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
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>
52 #include <linux/sched/isolation.h>
56 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
57 static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
58 enum rw_hint hint, struct writeback_control *wbc);
60 #define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
62 inline void touch_buffer(struct buffer_head *bh)
64 trace_block_touch_buffer(bh);
65 folio_mark_accessed(bh->b_folio);
67 EXPORT_SYMBOL(touch_buffer);
69 void __lock_buffer(struct buffer_head *bh)
71 wait_on_bit_lock_io(&bh->b_state, BH_Lock, TASK_UNINTERRUPTIBLE);
73 EXPORT_SYMBOL(__lock_buffer);
75 void unlock_buffer(struct buffer_head *bh)
77 clear_bit_unlock(BH_Lock, &bh->b_state);
78 smp_mb__after_atomic();
79 wake_up_bit(&bh->b_state, BH_Lock);
81 EXPORT_SYMBOL(unlock_buffer);
84 * Returns if the folio has dirty or writeback buffers. If all the buffers
85 * are unlocked and clean then the folio_test_dirty information is stale. If
86 * any of the buffers are locked, it is assumed they are locked for IO.
88 void buffer_check_dirty_writeback(struct folio *folio,
89 bool *dirty, bool *writeback)
91 struct buffer_head *head, *bh;
95 BUG_ON(!folio_test_locked(folio));
97 head = folio_buffers(folio);
101 if (folio_test_writeback(folio))
106 if (buffer_locked(bh))
109 if (buffer_dirty(bh))
112 bh = bh->b_this_page;
113 } while (bh != head);
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, i_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 i_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 = ((loff_t)block << bd_inode->i_blkbits) / PAGE_SIZE;
203 folio = __filemap_get_folio(bd_mapping, index, FGP_ACCESSED, 0);
207 spin_lock(&bd_mapping->i_private_lock);
208 head = folio_buffers(folio);
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->i_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);
285 folio_end_read(folio, folio_uptodate);
289 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
293 struct postprocess_bh_ctx {
294 struct work_struct work;
295 struct buffer_head *bh;
298 static void verify_bh(struct work_struct *work)
300 struct postprocess_bh_ctx *ctx =
301 container_of(work, struct postprocess_bh_ctx, work);
302 struct buffer_head *bh = ctx->bh;
305 valid = fsverity_verify_blocks(bh->b_folio, bh->b_size, bh_offset(bh));
306 end_buffer_async_read(bh, valid);
310 static bool need_fsverity(struct buffer_head *bh)
312 struct folio *folio = bh->b_folio;
313 struct inode *inode = folio->mapping->host;
315 return fsverity_active(inode) &&
317 folio->index < DIV_ROUND_UP(inode->i_size, PAGE_SIZE);
320 static void decrypt_bh(struct work_struct *work)
322 struct postprocess_bh_ctx *ctx =
323 container_of(work, struct postprocess_bh_ctx, work);
324 struct buffer_head *bh = ctx->bh;
327 err = fscrypt_decrypt_pagecache_blocks(bh->b_folio, bh->b_size,
329 if (err == 0 && need_fsverity(bh)) {
331 * We use different work queues for decryption and for verity
332 * because verity may require reading metadata pages that need
333 * decryption, and we shouldn't recurse to the same workqueue.
335 INIT_WORK(&ctx->work, verify_bh);
336 fsverity_enqueue_verify_work(&ctx->work);
339 end_buffer_async_read(bh, err == 0);
344 * I/O completion handler for block_read_full_folio() - pages
345 * which come unlocked at the end of I/O.
347 static void end_buffer_async_read_io(struct buffer_head *bh, int uptodate)
349 struct inode *inode = bh->b_folio->mapping->host;
350 bool decrypt = fscrypt_inode_uses_fs_layer_crypto(inode);
351 bool verify = need_fsverity(bh);
353 /* Decrypt (with fscrypt) and/or verify (with fsverity) if needed. */
354 if (uptodate && (decrypt || verify)) {
355 struct postprocess_bh_ctx *ctx =
356 kmalloc(sizeof(*ctx), GFP_ATOMIC);
361 INIT_WORK(&ctx->work, decrypt_bh);
362 fscrypt_enqueue_decrypt_work(&ctx->work);
364 INIT_WORK(&ctx->work, verify_bh);
365 fsverity_enqueue_verify_work(&ctx->work);
371 end_buffer_async_read(bh, uptodate);
375 * Completion handler for block_write_full_folio() - folios which are unlocked
376 * during I/O, and which have the writeback flag cleared upon I/O completion.
378 static void end_buffer_async_write(struct buffer_head *bh, int uptodate)
381 struct buffer_head *first;
382 struct buffer_head *tmp;
385 BUG_ON(!buffer_async_write(bh));
389 set_buffer_uptodate(bh);
391 buffer_io_error(bh, ", lost async page write");
392 mark_buffer_write_io_error(bh);
393 clear_buffer_uptodate(bh);
394 folio_set_error(folio);
397 first = folio_buffers(folio);
398 spin_lock_irqsave(&first->b_uptodate_lock, flags);
400 clear_buffer_async_write(bh);
402 tmp = bh->b_this_page;
404 if (buffer_async_write(tmp)) {
405 BUG_ON(!buffer_locked(tmp));
408 tmp = tmp->b_this_page;
410 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
411 folio_end_writeback(folio);
415 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
420 * If a page's buffers are under async readin (end_buffer_async_read
421 * completion) then there is a possibility that another thread of
422 * control could lock one of the buffers after it has completed
423 * but while some of the other buffers have not completed. This
424 * locked buffer would confuse end_buffer_async_read() into not unlocking
425 * the page. So the absence of BH_Async_Read tells end_buffer_async_read()
426 * that this buffer is not under async I/O.
428 * The page comes unlocked when it has no locked buffer_async buffers
431 * PageLocked prevents anyone starting new async I/O reads any of
434 * PageWriteback is used to prevent simultaneous writeout of the same
437 * PageLocked prevents anyone from starting writeback of a page which is
438 * under read I/O (PageWriteback is only ever set against a locked page).
440 static void mark_buffer_async_read(struct buffer_head *bh)
442 bh->b_end_io = end_buffer_async_read_io;
443 set_buffer_async_read(bh);
446 static void mark_buffer_async_write_endio(struct buffer_head *bh,
447 bh_end_io_t *handler)
449 bh->b_end_io = handler;
450 set_buffer_async_write(bh);
453 void mark_buffer_async_write(struct buffer_head *bh)
455 mark_buffer_async_write_endio(bh, end_buffer_async_write);
457 EXPORT_SYMBOL(mark_buffer_async_write);
461 * fs/buffer.c contains helper functions for buffer-backed address space's
462 * fsync functions. A common requirement for buffer-based filesystems is
463 * that certain data from the backing blockdev needs to be written out for
464 * a successful fsync(). For example, ext2 indirect blocks need to be
465 * written back and waited upon before fsync() returns.
467 * The functions mark_buffer_dirty_inode(), fsync_inode_buffers(),
468 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
469 * management of a list of dependent buffers at ->i_mapping->i_private_list.
471 * Locking is a little subtle: try_to_free_buffers() will remove buffers
472 * from their controlling inode's queue when they are being freed. But
473 * try_to_free_buffers() will be operating against the *blockdev* mapping
474 * at the time, not against the S_ISREG file which depends on those buffers.
475 * So the locking for i_private_list is via the i_private_lock in the address_space
476 * which backs the buffers. Which is different from the address_space
477 * against which the buffers are listed. So for a particular address_space,
478 * mapping->i_private_lock does *not* protect mapping->i_private_list! In fact,
479 * mapping->i_private_list will always be protected by the backing blockdev's
482 * Which introduces a requirement: all buffers on an address_space's
483 * ->i_private_list must be from the same address_space: the blockdev's.
485 * address_spaces which do not place buffers at ->i_private_list via these
486 * utility functions are free to use i_private_lock and i_private_list for
487 * whatever they want. The only requirement is that list_empty(i_private_list)
488 * be true at clear_inode() time.
490 * FIXME: clear_inode should not call invalidate_inode_buffers(). The
491 * filesystems should do that. invalidate_inode_buffers() should just go
492 * BUG_ON(!list_empty).
494 * FIXME: mark_buffer_dirty_inode() is a data-plane operation. It should
495 * take an address_space, not an inode. And it should be called
496 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
499 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
500 * list if it is already on a list. Because if the buffer is on a list,
501 * it *must* already be on the right one. If not, the filesystem is being
502 * silly. This will save a ton of locking. But first we have to ensure
503 * that buffers are taken *off* the old inode's list when they are freed
504 * (presumably in truncate). That requires careful auditing of all
505 * filesystems (do it inside bforget()). It could also be done by bringing
510 * The buffer's backing address_space's i_private_lock must be held
512 static void __remove_assoc_queue(struct buffer_head *bh)
514 list_del_init(&bh->b_assoc_buffers);
515 WARN_ON(!bh->b_assoc_map);
516 bh->b_assoc_map = NULL;
519 int inode_has_buffers(struct inode *inode)
521 return !list_empty(&inode->i_data.i_private_list);
525 * osync is designed to support O_SYNC io. It waits synchronously for
526 * all already-submitted IO to complete, but does not queue any new
527 * writes to the disk.
529 * To do O_SYNC writes, just queue the buffer writes with write_dirty_buffer
530 * as you dirty the buffers, and then use osync_inode_buffers to wait for
531 * completion. Any other dirty buffers which are not yet queued for
532 * write will not be flushed to disk by the osync.
534 static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
536 struct buffer_head *bh;
542 list_for_each_prev(p, list) {
544 if (buffer_locked(bh)) {
548 if (!buffer_uptodate(bh))
560 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
561 * @mapping: the mapping which wants those buffers written
563 * Starts I/O against the buffers at mapping->i_private_list, and waits upon
566 * Basically, this is a convenience function for fsync().
