)) {
1145 mapping = folio->mapping;
1147 __folio_mark_dirty(folio, mapping, 0);
1149 folio_memcg_unlock(folio);
1151 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1154 EXPORT_SYMBOL(mark_buffer_dirty);
1156 void mark_buffer_write_io_error(struct buffer_head *bh)
1158 struct super_block *sb;
1160 set_buffer_write_io_error(bh);
1161 /* FIXME: do we need to set this in both places? */
1162 if (bh->b_folio && bh->b_folio->mapping)
1163 mapping_set_error(bh->b_folio->mapping, -EIO);
1164 if (bh->b_assoc_map)
1165 mapping_set_error(bh->b_assoc_map, -EIO);
1167 sb = READ_ONCE(bh->b_bdev->bd_super);
1169 errseq_set(&sb->s_wb_err, -EIO);
1172 EXPORT_SYMBOL(mark_buffer_write_io_error);
1175 * Decrement a buffer_head's reference count. If all buffers against a page
1176 * have zero reference count, are clean and unlocked, and if the page is clean
1177 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1178 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1179 * a page but it ends up not being freed, and buffers may later be reattached).
1181 void __brelse(struct buffer_head * buf)
1183 if (atomic_read(&buf->b_count)) {
1187 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1189 EXPORT_SYMBOL(__brelse);
1192 * bforget() is like brelse(), except it discards any
1193 * potentially dirty data.
1195 void __bforget(struct buffer_head *bh)
1197 clear_buffer_dirty(bh);
1198 if (bh->b_assoc_map) {
1199 struct address_space *buffer_mapping = bh->b_folio->mapping;
1201 spin_lock(&buffer_mapping->private_lock);
1202 list_del_init(&bh->b_assoc_buffers);
1203 bh->b_assoc_map = NULL;
1204 spin_unlock(&buffer_mapping->private_lock);
1208 EXPORT_SYMBOL(__bforget);
1210 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1213 if (buffer_uptodate(bh)) {
1218 bh->b_end_io = end_buffer_read_sync;
1219 submit_bh(REQ_OP_READ, bh);
1221 if (buffer_uptodate(bh))
1229 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1230 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1231 * refcount elevated by one when they're in an LRU. A buffer can only appear
1232 * once in a particular CPU's LRU. A single buffer can be present in multiple
1233 * CPU's LRUs at the same time.
1235 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1236 * sb_find_get_block().
1238 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1239 * a local interrupt disable for that.
1242 #define BH_LRU_SIZE 16
1245 struct buffer_head *bhs[BH_LRU_SIZE];
1248 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1251 #define bh_lru_lock() local_irq_disable()
1252 #define bh_lru_unlock() local_irq_enable()
1254 #define bh_lru_lock() preempt_disable()
1255 #define bh_lru_unlock() preempt_enable()
1258 static inline void check_irqs_on(void)
1260 #ifdef irqs_disabled
1261 BUG_ON(irqs_disabled());
1266 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1267 * inserted at the front, and the buffer_head at the back if any is evicted.
1268 * Or, if already in the LRU it is moved to the front.
1270 static void bh_lru_install(struct buffer_head *bh)
1272 struct buffer_head *evictee = bh;
1280 * the refcount of buffer_head in bh_lru prevents dropping the
1281 * attached page(i.e., try_to_free_buffers) so it could cause
1282 * failing page migration.
1283 * Skip putting upcoming bh into bh_lru until migration is done.
1285 if (lru_cache_disabled()) {
1290 b = this_cpu_ptr(&bh_lrus);
1291 for (i = 0; i < BH_LRU_SIZE; i++) {
1292 swap(evictee, b->bhs[i]);
1293 if (evictee == bh) {
1305 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1307 static struct buffer_head *
1308 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1310 struct buffer_head *ret = NULL;
1315 for (i = 0; i < BH_LRU_SIZE; i++) {
1316 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1318 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1319 bh->b_size == size) {
1322 __this_cpu_write(bh_lrus.bhs[i],
1323 __this_cpu_read(bh_lrus.bhs[i - 1]));
1326 __this_cpu_write(bh_lrus.bhs[0], bh);
1338 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1339 * it in the LRU and mark it as accessed. If it is not present then return
1342 struct buffer_head *
1343 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1345 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1348 /* __find_get_block_slow will mark the page accessed */
1349 bh = __find_get_block_slow(bdev, block);
1357 EXPORT_SYMBOL(__find_get_block);
1360 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1361 * which corresponds to the passed block_device, block and size. The
1362 * returned buffer has its reference count incremented.
1364 * __getblk_gfp() will lock up the machine if grow_dev_page's
1365 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1367 struct buffer_head *
1368 __getblk_gfp(struct block_device *bdev, sector_t block,
1369 unsigned size, gfp_t gfp)
1371 struct buffer_head *bh = __find_get_block(bdev, block, size);
1375 bh = __getblk_slow(bdev, block, size, gfp);
1378 EXPORT_SYMBOL(__getblk_gfp);
1381 * Do async read-ahead on a buffer..
1383 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1385 struct buffer_head *bh = __getblk(bdev, block, size);
1387 bh_readahead(bh, REQ_RAHEAD);
1391 EXPORT_SYMBOL(__breadahead);
1394 * __bread_gfp() - reads a specified block and returns the bh
1395 * @bdev: the block_device to read from
1396 * @block: number of block
1397 * @size: size (in bytes) to read
1398 * @gfp: page allocation flag
1400 * Reads a specified block, and returns buffer head that contains it.
1401 * The page cache can be allocated from non-movable area
1402 * not to prevent page migration if you set gfp to zero.
1403 * It returns NULL if the block was unreadable.
