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(page_folio(bh->b_page), bh->b_size,
313 end_buffer_async_read(bh, valid);
317 static bool need_fsverity(struct buffer_head *bh)
319 struct page *page = bh->b_page;
320 struct inode *inode = page->mapping->host;
322 return fsverity_active(inode) &&
324 page->index < DIV_ROUND_UP(inode->i_size, PAGE_SIZE);
327 static void decrypt_bh(struct work_struct *work)
329 struct postprocess_bh_ctx *ctx =
330 container_of(work, struct postprocess_bh_ctx, work);
331 struct buffer_head *bh = ctx->bh;
334 err = fscrypt_decrypt_pagecache_blocks(page_folio(bh->b_page),
335 bh->b_size, bh_offset(bh));
336 if (err == 0 && need_fsverity(bh)) {
338 * We use different work queues for decryption and for verity
339 * because verity may require reading metadata pages that need
340 * decryption, and we shouldn't recurse to the same workqueue.
342 INIT_WORK(&ctx->work, verify_bh);
343 fsverity_enqueue_verify_work(&ctx->work);
346 end_buffer_async_read(bh, err == 0);
351 * I/O completion handler for block_read_full_folio() - pages
352 * which come unlocked at the end of I/O.
354 static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
356 struct inode *inode = bh->b_folio->mapping->host;
357 bool decrypt = fscrypt_inode_uses_fs_layer_crypto(inode);
358 bool verify = need_fsverity(bh);
360 /* Decrypt (with fscrypt) and/or verify (with fsverity) if needed. */
361 if (uptodate && (decrypt || verify)) {
362 struct postprocess_bh_ctx *ctx =
363 kmalloc(sizeof(*ctx), GFP_ATOMIC);
368 INIT_WORK(&ctx->work, decrypt_bh);
369 fscrypt_enqueue_decrypt_work(&ctx->work);
371 INIT_WORK(&ctx->work, verify_bh);
372 fsverity_enqueue_verify_work(&ctx->work);
378 end_buffer_async_read(bh, uptodate);
382 * Completion handler for block_write_full_page() - pages which are unlocked
383 * during I/O, and which have PageWriteback cleared upon I/O completion.
385 void end_buffer_async_write(struct buffer_head *bh, int uptodate)
388 struct buffer_head *first;
389 struct buffer_head *tmp;
392 BUG_ON(!buffer_async_write(bh));
396 set_buffer_uptodate(bh);
398 buffer_io_error(bh, ", lost async page write");
399 mark_buffer_write_io_error(bh);
400 clear_buffer_uptodate(bh);
401 folio_set_error(folio);
404 first = folio_buffers(folio);
405 spin_lock_irqsave(&first->b_uptodate_lock, flags);
407 clear_buffer_async_write(bh);
409 tmp = bh->b_this_page;
411 if (buffer_async_write(tmp)) {
412 BUG_ON(!buffer_locked(tmp));
415 tmp = tmp->b_this_page;
417 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
418 folio_end_writeback(folio);
422 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
425 EXPORT_SYMBOL(end_buffer_async_write);
428 * If a page's buffers are under async readin (end_buffer_async_read
429 * completion) then there is a possibility that another thread of
430 * control could lock one of the buffers after it has completed
431 * but while some of the other buffers have not completed. This
432 * locked buffer would confuse end_buffer_async_read() into not unlocking
433 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
434 * that this buffer is not under async I/O.
436 * The page comes unlocked when it has no locked buffer_async buffers
439 * PageLocked prevents anyone starting new async I/O reads any of
442 * PageWriteback is used to prevent simultaneous writeout of the same
445 * PageLocked prevents anyone from starting writeback of a page which is
446 * under read I/O (PageWriteback is only ever set against a locked page).
448 static void mark_buffer_async_read(struct buffer_head *bh)
450 bh->b_end_io = end_buffer_async_read_io;
451 set_buffer_async_read(bh);
454 static void mark_buffer_async_write_endio(struct buffer_head *bh,
455 bh_end_io_t *handler)
457 bh->b_end_io = handler;
458 set_buffer_async_write(bh);
461 void mark_buffer_async_write(struct buffer_head *bh)
463 mark_buffer_async_write_endio(bh, end_buffer_async_write);
465 EXPORT_SYMBOL(mark_buffer_async_write);
469 * fs/buffer.c contains helper functions for buffer-backed address space's
470 * fsync functions. A common requirement for buffer-based filesystems is
471 * that certain data from the backing blockdev needs to be written out for
472 * a successful fsync(). For example, ext2 indirect blocks need to be
473 * written back and waited upon before fsync() returns.
475 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
476 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
477 * management of a list of dependent buffers at ->i_mapping->private_list.
479 * Locking is a little subtle: try_to_free_buffers() will remove buffers
480 * from their controlling inode's queue when they are being freed. But
481 * try_to_free_buffers() will be operating against the *blockdev* mapping
482 * at the time, not against the S_ISREG file which depends on those buffers.
483 * So the locking for private_list is via the private_lock in the address_space
484 * which backs the buffers. Which is different from the address_space
485 * against which the buffers are listed. So for a particular address_space,
486 * mapping->private_lock does *not* protect mapping->private_list! In fact,
487 * mapping->private_list will always be protected by the backing blockdev's
490 * Which introduces a requirement: all buffers on an address_space's
491 * ->private_list must be from the same address_space: the blockdev's.
493 * address_spaces which do not place buffers at ->private_list via these
494 * utility functions are free to use private_lock and private_list for
495 * whatever they want. The only requirement is that list_empty(private_list)
496 * be true at clear_inode() time.
498 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
499 * filesystems should do that. invalidate_inode_buffers() should just go
500 * BUG_ON(!list_empty).
502 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
503 * take an address_space, not an inode. And it should be called
504 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
507 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
508 * list if it is already on a list. Because if the buffer is on a list,
509 * it *must* already be on the right one. If not, the filesystem is being
510 * silly. This will save a ton of locking. But first we have to ensure
511 * that buffers are taken *off* the old inode's list when they are freed
512 * (presumably in truncate). That requires careful auditing of all
513 * filesystems (do it inside bforget()). It could also be done by bringing
518 * The buffer's backing address_space's private_lock must be held
520 static void __remove_assoc_queue(struct buffer_head *bh)
522 list_del_init(&bh->b_assoc_buffers);
523 WARN_ON(!bh->b_assoc_map);
524 bh->b_assoc_map = NULL;
527 int inode_has_buffers(struct inode *inode)
529 return !list_empty(&inode->i_data.private_list);
533 * osync is designed to support O_SYNC io. It waits synchronously for
534 * all already-submitted IO to complete, but does not queue any new
535 * writes to the disk.
537 * To do O_SYNC writes, just queue the buffer writes with write_dirty_buffer
538 * as you dirty the buffers, and then use osync_inode_buffers to wait for
539 * completion. Any other dirty buffers which are not yet queued for
540 * write will not be flushed to disk by the osync.
542 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
544 struct buffer_head *bh;
550 list_for_each_prev(p, list) {
552 if (buffer_locked(bh)) {
556 if (!buffer_uptodate(bh))
567 void emergency_thaw_bdev(struct super_block *sb)
569 while (sb->s_bdev && !thaw_bdev(sb->s_bdev))
570 printk(KERN_WARNING "Emergency Thaw on %pg\n", sb->s_bdev);
574 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
575 * @mapping: the mapping which wants those buffers written
577 * Starts I/O against the buffers at mapping->private_list, and waits upon
580 * Basically, this is a convenience function for fsync().
581 * @mapping is a file or directory which needs those buffers to be written for
582 * a successful fsync().
584 int sync_mapping_buffers(struct address_space *mapping)
586 struct address_space *buffer_mapping = mapping->private_data;
588 if (buffer_mapping == NULL || list_empty(&mapping->private_list))
591 return fsync_buffers_list(&buffer_mapping->private_lock,
592 &mapping->private_list);
594 EXPORT_SYMBOL(sync_mapping_buffers);
597 * Called when we've recently written block `bblock', and it is known that
598 * `bblock' was for a buffer_boundary() buffer. This means that the block at
599 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
600 * dirty, schedule it for IO. So that indirects merge nicely with their data.
602 void write_boundary_block(struct block_device *bdev,
603 sector_t bblock, unsigned blocksize)
605 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
607 if (buffer_dirty(bh))
608 write_dirty_buffer(bh, 0);
613 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
615 struct address_space *mapping = inode->i_mapping;
616 struct address_space *buffer_mapping = bh->b_folio->mapping;
618 mark_buffer_dirty(bh);
619 if (!mapping->private_data) {
620 mapping->private_data = buffer_mapping;
622 BUG_ON(mapping->private_data != buffer_mapping);
624 if (!bh->b_assoc_map) {
625 spin_lock(&buffer_mapping->private_lock);
626 list_move_tail(&bh->b_assoc_buffers,
627 &mapping->private_list);
628 bh->b_assoc_map = mapping;
629 spin_unlock(&buffer_mapping->private_lock);
632 EXPORT_SYMBOL(mark_buffer_dirty_inode);
635 * Add a page to the dirty page list.