567 * @mapping is a file or directory which needs those buffers to be written for
568 * a successful fsync().
570 int sync_mapping_buffers(struct address_space *mapping)
572 struct address_space *buffer_mapping = mapping->i_private_data;
574 if (buffer_mapping == NULL || list_empty(&mapping->i_private_list))
577 return fsync_buffers_list(&buffer_mapping->i_private_lock,
578 &mapping->i_private_list);
580 EXPORT_SYMBOL(sync_mapping_buffers);
583 * generic_buffers_fsync_noflush - generic buffer fsync implementation
584 * for simple filesystems with no inode lock
586 * @file: file to synchronize
587 * @start: start offset in bytes
588 * @end: end offset in bytes (inclusive)
589 * @datasync: only synchronize essential metadata if true
591 * This is a generic implementation of the fsync method for simple
592 * filesystems which track all non-inode metadata in the buffers list
593 * hanging off the address_space structure.
595 int generic_buffers_fsync_noflush(struct file *file, loff_t start, loff_t end,
598 struct inode *inode = file->f_mapping->host;
602 err = file_write_and_wait_range(file, start, end);
606 ret = sync_mapping_buffers(inode->i_mapping);
607 if (!(inode->i_state & I_DIRTY_ALL))
609 if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
612 err = sync_inode_metadata(inode, 1);
617 /* check and advance again to catch errors after syncing out buffers */
618 err = file_check_and_advance_wb_err(file);
623 EXPORT_SYMBOL(generic_buffers_fsync_noflush);
626 * generic_buffers_fsync - generic buffer fsync implementation
627 * for simple filesystems with no inode lock
629 * @file: file to synchronize
630 * @start: start offset in bytes
631 * @end: end offset in bytes (inclusive)
632 * @datasync: only synchronize essential metadata if true
634 * This is a generic implementation of the fsync method for simple
635 * filesystems which track all non-inode metadata in the buffers list
636 * hanging off the address_space structure. This also makes sure that
637 * a device cache flush operation is called at the end.
639 int generic_buffers_fsync(struct file *file, loff_t start, loff_t end,
642 struct inode *inode = file->f_mapping->host;
645 ret = generic_buffers_fsync_noflush(file, start, end, datasync);
647 ret = blkdev_issue_flush(inode->i_sb->s_bdev);
650 EXPORT_SYMBOL(generic_buffers_fsync);
653 * Called when we've recently written block `bblock', and it is known that
654 * `bblock' was for a buffer_boundary() buffer. This means that the block at
655 * `bblock + 1' is probably a dirty indirect block. Hunt it down and, if it's
656 * dirty, schedule it for IO. So that indirects merge nicely with their data.
658 void write_boundary_block(struct block_device *bdev,
659 sector_t bblock, unsigned blocksize)
661 struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
663 if (buffer_dirty(bh))
664 write_dirty_buffer(bh, 0);
669 void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
671 struct address_space *mapping = inode->i_mapping;
672 struct address_space *buffer_mapping = bh->b_folio->mapping;
674 mark_buffer_dirty(bh);
675 if (!mapping->i_private_data) {
676 mapping->i_private_data = buffer_mapping;
678 BUG_ON(mapping->i_private_data != buffer_mapping);
680 if (!bh->b_assoc_map) {
681 spin_lock(&buffer_mapping->i_private_lock);
682 list_move_tail(&bh->b_assoc_buffers,
683 &mapping->i_private_list);
684 bh->b_assoc_map = mapping;
685 spin_unlock(&buffer_mapping->i_private_lock);
688 EXPORT_SYMBOL(mark_buffer_dirty_inode);
691 * Add a page to the dirty page list.
693 * It is a sad fact of life that this function is called from several places
694 * deeply under spinlocking. It may not sleep.
696 * If the page has buffers, the uptodate buffers are set dirty, to preserve
697 * dirty-state coherency between the page and the buffers. It the page does
698 * not have buffers then when they are later attached they will all be set
701 * The buffers are dirtied before the page is dirtied. There's a small race
702 * window in which a writepage caller may see the page cleanness but not the
703 * buffer dirtiness. That's fine. If this code were to set the page dirty
704 * before the buffers, a concurrent writepage caller could clear the page dirty
705 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
706 * page on the dirty page list.
708 * We use i_private_lock to lock against try_to_free_buffers while using the
709 * page's buffer list. Also use this to protect against clean buffers being
710 * added to the page after it was set dirty.
712 * FIXME: may need to call ->reservepage here as well. That's rather up to the
713 * address_space though.
715 bool block_dirty_folio(struct address_space *mapping, struct folio *folio)
717 struct buffer_head *head;
720 spin_lock(&mapping->i_private_lock);
721 head = folio_buffers(folio);
723 struct buffer_head *bh = head;
726 set_buffer_dirty(bh);
727 bh = bh->b_this_page;
728 } while (bh != head);
731 * Lock out page's memcg migration to keep PageDirty
732 * synchronized with per-memcg dirty page counters.
734 folio_memcg_lock(folio);
735 newly_dirty = !folio_test_set_dirty(folio);
736 spin_unlock(&mapping->i_private_lock);
739 __folio_mark_dirty(folio, mapping, 1);
741 folio_memcg_unlock(folio);
744 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
748 EXPORT_SYMBOL(block_dirty_folio);
751 * Write out and wait upon a list of buffers.
753 * We have conflicting pressures: we want to make sure that all
754 * initially dirty buffers get waited on, but that any subsequently
755 * dirtied buffers don't. After all, we don't want fsync to last
756 * forever if somebody is actively writing to the file.
758 * Do this in two main stages: first we copy dirty buffers to a
759 * temporary inode list, queueing the writes as we go. Then we clean
760 * up, waiting for those writes to complete.
762 * During this second stage, any subsequent updates to the file may end
763 * up refiling the buffer on the original inode's dirty list again, so
764 * there is a chance we will end up with a buffer queued for write but
765 * not yet completed on that list. So, as a final cleanup we go through
766 * the osync code to catch these locked, dirty buffers without requeuing
767 * any newly dirty buffers for write.
769 static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
771 struct buffer_head *bh;
772 struct list_head tmp;
773 struct address_space *mapping;
775 struct blk_plug plug;
777 INIT_LIST_HEAD(&tmp);
778 blk_start_plug(&plug);
781 while (!list_empty(list)) {
782 bh = BH_ENTRY(list->next);
783 mapping = bh->b_assoc_map;
784 __remove_assoc_queue(bh);
785 /* Avoid race with mark_buffer_dirty_inode() which does
786 * a lockless check and we rely on seeing the dirty bit */
788 if (buffer_dirty(bh) || buffer_locked(bh)) {
789 list_add(&bh->b_assoc_buffers, &tmp);
790 bh->b_assoc_map = mapping;
791 if (buffer_dirty(bh)) {
795 * Ensure any pending I/O completes so that
796 * write_dirty_buffer() actually writes the
797 * current contents - it is a noop if I/O is
798 * still in flight on potentially older
801 write_dirty_buffer(bh, REQ_SYNC);
804 * Kick off IO for the previous mapping. Note
805 * that we will not run the very last mapping,
806 * wait_on_buffer() will do that for us
807 * through sync_buffer().
816 blk_finish_plug(&plug);
819 while (!list_empty(&tmp)) {
820 bh = BH_ENTRY(tmp.prev);
822 mapping = bh->b_assoc_map;
823 __remove_assoc_queue(bh);
824 /* Avoid race with mark_buffer_dirty_inode() which does
825 * a lockless check and we rely on seeing the dirty bit */
827 if (buffer_dirty(bh)) {
828 list_add(&bh->b_assoc_buffers,
829 &mapping->i_private_list);
830 bh->b_assoc_map = mapping;
834 if (!buffer_uptodate(bh))
841 err2 = osync_buffers_list(lock, list);
849 * Invalidate any and all dirty buffers on a given inode. We are
850 * probably unmounting the fs, but that doesn't mean we have already
851 * done a sync(). Just drop the buffers from the inode list.
853 * NOTE: we take the inode's blockdev's mapping's i_private_lock. Which
854 * assumes that all the buffers are against the blockdev. Not true
857 void invalidate_inode_buffers(struct inode *inode)
859 if (inode_has_buffers(inode)) {
860 struct address_space *mapping = &inode->i_data;
861 struct list_head *list = &mapping->i_private_list;
862 struct address_space *buffer_mapping = mapping->i_private_data;
864 spin_lock(&buffer_mapping->i_private_lock);
865 while (!list_empty(list))
866 __remove_assoc_queue(BH_ENTRY(list->next));
867 spin_unlock(&buffer_mapping->i_private_lock);
870 EXPORT_SYMBOL(invalidate_inode_buffers);
873 * Remove any clean buffers from the inode's buffer list. This is called
874 * when we're trying to free the inode itself. Those buffers can pin it.
876 * Returns true if all buffers were removed.
878 int remove_inode_buffers(struct inode *inode)
882 if (inode_has_buffers(inode)) {
883 struct address_space *mapping = &inode->i_data;
884 struct list_head *list = &mapping->i_private_list;
885 struct address_space *buffer_mapping = mapping->i_private_data;
887 spin_lock(&buffer_mapping->i_private_lock);
888 while (!list_empty(list)) {
889 struct buffer_head *bh = BH_ENTRY(list->next);
890 if (buffer_dirty(bh)) {
894 __remove_assoc_queue(bh);
896 spin_unlock(&buffer_mapping->i_private_lock);
902 * Create the appropriate buffers when given a folio for data area and
903 * the size of each buffer.. Use the bh->b_this_page linked list to
904 * follow the buffers created. Return NULL if unable to create more
907 * The retry flag is used to differentiate async IO (paging, swapping)
908 * which may not fail from ordinary buffer allocations.