1405 struct buffer_head *
1406 __bread_gfp(struct block_device *bdev, sector_t block,
1407 unsigned size, gfp_t gfp)
1409 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1411 if (likely(bh) && !buffer_uptodate(bh))
1412 bh = __bread_slow(bh);
1415 EXPORT_SYMBOL(__bread_gfp);
1417 static void __invalidate_bh_lrus(struct bh_lru *b)
1421 for (i = 0; i < BH_LRU_SIZE; i++) {
1427 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1428 * This doesn't race because it runs in each cpu either in irq
1429 * or with preempt disabled.
1431 static void invalidate_bh_lru(void *arg)
1433 struct bh_lru *b = &get_cpu_var(bh_lrus);
1435 __invalidate_bh_lrus(b);
1436 put_cpu_var(bh_lrus);
1439 bool has_bh_in_lru(int cpu, void *dummy)
1441 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1444 for (i = 0; i < BH_LRU_SIZE; i++) {
1452 void invalidate_bh_lrus(void)
1454 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1456 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1459 * It's called from workqueue context so we need a bh_lru_lock to close
1460 * the race with preemption/irq.
1462 void invalidate_bh_lrus_cpu(void)
1467 b = this_cpu_ptr(&bh_lrus);
1468 __invalidate_bh_lrus(b);
1472 void set_bh_page(struct buffer_head *bh,
1473 struct page *page, unsigned long offset)
1476 BUG_ON(offset >= PAGE_SIZE);
1477 if (PageHighMem(page))
1479 * This catches illegal uses and preserves the offset:
1481 bh->b_data = (char *)(0 + offset);
1483 bh->b_data = page_address(page) + offset;
1485 EXPORT_SYMBOL(set_bh_page);
1488 * Called when truncating a buffer on a page completely.
1491 /* Bits that are cleared during an invalidate */
1492 #define BUFFER_FLAGS_DISCARD \
1493 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1494 1 << BH_Delay | 1 << BH_Unwritten)
1496 static void discard_buffer(struct buffer_head * bh)
1498 unsigned long b_state;
1501 clear_buffer_dirty(bh);
1503 b_state = READ_ONCE(bh->b_state);
1505 } while (!try_cmpxchg(&bh->b_state, &b_state,
1506 b_state & ~BUFFER_FLAGS_DISCARD));
1511 * block_invalidate_folio - Invalidate part or all of a buffer-backed folio.
1512 * @folio: The folio which is affected.
1513 * @offset: start of the range to invalidate
1514 * @length: length of the range to invalidate
1516 * block_invalidate_folio() is called when all or part of the folio has been
1517 * invalidated by a truncate operation.
1519 * block_invalidate_folio() does not have to release all buffers, but it must
1520 * ensure that no dirty buffer is left outside @offset and that no I/O
1521 * is underway against any of the blocks which are outside the truncation
1522 * point. Because the caller is about to free (and possibly reuse) those
1525 void block_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1527 struct buffer_head *head, *bh, *next;
1528 size_t curr_off = 0;
1529 size_t stop = length + offset;
1531 BUG_ON(!folio_test_locked(folio));
1534 * Check for overflow
1536 BUG_ON(stop > folio_size(folio) || stop < length);
1538 head = folio_buffers(folio);
1544 size_t next_off = curr_off + bh->b_size;
1545 next = bh->b_this_page;
1548 * Are we still fully in range ?
1550 if (next_off > stop)
1554 * is this block fully invalidated?
1556 if (offset <= curr_off)
1558 curr_off = next_off;
1560 } while (bh != head);
1563 * We release buffers only if the entire folio is being invalidated.
1564 * The get_block cached value has been unconditionally invalidated,
1565 * so real IO is not possible anymore.
1567 if (length == folio_size(folio))
1568 filemap_release_folio(folio, 0);
1572 EXPORT_SYMBOL(block_invalidate_folio);
1576 * We attach and possibly dirty the buffers atomically wrt
1577 * block_dirty_folio() via private_lock. try_to_free_buffers
1578 * is already excluded via the page lock.
1580 void create_empty_buffers(struct page *page,
1581 unsigned long blocksize, unsigned long b_state)
1583 struct buffer_head *bh, *head, *tail;
1585 head = alloc_page_buffers(page, blocksize, true);
1588 bh->b_state |= b_state;
1590 bh = bh->b_this_page;
1592 tail->b_this_page = head;
1594 spin_lock(&page->mapping->private_lock);
1595 if (PageUptodate(page) || PageDirty(page)) {
1598 if (PageDirty(page))
1599 set_buffer_dirty(bh);
1600 if (PageUptodate(page))
1601 set_buffer_uptodate(bh);
1602 bh = bh->b_this_page;
1603 } while (bh != head);
1605 attach_page_private(page, head);
1606 spin_unlock(&page->mapping->private_lock);
1608 EXPORT_SYMBOL(create_empty_buffers);
1611 * clean_bdev_aliases: clean a range of buffers in block device
1612 * @bdev: Block device to clean buffers in
1613 * @block: Start of a range of blocks to clean
1614 * @len: Number of blocks to clean
1616 * We are taking a range of blocks for data and we don't want writeback of any
1617 * buffer-cache aliases starting from return from this function and until the
1618 * moment when something will explicitly mark the buffer dirty (hopefully that
1619 * will not happen until we will free that block ;-) We don't even need to mark
1620 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1621 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1622 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1623 * would confuse anyone who might pick it with bread() afterwards...
1625 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1626 * writeout I/O going on against recently-freed buffers. We don't wait on that
1627 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1628 * need to. That happens here.
1630 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1632 struct inode *bd_inode = bdev->bd_inode;
1633 struct address_space *bd_mapping = bd_inode->i_mapping;
1634 struct folio_batch fbatch;
1635 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1638 struct buffer_head *bh;
1639 struct buffer_head *head;
1641 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1642 folio_batch_init(&fbatch);
1643 while (filemap_get_folios(bd_mapping, &index, end, &fbatch)) {
1644 count = folio_batch_count(&fbatch);
1645 for (i = 0; i < count; i++) {
1646 struct folio *folio = fbatch.folios[i];
1648 if (!folio_buffers(folio))
1651 * We use folio lock instead of bd_mapping->private_lock
1652 * to pin buffers here since we can afford to sleep and
1653 * it scales better than a global spinlock lock.