637 * It is a sad fact of life that this function is called from several places
638 * deeply under spinlocking. It may not sleep.
640 * If the page has buffers, the uptodate buffers are set dirty, to preserve
641 * dirty-state coherency between the page and the buffers. It the page does
642 * not have buffers then when they are later attached they will all be set
645 * The buffers are dirtied before the page is dirtied. There's a small race
646 * window in which a writepage caller may see the page cleanness but not the
647 * buffer dirtiness. That's fine. If this code were to set the page dirty
648 * before the buffers, a concurrent writepage caller could clear the page dirty
649 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
650 * page on the dirty page list.
652 * We use private_lock to lock against try_to_free_buffers while using the
653 * page's buffer list. Also use this to protect against clean buffers being
654 * added to the page after it was set dirty.
656 * FIXME: may need to call ->reservepage here as well. That's rather up to the
657 * address_space though.
659 bool block_dirty_folio(struct address_space *mapping, struct folio *folio)
661 struct buffer_head *head;
664 spin_lock(&mapping->private_lock);
665 head = folio_buffers(folio);
667 struct buffer_head *bh = head;
670 set_buffer_dirty(bh);
671 bh = bh->b_this_page;
672 } while (bh != head);
675 * Lock out page's memcg migration to keep PageDirty
676 * synchronized with per-memcg dirty page counters.
678 folio_memcg_lock(folio);
679 newly_dirty = !folio_test_set_dirty(folio);
680 spin_unlock(&mapping->private_lock);
683 __folio_mark_dirty(folio, mapping, 1);
685 folio_memcg_unlock(folio);
688 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
692 EXPORT_SYMBOL(block_dirty_folio);
695 * Write out and wait upon a list of buffers.
697 * We have conflicting pressures: we want to make sure that all
698 * initially dirty buffers get waited on, but that any subsequently
699 * dirtied buffers don't. After all, we don't want fsync to last
700 * forever if somebody is actively writing to the file.
702 * Do this in two main stages: first we copy dirty buffers to a
703 * temporary inode list, queueing the writes as we go. Then we clean
704 * up, waiting for those writes to complete.
706 * During this second stage, any subsequent updates to the file may end
707 * up refiling the buffer on the original inode's dirty list again, so
708 * there is a chance we will end up with a buffer queued for write but
709 * not yet completed on that list. So, as a final cleanup we go through
710 * the osync code to catch these locked, dirty buffers without requeuing
711 * any newly dirty buffers for write.
713 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
715 struct buffer_head *bh;
716 struct list_head tmp;
717 struct address_space *mapping;
719 struct blk_plug plug;
721 INIT_LIST_HEAD(&tmp);
722 blk_start_plug(&plug);
725 while (!list_empty(list)) {
726 bh = BH_ENTRY(list->next);
727 mapping = bh->b_assoc_map;
728 __remove_assoc_queue(bh);
729 /* Avoid race with mark_buffer_dirty_inode() which does
730 * a lockless check and we rely on seeing the dirty bit */
732 if (buffer_dirty(bh) || buffer_locked(bh)) {
733 list_add(&bh->b_assoc_buffers, &tmp);
734 bh->b_assoc_map = mapping;
735 if (buffer_dirty(bh)) {
739 * Ensure any pending I/O completes so that
740 * write_dirty_buffer() actually writes the
741 * current contents - it is a noop if I/O is
742 * still in flight on potentially older
745 write_dirty_buffer(bh, REQ_SYNC);
748 * Kick off IO for the previous mapping. Note
749 * that we will not run the very last mapping,
750 * wait_on_buffer() will do that for us
751 * through sync_buffer().
760 blk_finish_plug(&plug);
763 while (!list_empty(&tmp)) {
764 bh = BH_ENTRY(tmp.prev);
766 mapping = bh->b_assoc_map;
767 __remove_assoc_queue(bh);
768 /* Avoid race with mark_buffer_dirty_inode() which does
769 * a lockless check and we rely on seeing the dirty bit */
771 if (buffer_dirty(bh)) {
772 list_add(&bh->b_assoc_buffers,
773 &mapping->private_list);
774 bh->b_assoc_map = mapping;
778 if (!buffer_uptodate(bh))
785 err2 = osync_buffers_list(lock, list);
793 * Invalidate any and all dirty buffers on a given inode. We are
794 * probably unmounting the fs, but that doesn't mean we have already
795 * done a sync(). Just drop the buffers from the inode list.
797 * NOTE: we take the inode's blockdev's mapping's private_lock. Which
798 * assumes that all the buffers are against the blockdev. Not true
801 void invalidate_inode_buffers(struct inode *inode)
803 if (inode_has_buffers(inode)) {
804 struct address_space *mapping = &inode->i_data;
805 struct list_head *list = &mapping->private_list;
806 struct address_space *buffer_mapping = mapping->private_data;
808 spin_lock(&buffer_mapping->private_lock);
809 while (!list_empty(list))
810 __remove_assoc_queue(BH_ENTRY(list->next));
811 spin_unlock(&buffer_mapping->private_lock);
814 EXPORT_SYMBOL(invalidate_inode_buffers);
817 * Remove any clean buffers from the inode's buffer list. This is called
818 * when we're trying to free the inode itself. Those buffers can pin it.
820 * Returns true if all buffers were removed.
822 int remove_inode_buffers(struct inode *inode)
826 if (inode_has_buffers(inode)) {
827 struct address_space *mapping = &inode->i_data;
828 struct list_head *list = &mapping->private_list;
829 struct address_space *buffer_mapping = mapping->private_data;
831 spin_lock(&buffer_mapping->private_lock);
832 while (!list_empty(list)) {
833 struct buffer_head *bh = BH_ENTRY(list->next);
834 if (buffer_dirty(bh)) {
838 __remove_assoc_queue(bh);
840 spin_unlock(&buffer_mapping->private_lock);
846 * Create the appropriate buffers when given a folio for data area and
847 * the size of each buffer.. Use the bh->b_this_page linked list to
848 * follow the buffers created. Return NULL if unable to create more
851 * The retry flag is used to differentiate async IO (paging, swapping)
852 * which may not fail from ordinary buffer allocations.
854 struct buffer_head *folio_alloc_buffers(struct folio *folio, unsigned long size,
857 struct buffer_head *bh, *head;
858 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
860 struct mem_cgroup *memcg, *old_memcg;
865 /* The folio lock pins the memcg */
866 memcg = folio_memcg(folio);
867 old_memcg = set_active_memcg(memcg);
870 offset = folio_size(folio);
871 while ((offset -= size) >= 0) {
872 bh = alloc_buffer_head(gfp);
876 bh->b_this_page = head;
882 /* Link the buffer to its folio */
883 folio_set_bh(bh, folio, offset);
886 set_active_memcg(old_memcg);
889 * In case anything failed, we just free everything we got.
895 head = head->b_this_page;
896 free_buffer_head(bh);
902 EXPORT_SYMBOL_GPL(folio_alloc_buffers);
904 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
907 return folio_alloc_buffers(page_folio(page), size, retry);
909 EXPORT_SYMBOL_GPL(alloc_page_buffers);
912 link_dev_buffers(struct page *page, struct buffer_head *head)
914 struct buffer_head *bh, *tail;
919 bh = bh->b_this_page;
921 tail->b_this_page = head;
922 attach_page_private(page, head);
925 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
927 sector_t retval = ~((sector_t)0);
928 loff_t sz = bdev_nr_bytes(bdev);
931 unsigned int sizebits = blksize_bits(size);
932 retval = (sz >> sizebits);
938 * Initialise the state of a blockdev page's buffers.
941 init_page_buffers(struct page *page, struct block_device *bdev,
942 sector_t block, int size)
944 struct buffer_head *head = page_buffers(page);
945 struct buffer_head *bh = head;
946 int uptodate = PageUptodate(page);
947 sector_t end_block = blkdev_max_block(bdev, size);
950 if (!buffer_mapped(bh)) {
952 bh->b_private = NULL;
954 bh->b_blocknr = block;
956 set_buffer_uptodate(bh);
957 if (block < end_block)
958 set_buffer_mapped(bh);
961 bh = bh->b_this_page;
962 } while (bh != head);
965 * Caller needs to validate requested block against end of device.
971 * Create the page-cache page that contains the requested block.
973 * This is used purely for blockdev mappings.
976 grow_dev_page(struct block_device *bdev, sector_t block,
977 pgoff_t index, int size, int sizebits, gfp_t gfp)
979 struct inode *inode = bdev->bd_inode;
981 struct buffer_head *bh;
986 gfp_mask = mapping_gfp_constraint(inode->i_mapping, ~__GFP_FS) | gfp;
989 * XXX: __getblk_slow() can not really deal with failure and
990 * will endlessly loop on improvised global reclaim. Prefer
991 * looping in the allocator rather than here, at least that
992 * code knows what it's doing.