910 struct buffer_head *folio_alloc_buffers(struct folio *folio, unsigned long size,
913 struct buffer_head *bh, *head;
915 struct mem_cgroup *memcg, *old_memcg;
917 /* The folio lock pins the memcg */
918 memcg = folio_memcg(folio);
919 old_memcg = set_active_memcg(memcg);
922 offset = folio_size(folio);
923 while ((offset -= size) >= 0) {
924 bh = alloc_buffer_head(gfp);
928 bh->b_this_page = head;
934 /* Link the buffer to its folio */
935 folio_set_bh(bh, folio, offset);
938 set_active_memcg(old_memcg);
941 * In case anything failed, we just free everything we got.
947 head = head->b_this_page;
948 free_buffer_head(bh);
954 EXPORT_SYMBOL_GPL(folio_alloc_buffers);
956 struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
959 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT;
963 return folio_alloc_buffers(page_folio(page), size, gfp);
965 EXPORT_SYMBOL_GPL(alloc_page_buffers);
967 static inline void link_dev_buffers(struct folio *folio,
968 struct buffer_head *head)
970 struct buffer_head *bh, *tail;
975 bh = bh->b_this_page;
977 tail->b_this_page = head;
978 folio_attach_private(folio, head);
981 static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
983 sector_t retval = ~((sector_t)0);
984 loff_t sz = bdev_nr_bytes(bdev);
987 unsigned int sizebits = blksize_bits(size);
988 retval = (sz >> sizebits);
994 * Initialise the state of a blockdev folio's buffers.
996 static sector_t folio_init_buffers(struct folio *folio,
997 struct block_device *bdev, unsigned size)
999 struct buffer_head *head = folio_buffers(folio);
1000 struct buffer_head *bh = head;
1001 bool uptodate = folio_test_uptodate(folio);
1002 sector_t block = div_u64(folio_pos(folio), size);
1003 sector_t end_block = blkdev_max_block(bdev, size);
1006 if (!buffer_mapped(bh)) {
1007 bh->b_end_io = NULL;
1008 bh->b_private = NULL;
1010 bh->b_blocknr = block;
1012 set_buffer_uptodate(bh);
1013 if (block < end_block)
1014 set_buffer_mapped(bh);
1017 bh = bh->b_this_page;
1018 } while (bh != head);
1021 * Caller needs to validate requested block against end of device.
1027 * Create the page-cache folio that contains the requested block.
1029 * This is used purely for blockdev mappings.
1031 * Returns false if we have a failure which cannot be cured by retrying
1032 * without sleeping. Returns true if we succeeded, or the caller should retry.
1034 static bool grow_dev_folio(struct block_device *bdev, sector_t block,
1035 pgoff_t index, unsigned size, gfp_t gfp)
1037 struct inode *inode = bdev->bd_inode;
1038 struct folio *folio;
1039 struct buffer_head *bh;
1040 sector_t end_block = 0;
1042 folio = __filemap_get_folio(inode->i_mapping, index,
1043 FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp);
1047 bh = folio_buffers(folio);
1049 if (bh->b_size == size) {
1050 end_block = folio_init_buffers(folio, bdev, size);
1055 * Retrying may succeed; for example the folio may finish
1056 * writeback, or buffers may be cleaned. This should not
1057 * happen very often; maybe we have old buffers attached to
1058 * this blockdev's page cache and we're trying to change
1061 if (!try_to_free_buffers(folio)) {
1067 bh = folio_alloc_buffers(folio, size, gfp | __GFP_ACCOUNT);
1072 * Link the folio to the buffers and initialise them. Take the
1073 * lock to be atomic wrt __find_get_block(), which does not
1074 * run under the folio lock.
1076 spin_lock(&inode->i_mapping->i_private_lock);
1077 link_dev_buffers(folio, bh);
1078 end_block = folio_init_buffers(folio, bdev, size);
1079 spin_unlock(&inode->i_mapping->i_private_lock);
1081 folio_unlock(folio);
1083 return block < end_block;
1087 * Create buffers for the specified block device block's folio. If
1088 * that folio was dirty, the buffers are set dirty also. Returns false
1089 * if we've hit a permanent error.
1091 static bool grow_buffers(struct block_device *bdev, sector_t block,
1092 unsigned size, gfp_t gfp)
1097 * Check for a block which lies outside our maximum possible
1100 if (check_mul_overflow(block, (sector_t)size, &pos) || pos > MAX_LFS_FILESIZE) {
1101 printk(KERN_ERR "%s: requested out-of-range block %llu for device %pg\n",
1102 __func__, (unsigned long long)block,
1107 /* Create a folio with the proper size buffers */
1108 return grow_dev_folio(bdev, block, pos / PAGE_SIZE, size, gfp);
1111 static struct buffer_head *
1112 __getblk_slow(struct block_device *bdev, sector_t block,
1113 unsigned size, gfp_t gfp)
1115 /* Size must be multiple of hard sectorsize */
1116 if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1117 (size < 512 || size > PAGE_SIZE))) {
1118 printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1120 printk(KERN_ERR "logical block size: %d\n",
1121 bdev_logical_block_size(bdev));
1128 struct buffer_head *bh;
1130 bh = __find_get_block(bdev, block, size);
1134 if (!grow_buffers(bdev, block, size, gfp))
1140 * The relationship between dirty buffers and dirty pages:
1142 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1143 * the page is tagged dirty in the page cache.
1145 * At all times, the dirtiness of the buffers represents the dirtiness of
1146 * subsections of the page. If the page has buffers, the page dirty bit is
1147 * merely a hint about the true dirty state.
1149 * When a page is set dirty in its entirety, all its buffers are marked dirty
1150 * (if the page has buffers).
1152 * When a buffer is marked dirty, its page is dirtied, but the page's other
1155 * Also. When blockdev buffers are explicitly read with bread(), they
1156 * individually become uptodate. But their backing page remains not
1157 * uptodate - even if all of its buffers are uptodate. A subsequent
1158 * block_read_full_folio() against that folio will discover all the uptodate
1159 * buffers, will set the folio uptodate and will perform no I/O.
1163 * mark_buffer_dirty - mark a buffer_head as needing writeout
1164 * @bh: the buffer_head to mark dirty
1166 * mark_buffer_dirty() will set the dirty bit against the buffer, then set
1167 * its backing page dirty, then tag the page as dirty in the page cache
1168 * and then attach the address_space's inode to its superblock's dirty
1171 * mark_buffer_dirty() is atomic. It takes bh->b_folio->mapping->i_private_lock,
1172 * i_pages lock and mapping->host->i_lock.
1174 void mark_buffer_dirty(struct buffer_head *bh)
1176 WARN_ON_ONCE(!buffer_uptodate(bh));
1178 trace_block_dirty_buffer(bh);
1181 * Very *carefully* optimize the it-is-already-dirty case.
1183 * Don't let the final "is it dirty" escape to before we
1184 * perhaps modified the buffer.
1186 if (buffer_dirty(bh)) {
1188 if (buffer_dirty(bh))
1192 if (!test_set_buffer_dirty(bh)) {
1193 struct folio *folio = bh->b_folio;
1194 struct address_space *mapping = NULL;
1196 folio_memcg_lock(folio);
1197 if (!folio_test_set_dirty(folio)) {
1198 mapping = folio->mapping;
1200 __folio_mark_dirty(folio, mapping, 0);
1202 folio_memcg_unlock(folio);
1204 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1207 EXPORT_SYMBOL(mark_buffer_dirty);
1209 void mark_buffer_write_io_error(struct buffer_head *bh)
1211 set_buffer_write_io_error(bh);
1212 /* FIXME: do we need to set this in both places? */
1213 if (bh->b_folio && bh->b_folio->mapping)
1214 mapping_set_error(bh->b_folio->mapping, -EIO);
1215 if (bh->b_assoc_map) {
1216 mapping_set_error(bh->b_assoc_map, -EIO);
1217 errseq_set(&bh->b_assoc_map->host->i_sb->s_wb_err, -EIO);
1220 EXPORT_SYMBOL(mark_buffer_write_io_error);
1223 * Decrement a buffer_head's reference count. If all buffers against a page
1224 * have zero reference count, are clean and unlocked, and if the page is clean
1225 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1226 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1227 * a page but it ends up not being freed, and buffers may later be reattached).
1229 void __brelse(struct buffer_head * buf)
1231 if (atomic_read(&buf->b_count)) {
1235 WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1237 EXPORT_SYMBOL(__brelse);
1240 * bforget() is like brelse(), except it discards any
1241 * potentially dirty data.
1243 void __bforget(struct buffer_head *bh)
1245 clear_buffer_dirty(bh);
1246 if (bh->b_assoc_map) {
1247 struct address_space *buffer_mapping = bh->b_folio->mapping;
1249 spin_lock(&buffer_mapping->i_private_lock);
1250 list_del_init(&bh->b_assoc_buffers);
1251 bh->b_assoc_map = NULL;
1252 spin_unlock(&buffer_mapping->i_private_lock);
1256 EXPORT_SYMBOL(__bforget);
1258 static struct buffer_head *__bread_slow(struct buffer_head *bh)
1261 if (buffer_uptodate(bh)) {
1266 bh->b_end_io = end_buffer_read_sync;
1267 submit_bh(REQ_OP_READ, bh);
1269 if (buffer_uptodate(bh))
1277 * Per-cpu buffer LRU implementation. To reduce the cost of __find_get_block().
1278 * The bhs[] array is sorted - newest buffer is at bhs[0]. Buffers have their
1279 * refcount elevated by one when they're in an LRU. A buffer can only appear
1280 * once in a particular CPU's LRU. A single buffer can be present in multiple
1281 * CPU's LRUs at the same time.