1656 /* Recheck when the folio is locked which pins bhs */
1657 head = folio_buffers(folio);
1662 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1664 if (bh->b_blocknr >= block + len)
1666 clear_buffer_dirty(bh);
1668 clear_buffer_req(bh);
1670 bh = bh->b_this_page;
1671 } while (bh != head);
1673 folio_unlock(folio);
1675 folio_batch_release(&fbatch);
1677 /* End of range already reached? */
1678 if (index > end || !index)
1682 EXPORT_SYMBOL(clean_bdev_aliases);
1685 * Size is a power-of-two in the range 512..PAGE_SIZE,
1686 * and the case we care about most is PAGE_SIZE.
1688 * So this *could* possibly be written with those
1689 * constraints in mind (relevant mostly if some
1690 * architecture has a slow bit-scan instruction)
1692 static inline int block_size_bits(unsigned int blocksize)
1694 return ilog2(blocksize);
1697 static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1699 BUG_ON(!PageLocked(page));
1701 if (!page_has_buffers(page))
1702 create_empty_buffers(page, 1 << READ_ONCE(inode->i_blkbits),
1704 return page_buffers(page);
1708 * NOTE! All mapped/uptodate combinations are valid:
1710 * Mapped Uptodate Meaning
1712 * No No "unknown" - must do get_block()
1713 * No Yes "hole" - zero-filled
1714 * Yes No "allocated" - allocated on disk, not read in
1715 * Yes Yes "valid" - allocated and up-to-date in memory.
1717 * "Dirty" is valid only with the last case (mapped+uptodate).
1721 * While block_write_full_page is writing back the dirty buffers under
1722 * the page lock, whoever dirtied the buffers may decide to clean them
1723 * again at any time. We handle that by only looking at the buffer
1724 * state inside lock_buffer().
1726 * If block_write_full_page() is called for regular writeback
1727 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1728 * locked buffer. This only can happen if someone has written the buffer
1729 * directly, with submit_bh(). At the address_space level PageWriteback
1730 * prevents this contention from occurring.
1732 * If block_write_full_page() is called with wbc->sync_mode ==
1733 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1734 * causes the writes to be flagged as synchronous writes.
1736 int __block_write_full_page(struct inode *inode, struct page *page,
1737 get_block_t *get_block, struct writeback_control *wbc,
1738 bh_end_io_t *handler)
1742 sector_t last_block;
1743 struct buffer_head *bh, *head;
1744 unsigned int blocksize, bbits;
1745 int nr_underway = 0;
1746 blk_opf_t write_flags = wbc_to_write_flags(wbc);
1748 head = create_page_buffers(page, inode,
1749 (1 << BH_Dirty)|(1 << BH_Uptodate));
1752 * Be very careful. We have no exclusion from block_dirty_folio
1753 * here, and the (potentially unmapped) buffers may become dirty at
1754 * any time. If a buffer becomes dirty here after we've inspected it
1755 * then we just miss that fact, and the page stays dirty.
1757 * Buffers outside i_size may be dirtied by block_dirty_folio;
1758 * handle that here by just cleaning them.
1762 blocksize = bh->b_size;
1763 bbits = block_size_bits(blocksize);
1765 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1766 last_block = (i_size_read(inode) - 1) >> bbits;
1769 * Get all the dirty buffers mapped to disk addresses and
1770 * handle any aliases from the underlying blockdev's mapping.
1773 if (block > last_block) {
1775 * mapped buffers outside i_size will occur, because
1776 * this page can be outside i_size when there is a
1777 * truncate in progress.
1780 * The buffer was zeroed by block_write_full_page()
1782 clear_buffer_dirty(bh);
1783 set_buffer_uptodate(bh);
1784 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1786 WARN_ON(bh->b_size != blocksize);
1787 err = get_block(inode, block, bh, 1);
1790 clear_buffer_delay(bh);
1791 if (buffer_new(bh)) {
1792 /* blockdev mappings never come here */
1793 clear_buffer_new(bh);
1794 clean_bdev_bh_alias(bh);
1797 bh = bh->b_this_page;
1799 } while (bh != head);
1802 if (!buffer_mapped(bh))
1805 * If it's a fully non-blocking write attempt and we cannot
1806 * lock the buffer then redirty the page. Note that this can
1807 * potentially cause a busy-wait loop from writeback threads
1808 * and kswapd activity, but those code paths have their own
1809 * higher-level throttling.
1811 if (wbc->sync_mode != WB_SYNC_NONE) {
1813 } else if (!trylock_buffer(bh)) {
1814 redirty_page_for_writepage(wbc, page);
1817 if (test_clear_buffer_dirty(bh)) {
1818 mark_buffer_async_write_endio(bh, handler);
1822 } while ((bh = bh->b_this_page) != head);
1825 * The page and its buffers are protected by PageWriteback(), so we can
1826 * drop the bh refcounts early.
1828 BUG_ON(PageWriteback(page));
1829 set_page_writeback(page);
1832 struct buffer_head *next = bh->b_this_page;
1833 if (buffer_async_write(bh)) {
1834 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh, wbc);
1838 } while (bh != head);
1843 if (nr_underway == 0) {
1845 * The page was marked dirty, but the buffers were
1846 * clean. Someone wrote them back by hand with
1847 * write_dirty_buffer/submit_bh. A rare case.
1849 end_page_writeback(page);
1852 * The page and buffer_heads can be released at any time from
1860 * ENOSPC, or some other error. We may already have added some
1861 * blocks to the file, so we need to write these out to avoid
1862 * exposing stale data.
1863 * The page is currently locked and not marked for writeback
1866 /* Recovery: lock and submit the mapped buffers */
1868 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1869 !buffer_delay(bh)) {
1871 mark_buffer_async_write_endio(bh, handler);
1874 * The buffer may have been set dirty during
1875 * attachment to a dirty page.