994 gfp_mask |= __GFP_NOFAIL;
996 page = find_or_create_page(inode->i_mapping, index, gfp_mask);
998 BUG_ON(!PageLocked(page));
1000 if (page_has_buffers(page)) {
1001 bh = page_buffers(page);
1002 if (bh->b_size == size) {
1003 end_block = init_page_buffers(page, bdev,
1004 (sector_t)index << sizebits,
1008 if (!try_to_free_buffers(page_folio(page)))
1013 * Allocate some buffers for this page
1015 bh = alloc_page_buffers(page, size, true);
1018 * Link the page to the buffers and initialise them. Take the
1019 * lock to be atomic wrt __find_get_block(), which does not
1020 * run under the page lock.
1022 spin_lock(&inode->i_mapping->private_lock);
1023 link_dev_buffers(page, bh);
1024 end_block = init_page_buffers(page, bdev, (sector_t)index << sizebits,
1026 spin_unlock(&inode->i_mapping->private_lock);
1028 ret = (block < end_block) ? 1 : -ENXIO;
1036 * Create buffers for the specified block device block's page. If
1037 * that page was dirty, the buffers are set dirty also.
1040 grow_buffers(struct block_device *bdev, sector_t block, int size, gfp_t gfp)
1045 sizebits = PAGE_SHIFT - __ffs(size);
1046 index = block >> sizebits;
1049 * Check for a block which wants to lie outside our maximum possible
1050 * pagecache index. (this comparison is done using sector_t types).
1052 if (unlikely(index != block >> sizebits)) {
1053 printk(KERN_ERR "%s: requested out-of-range block %llu for "
1055 __func__, (unsigned long long)block,
1060 /* Create a page with the proper size buffers.. */
1061 return grow_dev_page(bdev, block, index, size, sizebits, gfp);
1064 static struct buffer_head *
1065 __getblk_slow(struct block_device *bdev, sector_t block,
1066 unsigned size, gfp_t gfp)
1068 /* Size must be multiple of hard sectorsize */
1069 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1070 (size < 512 || size > PAGE_SIZE))) {
1071 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1073 printk(KERN_ERR "logical block size: %d\n",
1074 bdev_logical_block_size(bdev));
1081 struct buffer_head *bh;
1084 bh = __find_get_block(bdev, block, size);
1088 ret = grow_buffers(bdev, block, size, gfp);
1095 * The relationship between dirty buffers and dirty pages:
1097 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1098 * the page is tagged dirty in the page cache.
1100 * At all times, the dirtiness of the buffers represents the dirtiness of
1101 * subsections of the page. If the page has buffers, the page dirty bit is
1102 * merely a hint about the true dirty state.
1104 * When a page is set dirty in its entirety, all its buffers are marked dirty
1105 * (if the page has buffers).
1107 * When a buffer is marked dirty, its page is dirtied, but the page's other
1110 * Also. When blockdev buffers are explicitly read with bread(), they
1111 * individually become uptodate. But their backing page remains not
1112 * uptodate - even if all of its buffers are uptodate. A subsequent
1113 * block_read_full_folio() against that folio will discover all the uptodate
1114 * buffers, will set the folio uptodate and will perform no I/O.
1118 * mark_buffer_dirty - mark a buffer_head as needing writeout
1119 * @bh: the buffer_head to mark dirty
1121 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1122 * its backing page dirty, then tag the page as dirty in the page cache
1123 * and then attach the address_space's inode to its superblock's dirty
1126 * mark_buffer_dirty() is atomic. It takes bh->b_folio->mapping->private_lock,
1127 * i_pages lock and mapping->host->i_lock.
1129 void mark_buffer_dirty(struct buffer_head *bh)
1131 WARN_ON_ONCE(!buffer_uptodate(bh));
1133 trace_block_dirty_buffer(bh);
1136 * Very *carefully* optimize the it-is-already-dirty case.
1138 * Don't let the final "is it dirty" escape to before we
1139 * perhaps modified the buffer.
1141 if (buffer_dirty(bh)) {
1143 if (buffer_dirty(bh))
1147 if (!test_set_buffer_dirty(bh)) {
1148 struct folio *folio = bh->b_folio;
1149 struct address_space *mapping = NULL;
1151 folio_memcg_lock(folio);
1152 if (!folio_test_set_dirty(folio)) {
1153 mapping = folio->mapping;
1155 __folio_mark_dirty(folio, mapping, 0);
1157 folio_memcg_unlock(folio);
1159 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1162 EXPORT_SYMBOL(mark_buffer_dirty);
1164 void mark_buffer_write_io_error(struct buffer_head *bh)
1166 struct super_block *sb;
1168 set_buffer_write_io_error(bh);
1169 /* FIXME: do we need to set this in both places? */
1170 if (bh->b_folio && bh->b_folio->mapping)
1171 mapping_set_error(bh->b_folio->mapping, -EIO);
1172 if (bh->b_assoc_map)
1173 mapping_set_error(bh->b_assoc_map, -EIO);
1175 sb = READ_ONCE(bh->b_bdev->bd_super);
1177 errseq_set(&sb->s_wb_err, -EIO);
1180 EXPORT_SYMBOL(mark_buffer_write_io_error);
1183 * Decrement a buffer_head's reference count. If all buffers against a page
1184 * have zero reference count, are clean and unlocked, and if the page is clean
1185 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1186 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1187 * a page but it ends up not being freed, and buffers may later be reattached).
1189 void __brelse(struct buffer_head * buf)
1191 if (atomic_read(&buf->b_count)) {
1195 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1197 EXPORT_SYMBOL(__brelse);
1200 * bforget() is like brelse(), except it discards any
1201 * potentially dirty data.
1203 void __bforget(struct buffer_head *bh)
1205 clear_buffer_dirty(bh);
1206 if (bh->b_assoc_map) {
1207 struct address_space *buffer_mapping = bh->b_folio->mapping;
1209 spin_lock(&buffer_mapping->private_lock);
1210 list_del_init(&bh->b_assoc_buffers);
1211 bh->b_assoc_map = NULL;
1212 spin_unlock(&buffer_mapping->private_lock);
1216 EXPORT_SYMBOL(__bforget);
1218 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1221 if (buffer_uptodate(bh)) {
1226 bh->b_end_io = end_buffer_read_sync;
1227 submit_bh(REQ_OP_READ, bh);
1229 if (buffer_uptodate(bh))
1237 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1238 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1239 * refcount elevated by one when they're in an LRU. A buffer can only appear
1240 * once in a particular CPU's LRU. A single buffer can be present in multiple
1241 * CPU's LRUs at the same time.
1243 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1244 * sb_find_get_block().
1246 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1247 * a local interrupt disable for that.
1250 #define BH_LRU_SIZE 16
1253 struct buffer_head *bhs[BH_LRU_SIZE];
1256 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1259 #define bh_lru_lock() local_irq_disable()
1260 #define bh_lru_unlock() local_irq_enable()
1262 #define bh_lru_lock() preempt_disable()
1263 #define bh_lru_unlock() preempt_enable()
1266 static inline void check_irqs_on(void)
1268 #ifdef irqs_disabled
1269 BUG_ON(irqs_disabled());
1274 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1275 * inserted at the front, and the buffer_head at the back if any is evicted.
1276 * Or, if already in the LRU it is moved to the front.
1278 static void bh_lru_install(struct buffer_head *bh)
1280 struct buffer_head *evictee = bh;
1288 * the refcount of buffer_head in bh_lru prevents dropping the
1289 * attached page(i.e., try_to_free_buffers) so it could cause
1290 * failing page migration.
1291 * Skip putting upcoming bh into bh_lru until migration is done.
1293 if (lru_cache_disabled()) {
1298 b = this_cpu_ptr(&bh_lrus);
1299 for (i = 0; i < BH_LRU_SIZE; i++) {
1300 swap(evictee, b->bhs[i]);
1301 if (evictee == bh) {
1313 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1315 static struct buffer_head *
1316 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1318 struct buffer_head *ret = NULL;
1323 for (i = 0; i < BH_LRU_SIZE; i++) {
1324 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1326 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1327 bh->b_size == size) {
1330 __this_cpu_write(bh_lrus.bhs[i],
1331 __this_cpu_read(bh_lrus.bhs[i - 1]));
1334 __this_cpu_write(bh_lrus.bhs[0], bh);
1346 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1347 * it in the LRU and mark it as accessed. If it is not present then return
1350 struct buffer_head *
1351 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1353 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1356 /* __find_get_block_slow will mark the page accessed */
1357 bh = __find_get_block_slow(bdev, block);
1365 EXPORT_SYMBOL(__find_get_block);
1368 * __getblk_gfp() will locate (and, if necessary, create) the buffer_head
1369 * which corresponds to the passed block_device, block and size. The
1370 * returned buffer has its reference count incremented.