1283 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1284 * sb_find_get_block().
1286 * The LRUs themselves only need locking against invalidate_bh_lrus. We use
1287 * a local interrupt disable for that.
1290 #define BH_LRU_SIZE 16
1293 struct buffer_head *bhs[BH_LRU_SIZE];
1296 static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1299 #define bh_lru_lock() local_irq_disable()
1300 #define bh_lru_unlock() local_irq_enable()
1302 #define bh_lru_lock() preempt_disable()
1303 #define bh_lru_unlock() preempt_enable()
1306 static inline void check_irqs_on(void)
1308 #ifdef irqs_disabled
1309 BUG_ON(irqs_disabled());
1314 * Install a buffer_head into this cpu's LRU. If not already in the LRU, it is
1315 * inserted at the front, and the buffer_head at the back if any is evicted.
1316 * Or, if already in the LRU it is moved to the front.
1318 static void bh_lru_install(struct buffer_head *bh)
1320 struct buffer_head *evictee = bh;
1328 * the refcount of buffer_head in bh_lru prevents dropping the
1329 * attached page(i.e., try_to_free_buffers) so it could cause
1330 * failing page migration.
1331 * Skip putting upcoming bh into bh_lru until migration is done.
1333 if (lru_cache_disabled() || cpu_is_isolated(smp_processor_id())) {
1338 b = this_cpu_ptr(&bh_lrus);
1339 for (i = 0; i < BH_LRU_SIZE; i++) {
1340 swap(evictee, b->bhs[i]);
1341 if (evictee == bh) {
1353 * Look up the bh in this cpu's LRU. If it's there, move it to the head.
1355 static struct buffer_head *
1356 lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1358 struct buffer_head *ret = NULL;
1363 if (cpu_is_isolated(smp_processor_id())) {
1367 for (i = 0; i < BH_LRU_SIZE; i++) {
1368 struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1370 if (bh && bh->b_blocknr == block && bh->b_bdev == bdev &&
1371 bh->b_size == size) {
1374 __this_cpu_write(bh_lrus.bhs[i],
1375 __this_cpu_read(bh_lrus.bhs[i - 1]));
1378 __this_cpu_write(bh_lrus.bhs[0], bh);
1390 * Perform a pagecache lookup for the matching buffer. If it's there, refresh
1391 * it in the LRU and mark it as accessed. If it is not present then return
1394 struct buffer_head *
1395 __find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1397 struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1400 /* __find_get_block_slow will mark the page accessed */
1401 bh = __find_get_block_slow(bdev, block);
1409 EXPORT_SYMBOL(__find_get_block);
1412 * bdev_getblk - Get a buffer_head in a block device's buffer cache.
1413 * @bdev: The block device.
1414 * @block: The block number.
1415 * @size: The size of buffer_heads for this @bdev.
1416 * @gfp: The memory allocation flags to use.
1418 * Return: The buffer head, or NULL if memory could not be allocated.
1420 struct buffer_head *bdev_getblk(struct block_device *bdev, sector_t block,
1421 unsigned size, gfp_t gfp)
1423 struct buffer_head *bh = __find_get_block(bdev, block, size);
1429 return __getblk_slow(bdev, block, size, gfp);
1431 EXPORT_SYMBOL(bdev_getblk);
1434 * Do async read-ahead on a buffer..
1436 void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1438 struct buffer_head *bh = bdev_getblk(bdev, block, size,
1439 GFP_NOWAIT | __GFP_MOVABLE);
1442 bh_readahead(bh, REQ_RAHEAD);
1446 EXPORT_SYMBOL(__breadahead);
1449 * __bread_gfp() - reads a specified block and returns the bh
1450 * @bdev: the block_device to read from
1451 * @block: number of block
1452 * @size: size (in bytes) to read
1453 * @gfp: page allocation flag
1455 * Reads a specified block, and returns buffer head that contains it.
1456 * The page cache can be allocated from non-movable area
1457 * not to prevent page migration if you set gfp to zero.
1458 * It returns NULL if the block was unreadable.
1460 struct buffer_head *
1461 __bread_gfp(struct block_device *bdev, sector_t block,
1462 unsigned size, gfp_t gfp)
1464 struct buffer_head *bh;
1466 gfp |= mapping_gfp_constraint(bdev->bd_inode->i_mapping, ~__GFP_FS);
1469 * Prefer looping in the allocator rather than here, at least that
1470 * code knows what it's doing.
1472 gfp |= __GFP_NOFAIL;
1474 bh = bdev_getblk(bdev, block, size, gfp);
1476 if (likely(bh) && !buffer_uptodate(bh))
1477 bh = __bread_slow(bh);
1480 EXPORT_SYMBOL(__bread_gfp);
1482 static void __invalidate_bh_lrus(struct bh_lru *b)
1486 for (i = 0; i < BH_LRU_SIZE; i++) {
1492 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1493 * This doesn't race because it runs in each cpu either in irq
1494 * or with preempt disabled.
1496 static void invalidate_bh_lru(void *arg)
1498 struct bh_lru *b = &get_cpu_var(bh_lrus);
1500 __invalidate_bh_lrus(b);
1501 put_cpu_var(bh_lrus);
1504 bool has_bh_in_lru(int cpu, void *dummy)
1506 struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1509 for (i = 0; i < BH_LRU_SIZE; i++) {
1517 void invalidate_bh_lrus(void)
1519 on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1);
1521 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1524 * It's called from workqueue context so we need a bh_lru_lock to close
1525 * the race with preemption/irq.
1527 void invalidate_bh_lrus_cpu(void)
1532 b = this_cpu_ptr(&bh_lrus);
1533 __invalidate_bh_lrus(b);
1537 void folio_set_bh(struct buffer_head *bh, struct folio *folio,
1538 unsigned long offset)
1540 bh->b_folio = folio;
1541 BUG_ON(offset >= folio_size(folio));
1542 if (folio_test_highmem(folio))
1544 * This catches illegal uses and preserves the offset:
1546 bh->b_data = (char *)(0 + offset);
1548 bh->b_data = folio_address(folio) + offset;
1550 EXPORT_SYMBOL(folio_set_bh);
1553 * Called when truncating a buffer on a page completely.
1556 /* Bits that are cleared during an invalidate */
1557 #define BUFFER_FLAGS_DISCARD \
1558 (1 << BH_Mapped | 1 << BH_New | 1 << BH_Req | \
1559 1 << BH_Delay | 1 << BH_Unwritten)
1561 static void discard_buffer(struct buffer_head * bh)
1563 unsigned long b_state;
1566 clear_buffer_dirty(bh);
1568 b_state = READ_ONCE(bh->b_state);
1570 } while (!try_cmpxchg(&bh->b_state, &b_state,
1571 b_state & ~BUFFER_FLAGS_DISCARD));
1576 * block_invalidate_folio - Invalidate part or all of a buffer-backed folio.
1577 * @folio: The folio which is affected.
1578 * @offset: start of the range to invalidate
1579 * @length: length of the range to invalidate
1581 * block_invalidate_folio() is called when all or part of the folio has been
1582 * invalidated by a truncate operation.
1584 * block_invalidate_folio() does not have to release all buffers, but it must
1585 * ensure that no dirty buffer is left outside @offset and that no I/O
1586 * is underway against any of the blocks which are outside the truncation
1587 * point. Because the caller is about to free (and possibly reuse) those
1590 void block_invalidate_folio(struct folio *folio, size_t offset, size_t length)
1592 struct buffer_head *head, *bh, *next;
1593 size_t curr_off = 0;
1594 size_t stop = length + offset;
1596 BUG_ON(!folio_test_locked(folio));
1599 * Check for overflow
1601 BUG_ON(stop > folio_size(folio) || stop < length);
1603 head = folio_buffers(folio);
1609 size_t next_off = curr_off + bh->b_size;
1610 next = bh->b_this_page;
1613 * Are we still fully in range ?
1615 if (next_off > stop)
1619 * is this block fully invalidated?
1621 if (offset <= curr_off)
1623 curr_off = next_off;
1625 } while (bh != head);
1628 * We release buffers only if the entire folio is being invalidated.
1629 * The get_block cached value has been unconditionally invalidated,
1630 * so real IO is not possible anymore.
1632 if (length == folio_size(folio))
1633 filemap_release_folio(folio, 0);
1637 EXPORT_SYMBOL(block_invalidate_folio);
1640 * We attach and possibly dirty the buffers atomically wrt
1641 * block_dirty_folio() via i_private_lock. try_to_free_buffers
1642 * is already excluded via the folio lock.
1644 struct buffer_head *create_empty_buffers(struct folio *folio,
1645 unsigned long blocksize, unsigned long b_state)
1647 struct buffer_head *bh, *head, *tail;
1648 gfp_t gfp = GFP_NOFS | __GFP_ACCOUNT | __GFP_NOFAIL;
1650 head = folio_alloc_buffers(folio, blocksize, gfp);
1653 bh->b_state |= b_state;
1655 bh = bh->b_this_page;
1657 tail->b_this_page = head;
1659 spin_lock(&folio->mapping->i_private_lock);
1660 if (folio_test_uptodate(folio) || folio_test_dirty(folio)) {
1663 if (folio_test_dirty(folio))
1664 set_buffer_dirty(bh);
1665 if (folio_test_uptodate(folio))
1666 set_buffer_uptodate(bh);
1667 bh = bh->b_this_page;
1668 } while (bh != head);
1670 folio_attach_private(folio, head);
1671 spin_unlock(&folio->mapping->i_private_lock);
1675 EXPORT_SYMBOL(create_empty_buffers);
1678 * clean_bdev_aliases: clean a range of buffers in block device
1679 * @bdev: Block device to clean buffers in
1680 * @block: Start of a range of blocks to clean
1681 * @len: Number of blocks to clean
1683 * We are taking a range of blocks for data and we don't want writeback of any
1684 * buffer-cache aliases starting from return from this function and until the
1685 * moment when something will explicitly mark the buffer dirty (hopefully that
1686 * will not happen until we will free that block ;-) We don't even need to mark
1687 * it not-uptodate - nobody can expect anything from a newly allocated buffer
1688 * anyway. We used to use unmap_buffer() for such invalidation, but that was
1689 * wrong. We definitely don't want to mark the alias unmapped, for example - it
1690 * would confuse anyone who might pick it with bread() afterwards...