1877 clear_buffer_dirty(bh);
1879 } while ((bh = bh->b_this_page) != head);
1881 BUG_ON(PageWriteback(page));
1882 mapping_set_error(page->mapping, err);
1883 set_page_writeback(page);
1885 struct buffer_head *next = bh->b_this_page;
1886 if (buffer_async_write(bh)) {
1887 clear_buffer_dirty(bh);
1888 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh, wbc);
1892 } while (bh != head);
1896 EXPORT_SYMBOL(__block_write_full_page);
1899 * If a page has any new buffers, zero them out here, and mark them uptodate
1900 * and dirty so they'll be written out (in order to prevent uninitialised
1901 * block data from leaking). And clear the new bit.
1903 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1905 unsigned int block_start, block_end;
1906 struct buffer_head *head, *bh;
1908 BUG_ON(!PageLocked(page));
1909 if (!page_has_buffers(page))
1912 bh = head = page_buffers(page);
1915 block_end = block_start + bh->b_size;
1917 if (buffer_new(bh)) {
1918 if (block_end > from && block_start < to) {
1919 if (!PageUptodate(page)) {
1920 unsigned start, size;
1922 start = max(from, block_start);
1923 size = min(to, block_end) - start;
1925 zero_user(page, start, size);
1926 set_buffer_uptodate(bh);
1929 clear_buffer_new(bh);
1930 mark_buffer_dirty(bh);
1934 block_start = block_end;
1935 bh = bh->b_this_page;
1936 } while (bh != head);
1938 EXPORT_SYMBOL(page_zero_new_buffers);
1941 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1942 const struct iomap *iomap)
1944 loff_t offset = block << inode->i_blkbits;
1946 bh->b_bdev = iomap->bdev;
1949 * Block points to offset in file we need to map, iomap contains
1950 * the offset at which the map starts. If the map ends before the
1951 * current block, then do not map the buffer and let the caller
1954 BUG_ON(offset >= iomap->offset + iomap->length);
1956 switch (iomap->type) {
1959 * If the buffer is not up to date or beyond the current EOF,
1960 * we need to mark it as new to ensure sub-block zeroing is
1961 * executed if necessary.
1963 if (!buffer_uptodate(bh) ||
1964 (offset >= i_size_read(inode)))
1967 case IOMAP_DELALLOC:
1968 if (!buffer_uptodate(bh) ||
1969 (offset >= i_size_read(inode)))
1971 set_buffer_uptodate(bh);
1972 set_buffer_mapped(bh);
1973 set_buffer_delay(bh);
1975 case IOMAP_UNWRITTEN:
1977 * For unwritten regions, we always need to ensure that regions
1978 * in the block we are not writing to are zeroed. Mark the
1979 * buffer as new to ensure this.
1982 set_buffer_unwritten(bh);
1985 if ((iomap->flags & IOMAP_F_NEW) ||
1986 offset >= i_size_read(inode))
1988 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
1990 set_buffer_mapped(bh);
1995 int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len,
1996 get_block_t *get_block, const struct iomap *iomap)
1998 unsigned from = pos & (PAGE_SIZE - 1);
1999 unsigned to = from + len;
2000 struct inode *inode = folio->mapping->host;
2001 unsigned block_start, block_end;
2004 unsigned blocksize, bbits;
2005 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2007 BUG_ON(!folio_test_locked(folio));
2008 BUG_ON(from > PAGE_SIZE);
2009 BUG_ON(to > PAGE_SIZE);
2012 head = create_page_buffers(&folio->page, inode, 0);
2013 blocksize = head->b_size;
2014 bbits = block_size_bits(blocksize);
2016 block = (sector_t)folio->index << (PAGE_SHIFT - bbits);
2018 for(bh = head, block_start = 0; bh != head || !block_start;
2019 block++, block_start=block_end, bh = bh->b_this_page) {
2020 block_end = block_start + blocksize;
2021 if (block_end <= from || block_start >= to) {
2022 if (folio_test_uptodate(folio)) {
2023 if (!buffer_uptodate(bh))
2024 set_buffer_uptodate(bh);
2029 clear_buffer_new(bh);
2030 if (!buffer_mapped(bh)) {
2031 WARN_ON(bh->b_size != blocksize);
2033 err = get_block(inode, block, bh, 1);
2037 iomap_to_bh(inode, block, bh, iomap);
2040 if (buffer_new(bh)) {
2041 clean_bdev_bh_alias(bh);
2042 if (folio_test_uptodate(folio)) {
2043 clear_buffer_new(bh);
2044 set_buffer_uptodate(bh);
2045 mark_buffer_dirty(bh);
2048 if (block_end > to || block_start < from)
2049 folio_zero_segments(folio,
2055 if (folio_test_uptodate(folio)) {
2056 if (!buffer_uptodate(bh))
2057 set_buffer_uptodate(bh);
2060 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2061 !buffer_unwritten(bh) &&
2062 (block_start < from || block_end > to)) {
2063 bh_read_nowait(bh, 0);
2068 * If we issued read requests - let them complete.
2070 while(wait_bh > wait) {
2071 wait_on_buffer(*--wait_bh);
2072 if (!buffer_uptodate(*wait_bh))
2076 page_zero_new_buffers(&folio->page, from, to);
2080 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2081 get_block_t *get_block)
2083 return __block_write_begin_int(page_folio(page), pos, len, get_block,
2086 EXPORT_SYMBOL(__block_write_begin);
2088 static int __block_commit_write(struct inode *inode, struct page *page,
2089 unsigned from, unsigned to)
2091 unsigned block_start, block_end;
2094 struct buffer_head *bh, *head;
2096 bh = head = page_buffers(page);
2097 blocksize = bh->b_size;
2101 block_end = block_start + blocksize;
2102 if (block_end <= from || block_start >= to) {
2103 if (!buffer_uptodate(bh))
2106 set_buffer_uptodate(bh);
2107 mark_buffer_dirty(bh);
2110 clear_buffer_new(bh);
2112 block_start = block_end;
2113 bh = bh->b_this_page;
2114 } while (bh != head);
2117 * If this is a partial write which happened to make all buffers
2118 * uptodate then we can optimize away a bogus read_folio() for
2119 * the next read(). Here we 'discover' whether the page went
2120 * uptodate as a result of this (potentially partial) write.