1372 * __getblk_gfp() will lock up the machine if grow_dev_page's
1373 * try_to_free_buffers() attempt is failing. FIXME, perhaps?
1375 struct buffer_head *
1376 __getblk_gfp(struct block_device *bdev, sector_t block,
1377 unsigned size, gfp_t gfp)
1379 struct buffer_head *bh = __find_get_block(bdev, block, size);
1383 bh = __getblk_slow(bdev, block, size, gfp);
1386 EXPORT_SYMBOL(__getblk_gfp);
1389 * Do async read-ahead on a buffer..
1391 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1393 struct buffer_head *bh = __getblk(bdev, block, size);
1395 bh_readahead(bh, REQ_RAHEAD);
1399 EXPORT_SYMBOL(__breadahead);
1402 * __bread_gfp() - reads a specified block and returns the bh
1403 * @bdev: the block_device to read from
1404 * @block: number of block
1405 * @size: size (in bytes) to read
1406 * @gfp: page allocation flag
1408 * Reads a specified block, and returns buffer head that contains it.
1409 * The page cache can be allocated from non-movable area
1410 * not to prevent page migration if you set gfp to zero.
1411 * It returns NULL if the block was unreadable.
1413 struct buffer_head *
1414 __bread_gfp(struct block_device *bdev, sector_t block,
1415 unsigned size, gfp_t gfp)
1417 struct buffer_head *bh = __getblk_gfp(bdev, block, size, gfp);
1419 if (likely(bh) && !buffer_uptodate(bh))
1420 bh = __bread_slow(bh);
1423 EXPORT_SYMBOL(__bread_gfp);
1425 static void __invalidate_bh_lrus(struct bh_lru *b)
1429 for (i = 0; i < BH_LRU_SIZE; i++) {
1435 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1436 * This doesn't race because it runs in each cpu either in irq
1437 * or with preempt disabled.
1439 static void invalidate_bh_lru(void *arg)
1441 struct bh_lru *b = &get_cpu_var(bh_lrus);
1443 __invalidate_bh_lrus(b);
1444 put_cpu_var(bh_lrus);
1447 bool has_bh_in_lru(int cpu, void *dummy)
1449 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1452 for (i = 0; i < BH_LRU_SIZE; i++) {
1460 void invalidate_bh_lrus(void)
1462 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1464 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1467 * It's called from workqueue context so we need a bh_lru_lock to close
1468 * the race with preemption/irq.
1470 void invalidate_bh_lrus_cpu(void)
1475 b = this_cpu_ptr(&bh_lrus);
1476 __invalidate_bh_lrus(b);
1480 void set_bh_page(struct buffer_head *bh,
1481 struct page *page, unsigned long offset)
1484 BUG_ON(offset >= PAGE_SIZE);
1485 if (PageHighMem(page))
1487 * This catches illegal uses and preserves the offset:
1489 bh->b_data = (char *)(0 + offset);
1491 bh->b_data = page_address(page) + offset;
1493 EXPORT_SYMBOL(set_bh_page);
1495 void folio_set_bh(struct buffer_head *bh, struct folio *folio,
1496 unsigned long offset)
1498 bh->b_folio = folio;
1499 BUG_ON(offset >= folio_size(folio));
1500 if (folio_test_highmem(folio))
1502 * This catches illegal uses and preserves the offset:
1504 bh->b_data = (char *)(0 + offset);
1506 bh->b_data = folio_address(folio) + offset;
1508 EXPORT_SYMBOL(folio_set_bh);
1511 * Called when truncating a buffer on a page completely.
1514 /* Bits that are cleared during an invalidate */
1515 #define BUFFER_FLAGS_DISCARD \
1516 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1517 1 << BH_Delay | 1 << BH_Unwritten)
1519 static void discard_buffer(struct buffer_head * bh)
1521 unsigned long b_state;
1524 clear_buffer_dirty(bh);
1526 b_state = READ_ONCE(bh->b_state);
1528 } while (!try_cmpxchg(&bh->b_state, &b_state,
1529 b_state & ~BUFFER_FLAGS_DISCARD));
1534 * block_invalidate_folio - Invalidate part or all of a buffer-backed folio.
1535 * @folio: The folio which is affected.
1536 * @offset: start of the range to invalidate
1537 * @length: length of the range to invalidate
1539 * block_invalidate_folio() is called when all or part of the folio has been
1540 * invalidated by a truncate operation.
1542 * block_invalidate_folio() does not have to release all buffers, but it must
1543 * ensure that no dirty buffer is left outside @offset and that no I/O
1544 * is underway against any of the blocks which are outside the truncation
1545 * point. Because the caller is about to free (and possibly reuse) those
1548 void block_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1550 struct buffer_head *head, *bh, *next;
1551 size_t curr_off = 0;
1552 size_t stop = length + offset;
1554 BUG_ON(!folio_test_locked(folio));
1557 * Check for overflow
1559 BUG_ON(stop > folio_size(folio) || stop < length);
1561 head = folio_buffers(folio);
1567 size_t next_off = curr_off + bh->b_size;
1568 next = bh->b_this_page;
1571 * Are we still fully in range ?
1573 if (next_off > stop)
1577 * is this block fully invalidated?
1579 if (offset <= curr_off)
1581 curr_off = next_off;
1583 } while (bh != head);
1586 * We release buffers only if the entire folio is being invalidated.
1587 * The get_block cached value has been unconditionally invalidated,
1588 * so real IO is not possible anymore.
1590 if (length == folio_size(folio))
1591 filemap_release_folio(folio, 0);
1595 EXPORT_SYMBOL(block_invalidate_folio);
1598 * We attach and possibly dirty the buffers atomically wrt
1599 * block_dirty_folio() via private_lock. try_to_free_buffers
1600 * is already excluded via the folio lock.
1602 void folio_create_empty_buffers(struct folio *folio, unsigned long blocksize,
1603 unsigned long b_state)
1605 struct buffer_head *bh, *head, *tail;
1607 head = folio_alloc_buffers(folio, blocksize, true);
1610 bh->b_state |= b_state;
1612 bh = bh->b_this_page;
1614 tail->b_this_page = head;
1616 spin_lock(&folio->mapping->private_lock);
1617 if (folio_test_uptodate(folio) || folio_test_dirty(folio)) {
1620 if (folio_test_dirty(folio))
1621 set_buffer_dirty(bh);
1622 if (folio_test_uptodate(folio))
1623 set_buffer_uptodate(bh);
1624 bh = bh->b_this_page;
1625 } while (bh != head);
1627 folio_attach_private(folio, head);
1628 spin_unlock(&folio->mapping->private_lock);
1630 EXPORT_SYMBOL(folio_create_empty_buffers);
1632 void create_empty_buffers(struct page *page,
1633 unsigned long blocksize, unsigned long b_state)
1635 folio_create_empty_buffers(page_folio(page), blocksize, b_state);
1637 EXPORT_SYMBOL(create_empty_buffers);
1640 * clean_bdev_aliases: clean a range of buffers in block device
1641 * @bdev: Block device to clean buffers in
1642 * @block: Start of a range of blocks to clean
1643 * @len: Number of blocks to clean
1645 * We are taking a range of blocks for data and we don't want writeback of any
1646 * buffer-cache aliases starting from return from this function and until the
1647 * moment when something will explicitly mark the buffer dirty (hopefully that
1648 * will not happen until we will free that block ;-) We don't even need to mark
1649 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1650 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1651 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1652 * would confuse anyone who might pick it with bread() afterwards...
1654 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1655 * writeout I/O going on against recently-freed buffers. We don't wait on that
1656 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1657 * need to. That happens here.
1659 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1661 struct inode *bd_inode = bdev->bd_inode;
1662 struct address_space *bd_mapping = bd_inode->i_mapping;
1663 struct folio_batch fbatch;
1664 pgoff_t index = block >> (PAGE_SHIFT - bd_inode->i_blkbits);
1667 struct buffer_head *bh;
1668 struct buffer_head *head;
1670 end = (block + len - 1) >> (PAGE_SHIFT - bd_inode->i_blkbits);
1671 folio_batch_init(&fbatch);
1672 while (filemap_get_folios(bd_mapping, &index, end, &fbatch)) {
1673 count = folio_batch_count(&fbatch);
1674 for (i = 0; i < count; i++) {
1675 struct folio *folio = fbatch.folios[i];
1677 if (!folio_buffers(folio))
1680 * We use folio lock instead of bd_mapping->private_lock
1681 * to pin buffers here since we can afford to sleep and
1682 * it scales better than a global spinlock lock.