1692 * Also.. Note that bforget() doesn't lock the buffer. So there can be
1693 * writeout I/O going on against recently-freed buffers. We don't wait on that
1694 * I/O in bforget() - it's more efficient to wait on the I/O only if we really
1695 * need to. That happens here.
1697 void clean_bdev_aliases(struct block_device *bdev, sector_t block, sector_t len)
1699 struct inode *bd_inode = bdev->bd_inode;
1700 struct address_space *bd_mapping = bd_inode->i_mapping;
1701 struct folio_batch fbatch;
1702 pgoff_t index = ((loff_t)block << bd_inode->i_blkbits) / PAGE_SIZE;
1705 struct buffer_head *bh;
1706 struct buffer_head *head;
1708 end = ((loff_t)(block + len - 1) << bd_inode->i_blkbits) / PAGE_SIZE;
1709 folio_batch_init(&fbatch);
1710 while (filemap_get_folios(bd_mapping, &index, end, &fbatch)) {
1711 count = folio_batch_count(&fbatch);
1712 for (i = 0; i < count; i++) {
1713 struct folio *folio = fbatch.folios[i];
1715 if (!folio_buffers(folio))
1718 * We use folio lock instead of bd_mapping->i_private_lock
1719 * to pin buffers here since we can afford to sleep and
1720 * it scales better than a global spinlock lock.
1723 /* Recheck when the folio is locked which pins bhs */
1724 head = folio_buffers(folio);
1729 if (!buffer_mapped(bh) || (bh->b_blocknr < block))
1731 if (bh->b_blocknr >= block + len)
1733 clear_buffer_dirty(bh);
1735 clear_buffer_req(bh);
1737 bh = bh->b_this_page;
1738 } while (bh != head);
1740 folio_unlock(folio);
1742 folio_batch_release(&fbatch);
1744 /* End of range already reached? */
1745 if (index > end || !index)
1749 EXPORT_SYMBOL(clean_bdev_aliases);
1751 static struct buffer_head *folio_create_buffers(struct folio *folio,
1752 struct inode *inode,
1753 unsigned int b_state)
1755 struct buffer_head *bh;
1757 BUG_ON(!folio_test_locked(folio));
1759 bh = folio_buffers(folio);
1761 bh = create_empty_buffers(folio,
1762 1 << READ_ONCE(inode->i_blkbits), b_state);
1767 * NOTE! All mapped/uptodate combinations are valid:
1769 * Mapped Uptodate Meaning
1771 * No No "unknown" - must do get_block()
1772 * No Yes "hole" - zero-filled
1773 * Yes No "allocated" - allocated on disk, not read in
1774 * Yes Yes "valid" - allocated and up-to-date in memory.
1776 * "Dirty" is valid only with the last case (mapped+uptodate).
1780 * While block_write_full_folio is writing back the dirty buffers under
1781 * the page lock, whoever dirtied the buffers may decide to clean them
1782 * again at any time. We handle that by only looking at the buffer
1783 * state inside lock_buffer().
1785 * If block_write_full_folio() is called for regular writeback
1786 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1787 * locked buffer. This only can happen if someone has written the buffer
1788 * directly, with submit_bh(). At the address_space level PageWriteback
1789 * prevents this contention from occurring.
1791 * If block_write_full_folio() is called with wbc->sync_mode ==
1792 * WB_SYNC_ALL, the writes are posted using REQ_SYNC; this
1793 * causes the writes to be flagged as synchronous writes.
1795 int __block_write_full_folio(struct inode *inode, struct folio *folio,
1796 get_block_t *get_block, struct writeback_control *wbc)
1800 sector_t last_block;
1801 struct buffer_head *bh, *head;
1803 int nr_underway = 0;
1804 blk_opf_t write_flags = wbc_to_write_flags(wbc);
1806 head = folio_create_buffers(folio, inode,
1807 (1 << BH_Dirty) | (1 << BH_Uptodate));
1810 * Be very careful. We have no exclusion from block_dirty_folio
1811 * here, and the (potentially unmapped) buffers may become dirty at
1812 * any time. If a buffer becomes dirty here after we've inspected it
1813 * then we just miss that fact, and the folio stays dirty.
1815 * Buffers outside i_size may be dirtied by block_dirty_folio;
1816 * handle that here by just cleaning them.
1820 blocksize = bh->b_size;
1822 block = div_u64(folio_pos(folio), blocksize);
1823 last_block = div_u64(i_size_read(inode) - 1, blocksize);
1826 * Get all the dirty buffers mapped to disk addresses and
1827 * handle any aliases from the underlying blockdev's mapping.
1830 if (block > last_block) {
1832 * mapped buffers outside i_size will occur, because
1833 * this folio can be outside i_size when there is a
1834 * truncate in progress.
1837 * The buffer was zeroed by block_write_full_folio()
1839 clear_buffer_dirty(bh);
1840 set_buffer_uptodate(bh);
1841 } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1843 WARN_ON(bh->b_size != blocksize);
1844 err = get_block(inode, block, bh, 1);
1847 clear_buffer_delay(bh);
1848 if (buffer_new(bh)) {
1849 /* blockdev mappings never come here */
1850 clear_buffer_new(bh);
1851 clean_bdev_bh_alias(bh);
1854 bh = bh->b_this_page;
1856 } while (bh != head);
1859 if (!buffer_mapped(bh))
1862 * If it's a fully non-blocking write attempt and we cannot
1863 * lock the buffer then redirty the folio. Note that this can
1864 * potentially cause a busy-wait loop from writeback threads
1865 * and kswapd activity, but those code paths have their own
1866 * higher-level throttling.
1868 if (wbc->sync_mode != WB_SYNC_NONE) {
1870 } else if (!trylock_buffer(bh)) {
1871 folio_redirty_for_writepage(wbc, folio);
1874 if (test_clear_buffer_dirty(bh)) {
1875 mark_buffer_async_write_endio(bh,
1876 end_buffer_async_write);
1880 } while ((bh = bh->b_this_page) != head);
1883 * The folio and its buffers are protected by the writeback flag,
1884 * so we can drop the bh refcounts early.
1886 BUG_ON(folio_test_writeback(folio));
1887 folio_start_writeback(folio);
1890 struct buffer_head *next = bh->b_this_page;
1891 if (buffer_async_write(bh)) {
1892 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh,
1893 inode->i_write_hint, wbc);
1897 } while (bh != head);
1898 folio_unlock(folio);
1902 if (nr_underway == 0) {
1904 * The folio was marked dirty, but the buffers were
1905 * clean. Someone wrote them back by hand with
1906 * write_dirty_buffer/submit_bh. A rare case.
1908 folio_end_writeback(folio);
1911 * The folio and buffer_heads can be released at any time from
1919 * ENOSPC, or some other error. We may already have added some
1920 * blocks to the file, so we need to write these out to avoid
1921 * exposing stale data.
1922 * The folio is currently locked and not marked for writeback
1925 /* Recovery: lock and submit the mapped buffers */
1927 if (buffer_mapped(bh) && buffer_dirty(bh) &&
1928 !buffer_delay(bh)) {
1930 mark_buffer_async_write_endio(bh,
1931 end_buffer_async_write);
1934 * The buffer may have been set dirty during
1935 * attachment to a dirty folio.
1937 clear_buffer_dirty(bh);
1939 } while ((bh = bh->b_this_page) != head);
1940 folio_set_error(folio);
1941 BUG_ON(folio_test_writeback(folio));
1942 mapping_set_error(folio->mapping, err);
1943 folio_start_writeback(folio);
1945 struct buffer_head *next = bh->b_this_page;
1946 if (buffer_async_write(bh)) {
1947 clear_buffer_dirty(bh);
1948 submit_bh_wbc(REQ_OP_WRITE | write_flags, bh,
1949 inode->i_write_hint, wbc);
1953 } while (bh != head);
1954 folio_unlock(folio);
1957 EXPORT_SYMBOL(__block_write_full_folio);
1960 * If a folio has any new buffers, zero them out here, and mark them uptodate
1961 * and dirty so they'll be written out (in order to prevent uninitialised
1962 * block data from leaking). And clear the new bit.
1964 void folio_zero_new_buffers(struct folio *folio, size_t from, size_t to)
1966 size_t block_start, block_end;
1967 struct buffer_head *head, *bh;
1969 BUG_ON(!folio_test_locked(folio));
1970 head = folio_buffers(folio);
1977 block_end = block_start + bh->b_size;
1979 if (buffer_new(bh)) {
1980 if (block_end > from && block_start < to) {
1981 if (!folio_test_uptodate(folio)) {
1984 start = max(from, block_start);
1985 xend = min(to, block_end);
1987 folio_zero_segment(folio, start, xend);
1988 set_buffer_uptodate(bh);
1991 clear_buffer_new(bh);
1992 mark_buffer_dirty(bh);
1996 block_start = block_end;
1997 bh = bh->b_this_page;
1998 } while (bh != head);
2000 EXPORT_SYMBOL(folio_zero_new_buffers);
2003 iomap_to_bh(struct inode *inode, sector_t block, struct buffer_head *bh,
2004 const struct iomap *iomap)
2006 loff_t offset = (loff_t)block << inode->i_blkbits;
2008 bh->b_bdev = iomap->bdev;
2011 * Block points to offset in file we need to map, iomap contains
2012 * the offset at which the map starts. If the map ends before the
2013 * current block, then do not map the buffer and let the caller
2016 if (offset >= iomap->offset + iomap->length)
2019 switch (iomap->type) {
2022 * If the buffer is not up to date or beyond the current EOF,
2023 * we need to mark it as new to ensure sub-block zeroing is
2024 * executed if necessary.