2123 SetPageUptodate(page);
2128 * block_write_begin takes care of the basic task of block allocation and
2129 * bringing partial write blocks uptodate first.
2131 * The filesystem needs to handle block truncation upon failure.
2133 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2134 struct page **pagep, get_block_t *get_block)
2136 pgoff_t index = pos >> PAGE_SHIFT;
2140 page = grab_cache_page_write_begin(mapping, index);
2144 status = __block_write_begin(page, pos, len, get_block);
2145 if (unlikely(status)) {
2154 EXPORT_SYMBOL(block_write_begin);
2156 int block_write_end(struct file *file, struct address_space *mapping,
2157 loff_t pos, unsigned len, unsigned copied,
2158 struct page *page, void *fsdata)
2160 struct inode *inode = mapping->host;
2163 start = pos & (PAGE_SIZE - 1);
2165 if (unlikely(copied < len)) {
2167 * The buffers that were written will now be uptodate, so
2168 * we don't have to worry about a read_folio reading them
2169 * and overwriting a partial write. However if we have
2170 * encountered a short write and only partially written
2171 * into a buffer, it will not be marked uptodate, so a
2172 * read_folio might come in and destroy our partial write.
2174 * Do the simplest thing, and just treat any short write to a
2175 * non uptodate page as a zero-length write, and force the
2176 * caller to redo the whole thing.
2178 if (!PageUptodate(page))
2181 page_zero_new_buffers(page, start+copied, start+len);
2183 flush_dcache_page(page);
2185 /* This could be a short (even 0-length) commit */
2186 __block_commit_write(inode, page, start, start+copied);
2190 EXPORT_SYMBOL(block_write_end);
2192 int generic_write_end(struct file *file, struct address_space *mapping,
2193 loff_t pos, unsigned len, unsigned copied,
2194 struct page *page, void *fsdata)
2196 struct inode *inode = mapping->host;
2197 loff_t old_size = inode->i_size;
2198 bool i_size_changed = false;
2200 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2203 * No need to use i_size_read() here, the i_size cannot change under us
2204 * because we hold i_rwsem.
2206 * But it's important to update i_size while still holding page lock:
2207 * page writeout could otherwise come in and zero beyond i_size.
2209 if (pos + copied > inode->i_size) {
2210 i_size_write(inode, pos + copied);
2211 i_size_changed = true;
2218 pagecache_isize_extended(inode, old_size, pos);
2220 * Don't mark the inode dirty under page lock. First, it unnecessarily
2221 * makes the holding time of page lock longer. Second, it forces lock
2222 * ordering of page lock and transaction start for journaling
2226 mark_inode_dirty(inode);
2229 EXPORT_SYMBOL(generic_write_end);
2232 * block_is_partially_uptodate checks whether buffers within a folio are
2235 * Returns true if all buffers which correspond to the specified part
2236 * of the folio are uptodate.
2238 bool block_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
2240 unsigned block_start, block_end, blocksize;
2242 struct buffer_head *bh, *head;
2245 head = folio_buffers(folio);
2248 blocksize = head->b_size;
2249 to = min_t(unsigned, folio_size(folio) - from, count);
2251 if (from < blocksize && to > folio_size(folio) - blocksize)
2257 block_end = block_start + blocksize;
2258 if (block_end > from && block_start < to) {
2259 if (!buffer_uptodate(bh)) {
2263 if (block_end >= to)
2266 block_start = block_end;
2267 bh = bh->b_this_page;
2268 } while (bh != head);
2272 EXPORT_SYMBOL(block_is_partially_uptodate);
2275 * Generic "read_folio" function for block devices that have the normal
2276 * get_block functionality. This is most of the block device filesystems.
2277 * Reads the folio asynchronously --- the unlock_buffer() and
2278 * set/clear_buffer_uptodate() functions propagate buffer state into the
2279 * folio once IO has completed.
2281 int block_read_full_folio(struct folio *folio, get_block_t *get_block)
2283 struct inode *inode = folio->mapping->host;
2284 sector_t iblock, lblock;
2285 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2286 unsigned int blocksize, bbits;
2288 int fully_mapped = 1;
2289 bool page_error = false;
2290 loff_t limit = i_size_read(inode);
2292 /* This is needed for ext4. */
2293 if (IS_ENABLED(CONFIG_FS_VERITY) && IS_VERITY(inode))
2294 limit = inode->i_sb->s_maxbytes;
2296 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2298 head = create_page_buffers(&folio->page, inode, 0);
2299 blocksize = head->b_size;
2300 bbits = block_size_bits(blocksize);
2302 iblock = (sector_t)folio->index << (PAGE_SHIFT - bbits);
2303 lblock = (limit+blocksize-1) >> bbits;
2309 if (buffer_uptodate(bh))
2312 if (!buffer_mapped(bh)) {
2316 if (iblock < lblock) {
2317 WARN_ON(bh->b_size != blocksize);
2318 err = get_block(inode, iblock, bh, 0);
2320 folio_set_error(folio);
2324 if (!buffer_mapped(bh)) {
2325 folio_zero_range(folio, i * blocksize,
2328 set_buffer_uptodate(bh);
2332 * get_block() might have updated the buffer
2335 if (buffer_uptodate(bh))
2339 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2342 folio_set_mappedtodisk(folio);
2346 * All buffers are uptodate - we can set the folio uptodate
2347 * as well. But not if get_block() returned an error.