1685 /* Recheck when the folio is locked which pins bhs */
1686 head = folio_buffers(folio);
1691 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1693 if (bh->b_blocknr >= block + len)
1695 clear_buffer_dirty(bh);
1697 clear_buffer_req(bh);
1699 bh = bh->b_this_page;
1700 } while (bh != head);
1702 folio_unlock(folio);
1704 folio_batch_release(&fbatch);
1706 /* End of range already reached? */
1707 if (index > end || !index)
1711 EXPORT_SYMBOL(clean_bdev_aliases);
1714 * Size is a power-of-two in the range 512..PAGE_SIZE,
1715 * and the case we care about most is PAGE_SIZE.
1717 * So this *could* possibly be written with those
1718 * constraints in mind (relevant mostly if some
1719 * architecture has a slow bit-scan instruction)
1721 static inline int block_size_bits(unsigned int blocksize)
1723 return ilog2(blocksize);
1726 static struct buffer_head *folio_create_buffers(struct folio *folio,
1727 struct inode *inode,
1728 unsigned int b_state)
1730 BUG_ON(!folio_test_locked(folio));
1732 if (!folio_buffers(folio))
1733 folio_create_empty_buffers(folio,
1734 1 << READ_ONCE(inode->i_blkbits),
1736 return folio_buffers(folio);
1740 * NOTE! All mapped/uptodate combinations are valid:
1742 * Mapped Uptodate Meaning
1744 * No No "unknown" - must do get_block()
1745 * No Yes "hole" - zero-filled
1746 * Yes No "allocated" - allocated on disk, not read in
1747 * Yes Yes "valid" - allocated and up-to-date in memory.
1749 * "Dirty" is valid only with the last case (mapped+uptodate).
1753 * While block_write_full_page is writing back the dirty buffers under
1754 * the page lock, whoever dirtied the buffers may decide to clean them
1755 * again at any time. We handle that by only looking at the buffer
1756 * state inside lock_buffer().
1758 * If block_write_full_page() is called for regular writeback
1759 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1760 * locked buffer. This only can happen if someone has written the buffer
1761 * directly, with submit_bh(). At the address_space level PageWriteback
1762 * prevents this contention from occurring.
1764 * If block_write_full_page() is called with wbc->sync_mode ==
1765 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1766 * causes the writes to be flagged as synchronous writes.
1768 int __block_write_full_page(struct inode *inode, struct page *page,
1769 get_block_t *get_block, struct writeback_control *wbc,
1770 bh_end_io_t *handler)
1774 sector_t last_block;
1775 struct buffer_head *bh, *head;
1776 unsigned int blocksize, bbits;
1777 int nr_underway = 0;
1778 blk_opf_t write_flags = wbc_to_write_flags(wbc);
1780 head = folio_create_buffers(page_folio(page), inode,
1781 (1 << BH_Dirty) | (1 << BH_Uptodate));
1784 * Be very careful. We have no exclusion from block_dirty_folio
1785 * here, and the (potentially unmapped) buffers may become dirty at
1786 * any time. If a buffer becomes dirty here after we've inspected it
1787 * then we just miss that fact, and the page stays dirty.
1789 * Buffers outside i_size may be dirtied by block_dirty_folio;
1790 * handle that here by just cleaning them.
1794 blocksize = bh->b_size;
1795 bbits = block_size_bits(blocksize);
1797 block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1798 last_block = (i_size_read(inode) - 1) >> bbits;
1801 * Get all the dirty buffers mapped to disk addresses and
1802 * handle any aliases from the underlying blockdev's mapping.
1805 if (block > last_block) {
1807 * mapped buffers outside i_size will occur, because
1808 * this page can be outside i_size when there is a
1809 * truncate in progress.
1812 * The buffer was zeroed by block_write_full_page()
1814 clear_buffer_dirty(bh);
1815 set_buffer_uptodate(bh);
1816 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1818 WARN_ON(bh->b_size != blocksize);
1819 err = get_block(inode, block, bh, 1);
1822 clear_buffer_delay(bh);
1823 if (buffer_new(bh)) {
1824 /* blockdev mappings never come here */
1825 clear_buffer_new(bh);
1826 clean_bdev_bh_alias(bh);
1829 bh = bh->b_this_page;
1831 } while (bh != head);
1834 if (!buffer_mapped(bh))
1837 * If it's a fully non-blocking write attempt and we cannot
1838 * lock the buffer then redirty the page. Note that this can
1839 * potentially cause a busy-wait loop from writeback threads
1840 * and kswapd activity, but those code paths have their own
1841 * higher-level throttling.
1843 if (wbc->sync_mode != WB_SYNC_NONE) {
1845 } else if (!trylock_buffer(bh)) {
1846 redirty_page_for_writepage(wbc, page);
1849 if (test_clear_buffer_dirty(bh)) {
1850 mark_buffer_async_write_endio(bh, handler);
1854 } while ((bh = bh->b_this_page) != head);
1857 * The page and its buffers are protected by PageWriteback(), so we can
1858 * drop the bh refcounts early.
1860 BUG_ON(PageWriteback(page));
1861 set_page_writeback(page);
1864 struct buffer_head *next = bh->b_this_page;
1865 if (buffer_async_write(bh)) {
1866 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh, wbc);
1870 } while (bh != head);
1875 if (nr_underway == 0) {
1877 * The page was marked dirty, but the buffers were
1878 * clean. Someone wrote them back by hand with
1879 * write_dirty_buffer/submit_bh. A rare case.
1881 end_page_writeback(page);
1884 * The page and buffer_heads can be released at any time from
1892 * ENOSPC, or some other error. We may already have added some
1893 * blocks to the file, so we need to write these out to avoid
1894 * exposing stale data.
1895 * The page is currently locked and not marked for writeback
1898 /* Recovery: lock and submit the mapped buffers */
1900 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1901 !buffer_delay(bh)) {
1903 mark_buffer_async_write_endio(bh, handler);
1906 * The buffer may have been set dirty during
1907 * attachment to a dirty page.
1909 clear_buffer_dirty(bh);
1911 } while ((bh = bh->b_this_page) != head);
1913 BUG_ON(PageWriteback(page));
1914 mapping_set_error(page->mapping, err);
1915 set_page_writeback(page);
1917 struct buffer_head *next = bh->b_this_page;
1918 if (buffer_async_write(bh)) {
1919 clear_buffer_dirty(bh);
1920 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh, wbc);
1924 } while (bh != head);
1928 EXPORT_SYMBOL(__block_write_full_page);
1931 * If a page has any new buffers, zero them out here, and mark them uptodate
1932 * and dirty so they'll be written out (in order to prevent uninitialised
1933 * block data from leaking). And clear the new bit.
1935 void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1937 unsigned int block_start, block_end;
1938 struct buffer_head *head, *bh;
1940 BUG_ON(!PageLocked(page));
1941 if (!page_has_buffers(page))
1944 bh = head = page_buffers(page);
1947 block_end = block_start + bh->b_size;
1949 if (buffer_new(bh)) {
1950 if (block_end > from && block_start < to) {
1951 if (!PageUptodate(page)) {
1952 unsigned start, size;
1954 start = max(from, block_start);
1955 size = min(to, block_end) - start;
1957 zero_user(page, start, size);
1958 set_buffer_uptodate(bh);
1961 clear_buffer_new(bh);
1962 mark_buffer_dirty(bh);
1966 block_start = block_end;
1967 bh = bh->b_this_page;
1968 } while (bh != head);
1970 EXPORT_SYMBOL(page_zero_new_buffers);
1973 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
1974 const struct iomap *iomap)
1976 loff_t offset = block << inode->i_blkbits;
1978 bh->b_bdev = iomap->bdev;
1981 * Block points to offset in file we need to map, iomap contains
1982 * the offset at which the map starts. If the map ends before the
1983 * current block, then do not map the buffer and let the caller
1986 BUG_ON(offset >= iomap->offset + iomap->length);
1988 switch (iomap->type) {
1991 * If the buffer is not up to date or beyond the current EOF,
1992 * we need to mark it as new to ensure sub-block zeroing is
1993 * executed if necessary.
1995 if (!buffer_uptodate(bh) ||
1996 (offset >= i_size_read(inode)))
1999 case IOMAP_DELALLOC:
2000 if (!buffer_uptodate(bh) ||
2001 (offset >= i_size_read(inode)))
2003 set_buffer_uptodate(bh);
2004 set_buffer_mapped(bh);
2005 set_buffer_delay(bh);
2007 case IOMAP_UNWRITTEN:
2009 * For unwritten regions, we always need to ensure that regions
2010 * in the block we are not writing to are zeroed. Mark the
2011 * buffer as new to ensure this.