2026 if (!buffer_uptodate(bh) ||
2027 (offset >= i_size_read(inode)))
2030 case IOMAP_DELALLOC:
2031 if (!buffer_uptodate(bh) ||
2032 (offset >= i_size_read(inode)))
2034 set_buffer_uptodate(bh);
2035 set_buffer_mapped(bh);
2036 set_buffer_delay(bh);
2038 case IOMAP_UNWRITTEN:
2040 * For unwritten regions, we always need to ensure that regions
2041 * in the block we are not writing to are zeroed. Mark the
2042 * buffer as new to ensure this.
2045 set_buffer_unwritten(bh);
2048 if ((iomap->flags & IOMAP_F_NEW) ||
2049 offset >= i_size_read(inode)) {
2051 * This can happen if truncating the block device races
2052 * with the check in the caller as i_size updates on
2053 * block devices aren't synchronized by i_rwsem for
2056 if (S_ISBLK(inode->i_mode))
2060 bh->b_blocknr = (iomap->addr + offset - iomap->offset) >>
2062 set_buffer_mapped(bh);
2070 int __block_write_begin_int(struct folio *folio, loff_t pos, unsigned len,
2071 get_block_t *get_block, const struct iomap *iomap)
2073 size_t from = offset_in_folio(folio, pos);
2074 size_t to = from + len;
2075 struct inode *inode = folio->mapping->host;
2076 size_t block_start, block_end;
2080 struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
2082 BUG_ON(!folio_test_locked(folio));
2083 BUG_ON(to > folio_size(folio));
2086 head = folio_create_buffers(folio, inode, 0);
2087 blocksize = head->b_size;
2088 block = div_u64(folio_pos(folio), blocksize);
2090 for (bh = head, block_start = 0; bh != head || !block_start;
2091 block++, block_start=block_end, bh = bh->b_this_page) {
2092 block_end = block_start + blocksize;
2093 if (block_end <= from || block_start >= to) {
2094 if (folio_test_uptodate(folio)) {
2095 if (!buffer_uptodate(bh))
2096 set_buffer_uptodate(bh);
2101 clear_buffer_new(bh);
2102 if (!buffer_mapped(bh)) {
2103 WARN_ON(bh->b_size != blocksize);
2105 err = get_block(inode, block, bh, 1);
2107 err = iomap_to_bh(inode, block, bh, iomap);
2111 if (buffer_new(bh)) {
2112 clean_bdev_bh_alias(bh);
2113 if (folio_test_uptodate(folio)) {
2114 clear_buffer_new(bh);
2115 set_buffer_uptodate(bh);
2116 mark_buffer_dirty(bh);
2119 if (block_end > to || block_start < from)
2120 folio_zero_segments(folio,
2126 if (folio_test_uptodate(folio)) {
2127 if (!buffer_uptodate(bh))
2128 set_buffer_uptodate(bh);
2131 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
2132 !buffer_unwritten(bh) &&
2133 (block_start < from || block_end > to)) {
2134 bh_read_nowait(bh, 0);
2139 * If we issued read requests - let them complete.
2141 while(wait_bh > wait) {
2142 wait_on_buffer(*--wait_bh);
2143 if (!buffer_uptodate(*wait_bh))
2147 folio_zero_new_buffers(folio, from, to);
2151 int __block_write_begin(struct page *page, loff_t pos, unsigned len,
2152 get_block_t *get_block)
2154 return __block_write_begin_int(page_folio(page), pos, len, get_block,
2157 EXPORT_SYMBOL(__block_write_begin);
2159 static void __block_commit_write(struct folio *folio, size_t from, size_t to)
2161 size_t block_start, block_end;
2162 bool partial = false;
2164 struct buffer_head *bh, *head;
2166 bh = head = folio_buffers(folio);
2167 blocksize = bh->b_size;
2171 block_end = block_start + blocksize;
2172 if (block_end <= from || block_start >= to) {
2173 if (!buffer_uptodate(bh))
2176 set_buffer_uptodate(bh);
2177 mark_buffer_dirty(bh);
2180 clear_buffer_new(bh);
2182 block_start = block_end;
2183 bh = bh->b_this_page;
2184 } while (bh != head);
2187 * If this is a partial write which happened to make all buffers
2188 * uptodate then we can optimize away a bogus read_folio() for
2189 * the next read(). Here we 'discover' whether the folio went
2190 * uptodate as a result of this (potentially partial) write.
2193 folio_mark_uptodate(folio);
2197 * block_write_begin takes care of the basic task of block allocation and
2198 * bringing partial write blocks uptodate first.
2200 * The filesystem needs to handle block truncation upon failure.
2202 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
2203 struct page **pagep, get_block_t *get_block)
2205 pgoff_t index = pos >> PAGE_SHIFT;
2209 page = grab_cache_page_write_begin(mapping, index);
2213 status = __block_write_begin(page, pos, len, get_block);
2214 if (unlikely(status)) {
2223 EXPORT_SYMBOL(block_write_begin);
2225 int block_write_end(struct file *file, struct address_space *mapping,
2226 loff_t pos, unsigned len, unsigned copied,
2227 struct page *page, void *fsdata)
2229 struct folio *folio = page_folio(page);
2230 size_t start = pos - folio_pos(folio);
2232 if (unlikely(copied < len)) {
2234 * The buffers that were written will now be uptodate, so
2235 * we don't have to worry about a read_folio reading them
2236 * and overwriting a partial write. However if we have
2237 * encountered a short write and only partially written
2238 * into a buffer, it will not be marked uptodate, so a
2239 * read_folio might come in and destroy our partial write.
2241 * Do the simplest thing, and just treat any short write to a
2242 * non uptodate folio as a zero-length write, and force the
2243 * caller to redo the whole thing.
2245 if (!folio_test_uptodate(folio))
2248 folio_zero_new_buffers(folio, start+copied, start+len);
2250 flush_dcache_folio(folio);
2252 /* This could be a short (even 0-length) commit */
2253 __block_commit_write(folio, start, start + copied);
2257 EXPORT_SYMBOL(block_write_end);
2259 int generic_write_end(struct file *file, struct address_space *mapping,
2260 loff_t pos, unsigned len, unsigned copied,
2261 struct page *page, void *fsdata)
2263 struct inode *inode = mapping->host;
2264 loff_t old_size = inode->i_size;
2265 bool i_size_changed = false;
2267 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2270 * No need to use i_size_read() here, the i_size cannot change under us
2271 * because we hold i_rwsem.
2273 * But it's important to update i_size while still holding page lock:
2274 * page writeout could otherwise come in and zero beyond i_size.
2276 if (pos + copied > inode->i_size) {
2277 i_size_write(inode, pos + copied);
2278 i_size_changed = true;
2285 pagecache_isize_extended(inode, old_size, pos);
2287 * Don't mark the inode dirty under page lock. First, it unnecessarily
2288 * makes the holding time of page lock longer. Second, it forces lock
2289 * ordering of page lock and transaction start for journaling
2293 mark_inode_dirty(inode);
2296 EXPORT_SYMBOL(generic_write_end);
2299 * block_is_partially_uptodate checks whether buffers within a folio are
2302 * Returns true if all buffers which correspond to the specified part
2303 * of the folio are uptodate.
2305 bool block_is_partially_uptodate(struct folio *folio, size_t from, size_t count)
2307 unsigned block_start, block_end, blocksize;
2309 struct buffer_head *bh, *head;
2312 head = folio_buffers(folio);
2315 blocksize = head->b_size;
2316 to = min_t(unsigned, folio_size(folio) - from, count);
2318 if (from < blocksize && to > folio_size(folio) - blocksize)
2324 block_end = block_start + blocksize;
2325 if (block_end > from && block_start < to) {
2326 if (!buffer_uptodate(bh)) {
2330 if (block_end >= to)
2333 block_start = block_end;
2334 bh = bh->b_this_page;
2335 } while (bh != head);
2339 EXPORT_SYMBOL(block_is_partially_uptodate);
2342 * Generic "read_folio" function for block devices that have the normal
2343 * get_block functionality. This is most of the block device filesystems.
2344 * Reads the folio asynchronously --- the unlock_buffer() and
2345 * set/clear_buffer_uptodate() functions propagate buffer state into the
2346 * folio once IO has completed.
2348 int block_read_full_folio(struct folio *folio, get_block_t *get_block)
2350 struct inode *inode = folio->mapping->host;
2351 sector_t iblock, lblock;
2352 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2355 int fully_mapped = 1;
2356 bool page_error = false;
2357 loff_t limit = i_size_read(inode);
2359 /* This is needed for ext4. */
2360 if (IS_ENABLED(CONFIG_FS_VERITY) && IS_VERITY(inode))
2361 limit = inode->i_sb->s_maxbytes;
2363 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
2365 head = folio_create_buffers(folio, inode, 0);
2366 blocksize = head->b_size;
2368 iblock = div_u64(folio_pos(folio), blocksize);
2369 lblock = div_u64(limit + blocksize - 1, blocksize);
2375 if (buffer_uptodate(bh))
2378 if (!buffer_mapped(bh)) {
2382 if (iblock < lblock) {
2383 WARN_ON(bh->b_size != blocksize);
2384 err = get_block(inode, iblock, bh, 0);
2386 folio_set_error(folio);
2390 if (!buffer_mapped(bh)) {
2391 folio_zero_range(folio, i * blocksize,
2394 set_buffer_uptodate(bh);
2398 * get_block() might have updated the buffer
2401 if (buffer_uptodate(bh))
2405 } while (i++, iblock++, (bh = bh->b_this_page) != head);
2408 folio_set_mappedtodisk(folio);
2412 * All buffers are uptodate or get_block() returned an
2413 * error when trying to map them - we can finish the read.