2350 folio_mark_uptodate(folio);
2351 folio_unlock(folio);
2355 /* Stage two: lock the buffers */
2356 for (i = 0; i < nr; i++) {
2359 mark_buffer_async_read(bh);
2363 * Stage 3: start the IO. Check for uptodateness
2364 * inside the buffer lock in case another process reading
2365 * the underlying blockdev brought it uptodate (the sct fix).
2367 for (i = 0; i < nr; i++) {
2369 if (buffer_uptodate(bh))
2370 end_buffer_async_read(bh, 1);
2372 submit_bh(REQ_OP_READ, bh);
2376 EXPORT_SYMBOL(block_read_full_folio);
2378 /* utility function for filesystems that need to do work on expanding
2379 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2380 * deal with the hole.
2382 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2384 struct address_space *mapping = inode->i_mapping;
2385 const struct address_space_operations *aops = mapping->a_ops;
2387 void *fsdata = NULL;
2390 err = inode_newsize_ok(inode, size);
2394 err = aops->write_begin(NULL, mapping, size, 0, &page, &fsdata);
2398 err = aops->write_end(NULL, mapping, size, 0, 0, page, fsdata);
2404 EXPORT_SYMBOL(generic_cont_expand_simple);
2406 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2407 loff_t pos, loff_t *bytes)
2409 struct inode *inode = mapping->host;
2410 const struct address_space_operations *aops = mapping->a_ops;
2411 unsigned int blocksize = i_blocksize(inode);
2413 void *fsdata = NULL;
2414 pgoff_t index, curidx;
2416 unsigned zerofrom, offset, len;
2419 index = pos >> PAGE_SHIFT;
2420 offset = pos & ~PAGE_MASK;
2422 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2423 zerofrom = curpos & ~PAGE_MASK;
2424 if (zerofrom & (blocksize-1)) {
2425 *bytes |= (blocksize-1);
2428 len = PAGE_SIZE - zerofrom;
2430 err = aops->write_begin(file, mapping, curpos, len,
2434 zero_user(page, zerofrom, len);
2435 err = aops->write_end(file, mapping, curpos, len, len,
2442 balance_dirty_pages_ratelimited(mapping);
2444 if (fatal_signal_pending(current)) {
2450 /* page covers the boundary, find the boundary offset */
2451 if (index == curidx) {
2452 zerofrom = curpos & ~PAGE_MASK;
2453 /* if we will expand the thing last block will be filled */
2454 if (offset <= zerofrom) {
2457 if (zerofrom & (blocksize-1)) {
2458 *bytes |= (blocksize-1);
2461 len = offset - zerofrom;
2463 err = aops->write_begin(file, mapping, curpos, len,
2467 zero_user(page, zerofrom, len);
2468 err = aops->write_end(file, mapping, curpos, len, len,
2480 * For moronic filesystems that do not allow holes in file.
2481 * We may have to extend the file.
2483 int cont_write_begin(struct file *file, struct address_space *mapping,
2484 loff_t pos, unsigned len,
2485 struct page **pagep, void **fsdata,
2486 get_block_t *get_block, loff_t *bytes)
2488 struct inode *inode = mapping->host;
2489 unsigned int blocksize = i_blocksize(inode);
2490 unsigned int zerofrom;
2493 err = cont_expand_zero(file, mapping, pos, bytes);
2497 zerofrom = *bytes & ~PAGE_MASK;
2498 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2499 *bytes |= (blocksize-1);
2503 return block_write_begin(mapping, pos, len, pagep, get_block);
2505 EXPORT_SYMBOL(cont_write_begin);
2507 int block_commit_write(struct page *page, unsigned from, unsigned to)
2509 struct inode *inode = page->mapping->host;
2510 __block_commit_write(inode,page,from,to);
2513 EXPORT_SYMBOL(block_commit_write);
2516 * block_page_mkwrite() is not allowed to change the file size as it gets
2517 * called from a page fault handler when a page is first dirtied. Hence we must
2518 * be careful to check for EOF conditions here. We set the page up correctly
2519 * for a written page which means we get ENOSPC checking when writing into
2520 * holes and correct delalloc and unwritten extent mapping on filesystems that
2521 * support these features.
2523 * We are not allowed to take the i_mutex here so we have to play games to
2524 * protect against truncate races as the page could now be beyond EOF. Because
2525 * truncate writes the inode size before removing pages, once we have the
2526 * page lock we can determine safely if the page is beyond EOF. If it is not
2527 * beyond EOF, then the page is guaranteed safe against truncation until we
2530 * Direct callers of this function should protect against filesystem freezing
2531 * using sb_start_pagefault() - sb_end_pagefault() functions.