2014 set_buffer_unwritten(bh);
2017 if ((iomap->flags & IOMAP_F_NEW) ||
2018 offset >= i_size_read(inode))
2020 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
2022 set_buffer_mapped(bh);
2027 int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len,
2028 get_block_t *get_block, const struct iomap *iomap)
2030 unsigned from = pos & (PAGE_SIZE - 1);
2031 unsigned to = from + len;
2032 struct inode *inode = folio->mapping->host;
2033 unsigned block_start, block_end;
2036 unsigned blocksize, bbits;
2037 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2039 BUG_ON(!folio_test_locked(folio));
2040 BUG_ON(from > PAGE_SIZE);
2041 BUG_ON(to > PAGE_SIZE);
2044 head = folio_create_buffers(folio, inode, 0);
2045 blocksize = head->b_size;
2046 bbits = block_size_bits(blocksize);
2048 block = (sector_t)folio->index << (PAGE_SHIFT - bbits);
2050 for(bh = head, block_start = 0; bh != head || !block_start;
2051 block++, block_start=block_end, bh = bh->b_this_page) {
2052 block_end = block_start + blocksize;
2053 if (block_end <= from || block_start >= to) {
2054 if (folio_test_uptodate(folio)) {
2055 if (!buffer_uptodate(bh))
2056 set_buffer_uptodate(bh);
2061 clear_buffer_new(bh);
2062 if (!buffer_mapped(bh)) {
2063 WARN_ON(bh->b_size != blocksize);
2065 err = get_block(inode, block, bh, 1);
2069 iomap_to_bh(inode, block, bh, iomap);
2072 if (buffer_new(bh)) {
2073 clean_bdev_bh_alias(bh);
2074 if (folio_test_uptodate(folio)) {
2075 clear_buffer_new(bh);
2076 set_buffer_uptodate(bh);
2077 mark_buffer_dirty(bh);
2080 if (block_end > to || block_start < from)
2081 folio_zero_segments(folio,
2087 if (folio_test_uptodate(folio)) {
2088 if (!buffer_uptodate(bh))
2089 set_buffer_uptodate(bh);
2092 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2093 !buffer_unwritten(bh) &&
2094 (block_start < from || block_end > to)) {
2095 bh_read_nowait(bh, 0);
2100 * If we issued read requests - let them complete.
2102 while(wait_bh > wait) {
2103 wait_on_buffer(*--wait_bh);
2104 if (!buffer_uptodate(*wait_bh))
2108 page_zero_new_buffers(&folio->page, from, to);
2112 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2113 get_block_t *get_block)
2115 return __block_write_begin_int(page_folio(page), pos, len, get_block,
2118 EXPORT_SYMBOL(__block_write_begin);
2120 static int __block_commit_write(struct inode *inode, struct page *page,
2121 unsigned from, unsigned to)
2123 unsigned block_start, block_end;
2126 struct buffer_head *bh, *head;
2128 bh = head = page_buffers(page);
2129 blocksize = bh->b_size;
2133 block_end = block_start + blocksize;
2134 if (block_end <= from || block_start >= to) {
2135 if (!buffer_uptodate(bh))
2138 set_buffer_uptodate(bh);
2139 mark_buffer_dirty(bh);
2142 clear_buffer_new(bh);
2144 block_start = block_end;
2145 bh = bh->b_this_page;
2146 } while (bh != head);
2149 * If this is a partial write which happened to make all buffers
2150 * uptodate then we can optimize away a bogus read_folio() for
2151 * the next read(). Here we 'discover' whether the page went
2152 * uptodate as a result of this (potentially partial) write.
2155 SetPageUptodate(page);
2160 * block_write_begin takes care of the basic task of block allocation and
2161 * bringing partial write blocks uptodate first.
2163 * The filesystem needs to handle block truncation upon failure.
2165 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2166 struct page **pagep, get_block_t *get_block)
2168 pgoff_t index = pos >> PAGE_SHIFT;
2172 page = grab_cache_page_write_begin(mapping, index);
2176 status = __block_write_begin(page, pos, len, get_block);
2177 if (unlikely(status)) {
2186 EXPORT_SYMBOL(block_write_begin);
2188 int block_write_end(struct file *file, struct address_space *mapping,
2189 loff_t pos, unsigned len, unsigned copied,
2190 struct page *page, void *fsdata)
2192 struct inode *inode = mapping->host;
2195 start = pos & (PAGE_SIZE - 1);
2197 if (unlikely(copied < len)) {
2199 * The buffers that were written will now be uptodate, so
2200 * we don't have to worry about a read_folio reading them
2201 * and overwriting a partial write. However if we have
2202 * encountered a short write and only partially written
2203 * into a buffer, it will not be marked uptodate, so a
2204 * read_folio might come in and destroy our partial write.
2206 * Do the simplest thing, and just treat any short write to a
2207 * non uptodate page as a zero-length write, and force the
2208 * caller to redo the whole thing.
2210 if (!PageUptodate(page))
2213 page_zero_new_buffers(page, start+copied, start+len);
2215 flush_dcache_page(page);
2217 /* This could be a short (even 0-length) commit */
2218 __block_commit_write(inode, page, start, start+copied);
2222 EXPORT_SYMBOL(block_write_end);
2224 int generic_write_end(struct file *file, struct address_space *mapping,
2225 loff_t pos, unsigned len, unsigned copied,
2226 struct page *page, void *fsdata)
2228 struct inode *inode = mapping->host;
2229 loff_t old_size = inode->i_size;
2230 bool i_size_changed = false;
2232 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2235 * No need to use i_size_read() here, the i_size cannot change under us
2236 * because we hold i_rwsem.
2238 * But it's important to update i_size while still holding page lock:
2239 * page writeout could otherwise come in and zero beyond i_size.
2241 if (pos + copied > inode->i_size) {
2242 i_size_write(inode, pos + copied);
2243 i_size_changed = true;
2250 pagecache_isize_extended(inode, old_size, pos);
2252 * Don't mark the inode dirty under page lock. First, it unnecessarily
2253 * makes the holding time of page lock longer. Second, it forces lock
2254 * ordering of page lock and transaction start for journaling
2258 mark_inode_dirty(inode);
2261 EXPORT_SYMBOL(generic_write_end);
2264 * block_is_partially_uptodate checks whether buffers within a folio are
2267 * Returns true if all buffers which correspond to the specified part
2268 * of the folio are uptodate.
2270 bool block_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
2272 unsigned block_start, block_end, blocksize;
2274 struct buffer_head *bh, *head;
2277 head = folio_buffers(folio);
2280 blocksize = head->b_size;
2281 to = min_t(unsigned, folio_size(folio) - from, count);
2283 if (from < blocksize && to > folio_size(folio) - blocksize)
2289 block_end = block_start + blocksize;
2290 if (block_end > from && block_start < to) {
2291 if (!buffer_uptodate(bh)) {
2295 if (block_end >= to)
2298 block_start = block_end;
2299 bh = bh->b_this_page;
2300 } while (bh != head);
2304 EXPORT_SYMBOL(block_is_partially_uptodate);
2307 * Generic "read_folio" function for block devices that have the normal
2308 * get_block functionality. This is most of the block device filesystems.
2309 * Reads the folio asynchronously --- the unlock_buffer() and
2310 * set/clear_buffer_uptodate() functions propagate buffer state into the
2311 * folio once IO has completed.
2313 int block_read_full_folio(struct folio *folio, get_block_t *get_block)
2315 struct inode *inode = folio->mapping->host;
2316 sector_t iblock, lblock;
2317 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2318 unsigned int blocksize, bbits;
2320 int fully_mapped = 1;
2321 bool page_error = false;
2322 loff_t limit = i_size_read(inode);
2324 /* This is needed for ext4. */
2325 if (IS_ENABLED(CONFIG_FS_VERITY) && IS_VERITY(inode))
2326 limit = inode->i_sb->s_maxbytes;
2328 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2330 head = folio_create_buffers(folio, inode, 0);
2331 blocksize = head->b_size;
2332 bbits = block_size_bits(blocksize);
2334 iblock = (sector_t)folio->index << (PAGE_SHIFT - bbits);
2335 lblock = (limit+blocksize-1) >> bbits;
2341 if (buffer_uptodate(bh))
2344 if (!buffer_mapped(bh)) {
2348 if (iblock < lblock) {
2349 WARN_ON(bh->b_size != blocksize);
2350 err = get_block(inode, iblock, bh, 0);
2352 folio_set_error(folio);
2356 if (!buffer_mapped(bh)) {
2357 folio_zero_range(folio, i * blocksize,
2360 set_buffer_uptodate(bh);
2364 * get_block() might have updated the buffer
2367 if (buffer_uptodate(bh))
2371 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2374 folio_set_mappedtodisk(folio);
2378 * All buffers are uptodate - we can set the folio uptodate
2379 * as well. But not if get_block() returned an error.
2382 folio_mark_uptodate(folio);
2383 folio_unlock(folio);
2387 /* Stage two: lock the buffers */
2388 for (i = 0; i < nr; i++) {
2391 mark_buffer_async_read(bh);
2395 * Stage 3: start the IO. Check for uptodateness
2396 * inside the buffer lock in case another process reading
2397 * the underlying blockdev brought it uptodate (the sct fix).