2415 folio_end_read(folio, !page_error);
2419 /* Stage two: lock the buffers */
2420 for (i = 0; i < nr; i++) {
2423 mark_buffer_async_read(bh);
2427 * Stage 3: start the IO. Check for uptodateness
2428 * inside the buffer lock in case another process reading
2429 * the underlying blockdev brought it uptodate (the sct fix).
2431 for (i = 0; i < nr; i++) {
2433 if (buffer_uptodate(bh))
2434 end_buffer_async_read(bh, 1);
2436 submit_bh(REQ_OP_READ, bh);
2440 EXPORT_SYMBOL(block_read_full_folio);
2442 /* utility function for filesystems that need to do work on expanding
2443 * truncates. Uses filesystem pagecache writes to allow the filesystem to
2444 * deal with the hole.
2446 int generic_cont_expand_simple(struct inode *inode, loff_t size)
2448 struct address_space *mapping = inode->i_mapping;
2449 const struct address_space_operations *aops = mapping->a_ops;
2451 void *fsdata = NULL;
2454 err = inode_newsize_ok(inode, size);
2458 err = aops->write_begin(NULL, mapping, size, 0, &page, &fsdata);
2462 err = aops->write_end(NULL, mapping, size, 0, 0, page, fsdata);
2468 EXPORT_SYMBOL(generic_cont_expand_simple);
2470 static int cont_expand_zero(struct file *file, struct address_space *mapping,
2471 loff_t pos, loff_t *bytes)
2473 struct inode *inode = mapping->host;
2474 const struct address_space_operations *aops = mapping->a_ops;
2475 unsigned int blocksize = i_blocksize(inode);
2477 void *fsdata = NULL;
2478 pgoff_t index, curidx;
2480 unsigned zerofrom, offset, len;
2483 index = pos >> PAGE_SHIFT;
2484 offset = pos & ~PAGE_MASK;
2486 while (index > (curidx = (curpos = *bytes)>>PAGE_SHIFT)) {
2487 zerofrom = curpos & ~PAGE_MASK;
2488 if (zerofrom & (blocksize-1)) {
2489 *bytes |= (blocksize-1);
2492 len = PAGE_SIZE - zerofrom;
2494 err = aops->write_begin(file, mapping, curpos, len,
2498 zero_user(page, zerofrom, len);
2499 err = aops->write_end(file, mapping, curpos, len, len,
2506 balance_dirty_pages_ratelimited(mapping);
2508 if (fatal_signal_pending(current)) {
2514 /* page covers the boundary, find the boundary offset */
2515 if (index == curidx) {
2516 zerofrom = curpos & ~PAGE_MASK;
2517 /* if we will expand the thing last block will be filled */
2518 if (offset <= zerofrom) {
2521 if (zerofrom & (blocksize-1)) {
2522 *bytes |= (blocksize-1);
2525 len = offset - zerofrom;
2527 err = aops->write_begin(file, mapping, curpos, len,
2531 zero_user(page, zerofrom, len);
2532 err = aops->write_end(file, mapping, curpos, len, len,
2544 * For moronic filesystems that do not allow holes in file.
2545 * We may have to extend the file.
2547 int cont_write_begin(struct file *file, struct address_space *mapping,
2548 loff_t pos, unsigned len,
2549 struct page **pagep, void **fsdata,
2550 get_block_t *get_block, loff_t *bytes)
2552 struct inode *inode = mapping->host;
2553 unsigned int blocksize = i_blocksize(inode);
2554 unsigned int zerofrom;
2557 err = cont_expand_zero(file, mapping, pos, bytes);
2561 zerofrom = *bytes & ~PAGE_MASK;
2562 if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2563 *bytes |= (blocksize-1);
2567 return block_write_begin(mapping, pos, len, pagep, get_block);
2569 EXPORT_SYMBOL(cont_write_begin);
2571 void block_commit_write(struct page *page, unsigned from, unsigned to)
2573 struct folio *folio = page_folio(page);
2574 __block_commit_write(folio, from, to);
2576 EXPORT_SYMBOL(block_commit_write);
2579 * block_page_mkwrite() is not allowed to change the file size as it gets
2580 * called from a page fault handler when a page is first dirtied. Hence we must
2581 * be careful to check for EOF conditions here. We set the page up correctly
2582 * for a written page which means we get ENOSPC checking when writing into
2583 * holes and correct delalloc and unwritten extent mapping on filesystems that
2584 * support these features.
2586 * We are not allowed to take the i_mutex here so we have to play games to
2587 * protect against truncate races as the page could now be beyond EOF. Because
2588 * truncate writes the inode size before removing pages, once we have the
2589 * page lock we can determine safely if the page is beyond EOF. If it is not
2590 * beyond EOF, then the page is guaranteed safe against truncation until we
2593 * Direct callers of this function should protect against filesystem freezing
2594 * using sb_start_pagefault() - sb_end_pagefault() functions.
2596 int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2597 get_block_t get_block)
2599 struct folio *folio = page_folio(vmf->page);
2600 struct inode *inode = file_inode(vma->vm_file);
2606 size = i_size_read(inode);
2607 if ((folio->mapping != inode->i_mapping) ||
2608 (folio_pos(folio) >= size)) {
2609 /* We overload EFAULT to mean page got truncated */
2614 end = folio_size(folio);
2615 /* folio is wholly or partially inside EOF */
2616 if (folio_pos(folio) + end > size)
2617 end = size - folio_pos(folio);
2619 ret = __block_write_begin_int(folio, 0, end, get_block, NULL);
2623 __block_commit_write(folio, 0, end);
2625 folio_mark_dirty(folio);
2626 folio_wait_stable(folio);
2629 folio_unlock(folio);
2632 EXPORT_SYMBOL(block_page_mkwrite);
2634 int block_truncate_page(struct address_space *mapping,
2635 loff_t from, get_block_t *get_block)
2637 pgoff_t index = from >> PAGE_SHIFT;
2640 size_t offset, length, pos;
2641 struct inode *inode = mapping->host;
2642 struct folio *folio;
2643 struct buffer_head *bh;
2646 blocksize = i_blocksize(inode);
2647 length = from & (blocksize - 1);
2649 /* Block boundary? Nothing to do */
2653 length = blocksize - length;
2654 iblock = ((loff_t)index * PAGE_SIZE) >> inode->i_blkbits;
2656 folio = filemap_grab_folio(mapping, index);
2658 return PTR_ERR(folio);
2660 bh = folio_buffers(folio);
2662 bh = create_empty_buffers(folio, blocksize, 0);
2664 /* Find the buffer that contains "offset" */
2665 offset = offset_in_folio(folio, from);
2667 while (offset >= pos) {
2668 bh = bh->b_this_page;
2673 if (!buffer_mapped(bh)) {
2674 WARN_ON(bh->b_size != blocksize);
2675 err = get_block(inode, iblock, bh, 0);
2678 /* unmapped? It's a hole - nothing to do */
2679 if (!buffer_mapped(bh))
2683 /* Ok, it's mapped. Make sure it's up-to-date */
2684 if (folio_test_uptodate(folio))
2685 set_buffer_uptodate(bh);
2687 if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2688 err = bh_read(bh, 0);
2689 /* Uhhuh. Read error. Complain and punt. */
2694 folio_zero_range(folio, offset, length);
2695 mark_buffer_dirty(bh);
2698 folio_unlock(folio);
2703 EXPORT_SYMBOL(block_truncate_page);
2706 * The generic ->writepage function for buffer-backed address_spaces
2708 int block_write_full_folio(struct folio *folio, struct writeback_control *wbc,
2711 struct inode * const inode = folio->mapping->host;
2712 loff_t i_size = i_size_read(inode);
2714 /* Is the folio fully inside i_size? */
2715 if (folio_pos(folio) + folio_size(folio) <= i_size)
2716 return __block_write_full_folio(inode, folio, get_block, wbc);
2718 /* Is the folio fully outside i_size? (truncate in progress) */
2719 if (folio_pos(folio) >= i_size) {
2720 folio_unlock(folio);
2721 return 0; /* don't care */
2725 * The folio straddles i_size. It must be zeroed out on each and every
2726 * writepage invocation because it may be mmapped. "A file is mapped
2727 * in multiples of the page size. For a file that is not a multiple of
2728 * the page size, the remaining memory is zeroed when mapped, and
2729 * writes to that region are not written out to the file."