2533 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2534 get_block_t get_block)
2536 struct page *page = vmf->page;
2537 struct inode *inode = file_inode(vma->vm_file);
2543 size = i_size_read(inode);
2544 if ((page->mapping != inode->i_mapping) ||
2545 (page_offset(page) > size)) {
2546 /* We overload EFAULT to mean page got truncated */
2551 /* page is wholly or partially inside EOF */
2552 if (((page->index + 1) << PAGE_SHIFT) > size)
2553 end = size & ~PAGE_MASK;
2557 ret = __block_write_begin(page, 0, end, get_block);
2559 ret = block_commit_write(page, 0, end);
2561 if (unlikely(ret < 0))
2563 set_page_dirty(page);
2564 wait_for_stable_page(page);
2570 EXPORT_SYMBOL(block_page_mkwrite);
2572 int block_truncate_page(struct address_space *mapping,
2573 loff_t from, get_block_t *get_block)
2575 pgoff_t index = from >> PAGE_SHIFT;
2576 unsigned offset = from & (PAGE_SIZE-1);
2579 unsigned length, pos;
2580 struct inode *inode = mapping->host;
2582 struct buffer_head *bh;
2585 blocksize = i_blocksize(inode);
2586 length = offset & (blocksize - 1);
2588 /* Block boundary? Nothing to do */
2592 length = blocksize - length;
2593 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2595 page = grab_cache_page(mapping, index);
2599 if (!page_has_buffers(page))
2600 create_empty_buffers(page, blocksize, 0);
2602 /* Find the buffer that contains "offset" */
2603 bh = page_buffers(page);
2605 while (offset >= pos) {
2606 bh = bh->b_this_page;
2611 if (!buffer_mapped(bh)) {
2612 WARN_ON(bh->b_size != blocksize);
2613 err = get_block(inode, iblock, bh, 0);
2616 /* unmapped? It's a hole - nothing to do */
2617 if (!buffer_mapped(bh))
2621 /* Ok, it's mapped. Make sure it's up-to-date */
2622 if (PageUptodate(page))
2623 set_buffer_uptodate(bh);
2625 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2626 err = bh_read(bh, 0);
2627 /* Uhhuh. Read error. Complain and punt. */
2632 zero_user(page, offset, length);
2633 mark_buffer_dirty(bh);
2641 EXPORT_SYMBOL(block_truncate_page);
2644 * The generic ->writepage function for buffer-backed address_spaces
2646 int block_write_full_page(struct page *page, get_block_t *get_block,
2647 struct writeback_control *wbc)
2649 struct inode * const inode = page->mapping->host;
2650 loff_t i_size = i_size_read(inode);
2651 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2654 /* Is the page fully inside i_size? */
2655 if (page->index < end_index)
2656 return __block_write_full_page(inode, page, get_block, wbc,
2657 end_buffer_async_write);
2659 /* Is the page fully outside i_size? (truncate in progress) */
2660 offset = i_size & (PAGE_SIZE-1);
2661 if (page->index >= end_index+1 || !offset) {
2663 return 0; /* don't care */
2667 * The page straddles i_size. It must be zeroed out on each and every
2668 * writepage invocation because it may be mmapped. "A file is mapped
2669 * in multiples of the page size. For a file that is not a multiple of
2670 * the page size, the remaining memory is zeroed when mapped, and
2671 * writes to that region are not written out to the file."
2673 zero_user_segment(page, offset, PAGE_SIZE);
2674 return __block_write_full_page(inode, page, get_block, wbc,
2675 end_buffer_async_write);
2677 EXPORT_SYMBOL(block_write_full_page);
2679 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2680 get_block_t *get_block)
2682 struct inode *inode = mapping->host;
2683 struct buffer_head tmp = {
2684 .b_size = i_blocksize(inode),
2687 get_block(inode, block, &tmp, 0);
2688 return tmp.b_blocknr;
2690 EXPORT_SYMBOL(generic_block_bmap);
2692 static void end_bio_bh_io_sync(struct bio *bio)
2694 struct buffer_head *bh = bio->bi_private;
2696 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2697 set_bit(BH_Quiet, &bh->b_state);
2699 bh->b_end_io(bh, !bio->bi_status);
2703 static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
2704 struct writeback_control *wbc)
2706 const enum req_op op = opf & REQ_OP_MASK;
2709 BUG_ON(!buffer_locked(bh));
2710 BUG_ON(!buffer_mapped(bh));
2711 BUG_ON(!bh->b_end_io);
2712 BUG_ON(buffer_delay(bh));
2713 BUG_ON(buffer_unwritten(bh));
2716 * Only clear out a write error when rewriting
2718 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
2719 clear_buffer_write_io_error(bh);
2721 if (buffer_meta(bh))
2723 if (buffer_prio(bh))
2726 bio = bio_alloc(bh->b_bdev, 1, opf, GFP_NOIO);
2728 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
2730 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2732 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
2733 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
2735 bio->bi_end_io = end_bio_bh_io_sync;
2736 bio->bi_private = bh;
2738 /* Take care of bh's that straddle the end of the device */
2742 wbc_init_bio(wbc, bio);
2743 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
2749 void submit_bh(blk_opf_t opf, struct buffer_head *bh)
2751 submit_bh_wbc(opf, bh, NULL);
2753 EXPORT_SYMBOL(submit_bh);
2755 void write_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2758 if (!test_clear_buffer_dirty(bh)) {
2762 bh->b_end_io = end_buffer_write_sync;
2764 submit_bh(REQ_OP_WRITE | op_flags, bh);
2766 EXPORT_SYMBOL(write_dirty_buffer);
2769 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2770 * and then start new I/O and then wait upon it. The caller must have a ref on
2773 int __sync_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2775 WARN_ON(atomic_read(&bh->b_count) < 1);
2777 if (test_clear_buffer_dirty(bh)) {
2779 * The bh should be mapped, but it might not be if the
2780 * device was hot-removed. Not much we can do but fail the I/O.
2782 if (!buffer_mapped(bh)) {
2788 bh->b_end_io = end_buffer_write_sync;
2789 submit_bh(REQ_OP_WRITE | op_flags, bh);
2791 if (!buffer_uptodate(bh))
2798 EXPORT_SYMBOL(__sync_dirty_buffer);
2800 int sync_dirty_buffer(struct buffer_head *bh)
2802 return __sync_dirty_buffer(bh, REQ_SYNC);
2804 EXPORT_SYMBOL(sync_dirty_buffer);
2807 * try_to_free_buffers() checks if all the buffers on this particular folio
2808 * are unused, and releases them if so.
2810 * Exclusion against try_to_free_buffers may be obtained by either
2811 * locking the folio or by holding its mapping's private_lock.
2813 * If the folio is dirty but all the buffers are clean then we need to
2814 * be sure to mark the folio clean as well. This is because the folio
2815 * may be against a block device, and a later reattachment of buffers
2816 * to a dirty folio will set *all* buffers dirty. Which would corrupt
2817 * filesystem data on the same device.
2819 * The same applies to regular filesystem folios: if all the buffers are
2820 * clean then we set the folio clean and proceed. To do that, we require
2821 * total exclusion from block_dirty_folio(). That is obtained with
2824 * try_to_free_buffers() is non-blocking.