2399 for (i = 0; i < nr; i++) {
2401 if (buffer_uptodate(bh))
2402 end_buffer_async_read(bh, 1);
2404 submit_bh(REQ_OP_READ, bh);
2408 EXPORT_SYMBOL(block_read_full_folio);
2410 /* utility function for filesystems that need to do work on expanding
2411 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2412 * deal with the hole.
2414 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2416 struct address_space *mapping = inode->i_mapping;
2417 const struct address_space_operations *aops = mapping->a_ops;
2419 void *fsdata = NULL;
2422 err = inode_newsize_ok(inode, size);
2426 err = aops->write_begin(NULL, mapping, size, 0, &page, &fsdata);
2430 err = aops->write_end(NULL, mapping, size, 0, 0, page, fsdata);
2436 EXPORT_SYMBOL(generic_cont_expand_simple);
2438 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2439 loff_t pos, loff_t *bytes)
2441 struct inode *inode = mapping->host;
2442 const struct address_space_operations *aops = mapping->a_ops;
2443 unsigned int blocksize = i_blocksize(inode);
2445 void *fsdata = NULL;
2446 pgoff_t index, curidx;
2448 unsigned zerofrom, offset, len;
2451 index = pos >> PAGE_SHIFT;
2452 offset = pos & ~PAGE_MASK;
2454 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2455 zerofrom = curpos & ~PAGE_MASK;
2456 if (zerofrom & (blocksize-1)) {
2457 *bytes |= (blocksize-1);
2460 len = PAGE_SIZE - zerofrom;
2462 err = aops->write_begin(file, mapping, curpos, len,
2466 zero_user(page, zerofrom, len);
2467 err = aops->write_end(file, mapping, curpos, len, len,
2474 balance_dirty_pages_ratelimited(mapping);
2476 if (fatal_signal_pending(current)) {
2482 /* page covers the boundary, find the boundary offset */
2483 if (index == curidx) {
2484 zerofrom = curpos & ~PAGE_MASK;
2485 /* if we will expand the thing last block will be filled */
2486 if (offset <= zerofrom) {
2489 if (zerofrom & (blocksize-1)) {
2490 *bytes |= (blocksize-1);
2493 len = offset - zerofrom;
2495 err = aops->write_begin(file, mapping, curpos, len,
2499 zero_user(page, zerofrom, len);
2500 err = aops->write_end(file, mapping, curpos, len, len,
2512 * For moronic filesystems that do not allow holes in file.
2513 * We may have to extend the file.
2515 int cont_write_begin(struct file *file, struct address_space *mapping,
2516 loff_t pos, unsigned len,
2517 struct page **pagep, void **fsdata,
2518 get_block_t *get_block, loff_t *bytes)
2520 struct inode *inode = mapping->host;
2521 unsigned int blocksize = i_blocksize(inode);
2522 unsigned int zerofrom;
2525 err = cont_expand_zero(file, mapping, pos, bytes);
2529 zerofrom = *bytes & ~PAGE_MASK;
2530 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2531 *bytes |= (blocksize-1);
2535 return block_write_begin(mapping, pos, len, pagep, get_block);
2537 EXPORT_SYMBOL(cont_write_begin);
2539 int block_commit_write(struct page *page, unsigned from, unsigned to)
2541 struct inode *inode = page->mapping->host;
2542 __block_commit_write(inode,page,from,to);
2545 EXPORT_SYMBOL(block_commit_write);
2548 * block_page_mkwrite() is not allowed to change the file size as it gets
2549 * called from a page fault handler when a page is first dirtied. Hence we must
2550 * be careful to check for EOF conditions here. We set the page up correctly
2551 * for a written page which means we get ENOSPC checking when writing into
2552 * holes and correct delalloc and unwritten extent mapping on filesystems that
2553 * support these features.
2555 * We are not allowed to take the i_mutex here so we have to play games to
2556 * protect against truncate races as the page could now be beyond EOF. Because
2557 * truncate writes the inode size before removing pages, once we have the
2558 * page lock we can determine safely if the page is beyond EOF. If it is not
2559 * beyond EOF, then the page is guaranteed safe against truncation until we
2562 * Direct callers of this function should protect against filesystem freezing
2563 * using sb_start_pagefault() - sb_end_pagefault() functions.
2565 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2566 get_block_t get_block)
2568 struct page *page = vmf->page;
2569 struct inode *inode = file_inode(vma->vm_file);
2575 size = i_size_read(inode);
2576 if ((page->mapping != inode->i_mapping) ||
2577 (page_offset(page) > size)) {
2578 /* We overload EFAULT to mean page got truncated */
2583 /* page is wholly or partially inside EOF */
2584 if (((page->index + 1) << PAGE_SHIFT) > size)
2585 end = size & ~PAGE_MASK;
2589 ret = __block_write_begin(page, 0, end, get_block);
2591 ret = block_commit_write(page, 0, end);
2593 if (unlikely(ret < 0))
2595 set_page_dirty(page);
2596 wait_for_stable_page(page);
2602 EXPORT_SYMBOL(block_page_mkwrite);
2604 int block_truncate_page(struct address_space *mapping,
2605 loff_t from, get_block_t *get_block)
2607 pgoff_t index = from >> PAGE_SHIFT;
2608 unsigned offset = from & (PAGE_SIZE-1);
2611 unsigned length, pos;
2612 struct inode *inode = mapping->host;
2614 struct buffer_head *bh;
2617 blocksize = i_blocksize(inode);
2618 length = offset & (blocksize - 1);
2620 /* Block boundary? Nothing to do */
2624 length = blocksize - length;
2625 iblock = (sector_t)index << (PAGE_SHIFT - inode->i_blkbits);
2627 page = grab_cache_page(mapping, index);
2632 if (!page_has_buffers(page))
2633 create_empty_buffers(page, blocksize, 0);
2635 /* Find the buffer that contains "offset" */
2636 bh = page_buffers(page);
2638 while (offset >= pos) {
2639 bh = bh->b_this_page;
2645 if (!buffer_mapped(bh)) {
2646 WARN_ON(bh->b_size != blocksize);
2647 err = get_block(inode, iblock, bh, 0);
2650 /* unmapped? It's a hole - nothing to do */
2651 if (!buffer_mapped(bh))
2655 /* Ok, it's mapped. Make sure it's up-to-date */
2656 if (PageUptodate(page))
2657 set_buffer_uptodate(bh);
2659 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2660 err = bh_read(bh, 0);
2661 /* Uhhuh. Read error. Complain and punt. */
2666 zero_user(page, offset, length);
2667 mark_buffer_dirty(bh);
2676 EXPORT_SYMBOL(block_truncate_page);
2679 * The generic ->writepage function for buffer-backed address_spaces
2681 int block_write_full_page(struct page *page, get_block_t *get_block,
2682 struct writeback_control *wbc)
2684 struct inode * const inode = page->mapping->host;
2685 loff_t i_size = i_size_read(inode);
2686 const pgoff_t end_index = i_size >> PAGE_SHIFT;
2689 /* Is the page fully inside i_size? */
2690 if (page->index < end_index)
2691 return __block_write_full_page(inode, page, get_block, wbc,
2692 end_buffer_async_write);
2694 /* Is the page fully outside i_size? (truncate in progress) */
2695 offset = i_size & (PAGE_SIZE-1);
2696 if (page->index >= end_index+1 || !offset) {
2698 return 0; /* don't care */
2702 * The page straddles i_size. It must be zeroed out on each and every
2703 * writepage invocation because it may be mmapped. "A file is mapped
2704 * in multiples of the page size. For a file that is not a multiple of
2705 * the page size, the remaining memory is zeroed when mapped, and
2706 * writes to that region are not written out to the file."