2731 folio_zero_segment(folio, offset_in_folio(folio, i_size),
2733 return __block_write_full_folio(inode, folio, get_block, wbc);
2736 sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2737 get_block_t *get_block)
2739 struct inode *inode = mapping->host;
2740 struct buffer_head tmp = {
2741 .b_size = i_blocksize(inode),
2744 get_block(inode, block, &tmp, 0);
2745 return tmp.b_blocknr;
2747 EXPORT_SYMBOL(generic_block_bmap);
2749 static void end_bio_bh_io_sync(struct bio *bio)
2751 struct buffer_head *bh = bio->bi_private;
2753 if (unlikely(bio_flagged(bio, BIO_QUIET)))
2754 set_bit(BH_Quiet, &bh->b_state);
2756 bh->b_end_io(bh, !bio->bi_status);
2760 static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
2761 enum rw_hint write_hint,
2762 struct writeback_control *wbc)
2764 const enum req_op op = opf & REQ_OP_MASK;
2767 BUG_ON(!buffer_locked(bh));
2768 BUG_ON(!buffer_mapped(bh));
2769 BUG_ON(!bh->b_end_io);
2770 BUG_ON(buffer_delay(bh));
2771 BUG_ON(buffer_unwritten(bh));
2774 * Only clear out a write error when rewriting
2776 if (test_set_buffer_req(bh) && (op == REQ_OP_WRITE))
2777 clear_buffer_write_io_error(bh);
2779 if (buffer_meta(bh))
2781 if (buffer_prio(bh))
2784 bio = bio_alloc(bh->b_bdev, 1, opf, GFP_NOIO);
2786 fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
2788 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2789 bio->bi_write_hint = write_hint;
2791 __bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
2793 bio->bi_end_io = end_bio_bh_io_sync;
2794 bio->bi_private = bh;
2796 /* Take care of bh's that straddle the end of the device */
2800 wbc_init_bio(wbc, bio);
2801 wbc_account_cgroup_owner(wbc, bh->b_page, bh->b_size);
2807 void submit_bh(blk_opf_t opf, struct buffer_head *bh)
2809 submit_bh_wbc(opf, bh, WRITE_LIFE_NOT_SET, NULL);
2811 EXPORT_SYMBOL(submit_bh);
2813 void write_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2816 if (!test_clear_buffer_dirty(bh)) {
2820 bh->b_end_io = end_buffer_write_sync;
2822 submit_bh(REQ_OP_WRITE | op_flags, bh);
2824 EXPORT_SYMBOL(write_dirty_buffer);
2827 * For a data-integrity writeout, we need to wait upon any in-progress I/O
2828 * and then start new I/O and then wait upon it. The caller must have a ref on
2831 int __sync_dirty_buffer(struct buffer_head *bh, blk_opf_t op_flags)
2833 WARN_ON(atomic_read(&bh->b_count) < 1);
2835 if (test_clear_buffer_dirty(bh)) {
2837 * The bh should be mapped, but it might not be if the
2838 * device was hot-removed. Not much we can do but fail the I/O.
2840 if (!buffer_mapped(bh)) {
2846 bh->b_end_io = end_buffer_write_sync;
2847 submit_bh(REQ_OP_WRITE | op_flags, bh);
2849 if (!buffer_uptodate(bh))
2856 EXPORT_SYMBOL(__sync_dirty_buffer);
2858 int sync_dirty_buffer(struct buffer_head *bh)
2860 return __sync_dirty_buffer(bh, REQ_SYNC);
2862 EXPORT_SYMBOL(sync_dirty_buffer);
2865 * try_to_free_buffers() checks if all the buffers on this particular folio
2866 * are unused, and releases them if so.
2868 * Exclusion against try_to_free_buffers may be obtained by either
2869 * locking the folio or by holding its mapping's i_private_lock.
2871 * If the folio is dirty but all the buffers are clean then we need to
2872 * be sure to mark the folio clean as well. This is because the folio
2873 * may be against a block device, and a later reattachment of buffers
2874 * to a dirty folio will set *all* buffers dirty. Which would corrupt
2875 * filesystem data on the same device.
2877 * The same applies to regular filesystem folios: if all the buffers are
2878 * clean then we set the folio clean and proceed. To do that, we require
2879 * total exclusion from block_dirty_folio(). That is obtained with
2882 * try_to_free_buffers() is non-blocking.
2884 static inline int buffer_busy(struct buffer_head *bh)
2886 return atomic_read(&bh->b_count) |
2887 (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
2891 drop_buffers(struct folio *folio, struct buffer_head **buffers_to_free)
2893 struct buffer_head *head = folio_buffers(folio);
2894 struct buffer_head *bh;
2898 if (buffer_busy(bh))
2900 bh = bh->b_this_page;
2901 } while (bh != head);
2904 struct buffer_head *next = bh->b_this_page;
2906 if (bh->b_assoc_map)
2907 __remove_assoc_queue(bh);
2909 } while (bh != head);
2910 *buffers_to_free = head;
2911 folio_detach_private(folio);
2917 bool try_to_free_buffers(struct folio *folio)
2919 struct address_space * const mapping = folio->mapping;
2920 struct buffer_head *buffers_to_free = NULL;
2923 BUG_ON(!folio_test_locked(folio));
2924 if (folio_test_writeback(folio))
2927 if (mapping == NULL) { /* can this still happen? */
2928 ret = drop_buffers(folio, &buffers_to_free);
2932 spin_lock(&mapping->i_private_lock);
2933 ret = drop_buffers(folio, &buffers_to_free);
2936 * If the filesystem writes its buffers by hand (eg ext3)
2937 * then we can have clean buffers against a dirty folio. We
2938 * clean the folio here; otherwise the VM will never notice
2939 * that the filesystem did any IO at all.
2941 * Also, during truncate, discard_buffer will have marked all
2942 * the folio's buffers clean. We discover that here and clean
2945 * i_private_lock must be held over this entire operation in order
2946 * to synchronise against block_dirty_folio and prevent the
2947 * dirty bit from being lost.
2950 folio_cancel_dirty(folio);
2951 spin_unlock(&mapping->i_private_lock);
2953 if (buffers_to_free) {
2954 struct buffer_head *bh = buffers_to_free;
2957 struct buffer_head *next = bh->b_this_page;
2958 free_buffer_head(bh);
2960 } while (bh != buffers_to_free);
2964 EXPORT_SYMBOL(try_to_free_buffers);
2967 * Buffer-head allocation
2969 static struct kmem_cache *bh_cachep __ro_after_init;
2972 * Once the number of bh's in the machine exceeds this level, we start
2973 * stripping them in writeback.
2975 static unsigned long max_buffer_heads __ro_after_init;
2977 int buffer_heads_over_limit;
2979 struct bh_accounting {
2980 int nr; /* Number of live bh's */
2981 int ratelimit; /* Limit cacheline bouncing */
2984 static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
2986 static void recalc_bh_state(void)
2991 if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
2993 __this_cpu_write(bh_accounting.ratelimit, 0);
2994 for_each_online_cpu(i)
2995 tot += per_cpu(bh_accounting, i).nr;
2996 buffer_heads_over_limit = (tot > max_buffer_heads);
2999 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3001 struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3003 INIT_LIST_HEAD(&ret->b_assoc_buffers);
3004 spin_lock_init(&ret->b_uptodate_lock);
3006 __this_cpu_inc(bh_accounting.nr);
3012 EXPORT_SYMBOL(alloc_buffer_head);
3014 void free_buffer_head(struct buffer_head *bh)
3016 BUG_ON(!list_empty(&bh->b_assoc_buffers));
3017 kmem_cache_free(bh_cachep, bh);
3019 __this_cpu_dec(bh_accounting.nr);
3023 EXPORT_SYMBOL(free_buffer_head);
3025 static int buffer_exit_cpu_dead(unsigned int cpu)
3028 struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3030 for (i = 0; i < BH_LRU_SIZE; i++) {
3034 this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3035 per_cpu(bh_accounting, cpu).nr = 0;
3040 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3041 * @bh: struct buffer_head
3043 * Return true if the buffer is up-to-date and false,
3044 * with the buffer locked, if not.
3046 int bh_uptodate_or_lock(struct buffer_head *bh)
3048 if (!buffer_uptodate(bh)) {
3050 if (!buffer_uptodate(bh))
3056 EXPORT_SYMBOL(bh_uptodate_or_lock);
3059 * __bh_read - Submit read for a locked buffer
3060 * @bh: struct buffer_head
3061 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3062 * @wait: wait until reading finish
3064 * Returns zero on success or don't wait, and -EIO on error.
3066 int __bh_read(struct buffer_head *bh, blk_opf_t op_flags, bool wait)
3070 BUG_ON(!buffer_locked(bh));
3073 bh->b_end_io = end_buffer_read_sync;
3074 submit_bh(REQ_OP_READ | op_flags, bh);
3077 if (!buffer_uptodate(bh))
3082 EXPORT_SYMBOL(__bh_read);
3085 * __bh_read_batch - Submit read for a batch of unlocked buffers
3086 * @nr: entry number of the buffer batch
3087 * @bhs: a batch of struct buffer_head
3088 * @op_flags: appending REQ_OP_* flags besides REQ_OP_READ
3089 * @force_lock: force to get a lock on the buffer if set, otherwise drops any
3090 * buffer that cannot lock.
3092 * Returns zero on success or don't wait, and -EIO on error.
3094 void __bh_read_batch(int nr, struct buffer_head *bhs[],
3095 blk_opf_t op_flags, bool force_lock)
3099 for (i = 0; i < nr; i++) {
3100 struct buffer_head *bh = bhs[i];
3102 if (buffer_uptodate(bh))
3108 if (!trylock_buffer(bh))
3111 if (buffer_uptodate(bh)) {
3116 bh->b_end_io = end_buffer_read_sync;
3118 submit_bh(REQ_OP_READ | op_flags, bh);
3121 EXPORT_SYMBOL(__bh_read_batch);
3123 void __init buffer_init(void)
3125 unsigned long nrpages;
3128 bh_cachep = KMEM_CACHE(buffer_head,
3129 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC);
3131 * Limit the bh occupancy to 10% of ZONE_NORMAL
3133 nrpages = (nr_free_buffer_pages() * 10) / 100;
3134 max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3135 ret = cpuhp_setup_state_nocalls(CPUHP_FS_BUFF_DEAD, "fs/buffer:dead",
3136 NULL, buffer_exit_cpu_dead);