2826 static inline int buffer_busy(struct buffer_head *bh)
2828 return atomic_read(&bh->b_count) |
2829 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2833 drop_buffers(struct folio *folio, struct buffer_head **buffers_to_free)
2835 struct buffer_head *head = folio_buffers(folio);
2836 struct buffer_head *bh;
2840 if (buffer_busy(bh))
2842 bh = bh->b_this_page;
2843 } while (bh != head);
2846 struct buffer_head *next = bh->b_this_page;
2848 if (bh->b_assoc_map)
2849 __remove_assoc_queue(bh);
2851 } while (bh != head);
2852 *buffers_to_free = head;
2853 folio_detach_private(folio);
2859 bool try_to_free_buffers(struct folio *folio)
2861 struct address_space * const mapping = folio->mapping;
2862 struct buffer_head *buffers_to_free = NULL;
2865 BUG_ON(!folio_test_locked(folio));
2866 if (folio_test_writeback(folio))
2869 if (mapping == NULL) { /* can this still happen? */
2870 ret = drop_buffers(folio, &buffers_to_free);
2874 spin_lock(&mapping->private_lock);
2875 ret = drop_buffers(folio, &buffers_to_free);
2878 * If the filesystem writes its buffers by hand (eg ext3)
2879 * then we can have clean buffers against a dirty folio. We
2880 * clean the folio here; otherwise the VM will never notice
2881 * that the filesystem did any IO at all.
2883 * Also, during truncate, discard_buffer will have marked all
2884 * the folio's buffers clean. We discover that here and clean
2887 * private_lock must be held over this entire operation in order
2888 * to synchronise against block_dirty_folio and prevent the
2889 * dirty bit from being lost.
2892 folio_cancel_dirty(folio);
2893 spin_unlock(&mapping->private_lock);
2895 if (buffers_to_free) {
2896 struct buffer_head *bh = buffers_to_free;
2899 struct buffer_head *next = bh->b_this_page;
2900 free_buffer_head(bh);
2902 } while (bh != buffers_to_free);
2906 EXPORT_SYMBOL(try_to_free_buffers);
2909 * Buffer-head allocation
2911 static struct kmem_cache *bh_cachep __read_mostly;
2914 * Once the number of bh's in the machine exceeds this level, we start
2915 * stripping them in writeback.
2917 static unsigned long max_buffer_heads;
2919 int buffer_heads_over_limit;
2921 struct bh_accounting {
2922 int nr; /* Number of live bh's */
2923 int ratelimit; /* Limit cacheline bouncing */
2926 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2928 static void recalc_bh_state(void)
2933 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
2935 __this_cpu_write(bh_accounting.ratelimit, 0);
2936 for_each_online_cpu(i)
2937 tot += per_cpu(bh_accounting, i).nr;
2938 buffer_heads_over_limit = (tot > max_buffer_heads);
2941 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
2943 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
2945 INIT_LIST_HEAD(&ret->b_assoc_buffers);
2946 spin_lock_init(&ret->b_uptodate_lock);
2948 __this_cpu_inc(bh_accounting.nr);
2954 EXPORT_SYMBOL(alloc_buffer_head);
2956 void free_buffer_head(struct buffer_head *bh)
2958 BUG_ON(!list_empty(&bh->b_assoc_buffers));
2959 kmem_cache_free(bh_cachep, bh);
2961 __this_cpu_dec(bh_accounting.nr);
2965 EXPORT_SYMBOL(free_buffer_head);
2967 static int buffer_exit_cpu_dead(unsigned int cpu)
2970 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
2972 for (i = 0; i < BH_LRU_SIZE; i++) {
2976 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
2977 per_cpu(bh_accounting, cpu).nr = 0;
2982 * bh_uptodate_or_lock - Test whether the buffer is uptodate
2983 * @bh: struct buffer_head
2985 * Return true if the buffer is up-to-date and false,
2986 * with the buffer locked, if not.
2988 int bh_uptodate_or_lock(struct buffer_head *bh)
2990 if (!buffer_uptodate(bh)) {
2992 if (!buffer_uptodate(bh))
2998 EXPORT_SYMBOL(bh_uptodate_or_lock);
3001 * __bh_read - Submit read for a locked buffer
3002 * @bh: struct buffer_head
3003 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3004 * @wait: wait until reading finish
3006 * Returns zero on success or don't wait, and -EIO on error.
3008 int __bh_read(struct buffer_head *bh, blk_opf_t op_flags, bool wait)
3012 BUG_ON(!buffer_locked(bh));
3015 bh->b_end_io = end_buffer_read_sync;
3016 submit_bh(REQ_OP_READ | op_flags, bh);
3019 if (!buffer_uptodate(bh))
3024 EXPORT_SYMBOL(__bh_read);
3027 * __bh_read_batch - Submit read for a batch of unlocked buffers
3028 * @nr: entry number of the buffer batch
3029 * @bhs: a batch of struct buffer_head
3030 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3031 * @force_lock: force to get a lock on the buffer if set, otherwise drops any
3032 * buffer that cannot lock.
3034 * Returns zero on success or don't wait, and -EIO on error.
3036 void __bh_read_batch(int nr, struct buffer_head *bhs[],
3037 blk_opf_t op_flags, bool force_lock)
3041 for (i = 0; i < nr; i++) {
3042 struct buffer_head *bh = bhs[i];
3044 if (buffer_uptodate(bh))
3050 if (!trylock_buffer(bh))
3053 if (buffer_uptodate(bh)) {
3058 bh->b_end_io = end_buffer_read_sync;
3060 submit_bh(REQ_OP_READ | op_flags, bh);
3063 EXPORT_SYMBOL(__bh_read_batch);
3065 void __init buffer_init(void)
3067 unsigned long nrpages;
3070 bh_cachep = kmem_cache_create("buffer_head",
3071 sizeof(struct buffer_head), 0,
3072 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3077 * Limit the bh occupancy to 10% of ZONE_NORMAL
3079 nrpages = (nr_free_buffer_pages() * 10) / 100;
3080 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3081 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3082 NULL, buffer_exit_cpu_dead);