2708 zero_user_segment(page, offset, PAGE_SIZE);
2709 return __block_write_full_page(inode, page, get_block, wbc,
2710 end_buffer_async_write);
2712 EXPORT_SYMBOL(block_write_full_page);
2714 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2715 get_block_t *get_block)
2717 struct inode *inode = mapping->host;
2718 struct buffer_head tmp = {
2719 .b_size = i_blocksize(inode),
2722 get_block(inode, block, &tmp, 0);
2723 return tmp.b_blocknr;
2725 EXPORT_SYMBOL(generic_block_bmap);
2727 static void end_bio_bh_io_sync(struct bio *bio)
2729 struct buffer_head *bh = bio->bi_private;
2731 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2732 set_bit(BH_Quiet, &bh->b_state);
2734 bh->b_end_io(bh, !bio->bi_status);
2738 static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
2739 struct writeback_control *wbc)
2741 const enum req_op op = opf & REQ_OP_MASK;
2744 BUG_ON(!buffer_locked(bh));
2745 BUG_ON(!buffer_mapped(bh));
2746 BUG_ON(!bh->b_end_io);
2747 BUG_ON(buffer_delay(bh));
2748 BUG_ON(buffer_unwritten(bh));
2751 * Only clear out a write error when rewriting
2753 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
2754 clear_buffer_write_io_error(bh);
2756 if (buffer_meta(bh))
2758 if (buffer_prio(bh))
2761 bio = bio_alloc(bh->b_bdev, 1, opf, GFP_NOIO);
2763 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
2765 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2767 bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
2768 BUG_ON(bio->bi_iter.bi_size != bh->b_size);
2770 bio->bi_end_io = end_bio_bh_io_sync;
2771 bio->bi_private = bh;
2773 /* Take care of bh's that straddle the end of the device */
2777 wbc_init_bio(wbc, bio);
2778 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
2784 void submit_bh(blk_opf_t opf, struct buffer_head *bh)
2786 submit_bh_wbc(opf, bh, NULL);
2788 EXPORT_SYMBOL(submit_bh);
2790 void write_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2793 if (!test_clear_buffer_dirty(bh)) {
2797 bh->b_end_io = end_buffer_write_sync;
2799 submit_bh(REQ_OP_WRITE | op_flags, bh);
2801 EXPORT_SYMBOL(write_dirty_buffer);
2804 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2805 * and then start new I/O and then wait upon it. The caller must have a ref on
2808 int __sync_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2810 WARN_ON(atomic_read(&bh->b_count) < 1);
2812 if (test_clear_buffer_dirty(bh)) {
2814 * The bh should be mapped, but it might not be if the
2815 * device was hot-removed. Not much we can do but fail the I/O.
2817 if (!buffer_mapped(bh)) {
2823 bh->b_end_io = end_buffer_write_sync;
2824 submit_bh(REQ_OP_WRITE | op_flags, bh);
2826 if (!buffer_uptodate(bh))
2833 EXPORT_SYMBOL(__sync_dirty_buffer);
2835 int sync_dirty_buffer(struct buffer_head *bh)
2837 return __sync_dirty_buffer(bh, REQ_SYNC);
2839 EXPORT_SYMBOL(sync_dirty_buffer);
2842 * try_to_free_buffers() checks if all the buffers on this particular folio
2843 * are unused, and releases them if so.
2845 * Exclusion against try_to_free_buffers may be obtained by either
2846 * locking the folio or by holding its mapping's private_lock.
2848 * If the folio is dirty but all the buffers are clean then we need to
2849 * be sure to mark the folio clean as well. This is because the folio
2850 * may be against a block device, and a later reattachment of buffers
2851 * to a dirty folio will set *all* buffers dirty. Which would corrupt
2852 * filesystem data on the same device.
2854 * The same applies to regular filesystem folios: if all the buffers are
2855 * clean then we set the folio clean and proceed. To do that, we require
2856 * total exclusion from block_dirty_folio(). That is obtained with
2859 * try_to_free_buffers() is non-blocking.
2861 static inline int buffer_busy(struct buffer_head *bh)
2863 return atomic_read(&bh->b_count) |
2864 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2868 drop_buffers(struct folio *folio, struct buffer_head **buffers_to_free)
2870 struct buffer_head *head = folio_buffers(folio);
2871 struct buffer_head *bh;
2875 if (buffer_busy(bh))
2877 bh = bh->b_this_page;
2878 } while (bh != head);
2881 struct buffer_head *next = bh->b_this_page;
2883 if (bh->b_assoc_map)
2884 __remove_assoc_queue(bh);
2886 } while (bh != head);
2887 *buffers_to_free = head;
2888 folio_detach_private(folio);
2894 bool try_to_free_buffers(struct folio *folio)
2896 struct address_space * const mapping = folio->mapping;
2897 struct buffer_head *buffers_to_free = NULL;
2900 BUG_ON(!folio_test_locked(folio));
2901 if (folio_test_writeback(folio))
2904 if (mapping == NULL) { /* can this still happen? */
2905 ret = drop_buffers(folio, &buffers_to_free);
2909 spin_lock(&mapping->private_lock);
2910 ret = drop_buffers(folio, &buffers_to_free);
2913 * If the filesystem writes its buffers by hand (eg ext3)
2914 * then we can have clean buffers against a dirty folio. We
2915 * clean the folio here; otherwise the VM will never notice
2916 * that the filesystem did any IO at all.
2918 * Also, during truncate, discard_buffer will have marked all
2919 * the folio's buffers clean. We discover that here and clean
2922 * private_lock must be held over this entire operation in order
2923 * to synchronise against block_dirty_folio and prevent the
2924 * dirty bit from being lost.
2927 folio_cancel_dirty(folio);
2928 spin_unlock(&mapping->private_lock);
2930 if (buffers_to_free) {
2931 struct buffer_head *bh = buffers_to_free;
2934 struct buffer_head *next = bh->b_this_page;
2935 free_buffer_head(bh);
2937 } while (bh != buffers_to_free);
2941 EXPORT_SYMBOL(try_to_free_buffers);
2944 * Buffer-head allocation
2946 static struct kmem_cache *bh_cachep __read_mostly;
2949 * Once the number of bh's in the machine exceeds this level, we start
2950 * stripping them in writeback.
2952 static unsigned long max_buffer_heads;
2954 int buffer_heads_over_limit;
2956 struct bh_accounting {
2957 int nr; /* Number of live bh's */
2958 int ratelimit; /* Limit cacheline bouncing */
2961 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2963 static void recalc_bh_state(void)
2968 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
2970 __this_cpu_write(bh_accounting.ratelimit, 0);
2971 for_each_online_cpu(i)
2972 tot += per_cpu(bh_accounting, i).nr;
2973 buffer_heads_over_limit = (tot > max_buffer_heads);
2976 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
2978 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
2980 INIT_LIST_HEAD(&ret->b_assoc_buffers);
2981 spin_lock_init(&ret->b_uptodate_lock);
2983 __this_cpu_inc(bh_accounting.nr);
2989 EXPORT_SYMBOL(alloc_buffer_head);
2991 void free_buffer_head(struct buffer_head *bh)
2993 BUG_ON(!list_empty(&bh->b_assoc_buffers));
2994 kmem_cache_free(bh_cachep, bh);
2996 __this_cpu_dec(bh_accounting.nr);
3000 EXPORT_SYMBOL(free_buffer_head);
3002 static int buffer_exit_cpu_dead(unsigned int cpu)
3005 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3007 for (i = 0; i < BH_LRU_SIZE; i++) {
3011 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3012 per_cpu(bh_accounting, cpu).nr = 0;
3017 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3018 * @bh: struct buffer_head
3020 * Return true if the buffer is up-to-date and false,
3021 * with the buffer locked, if not.
3023 int bh_uptodate_or_lock(struct buffer_head *bh)
3025 if (!buffer_uptodate(bh)) {
3027 if (!buffer_uptodate(bh))
3033 EXPORT_SYMBOL(bh_uptodate_or_lock);
3036 * __bh_read - Submit read for a locked buffer
3037 * @bh: struct buffer_head
3038 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3039 * @wait: wait until reading finish
3041 * Returns zero on success or don't wait, and -EIO on error.
3043 int __bh_read(struct buffer_head *bh, blk_opf_t op_flags, bool wait)
3047 BUG_ON(!buffer_locked(bh));
3050 bh->b_end_io = end_buffer_read_sync;
3051 submit_bh(REQ_OP_READ | op_flags, bh);
3054 if (!buffer_uptodate(bh))
3059 EXPORT_SYMBOL(__bh_read);
3062 * __bh_read_batch - Submit read for a batch of unlocked buffers
3063 * @nr: entry number of the buffer batch
3064 * @bhs: a batch of struct buffer_head
3065 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3066 * @force_lock: force to get a lock on the buffer if set, otherwise drops any
3067 * buffer that cannot lock.
3069 * Returns zero on success or don't wait, and -EIO on error.
3071 void __bh_read_batch(int nr, struct buffer_head *bhs[],
3072 blk_opf_t op_flags, bool force_lock)
3076 for (i = 0; i < nr; i++) {
3077 struct buffer_head *bh = bhs[i];
3079 if (buffer_uptodate(bh))
3085 if (!trylock_buffer(bh))
3088 if (buffer_uptodate(bh)) {
3093 bh->b_end_io = end_buffer_read_sync;
3095 submit_bh(REQ_OP_READ | op_flags, bh);
3098 EXPORT_SYMBOL(__bh_read_batch);
3100 void __init buffer_init(void)
3102 unsigned long nrpages;
3105 bh_cachep = kmem_cache_create("buffer_head",
3106 sizeof(struct buffer_head), 0,
3107 (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3112 * Limit the bh occupancy to 10% of ZONE_NORMAL
3114 nrpages = (nr_free_buffer_pages() * 10) / 100;
3115 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3116 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3117 NULL, buffer_exit_cpu_dead);